This content has been downloaded from IOPscience. Please scroll down to see the full text.

Download details:

IP Address: 129.94.8.6
This content was downloaded on 07/11/2023 at 10:40

Please note that terms and conditions apply.

You may also like:

Biosensors for Virus Detection

Investigating the Effects of Slurry Preparation Protocol on Electrochemical Performance of Carbon

Based Flowable Electrodes
Bilen Akuzum, Lutfi Agartan, Yury Gogotsi et al.

10th anniversary of attosecond pulses

Reinhard Kienberger, Zenghu Chang and Chang Hee Nam

Complex diagnostics of concrete lining technical condition

Serhii Skipochka, Viktor Serhiienko, Volodymyr Amelin et al.

Influence of Cobalt and Sodium Doping on MnO/CNT Composite Anode Materials for Li-Ion Batteries
Alessandro Palmieri, Raana Kashfi-Sadabad, Sajad Yazdani et al.
Modern Optimization Methods for
Science, Engineering and
Technology
Modern Optimization Methods for
Science, Engineering and
Technology
Edited by
G R Sinha
Myanmar Institute of Information Technology Mandalay, Myanmar

IOP Publishing, Bristol, UK

ª IOP Publishing Ltd 2020

All rights reserved. No part of this publication may be reproduced, stored in a retrieval system
or transmitted in any form or by any means, electronic, mechanical, photocopying, recording
or otherwise, without the prior permission of the publisher, or as expressly permitted by law or
under terms agreed with the appropriate rights organization. Multiple copying is permitted in
accordance with the terms of licences issued by the Copyright Licensing Agency, the Copyright
Clearance Centre and other reproduction rights organizations.

Permission to make use of IOP Publishing content other than as set out above may be sought
at [email protected].

G R Sinha has asserted his right to be identified as the author of this work in accordance with
sections 77 and 78 of the Copyright, Designs and Patents Act 1988.

ISBN 978-0-7503-2404-5 (ebook)

ISBN 978-0-7503-2402-1 (print)
ISBN 978-0-7503-2403-8 (mobi)

DOI 10.1088/978-0-7503-2404-5

Version: 20191101

IOP ebooks

British Library Cataloguing-in-Publication Data: A catalogue record for this book is available
from the British Library.

Published by IOP Publishing, wholly owned by The Institute of Physics, London

IOP Publishing, Temple Circus, Temple Way, Bristol, BS1 6HG, UK

US Office: IOP Publishing, Inc., 190 North Independence Mall West, Suite 601, Philadelphia,
PA 19106, USA
Dedicated to my late grandparents, my teachers and Revered Swami Vivekananda.
 Contents

Preface xvi
Acknowledgements xvii
Editor biography xviii
List of contributors xx

1 Introduction and background to optimization theory 1-1

1.1 Historical development 1-1
1.1.1 Robustness and optimization 1-2
1.2 Definition and elements of optimization 1-3
1.2.1 Design variables and parameters 1-4
1.2.2 Objectives 1-4
1.2.3 Constraints and bounds 1-5
1.3 Optimization problems and methods 1-6
1.3.1 Workflow of optimization methods 1-6
1.3.2 Classification of optimization methods 1-8
1.4 Design and structural optimization methods 1-9
1.4.1 Structural optimization 1-9
1.4.2 Design optimization 1-11
1.5 Optimization for signal processing and control applications 1-11
1.5.1 Signal processing optimization 1-12
1.5.2 Communication and control optimization 1-13
1.6 Design vectors, matrices, vector spaces, geometry and transforms 1-13
1.6.1 Linear algebra, matrices and design vectors 1-14
1.6.2 Vector spaces 1-15
1.6.3 Geometry, transforms, binary and fuzzy logic 1-15
References 1-17

2 Linear programming 2-1

2.1 Introduction 2-1
2.2 Applicability of LPP 2-3
2.2.1 The product mix problem 2-3
2.2.2 Diet problem 2-4
2.2.3 Transportation problem 2-4

vii
 Modern Optimization Methods for Science, Engineering and Technology

2.2.4 Portfolio optimization 2-8

2.3 The simplex method 2-10
2.4 Artificial variable techniques 2-12
2.5 Duality 2-14
2.6 Sensitivity analysis 2-15
2.7 Network models 2-17
2.7.1 Shortest path problem 2-17
2.8 Dual simplex method 2-18
2.9 Software packages to solve LPP 2-19
Further reading 2-19

3 Multivariable optimization methods for risk assessment of the 3-1

business processes of manufacturing enterprises
3.1 Introduction 3-1
3.2 A mathematical model of a business process 3-5
3.3 The market and specific risks, the features of their account 3-6
3.4 Measurement of the risk of using the discount rate, expert 3-12
assessments and indicators of sensitivity
3.5 Conclusion 3-24
References 3-24

4 Nonlinear optimization methods—overview and future scope 4-1

4.1 Introduction 4-2
4.1.1 Optimization 4-2
4.1.2 NLP 4-4
4.1.3 Nonlinear optimization problem and models 4-5
4.2 Convex analysis 4-6
4.2.1 Sets and functions 4-6
4.2.2 Convex cone 4-7
4.2.3 Concave function 4-7
4.2.4 Nonlinear optimization: the interior-point approach 4-7
4.3 Applications of nonlinear optimizations techniques 4-10
4.3.1 LOQO: an interior-point code for NLP 4-10
4.3.2 Digital audio filter 4-10
4.4 Future research scope 4-11
References 4-11

viii
 Modern Optimization Methods for Science, Engineering and Technology

5 Implementing the traveling salesman problem using 5-1

a modified ant colony optimization algorithm
5.1 ACO and candidate list 5-1
5.2 Description of candidate lists 5-2
5.3 Reasons for the tuning parameter 5-3
5.4 The improved ACO algorithm 5-3
5.4.1 Dynamic candidate set based on nearest neighbors 5-6
5.4.2 Heuristic parameter updating 5-8
5.5 Improvement strategy 5-10
5.5.1 2-Opt local search 5-10
5.6 Procedure of IACO 5-11
5.7 Flow of IACO 5-12
5.8 IACO for solving the TSP 5-12
5.9 Implementing the IACO algorithm 5-15
5.10 Experiment and performance evaluation 5-19
5.10.1 Evaluation criteria 5-20
5.10.2 Path evaluation model 5-20
5.10.3 Evaluation of solution quality 5-21
5.11 TSPLIB and experimental results 5-21
5.11.1 Experiment 1 (analysis of tour length results) 5-22
5.11.2 Experiment 2 (comparison of convergence speed) 5-25
5.12 Comparison experiment 5-27
5.13 Analysis on varying number of ants 5-34
5.13.1 Analysis of ants starting at different cities versus the same city 5-34
5.13.2 Analysis on an increasing number of ants versus number of 5-36
iterations
5.14 IACO comparison results 5-40
5.15 Conclusions 5-41
References 5-42

6 Application of a particle swarm optimization technique 6-1

in a motor imagery classification problem
6.1 Introduction 6-2
6.1.1 Literature review 6-4
6.1.2 Motivation and requirements 6-6
6.2 Particle swarm optimization 6-7
6.2.1 The mathematical model of PSO 6-8

ix
 Modern Optimization Methods for Science, Engineering and Technology

6.2.2 Constraint-based optimization 6-10

6.3 Proposed method 6-11
6.3.1 Materials and methods 6-12
6.3.2 Classification 6-14
6.4 Results 6-20
6.5 Conclusion 6-22
References 6-23

7 Multi-criterion and topology optimization using Lie symmetries 7-1

for differential equations
7.1 Introduction 7-2
7.2 Fundamentals of topological manifolds 7-3
7.2.1 Analytic manifolds 7-3
7.2.2 Lie groups and vector fields 7-4
7.3 Differential equations, groups and the jet space 7-7
7.3.1 Prolongation of group action and vector fields 7-8
7.3.2 Total derivatives of vector fields and general 7-8
prolongation formula
7.3.3 Criterion of maximal rank and infinitesimal 7-11
invariance for differential equations
7.3.4 Differential equations and symmetry groups 7-11
7.3.5 Differential invariants and the group invariant solutions 7-13
7.4 Classification of the group invariant solutions and optimal solutions 7-14
7.4.1 Adjoint representation for the cKdV and optimization 7-14
of the group generators
7.4.2 Calculation of the optimal group invariant solutions 7-18
for the cKdV
7.5 Concluding remarks 7-20
References 7-20

8 Learning classifier system 8-1

8.1 Introduction 8-1
8.2 Background 8-2
8.3 Classification learner tools 8-3
8.3.1 MATLAB®: classification learner app 8-3
8.3.2 BigML® 8-4
8.3.3 Microsoft® AzureML® 8-4

x
 Modern Optimization Methods for Science, Engineering and Technology

8.4 Sample dataset 8-4

8.4.1 Splitting the dataset 8-5
8.5 Learning classifier algorithms 8-6
8.5.1 Logistic regression classifiers 8-8
8.5.2 Decision tree classifiers 8-12
8.5.3 Discriminant analysis classifiers 8-15
8.5.4 Support vector machine classifiers 8-16
8.5.5 Nearest neighbor classifiers 8-17
8.5.6 Ensemble classifiers 8-18
8.6 Performance 8-18
8.6.1 Confusion matrix 8-20
8.6.2 Receiver operating characteristic 8-25
8.6.3 Parallel plot 8-27
8.7 Conclusion 8-28
Acknowledgments 8-29
References 8-29

9 A case study on the implementation of six sigma 9-1

tools for process improvement
9.1 Introduction 9-2
9.1.1 Generation and cleaning of BF gas 9-2
9.2 Problem overview 9-3
9.3 Project phase summaries 9-4
9.3.1 Definition 9-4
9.3.2 Measurement 9-5
9.3.3 Analyze and improvement 9-15
9.3.4 Control 9-20
9.4 Conclusion 9-20
9.4.1 Financial benefits 9-20
9.4.2 Non-financial benefits 9-20

10 Performance evaluations and measures 10-1

10.1 Performance measurement models 10-1
10.1.1 Fuzzy sets 10-2
10.2 AHP and fuzzy AHP 10-3
10.2.1 Fuzzy AHP 10-4
10.2.2 Linear programming method 10-4

xi
 Modern Optimization Methods for Science, Engineering and Technology

10.3 Performance measurement in the production approach 10-5

10.3.1 Free disposability hull 10-6
10.4 Data envelopment analysis 10-6
10.4.1 CCR model 10-7
10.4.2 BCC model 10-13
10.4.3 Other models 10-14
10.5 R as a tool for DEA 10-16
References 10-17

11 Evolutionary techniques in the design of PID controllers 11-1

11.1 The PID controller 11-2
11.1.1 Design procedure 11-3
11.1.2 Method 1: PID controller design using PSO 11-5
11.1.3 Method 2: PID controller design using BBBC 11-13
11.2 FOPID controller 11-17
11.2.1 Statement of the problem 11-18
11.2.2 BBBC aided tuning of FOPID controller parameters 11-18
11.2.3 Illustrative examples 11-18
11.3 Conclusion 11-22
References 11-26

12 A variational approach to substantial efficiency for 12-1

linear multi-objective optimization problems with
implications for market problems
12.1 Introduction 12-1
12.2 Background 12-5
12.3 A review of substantial efficiency 12-8
12.4 New results and examples 12-9
12.5 Conclusion 12-24
References 12-25

13 A machine learning approach for engineering 13-1

optimization tasks
13.1 Optimization: classification hierarchy 13-2
13.2 Optimization problems in machine learning 13-5

xii
 Modern Optimization Methods for Science, Engineering and Technology

13.3 Optimization in supervised learning 13-6

13.3.1 Bayesian optimization 13-7
13.3.2 Bayesian optimization for weight computation: a case study 13-8
13.3.3 Bayesian optimal classification: a case study 13-9
13.3.4 Bayesian optimization via binary classification: a case study 13-16
13.4 Optimization for feature selection 13-18
13.4.1 Feature extraction using precedence relations: a case study 13-20
13.4.2 Feature extraction via ensemble pruning: a case study 13-23
13.4.3 Feature-vector ranking metrics 13-25
References 13-26

14 Simulation of the formation process of spatial fine 14-1

structures in environmental safety management
systems and optimization of the parameters of
dispersive devices
14.1 The use of spatial finely dispersed multiphase structures 14-2
in ensuring ecological and technogenic safety
14.1.1 Analysis of recent research and publications 14-2
14.1.2 Statement of the problem and its solution 14-4
14.2 Physical and mathematical simulation of the creation process 14-5
of spatial finely dispersed structures
14.2.1 Gas phase study and mathematical model description 14-5
14.2.2 Dispersed phase study and mathematical model description 14-8
14.2.3 Mathematical model of interfacial interaction 14-10
14.3 Numerical simulation of the formation of spatial dispersed 14-11
structures and the determination of the most effective ways
of supplying fluid to eliminate various hazards
14.3.1 Ensuring numerical solution stability, convergence 14-11
and accuracy
14.3.2 Description of the numerical integration method 14-12
of the dispersed phase equations
14.3.3 Results of numerical simulation of a spatial finely dispersed 14-14
structure creation process which suppresses dust
14.3.4 Results of numerical simulation of the spatial finely 14-24
dispersed structure creation process, which instantly
reduces the gas stream temperature
14.4 General conclusions 14-36
References 14-36

xiii
 Modern Optimization Methods for Science, Engineering and Technology

15 Future directions: IoT, robotics and AI based applications 15-1

15.1 Introduction 15-2
15.1.1 The impact of AI and robotics in medicine and healthcare 15-3
15.1.2 Advances in AI technology and their impact on 15-4
the workforce
15.1.3 AI technologies and human intelligence 15-6
15.2 Cloud robotics, remote brains and their implications 15-7
15.2.1 Cloud computing and the RoboEarth project 15-9
15.2.2 The DAvinCi platform as a service (PaaS) surgical robot 15-9
15.3 AI and innovations in industry 15-10
15.3.1 Watson Analytics and data science 15-11
15.4 Innovative solutions for a smart society using AI, robotics and 15-11
the IoT
15.4.1 Cyber-physical systems (CPSs) 15-12
15.4.2 IoT architecture, its enabling technologies, security 15-14
and privacy, and applications
15.4.3 The Internet of robotic things (IoRT) and Industry 4.0 15-16
15.4.4 Cloud robotics and Industry 4.0 15-17
15.4.5 Opportunities, challenges and future directions 15-18
15.5 The human 4.0 or the Internet of skills (IoS) and the tactile 15-20
Internet (zero delay Internet)
15.6 Future directions in robotics, AI and the IoT 15-20
References 15-23

16 Efficacy of genetic algorithms for computationally intractable 16-1

problems
16.1 Introduction 16-2
16.2 Genetic algorithm implementation 16-3
16.3 Convergence analysis of the genetic algorithm 16-12
16.4 Key factors 16-14
16.4.1 Exploitation and exploration 16-14
16.4.2 Constrained optimization 16-15
16.4.3 Multimodal optimization 16-16
16.4.4 Multi-objective optimization 16-17
16.5 Concluding remarks 16-18
References 16-18

xiv
 Modern Optimization Methods for Science, Engineering and Technology

17 A novel approach for QoS optimization in 17-1

4G cellular networks
17.1 Mobile generations 17-1
17.2 OFDMA networks 17-2
17.2.1 Limitations of FDMA, TDMA and WCDMA networks 17-3
17.2.2 Features of OFDMA networks 17-3
17.2.3 Quality of service in OFDMA networks 17-5
17.2.4 QoS improvement techniques in OFDMA networks 17-6
17.3 Simulation model and parameters 17-9
17.3.1 Simulation topology 17-9
17.3.2 Performance metrics 17-10
17.4 Adaptive rate scheduling in OFDMA networks 17-10
17.4.1 Introduction 17-10
17.4.2 Adaptive rate scheduling algorithm 17-11
17.4.3 Average scheduling delay estimation for the ARS scheme 17-13
17.5 Conclusions 17-13
References 17-13

xv
 Preface

Optimization is generally defined as the process by which an optimum solution of a

problem is achieved. Optimization methods are designed to provide the best possible
solution or values in engineering problems or system design. These methods are
intended to improve the performance of a system in terms of several performance
evaluation factors, such as cost, time, computational complexity, raw materials, etc.
Genetic algorithms are one of the most popular areas of study and are based on an
optimization theory which works by utilizing the concept of evolution and natural
selection. Achieving better solutions and improving the performance of existing
system designs is an ongoing and continuous process on which scientists, engineers,
mathematicians, philosophers and researchers have been working for many years.
Optimization techniques are widely used in a wide range of applications, such as
robotics, artificial intelligence (AI) based applications, the chemical, electrical and
manufacturing industries, and many others.
This book focuses on the following: an introduction and background; linear
programming; multivariable methods for risk assessment; an overview of nonlinear
methods; implementation of the traveling salesman problem using modified ant
colony optimization; the application of particle swarm optimization; multi-criterion
and topology optimization; learning classifiers; case studies on six sigma real-time
steel industry applications; performance measures and evaluation; multi-objective
optimization problems; machine learning approaches; genetic algorithms and their
application; QoS optimization in cellular networks; and the future directions of
optimization methods and applications.
The purpose of this book is to present the fundamentals, background and
theoretical concepts of optimization principles in a comprehensive manner, along
with potential applications and implementation strategies. This book covers case
studies, real-time applications, development objectives and research directions, in
addition to the basic fundamentals. The book will be very useful for a wide spectrum
of readers, such as research scholars, academics and industry professionals, in
particular for those who are working on solving optimization issues, challenges and
problems.

xvi
 Acknowledgements

I begin by expressing my sincere thanks to my wife Shubhra, my daughter Samprati

and my wonderful parents for their great support and encouragement throughout
the completion of this important book. This book is the outcome of focused and
sincere efforts that could be given to the book only due to the great support of my
family.
I am grateful to my teachers who have left no stone unturned in empowering and
enlightening me, in particular Shri Bhagwati Prasad Verma who is like a godfather
to me. I extend my heartfelt thanks to the Ramakrishna Mission order and Revered
Swami Satyaroopananda of Ramakrishna Mission, Raipur, India.
I extend my sincere thanks to all the contributors for writing on the relevant
theoretical background and real-time applications of optimization methods and
entrusting upon me the role of editor.
I also wish to thank all my friends, well-wishers and all those who keep me
motivated to do more and more, better and better (as is the objective of any
optimization method).
My reverence with folded hands to Swami Vivekananda who has been the source
of inspiration for all my work and achievements.
Last, but most importantly, I express my humble thanks to Dr John Navas,
Senior Commissioning Manager of IOP Publishing for his great support, necessary
help, appreciation and quick responses. It has been wonderful experience working
with John. My Thanks to Daniel Heatley of IOP for the necessary support and many
others in the IOP team who helped me directly or indirectly. I also wish to thank IOP
Publishing for giving me this opportunity to contribute on a relevant topic with a
reputed publisher.

xvii
 Editor biography

G R Sinha
Dr G R Sinha is Adjunct Professor at the Institute of Information
Technology (IIIT) Bangalore and is currently deputed as a Professor
at Myanmar Institute of Information Technology (MIIT),
Mandalay, Myanmar. He obtained his BE (electronics engineering)
and MTech (computer technology) with a Gold Medal from the
National Institute of Technology, Raipur, India. He received his
PhD in electronics and telecommunications engineering from
Chhattisgarh Swami Vivekanand Technical University (CSVTU), Bhilai, India.
He has published 227 research papers in various international and national
journals and conferences. He is an active reviewer and editorial member of more
than 12 reputed international journals such IEEE’s Transactions on Image
Processing, Elsevier’s Computer Methods and Programs in Biomedicine, etc. He
has been Dean of Faculty and Executive Council Member of CSVTU India and is
currently a member of the Senate of MIIT. Dr Sinha has been appointed as an ACM
Distinguished Speaker in the field of DSP for the years 2017–20. He has also been
appointed as an Expert Member for the Vocational Training Program by Tata
Institute of Social Sciences (TISS) for two years (2017–19). He has been the
Chhattisgarh Representative of the IEEE MP Sub-Section Executive Council for
the last three years. He has served as a Distinguished Speaker in Digital Image
Processing for the Computer Society of India (2015). He also served as
Distinguished IEEE Lecturer on the IEEE India council for the Bombay section.
He has been a Senior Member of IEEE for many years.
He is the recipient of many awards, such as the TCS Award 2014 for Outstanding
Contributions in the Campus Commune of TCS, R B Patil ISTE National Award
2013 for Promising Teacher by ISTE New Delhi, Emerging Chhattisgarh Award
2013, Engineer of the Year Award 2011, Young Engineer Award 2008, Young
Scientist Award 2005, IEI Expert Engineer Award 2007, ISCA Young Scientist
Award 2006, and the nomination and awarding of the Deshbandhu Merit
Scholarship for five years. He has authored six books, including Biometrics
published by Wiley India, a subsidiary of John Wiley, and Medical Image
Processing, published by Prentice Hall of India. He is a consultant for various skill
development initiatives of NSDC, Government of India. He is a regular referee of
project grants under the DST-EMR scheme and several other schemes of the
Government of India. He has delivered many keynote/invited talks and chaired
many technical sessions at international conferences in Singapore, Myanmar,
Bangalore, Mumbai, Trivandrum, Hyderabad, Mysore, Allahabad, Nagercoil,
Nagpur, Kolaghat, Yangon, Meikhtila and many other places. His special session
on ‘Deep Learning in Biometrics’ was included in the IEEE International
Conference on Image Processing in 2017. He is a Fellow of IETE New Delhi and
a member of international professional societies such as IEEE, ACM and many

xviii
 Modern Optimization Methods for Science, Engineering and Technology

other national professional bodies such as ISTE, CSI, ISCA and IEI. He is a
member of various committees of the university and has been Vice President of the
Computer Society of India for the Bhilai chapter for two consecutive years. He has
supervised eight PhD scholars and 15 MTech scholars. His research interests include
image processing and computer vision, optimization methods, employability skills,
outcome based education (OBE), etc.

xix
 List of contributors

Sirajuddin Ahmed
Jamia Millia Islamia
New Delhi
India

Rajesh Chamorshikar
Bhilai Steel Plant
Bhilai
Chhattisgarh
India

Siddharth Choubey
SSTC-SSGI
CSVTU
Bhilai
Chhattisgarh
India
Abha Choubey
SSTC-SSGI
CSVTU
Bhilai
Chhattisgarh
India
Sien Deng
Department of Mathematical Sciences
Northern Illinois University
DeKalb, IL
USA
Santosh R Desai
Electronics and Instrumentation Engineering
BMS College of Engineering
Basavangudi
Bangalore
India
Somesh Kumar Dewangan
SSTC-SSGI
CSVTU
Bhilai
Chhattisgarh
India

xx
 Modern Optimization Methods for Science, Engineering and Technology

Vladimir Gorbunov
Moscow Institute of Electronic Technology
Moscow
Russia

Shankru Guggari
Department of Computer Applications
BMS College of Engineering
Bengaluru
Karnataka
India

Sailesh Kumar Gupta

Darjeeling Government College
Darjeeling
West Bengal
India

Glenn Harris
Department of Mathematical Sciences
Northern Illinois University
DeKalb, IL
USA

Zar Chi Su Su Hlaing

University of Computer Studies (Magway)
Magway
Myanmar

Nadeem Ahmad Khan

Jamia Millia Islamia
New Delhi
India

Vandana Khare
CMR College of Engineering and Technology
Hyderabad
Telangana
India

Myo Khaing
University of Computer Studies (Magway)
Magway
Myanmar

xxi
 Modern Optimization Methods for Science, Engineering and Technology

Ajay Kulkarni
Medi-Caps University
Indore
Madhya Pradesh
India

Rahul Kumar
Department of Information Technology
National Institute of Technology Raipur
Raipur
Chhattisgarh
India

Bonya Mukherjee
Bhilai Steel Plant
Bhilai
Chhattisgarh
India

Kapil Kumar Nagwanshi

MPSTME Shirpur Campus
SVKM’s NMIMS University
Mumbai
Maharashtra
India

Pushkala Narasimhan
PG Department of Commerce
NMKRV College for Women
Bangalore
Karnataka
India

Jyotiprakash Patra
SSTC-SSGI
CSVTU
Bhilai
Chhattisgarh
India

Jyothi Pillai
Department of Information Technology
Bhilai Institute of Technology
Durg
Chhattisgarh
India

xxii
 Modern Optimization Methods for Science, Engineering and Technology

Rajendra Prasad
Indian Institute of Technology Roorkee
Roorkee
Uttarakhand
India

Sachin Puntambekar
Medi-Caps University
Indore
Madhya Pradesh
India

Subrahmanian Ramani
Bhilai Steel Plant
Bhilai
Chhattisgarh
India

K C Raveendranathan
Rajadhani Institute of Engineering and Technology
Thiruvananthapuram
Kerala
India

Arpana Rawal
Bhilai Institute of Technology
Durg
Chhattisgarh
India

Mridu Sahu
Department of Information Technology
National Institute of Technology Raipur
Raipur
Chhattisgarh
India

Mamta Singh
Sai College
Bhilai
Chhattisgarh
India
G R Sinha
Myanmar Institute of Information Technology
Mandalay
Myanmar

xxiii
 Modern Optimization Methods for Science, Engineering and Technology

Kostiantyn Tkachuk
Igor Sikorsky Kyiv Polytechnic Institute
Kiev
Ukraine

Oksana Tverda
Igor Sikorsky Kyiv Polytechnic Institute
Kiev
Ukraine

Sergij Vambol
State Ecological Academy of Postgraduate Education and Management
Berdyansk State Pedagogical University
Berdyansk
Ukraine

Viola Vambol
Berdyansk State Pedagogical University
Berdyansk
Ukraine

K A Venkatesh
Myanmar Institute of Information Technology
Mandalay
Myanmar

xxiv
 IOP Publishing

Modern Optimization Methods for Science, Engineering and

Technology
G R Sinha

Chapter 1
Introduction and background to
optimization theory
G R Sinha

This chapter presents an overview and brief background of optimization methods

which are very popular in almost all applications of science, engineering, technology
and mathematics. The historical background of optimization is studied, making a
distinction between optimization and robustness. A few important definitions such
as design variables, constraints and optimization objectives are highlighted followed
by optimization problems and methods. Structural, control, signal processing and
other important optimization methods are introduced, and finally the importance of
the mathematical background which is necessary for implementing optimization
methods for various tasks is briefly presented.

1.1 Historical development

Optimization is a much talked about topic that has long been challenging. Much
literature suggests that the concept of optimization was used in 100 BC for
calculating the appropriate distance between two points. There are several such
literature sources, articles and stories about optimization theory and practice. In the
developments and advancements in any area of science, engineering, technology,
mathematics, philosophy and many others, an attempt is always made to achieve
better results, products or outcomes with better performance. The improvement of
performance over existing methods, solutions, products, devices or any advances of
science, technology, engineering and mathematics (STEM), is always studied by
numerous scholars and research groups. Research continues in all areas of STEM
for further improvement of the performance of existing work or discovering some
better method. In fact, robustness and optimization are two important terms which
can be considered as analogous because both are attempts to continuously improve.

doi:10.1088/978-0-7503-2404-5ch1 1-1 ª IOP Publishing Ltd 2020

 Modern Optimization Methods for Science, Engineering and Technology

Wikipedia suggests that the optimization problem was studied in the seventeenth
and eighteenth centuries, including with regard to Kepler’s law, Bernoulli’s theorem,
Leibniz’s calculus of variations, etc. In calculus, probability theory, mechanics,
chemical equilibrium and many other fields, optimization theorems and principles
evolved in the nineteenth and twentieth centuries. In the twentieth century, the
traveling salesman problem, which is one of the most popular among optimization
problems, was studied and worked upon. Now, optimization theory is not only
limited to STEM and allied branches but the concept is widely used in all fields, such
as financial management, economics, life sciences, genetics, population studies and
several others. We can use the evolution of human beings as an example, with
human ancestors being more ape-like. Evolution then continued and human beings
can be seen in the present context as the best possible form of the species of its kind
that natural selection could produce. The human species is now educated and is
trying become even faster through computing speed, although the thinking capacity
of the brain is unlimited. Thought processes and their outcomes are being improved
continuously. This is a simple case of optimization. A person can become a better
intellectual or individual using some methods, thought processes or techniques—
employing some type of optimization concept to perform better.
Let us understand optimization problems with the help of an example. A system is
defined by the following equations:
M = ax1 + bx2 (1.1)

P = Mx1 + c, (1.2)
where x1 and x2 are two input variables, a and b are constants, M is an intermediate
output, c is one more constant and P is the final output or response of the system.
Now, with a given set of conditions we are interested in obtaining the optimum
value of P, which can be achieved in two different ways:
• Optimizing system performance by optimizing the internal parameters and
constants which are characteristics of the system.
• Optimizing by adjusting or optimizing all individual variables or selected
variables present in the system.

Optimization is expected to produce the best possible final value otherwise some
more optimization methods need to be applied to further improve the results—there
arises the need for a robust approach that can suit all requirements and conditions.
Another situation for using optimization in the above example may be due to some
error introduced in the final value. If the error in the final value is more than the
tolerance limit (the allowed limit) then the error can be minimized by using a
suitable optimization method.

1.1.1 Robustness and optimization

Optimization and robustness are strongly related in terms of a common objective,
which is to achieve some improvement. Optimization always aims to improve

1-2
 Modern Optimization Methods for Science, Engineering and Technology

something, some parameter, so that a method, system or device performs better. The
term robustness relates to attempting a process more and more times since the solution
can never be optimum, rather it can be optimal. So, researchers will be always
engaging in experiments, simulations, modeling and study so that a better result or
performance is obtained. Robustness has long been a challenging research problem in
all fields of science and technology irrespective of application. This could be in the field
of automobiles, manufacturing, computing, or information and communication
technology (ICT) enabled services, all of which need continuous improvement. If we
are to discuss any particular example of robustness then let us discuss image processing
and computer vision based applications [1, 2]. This is an important application because
computer vision is used in almost all modern advancements. The prominent authors of
image processing admit that there is no general theory in image processing, which
means that whatever is done in a novel manner becomes new in the area of image
processing. Obtaining a robust approach for any components of image processing such
as image de-noising, segmentation or pattern matching is challenging, and it is very
difficult to claim that the obtained solution is robust one. The robustness depends on
optimizing one or a few or all the parameters involved in the research. The extensive
research on image processing suggests that robustness is a permanent research
problem, a few research contributions can be seen in [1–6] where optimization of
performance was attempted but not as a robust approach.
Combining the two important terms, robust optimization is an emerging area of
study and research where robustness is investigated among a number of optimization
problems and solutions in real-time industries and practice. The optimization process
employs a number of methods for improving the performance of a particular task
and different optimization methods produce different performance metrics and
therefore robustness needs to be investigated among the optimization methods.
The best possible optimization method that can result in optimum performance of a
system will be referred to as a robust optimization method. If the method is robust
then it should also be used for a variety of tasks irrespective of different performance
evaluation parameters.

1.2 Definition and elements of optimization

Oxford Dictionary defines optimization as a process or a method that can make
something perfect and effective1. Something here may include a design, system or
decision which becomes better and better using the methodology called optimiza-
tion. Cambridge Dictionary defines optimization as process that makes something as
good as possible, or as effective as possible2. We can refer to many other standard
definitions and find one thing common, which is the improvement in some process,
method, design or decision to make something more effective.
Although optimization is defined as a process or methodology, it is actually made
up of several components, such as decision variables, constraints and objectives.

1
The definition is taken from: https://www.lexico.com/en/definition/optimization and slightly modified.
2
The definition is taken from https://dictionary.cambridge.org/dictionary/english/optimization and slightly
modified.

1-3
 Modern Optimization Methods for Science, Engineering and Technology

Variables are the most important and guiding factors deciding the result being
produced by the optimization and constraints are certain conditions under which
optimization achieves its desired goal (its objective).

1.2.1 Design variables and parameters

Variables are design parameters used in any system which are actually responsible
for obtaining a desired output or a product. Generally, the design variables are set of
one-dimensional or multi-dimensional vectors, such as
X = (x1, x2 , x3, … , xn), (1.3)
where x1, x2, x3, …, are n number of variables and constitute together a single vector
X. Here, the vector X is one-dimensional but it can be two-dimensional, three-
dimensional or more.
These variables are controlled, manipulated and varied in the process of
optimization. In some cases, only one variable is varied or controlled and perform-
ance improvement results and optimization works, but in other cases a greater
number of variables may also change in order to achieve the desired objective in the
system design. In the applications where the theorem of equilibrium or inequalities
is used to control and obtain equilibrium, many variables are used. If optimization is
desired, then there is a wide range of possibilities to optimize the behavior of a
mechanical system [7]. We sometimes use the terms parameters and variables
interchangeably, but there is a slight difference between the two. Variables are
vectors or values which actually vary in the system, whereas parameters may be
some standard operating parameters, constants or the attributes that interrelate the
design variables. So, both of them are attributes, and their small difference lies in the
role of parameters in obtaining some relationship between the variables.

1.2.2 Objectives
An objective means a goal to be achieved. Every STEM performance improvement
will have a certain goal to achieve, maybe an error to minimize at some specific level.
In all such cases, optimization is used with the important purpose to achieve the goal
that is set by the system response or method. Thus, the optimization method is
assigned a minimum goal that is required to be achieved. As an example, a few
images are presented in figure 1.1 which are the results of breast cancer segmentation
and detection. The results were obtained using an optimized method that combined
gray level clustering enhancement and adaptive threshold segmentation (GLCE-
ATS). If we look at four different segmentation results which all highlight breast
cancer masses, they appear the same and it is very difficult to determine the cases in
which more cancerous elements (malignant elements) are present. The optimization
method has the objective to achieve the maximum number of malignant masses
present in the breast mammograms [8].
The objective of optimization is also to minimize the error or value of a
suitable function e(x) or f(x), y = minimize {f(x)} or minimize {e(x)}.

1-4
 Modern Optimization Methods for Science, Engineering and Technology

(a) (b)

(c) (d)
Figure 1.1. Results of breast cancer detection, with different results for different optimization parameters used
in the GLCE-ATS method.

1.2.3 Constraints and bounds

Constraints are some limitations and conditions under which the optimization
technique or method works. The method successfully improves the system perform-
ance provided a certain set of constraints are satisfied. The limitations include some
bounds as well, which may be upper bounds, lower bounds or intermediate.
The optimization operates on a set of data or samples subjected to some
limitations that may depend on certain characteristics or features of the system
being optimized, or some external factors affecting the system design or perform-
ance. The bounds are generally based on design variables but may also depend on
parameters. For example,
Y = minimize {e(x )} under x2 < x0 and x1 > xp . (1.4)

1-5
 Modern Optimization Methods for Science, Engineering and Technology

Here the goal is to minimize the value of the error signal. The minimum or
optimum value has to be obtained as Y and the optimization has to be achieved
under the conditions given in equation (1.4). These conditions are bounds which are
apparently lower and upper bounds in the case of x2 and x1, respectively. The error
signal or value needs to be minimized between x1 and x2. In different types of
optimization problems and applications, various types of bounds and constraints are
used. The optimization methods can be generalized but the constraints are specific to
the applications or problems being addressed. Therefore, the constraints are
subjective, whereas the optimization methods are objective for the applications. In
[7], the solution to achieving equilibrium requires a number of constraints in the
implementation of optimization in the work.

1.3 Optimization problems and methods

In applications related to STEM or allied fields, there are a variety of problems that
need to be addressed in terms of improvement of system design and performance, or
minimization of error, reduction of labor and cost, etc. These problems are solved
using suitable optimization methods [9]. In the determination of optimization
problems, several standard aspects are used, such as:
1. Optimal input number of arguments.
2. Maximum and minimum values of functions.
3. Characteristics of the system.
4. Constraints and bounds.

The number of arguments or attributes plays an important role in choosing the

optimization problem. In a few cases, the optimization performs in such a manner
that the computational complexity decreases with the increase in the number of
input arguments, whereas in some other cases the complexity is also increased. In
fact, the goal should also be to reduce the computational complexity in addition to
improved performance or system design. The number of arguments is to be selected
wisely so that optimization produces the best possible result while expending the
least possible time and fewest operations. The bounds are limits to certain
conditions, which are also important points to be kept in mind while designing an
optimization method and implementing it. So, the problems are governed by the
upper and lower limits of constraints and bounds, and also by the minimum and
maximum possible values of the functions being optimized [7, 9]. The targets for
optimization are also decided on the basis of specific characteristics of system
response, design or behavior.
The number of methods for optimization is large and these can be employed for
solving a variety of problems.

1.3.1 Workflow of optimization methods

Any optimization problem and its possible solution has some typical workflow
diagram, one such workflow for the execution of any general-purpose optimization
method is shown in figure 1.2. This figure highlights the major steps that are

1-6
 Modern Optimization Methods for Science, Engineering and Technology

Need for Optimization

Selection of Optimization of Problem

Selection of Design Variables

Set Optimization Objective

Constraints and Bounds

Selection of suitable Optimization

Method/algorithm

Obtain Best Possible Solution

Figure 1.2. Major steps of the workflow of a typical optimization method.

executed in the implementation of any general-purpose optimization method. The

process begins with the need for optimization, which is the very first step in any
application. The need for optimization helps in the identification of possible
optimization problems. There may be multiple problems in an application that
could be optimized using a single method or multi-method based optimization. Once
the selection of problems is finally made, the design variables are selected which
further determine the main objective of employing a suitable optimization method.
The optimization has to be carried out under certain constraints and boundary
values, and therefore before we actually choose an appropriate method, we set
various design constraints and limitations.
Then, the optimization method is decided based on the design variables, require-
ments, optimization problem, constraints and bounds. The results of the
suitable optimization method are assessed as to whether they are the best possible
results which were expected or not. If not, then the workflow of the implementation
may have several reverse operations (the reverse direction arrows in the diagram are
not shown as they depend on what is actually needed). If the design variables need
some modification or manipulation, or similarly the constraints require some
changes along with set-in limit values, these aspects are corrected so as to achieve
the best possible response or outcome produced by the optimization method. Thus,
it can be seen in the flow diagram that the optimization problem is chosen in the

1-7
 Modern Optimization Methods for Science, Engineering and Technology

beginning as per the need and the optimization method is selected after certain steps,
such as defining design variables, constraints and bounds.

1.3.2 Classification of optimization methods

As we have discussed, the optimization method is decided based on several factors,
and constraints may be linear as well as non-linear. Accordingly, the methods are
broadly categories based on:
• Linear constraint terms and bounds.
• Non-linear constraint terms and bounds.

The methods of optimization will fall under different categories, but the basis of
choosing a suitable method will be the above two factors, that is, the constraints and
bounds being either linear or non-linear. The optimization methods are further
classified as:
• Single variable or multi-variable methods.
• Constrained or non-constrained methods.
• Linear or non-linear optimization methods.
• Single objective or multi-objective optimization methods.
• Specialized or general-purpose methods.
• Traditional or non-traditional methods.

The classification of optimization methods is also based on the characteristics or

properties of the objective function of optimization [9]. For example, if control
variables are continuous and real numbers, then the optimization method is called
continuous optimization. Similarly, for integers it is known as discrete optimization
and for a combination of continuous real numbers and integers the optimization is
known as mixed integer. Based on different types of design variables, quadratic,
linear and non-linear, the optimization method is referred to accordingly as
quadratic, linear or non-linear. If the optimization problem is subject to some
limitations or constraints then the optimization method is called a constrained
optimization method and if there are no such constraints, the method is called an
unconstrained optimization method.
Conventional or traditional methods of optimization include mainly linear
programming, non-linear programming, separable programming and integer
programming optimization methods. The methods which are based on evolu-
tionary concepts are referred to as evolutionary methods of optimization, which
include a few main methods such as the heuristic search, genetic algorithm,
evolutionary programming, particle swarm optimization (PSO), simulated anneal-
ing, bacterial foraging and differential equation based methods. Examples of
unconstrained optimization methods are the gradient method, Newton’s method,
quasi-Newton method, conjugate directive method, etc. Another emerging area of
optimization is global optimization methods, that include some probabilistic and
deterministic methods. These methods are based on probability theory and similar
mathematical fundamentals. Global optimization methods are widely used in

1-8
 Modern Optimization Methods for Science, Engineering and Technology

numerical analysis and feature based analysis in various signal processing and
control applications.
In the fast changing dynamic environments of various emerging applications of
signal processing, control, transport, etc, optimization is employed where the
methods, models and problems are dynamic in nature. Telecommunications,
artificial intelligence based advancements, financial analysis and several others fields
are rapidly changing with a wide range of dynamics wherein the problems and
models of optimization used also change rapidly. In such applications, the methods
which are traditionally used also need modification and dynamic updates [10]. There
are a number of optimization methods which are recommended for such applica-
tions, a few of which are: ant colony optimization, evolutionary optimization,
genetic algorithms, neural networks and swarm intelligence based methods, etc. In
the context of dynamic optimization, the quantitative measures of performance to
assess how the optimization method is working become essential. Such measures are
required in all optimization methods, but in a dynamically changing environment
the assessment or evaluation is particularly important [1–3, 9, 10]. Some of the
important measures which are widely used, irrespective of type of optimization
problem and application, are:
• Cumulative fitness and mean fitness.
• Off-line error.
• Robustness.
• Diversity.
• Standard deviation.
• Time.
• Computational complexity.
• Average error.
• Average best function value.
• Current best evolution.

1.4 Design and structural optimization methods

This is an era of technological and industrial revolutions that keep on changing
dynamically. In order to meet the ever-changing needs of technology enabled
industries, optimization is required to achieve better design and structural performance.
For example, numerous studies suggest that airfoil design plays an important role in
the aviation sector and a number of significant research contributions can be found
in the literature in which airfoils, blades and others components of flight are designed in
such an optimal manner that the performance of airplanes improves considerably. The
structure of any system mainly depends on the shape, size and other similar attributes,
and optimization works on optimizing one or more of these attributes.

1.4.1 Structural optimization

Structural optimization deals directly with the variables related to the structure of
the system or a product, such as size, shape, area, perimeter, volume, etc. If the

1-9
 Modern Optimization Methods for Science, Engineering and Technology

Figure 1.3. A typical structural design of the outer surface of an airplane.

comfort level has to be increased in the field of ergonomics, then the optimization
utilizes the structural properties of the system or device whose ergonomics require
improvement. Figure 1.3 shows a simple and typical structural diagram of the outer
surface of an airplane where the airfoil or blade design and their optimization are
important factors in the performance of the flight. This is just an symbolic diagram,
as there are thousands of small, medium and high level structures present in an
airplane.
Structural optimization is divided into various types based on the structural
attributes. A few major structural optimization methods or techniques are as
follows:
1. Shape optimization: This is based on the different shapes of systems or
products.
2. Area and volume optimization: Area and volume decide the performance of a
huge number of products in the transport and control based applications,
and therefore the performance could be improved by optimizing either the
area or volume or both or some other similar variables.
3. Size optimization: This may include length, width and other similar dimen-
sional features in the optimization of the system’s structural performance.
4. Topological optimization: Since the topology represents the overall intercon-
nection of various components in a system and how those components are
structured together determines the system’s performance, in particular in the
automobile and similar sectors, and the role of optimization is to improve the
topology of the system architecture [11].
The topological optimization of structures such as seismic loads, struc-
tural optimization for steel plants and reliability optimization are examples
of structural optimization methods. Topological optimization is imple-
mented following a certain workflow of operations; one such flow diagram
can be seen in figure 1.4. The need for optimization arises and structural
properties are studied. The optimization problem is identified and a
suitable method is implemented that utilizes the structural attributes in the
process of optimization. The optimization results in some post-processing
with the application of an optimization method which is sensitive to the
structural features of the system.

1-10
 Modern Optimization Methods for Science, Engineering and Technology

Need for Optimization based on structure of

the system

Optimization of Problem

Structural
Optimization Method Analysis

Application of
Optimization

Figure 1.4. Topological optimization design flow.

1.4.2 Design optimization

Design optimization is not entirely different from structural optimization because
the design of a system and its structural properties are interrelated. The improve-
ment in system performance, reduction in cost and other benefits of optimization can
be achieved using structural optimization methods. The design optimization
techniques is mainly associated with computer-aided design (CAD) for various
applications. Software tools are embedded within the CAD systems and any design
modification or manipulation can easily be done so as to obtain the desired system
characteristics.

1.5 Optimization for signal processing and control applications

Signal processing has become an essential component of almost all modern
applications and implementations of STEM and allied fields, and in all such
applications signal processing and certain signal operations are decisive factors for
the performance of application based systems. We can mention communication
systems, robotics, self-driving cars, remote sensing, navigational aids, global
positioning systems, aerodynamics, mechanics, geographic information systems,
weather forecasting, financial analysis, mobile communications, satellite communi-
cations, etc, in which signal processing and some suitable mathematical operations
are very important factors in the assessment of the performance of the systems. If the
performance improvement of the system is intended for such applications, then
applying optimization methods in some or all signal processing variables can help in
improving performance.
For example, in the area of biometrics and CAD systems, there are a significant
number of research publications, but still no robust approach is available and
therefore optimization can help in optimizing some parameters so that the system
behaves in a robust way.

1-11
 Modern Optimization Methods for Science, Engineering and Technology

1.5.1 Signal processing optimization

Signal processing involves digital signal processing (DSP) or DSP processors where
numerous signal processing operations are performed. We need to understand the
scope of optimization in signal processing applications by analyzing DSP, since DSP
is the core part of signal processing and therefore let us examine the main
components of DSP. DSP includes some major steps:
1. Data acquisition and sampling.
2. Quantization.
3. Encoding.
4. Signal processing operations.

We have to understand where the optimization could be applied so that signal

processing becomes better and contributes to the overall improvement of system
performance in which signal processing is present. If there is scope to apply a
suitable optimization method in sample processing, sampling rate, or a data
acquisition step then this has to be identified. Quantization plays an important
role in DSP and therefore optimization can also be implemented with a quantization
process. Encoding is the step that converts the data into a digital format which is
used in all DSP, and thus optimization can be applied in the process of encoding as
well. In all such optimization processes, the following types of optimization could be
broadly required:
• Window size optimization: In applications where DSP is required, several
filtering operations will also be required, which further employ certain types
of windowing operations. In DSP applications, several types of window
functions are used, but even after choosing a suitable window function, its
operation could be improved by using some optimization methods. Window
size along with execution time, memory requirements and threshold value can
also be optimized so that the windowing effect ultimately contributes to the
system in optimal ways.
• System throughput optimization: Signal processing is implemented using
computers and a number of operations are associated with signal processing.
DSP implementation affects memory requirements, central processing unit
(CPU) time, memory utilization, parallelism, CPU scheduling, etc. The goal
of optimization at the throughput level is to obtain more and better output
from computing systems.

Best candidate selection is a factor that plays an important role in operation

research, where the core of the performance lies in signal processing. How to choose
an appropriate sorting method for best candidate selection in operation research
tasks matters a lot in achieving reduced complexity and shorter computation times.
Sometimes it is very difficult to achieve both reduced complexity and shorter
computation times, and one needs to sacrifice something to achieve better values
elsewhere. This balance, which is essential, can be obtained using an optimization
method [12].

1-12
 Modern Optimization Methods for Science, Engineering and Technology

1.5.2 Communication and control optimization

We have seen that signal processing flow is communicated to various stages of the
system design and that the control of signal flow also affects the system’s perform-
ance. Therefore, we need to also understand the optimization concept for commu-
nication and control related operations. There are many optimization methods in
the literature that are suitable for the applications, but these methods are all subject
to one concept in optimization theory which is a common factor in all communi-
cation related applications, namely the convex theory of optimization. One example
of such convex optimization for communication and signal processing can be seen in
[13], which highlights the concepts of the theory and some examples of optimization
methods utilizing convex theory as central theme of optimization.
The main aim of using convex optimization in signal processing and communi-
cation is to deal with convex constraints that are obtained from convex mathemat-
ical functions. Conic optimization and interior point method optimization are
examples of convex theory based optimization. Convex itself means to achieve a
suitable objective function that can help in obtaining a local optimal function in a
problem. This optimization concept is used in obtaining both global and local
optima. Several engineering applications employ convex optimization methods.
Linear sub-space, second order cone and Affine sub-space are a few major examples
of the convex functions used for optimization tasks [13]. In fact, unconstrained
methods can also be converted into some finite constrained methods by using the
convex concept. Analog band pulse amplitude modulation and filter design are the
cases where convex optimization is useful in communication related applications.

1.6 Design vectors, matrices, vector spaces, geometry and transforms

In applications related to search engine optimization [14], or any STEM application,
the implementation of optimization requires a strong mathematical background
because optimization uses variables, operations, constraints, minima and maxima
functions, etc. Therefore, system design optimization in any applications of
optimization needs sufficient consideration of the vectors, matrices and vector
spaces that support the mathematical implementation of the optimization method.
This section provides a brief overview of these concepts, because all the methods
discussed and implemented in the other chapters use realistic and actual mathemat-
ical functions and variables.
The advances in technology, evolution of the internet, increased use of computing
devices and large amounts of data are some of the factors that require optimization
of various types of system designs and operations. In the flow diagram of
implementation of optimization, we have discussed that the process depends on
design variables, constraints and some final conditions to check if the optimization is
working or not. All of these variables and their operations make use of linear
algebra, matrices and vector spaces [15] for different computations involved in the
system, which are presented in this section.

1-13
 Modern Optimization Methods for Science, Engineering and Technology

1.6.1 Linear algebra, matrices and design vectors

Linear algebra is the basis for linear optimization methods and the algebra utilizes
linear equations. A typical example of set of linear equations is
X1 + 2Y1 + 1.5 Z1 = P1 (1.5)

X2 + 2Y2 + 3Z2 = P2 (1.6)

X3 + 1.5Y3 + 5Z3 = P3, (1.7)

where X1, X2, X3, Y1, Y2, Y3, Z1, Z2, Z3 are variables and P1, P2 and P3 are
constants. The equations are linear equations and the values of P1, P2 and P3 may be
any real values. The representation as shown by the three equations can also be
represented as a vector:
• X = [X1 × 2 × 3]; Y = [Y1 Y2 Y3]; Z = [Z1 Z2 Z3]; and P = [P1 P2 P3].
• The coefficients used in the three equations can also be expressed as vectors:
A1 = [1 1 1]; B = [2 2 1.5]; and C = [1.5 3 5].

We can also write the above equations as

⎡1 2 1.5⎤⎡ X ⎤ ⎡ P1 ⎤
⎢1 2 3 ⎥⎢Y ⎥ = ⎢ P ⎥ . (1.8)
⎢⎣ ⎥ ⎢ 2⎥
1 1.5 5 ⎦⎢⎣ Z ⎥⎦ ⎣ P3 ⎦

There are many other ways in which vector representation can replace the set of
linear equations. This representation of equations is more appropriately called a
matrix representation. When matrix representation comes into the analysis of any
optimization theory then all of its properties, such as determinant, rank, character-
istic equation, eigenvalues, eigenvectors, etc, become very important to understand.
Since the matrix and its properties at this level fall under the elementary knowledge
of matrices we will not discuss them further here. Thus, linear optimization can
work on:
• Linear equations.
• Vector representations of linear equations.
• Matrix representations.

Matrix representations provide an efficient way to express the operations and also
better help in the analysis of systems compared to the other approaches as far as the
implementation of linear optimization is concerned [16]. Linear representation based
optimization includes:
• The steepest descent method.
• The conjugate gradient method [17].
• The normal equation method.

1-14
 Modern Optimization Methods for Science, Engineering and Technology

x x+y

Figure 1.5. Vector addition.

1.6.2 Vector spaces

Vector space represents a set function whose elements are called vectors and vector
space also includes another field which is a scalar quantity (generally a real number).
Two important operations are characterized by vector space operation, plus (+) and
multiplication (·), which are vector addition and scalar multiplication, respectively.
If x and y are vectors of vector space X then x + y will also belong to the vector space
X. The following are the important principles of vector addition:
• Cumulative law: x + y = y + x, for all x, y in X.
• Associative law: x + (y + z) = (x + y) + z, where x, y and z are vectors in
vector space X.
• Additive identity and additive inverse: If a vector is added with additive
identity then it remains the same vector, x + 0 = x, and if the vector is added
with additive inverse then it results in an additive identity: x + x′ = 0.

Similarly, scalar multiplication satisfies the following:

• Distributive law: p·(x + y) = p·x + p·y and (p + q)·x = p·x + q·x where p ad q
are real numbers.
• Associative law: p·(q·x) = (p·q)·x.
• Unitary law: 1·x = x for all x in X.

Any vector space can also have its vector sub-spaces. The concepts of vector space
and sub-spaces are utilized in the optimization methods for linear optimization in
several applications [18–24]. Vector space includes vector addition which is a
common mathematical tool in vector algebra. This can be seen in figure 1.5.

1.6.3 Geometry, transforms, binary and fuzzy logic

Mathematical geometry and transforms are used in a number of optimization
problems and solutions. Different types of operations such as rotation, scaling, shift,
resizing and reformatting all come under the geometry and transforms that are used

1-15
 Modern Optimization Methods for Science, Engineering and Technology

Figure 1.6. Rotation by a certain angle.

Hot

Cold

Figure 1.7. Two possible values of temperature as hot and cold.

some steps of optimization methods. In image processing, for example, rotation is an

important translation operation which can be performed to any extent; one such
example is shown in figure 1.6.
There are several examples of operations which can be included in this section,
but the idea of showing one example is to just highlight what transform and
geometry operations mean.
Now in the age of DSP, binary operations are essential components in all stages
of DSP. When it comes to decision making and path planning control then a new
logic is preferred in all STEM applications, which is called fuzzy logic. The concept
of fuzzy logic is now utilized in almost all applications, including optimization
methods. Neural network and other machine leaning methods are used for various
feature extraction and training methods for samples, but a hybrid approach is
recommended for better performance of neuro-fuzzy logic to reduce the uncertainty.
Figure 1.7 shows the temperature of an object as hot and cold.
The problem arises when we want to know the degree of being hot or cold, which
means how hot or cold the object is. Here, the binary concept produces some
amount of uncertainty which is addressed by using fuzzy logic. This logic is
characterized by using some suitable membership function, for example in the
current case of temperature, if the membership function is defined properly then the
degree of hotness or coldness can be determined (see figure 1.8).

1-16
 Modern Optimization Methods for Science, Engineering and Technology

Hot

0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0

Membership value (function), m

Figure 1.8. Membership function in defining a fuzzy variable.

If we describe the temperature with a suitable value of membership function, say

the maximum temperature is 1100 °C and the minimum is 100 °C, then if the
membership function (value) is 0.3 it means that the temperature of the object is 100
(minimum) + 0.3 × (1100 – 100) = 400 °C. So, the object is hot but the certainty is
there as to what extent the object is hot. The concept of fuzzy logic is very popular
among scientists and researchers for various science and engineering applications.

References
[1] Sinha G R, Raju K S, Patra R, Aye D W and Khin D T 2018 Research studies on human
cognitive ability Int. J. Intell. Defen. Supp. Syst. 5 298–304
[2] Sinha H, Meshram M R and Sinha G R 2018 BER performance analysis of MIMO-OFDM
over wireless channel Int. J. Pure Appl. Math. 118 195–206
[3] Patel B C and Sinha G R 2011 Comparative performance evaluation of segmentation
methods in breast cancer images Int. J. Mach. Intell. 3 130–3
[4] Sinha G R 2015 Fuzzy Based Medical Image Processing Advances in Medical Technologies
and Clinical Practice (AMTCP) Book Series (Hershey, PA: IGI Global)
[5] Patel B C and Sinha G R 2010 Early detection of breast cancer using self-similar fractal
method Int. J. Comput. Appl. 10 39–43
[6] Sinha K and Sinha G R 2013 Comparative analysis of optimized K-means and C-means
clustering methods for segmentation of brain MRI images for tumor extraction Proc. of Int.
Conf. on Emerging Research in Computing, Information, Communication and Applications
(Amsterdam: Elsevier), pp 619–25
[7] Prepoka A 1980 On the development of optimization theory Am. Math. Month. 87 527–42
[8] Patel B C and Sinha G R 2018 Mass segmentation and feature extraction of mammographic
breast images in computer-aided diagnosis PhD Thesis Chhattisgarh Swami Vivekanand
Technical University Bhilai
[9] Chong E K P and Zak S H 2001 An Introduction to Optimization (New York: Wiley)
[10] Cruz C, González J R and Pelta D A 2011 Optimization in dynamic environments: a survey
on problems, methods and measures Soft Comput. 15 1427–48
[11] Labanada S R 2015 Mathematical programming methods for large-scale topology opti-
mization problems PhD Thesis Technical University of Denmark
[12] Ahmad H A 2012 The best candidates method for solving optimization problems J. Comput.
Sci. 8 711–5
[13] Luo Z Q and Yu W 2006 An introduction to convex optimization for communications and
signal processing IEEE J. Select. Area Commun. 24 1426–38
[14] Bezhovski Z 2015 The historical development of search engine optimization Inform. Know.
Manage. 5 91–6

1-17
 Modern Optimization Methods for Science, Engineering and Technology

[15] Luenberger D G 1969 Optimization by Vector Space Methods (New York: Wiley)
[16] Absil P A, Mahony R and Sepulchre R 2008 Optimization Algorithms on Matrix Manifolds
(Princeton, NJ: Princeton University Press)
[17] Abrudan T, Eriksson J and Koivunen V 2009 Conjugate gradient algorithm for optimization
under unitary matrix constraint Signal Process. 89 1704–14
[18] Ehrgott M and Gandibleux X 2003 Multiple Criteria Optimization: State of the Art
Annotated Bibliographic Surveys (New York: Kluwer/Academic)
[19] Boumal N 2014 Optimization and estimation on manifolds PhD Princeton University https://
web.math.princeton.edu/~nboumal/papers/boumal_optimization_and_estimation_on_mani-
folds_phd_thesis.pdf
[20] Hartley R I and Kahl F 2009 Global optimization through rotation space search Int. J.
Comput. Vis. 82 64–79
[21] Gallier J and Quaintance J 2018 Fundamentals of linear algebra and optimization Technical
Report University of Pennsylvania
[22] Nocedal J and Wright S J 2006 Numerical Optimization 2nd ed (Berlin: Springer)
[23] Dubourg V 2011 Adaptive surrogate models for reliability analysis and reliability-based
design optimization PhD Thesis Université Blaise Pascal https://www.phimeca.com/IMG/
pdf/these_dubourg_2011.pdf
[24] Simon D 2013 Evolutionary Optimization Algorithms: Biologically-Inspired and Population-
Based Approaches to Computer Intelligence (Hoboken, NJ: Wiley)

1-18
 IOP Publishing

Modern Optimization Methods for Science, Engineering and

Technology
G R Sinha

Chapter 2
Linear programming
K A Venkatesh

Optimization is the key for success in any field. In general, every resource, such as
time, availability of skilled work force, etc, is finite; allocating scarce resources to the
requested needs in an optimal way is one of the important tasks in every operation.
Linear programming is one of the mathematical programming models which has
roots in ancient mathematics, namely algebra, in particular the solving of systems of
simultaneous linear equations, linear algebra and specifically matrix operations. The
potential applications of the linear programming problem (LPP) are enormous:
supply chain management, organ donor and recipient matching, communications
network design, aviation industries, financial engineering, network optimization,
smart grids, decision making and so on. LPP is a special case of optimization
problems, which look for maximum or minimum values for a single objective
function or multiple objective functions with a set of constraints; the constraints may
be equations or inequalities. In LPP, the variables in the objective function and
constraints are linearly related.

2.1 Introduction
The study of modeling with linear equations began its journey with the birth of
algebra. As a natural extension, many real-world problems can be modeled as a
system of simultaneous linear equations. Let us begin with the example of a high
school level mathematics problem.

Example 2.1. Three pens and two pencils cost $190 and two pens and three pencils
cost $180. Find the cost of each pen and pencil.
To model as a system, we use the letters x , y to denote the unknown quantities
and these are known as variables. The above problem can be written as

doi:10.1088/978-0-7503-2404-5ch2 2-1 ª IOP Publishing Ltd 2020

 Modern Optimization Methods for Science, Engineering and Technology

3x + 2y = 190
2x + 3y = 180.
The above system in matrix notion, can be written as AX = b, where A is the
coefficient matrix of order 2(n ), and X and b are column vectors of order 2 × 1
(n × 1). To solve the system, there are many methods from the theory of matrices. In
general, the methods are classified into two categories, namely direct methods and
iterative methods. Methods such as Cramer’s, Gauss elimination, Jordon and matrix
decomposition are examples of direct methods and Gauss–Jacobi, Gauss–Seidel and
SOR methods are examples of iterative methods. All the mentioned methods work
well as long as the coefficient matrix is a square one. Suppose that the coefficient
matrix is a rectangular one, that is the system has a greater number of variables than
the number of equations, we cannot then solve in the usual sense. Suppose the
system has k variables in m equations (k > m ), set k – m variables as zero then solve
for the remaining variables. Such solutions are called basic solutions. The number of
basic solutions is ( k ).
m
A single entity which consists of a linear expression, called the objective function,
with finitely many linear inequalities, called linear constraints and non-negative
restrictions, is called the linear programming problem (LPP). The general form of
the LPP is as follows.
Objective function:
n
Optimize Z = ∑i=1cixi . (2.1)

Subject to the constraints

AX {⩽, =, ⩾}b (2.2)

X ⩾ 0, (2.3)
where A is an m × n matrix, and X and b are column vectors of order m × 1.
The basic solutions which satisfy the given set of constraints are called basic
feasible solutions. The basic feasible solution which optimizes the objective function
is called the optimal basic feasible solution.
All LPPs fall into one of three categories: feasible and bounded, unbounded, and
infeasible. If the given LPP is feasible and bounded then it has at least one optimal
solution.
Prior to Dantzig’s simplex method to solve the LPP, there was an enumerative
algorithm to find a feasible solution to the given set of constraints, but this method
was not a viable model because it inherently generates additional inequalities. In
1963, Dantzig developed the simplex method and led the development of a new
branch of study, called LPP.
The simplex method principle is that rather than working with inequalities, you
transform the inequalities into equalities by introducing slack variables (if the sign of
the inequality is ⩽) or surplus (if the sign of the inequality is ⩾) variables. This

2-2
 Modern Optimization Methods for Science, Engineering and Technology

augmented form of LPP is called the standard form. If the type of objective function
is maximization, then convert all constraints which are of ⩾ type.

2.2 Applicability of LPP

From the time of the development of the simplex method and with advancements in
technology, a lot of applications of LPP have evolved, in particular, a major portion
of management science for decision making and analytics is based on LPP. In this
section we look at some classical examples such as the product mix and diet
problems, and the applicability of LPP in transportation problems, transshipment
problems and portfolio optimization.

2.2.1 The product mix problem

The product mix problem is one of the common problems in any book on LPPs. The
important characteristic of a product mix problem is the available finite amount of
resources; one has to supply these resources to competing products. Suppose there
are n products and m resources, the requirement and availability can be given in an
m × n table as m rows and n columns of the corresponding coefficients. The values in
a row of the table are the coefficients of a constraint in the LPP. In addition, the
table shows the profit per unit and associated with each available resource. The aim
is to identify the number of units to be produced to maximize the profit using the
available resources.

Example 2.2. A manufacturer produces three products P1, P2 and P3 in one of its
plants. Each product requires a certain amount of time on each of two machines,
given in table 2.1 (h unit−1).
The goal of the problem is to identify the optimal number of units of P1, P2, P3 to
produce so that the earned profit is maximized. Let x , y, z be the number of units of
P1, P2, P3 to be produced, respectively.
From the given scenario, the objective is to maximize the profit. Therefore the
objective function is given by
Maximize Z = 3x + 3y + 3z

Table 2.1. Product mix.

Machines
Profit per unit
Product M1 M2

P1 7 8 $3
P2 8 12 $4
P3 5 7 $2
Available m/c time (h week−1) 40 35

2-3
 Modern Optimization Methods for Science, Engineering and Technology

Subject to the constraints (machines)

7x + 8y + 5z ⩽ 40 (M1)
8x + 12y + 7z ⩽ 35 (M2)
x , y , z ⩾ 0.

2.2.2 Diet problem

This is another common and important problem in LPP. The goal of the diet
problem is to find the optimal combination of food products available such that the
daily requirement of the nutritional component is met with minimal cost. The diet
problem is given as: suppose there are k foods available say F1, F2, …, Fk in the
market that has l nutritional components say N1, N2, …, Nl. Let dj be the daily
requirement of Nj and Ci be the cost price per unit of the food Fi. Let mij be the
amount of nutrition Nj in food Fi.
Let xi be the unit of food Fi to be purchased to meet the objective. The objective
function is given by
k
Minimize ∑ ci xi .
i =1

Subject to the constraints

k
∑i=1mij xi ⩾ dj for j = 1, 2, … , l
xi ⩾ 0, ∀ i .

2.2.3 Transportation problem

A chain of stores in any city needs to establish the brand and the goodwill of the
customers. The daily problem in such a scenario is to identify the routes to transport
the commodities from the various warehouses to their retail stores. Suppose there
are m warehouses say W1, W2, …, Wm and n retail stores say R1, R2, …, Rn and the
unit cost of transporting the commodity from the warehouse Wi to the retail store Rj
is denoted by Cij. The commodities must be supplied as per the demand at the retail
store and the availability at the warehouse. If the sum of the availability/supply is
equal to the sum of the demand then it is called a balanced transportation problem.
In the case that the sum of the demand is not equal to the sum of supply, then the
transportation problem is called an unbalanced transportation problem. The
unbalanced transportation problem can be solved after converting it to a trans-
portation problem by adding a dummy row(s)/column(s) with cost 0 in each of its
cells. A balanced transportation problem and its entries are shown in table 2.2. The
transportation model is also represented as a complete bi-partite graph. The
transportation problem can be formulated as an LPP.

2-4
 Modern Optimization Methods for Science, Engineering and Technology

Table 2.2. Transportation cost table.

Warehouse/retail store R1 R2 R3 … … Rn Supply

W1 C11 C12 C13 … … C1n b1

W2 C21 C22 C23 … … C2n b2
W3 C31 C32 C33 … … C3n b3
… … … … … …
… … … … … …
Wm Cm1 Cm2 Cm3 … … Cmn bm
Demand a1 a2 a3 an

The objective function of the transportation problem is

m n
Minimize Z = ∑i=1∑ j=1Cijxij .

Subject to the constraints

n
∑ j=1Cijxij ⩽ bi for i = 1, 2, … , m (constraints due to supply)
m
∑i=1Cijxij ⩾ aj for j = 1, 2, … , n (constraints due to demand)
xij ⩾ 0, 1 ⩽ i ⩽ m , 1 ⩽ j ⩽ n .

Note that here the warehouses and retail stores are known as pure supply nodes and
pure consumption nodes. The transportation problem can be represented by a
complete bi-partite graph. In certain scenarios, some nodes can be supply as well as
consumption nodes, such models are called transshipment models and the nodes
which are supply and demand nodes are called transshipment nodes.

Example 2.3. A firm has three warehouses W1, W2 and W3 and four retail outlets
R1, R2, R3 and R4. The unit transportation cost of shifting from Wi to Rj and the
maximum possible supply and the demand requirements are given in the table 2.3.
The company wishes to minimize the transportation cost.

Solution: total demand is 5700 = total supply and hence the problem is balanced.
The corresponding LPP model is

Mininmize Z = 6x11 + 4x12 + 8x13 + 4x14 + 7x21 + 21x22 + 4x23

+ 6x24 + 10x31 + 4x32 + 9x32 + 7x34.

Subject to constraints:

2-5
 Modern Optimization Methods for Science, Engineering and Technology

Table 2.3. Transportation cost table.

R1 R2 R3 R4 Supply

W1 6 4 8 4 2000
W2 7 21 4 6 1200
W3 10 4 9 7 2500
Demand 1500 2000 1000 1200

Table 2.4. Result for the transportation cost problem.

R1 R2 R3 R4 Supply
W1 1300 0 0 700 2000 Leq 2000
W2 200 0 1000 0 1200 Leq 1200
W3 0 2000 0 500 2500 Leq 2500
1500 2000 1000 1200

Geq Geq Geq Geq

Demand 1500 2000 1000 1200 27 500

Supply constraints
6x11 + 4x12 + 8x13 + 4x14 ⩽ 2000
7x21 + 21x22 + 4x23 + 6x24 ⩽ 1200
10x31 + 4x32 + 9x32 + 7x34 ⩽ 2500.
Demand constraints
6x11 + 7x21 + 10x31 ⩾ 1500
4x12 + 21x22 + 4x32 ⩾ 2000
8x13 + 4x23 + 9x33 + 4x14 ⩾ 1000
4x14 + 6x24 + 7x34 ⩾ 1200
xij ⩾ 0; 1 ⩽ i ⩽ 3 & 1 ⩽ j ⩽ 4.

Using the Excel solver we obtain the optimal values of xij and the minimum
transportation cost and the same is presented in table 2.4.
The minimal transportation cost is 27 500.
The shipment details are:
• W1 to R1 is 1300 units and W1 to R4 is 700 units.
• W2 to R1 is 200 units and W2 to R3 is 1000 units.
• W3 to R2 is 2000 units and W3 to R4 is 500 units.

2-6
 Modern Optimization Methods for Science, Engineering and Technology

The variations of the transportation problem are:

1. Unbalanced transportation.
2. Maximization of the objective function.
3. Unacceptable routes.

Example 2.4. (Transshipment problem) A firm has two plants P1 and P2 which
produce switch gears; two carry and forward (C&F) agents (C1, C2 and C3) and
dealers (D1, D2 and D3). P1 and P2 produce 1500 switch gears and 2500 switch gears,
respectively, and the demands at the dealers are 1900, 1200 and 900, respectively. All
goods can reach the dealers through any of the two C&F agents. The unit
transportation costs between nodes are shown in the network diagram (figure 2.1).
The objective of the organization is to minimize the transportation cost.
Here P1 and P2 are pure supply nodes, C1, C2, D3 are transshipment nodes and
D1, D2 are pure demand nodes. Using the buffer, we will convert this problem into a
transportation problem.
Now we have nodes P1, P2, C1, C2, D3 as supply nodes and C1, C2, D1, D2, D3 are
demand nodes. The supply at the transshipment nodes equals the buffer, and
demand at the transshipment nodes equals demand plus buffer, where the buffer can
be total supply or total demand. The objective function is the same as in the
transportation problem. The transshipment cost table is presented in table 2.5.
The assignment problem is a special case of the transportation problem. There are
m jobs and n bidders who are available to complete the jobs. Assume that n ⩾ m or
vice versa. Every job is assigned to only one bidder and every bidder will do only one
job. Hence this can be modeled as square matrices and hence in both cases we add
dummy rows/columns as per the need and assign unity as the supply and demand for
both nodes. Let Cij be the cost quoted by the ith bidder for doing the jth job. While
formulating this scenario, as a linear programming model, change the inequality in
both constraints as equality and solve as usual.

Figure 2.1. Transshipment network model.

2-7
 Modern Optimization Methods for Science, Engineering and Technology

Table 2.5. Transshipment cost table.

Supply/demand nodes C1 C2 D1 D2 D3 Supply

P1 3 2 * * * 1500
P2 1 4 2500
C1 0 * 5 2 3 B
C2 6 0 4 3 8 B
D3 * * * 4 0 B
Demand B B 1900 1200 900 + B B = Buffer
Replace (*) with a sufficiently large value and solve as in the transportation problem.

Table 2.6. Assignment cost.

Jobs
Contractors
J1 J2 J3

Contractor C1 123 78 67
Contractor C2 158 65 75
Contractor C3 200 50 35
Contractor C4 100 30 40
Contractor C5 225 80 68

Example 2.5. Assignment problem.

Mandalay City Corporation called for tenders to do three jobs from five authorized
contractors. All eight contractors are eligible to perform all the jobs, but every
contractor will be assigned only one job and every job will be assigned to only one
contractor, as per the city corporation norms. Let Cij be the quote by the ith
contractor to perform the jth job. All jobs must be allocated with a minimal cost.
The quote submitted by the contractors is given in table 2.6.

To convert the assignment problem to a transportation problem, put unity as the

demand and supply, then solve as in the transportation problem (table 2.7).
Clearly this problem is a unbalanced problem, so add two dummy columns as J4
and J5 with 0 as values in each cell of J4 and J5 columns, to make the problem a
balanced one.
The optimal cost is 188 and J1 is assigned to contractor C1, J2 is assigned to
contractor C4 and J3 is assigned to contractor C3.
The solution obtained using the Excel solver is given in table 2.8.

2.2.4 Portfolio optimization

The mundane job of the fund manager is to identify the various instruments to invest
in, in the right proportions as per the needs of the investor/firm. The objective may
be minimizing the risk or maximizing the return on investment.

2-8
 Modern Optimization Methods for Science, Engineering and Technology

Table 2.7. Assignment problem as a transportation problem.

Jobs
Contractors Supply
J1 J2 J3 J4 J5

Contractor C1 123 78 67 0 0 1
Contractor C2 158 65 75 0 0 1
Contractor C3 200 50 35 0 0 1
Contractor C4 100 30 40 0 0 1
Contractor C5 225 80 68 0 0 1
Demand 1 1 1 1 1

Table 2.8. Answer table for example 2.5.

Contractors Jobs Supply

J1 J2 J3 J4 J5
Contractor C1 10 0 0 0 1 Eq 1
Contractor C2 00 0 1 0 1 Eq 1
Contractor C3 00 1 0 0 1 Eq 1
Contractor C4 01 0 0 0 1 Eq 1
Contractor C5 00 0 0 1 1 Eq 1
Demand 11 1 1 1 188
Eq Eq Eq Eq Eq
1 1 1 1 1

An investment management firm assigned one of its fund managers to select a

portfolio to invest $1 000 000 as per the firm’s policy. The firm’s policy states that all
new investments must be made in the information technology sector, insurance
sector and banking sector in the coming quarters. The diligence report of the firm
identified the investment opportunities in certain companies and the expected annual
rate of return as given in table 2.9.
Moreover, the restriction based on the diligence report is as follows:
• The investment in any industry sector should not exceed 50% of the funds
available.
• Investment made in the banking industry must be at least 20% of the total
investment in the IT sector.
• Investment in Info Systems will yield a high return but the risk is high and
hence the investment in Info Systems must be 55% of the total investment in
the insurance sector. Find the proportion of the amount invested in each
sector which would yield the maximum return.

2-9
 Modern Optimization Methods for Science, Engineering and Technology

Table 2.9. Predicted annual return.

# Firm Rate of return (%)

1 Info Systems 6
2 BTS 11
3 Advanced Systems 9
4 Pearless Insurance 5.50
5 Loyal Insurance 6
6 Commercial Bank 4.30
7 Agro Bank 4.10

Let us formulate the given scenario as LPP. Let I, B, A, P, L, CB and AB be the

decision variables to represent the proportions of the amount to be invested in the
respective firms.
The objective function is

Maximize Z = 0.11I + 0.06B + 0.09A + 0.055P + 0.06L + 0.043CB + 0.041AB .

Constraints:
I + B + A + P + L + CB + AB = 1000000
I + B + A ⩽ 500000
CB + AB ⩽ 500000
P + L ⩽ 500000
CB + AB ⩽ 0.2(I + B + A)
I ⩽ 0.55(CB + AB )
I , B, A , P , L , CB , AB ⩾ 0.

2.3 The simplex method

This method was the first to obtain the solution of the system of linear inequalities in
a large scale and efficient method. We deploy the Excel add-in Solver to solve the
given LPP in this chapter. The fundamental idea of this method is to start with a
basic feasible solution and obtain the optimal basic feasible solution by moving from
one boundary point to other adjacent boundary points. In a nut-shell, the simplex
method can be described as ‘moving from feasibility to optimality’.
The various steps in the simplex method (in the case of the maximization problem) are:
1. Convert into the standard form (converting all constraints into ⩽ type).
2. Introduce the slack variables (to change inequalities into equalities).
3. Set up the initial table (with the objective function as Z – rhs = 0).
4. Check for the optimality (in the z-row all values are ⩾ 0).

2-10
 Modern Optimization Methods for Science, Engineering and Technology

5. Identify the entering/pivot variable (the variable that has the most negative
value in the z-row, the denominator should be positive).
6. Identify the leaving variable (among the ratios between the solution column
values and the entering variable column values, choose the minimal one).
7. Create the new table as in the Gauss–Jordon method for the next iteration
8. If not, repeat step 4–6 until optimality is reached.

Let us consider an example of an LPP.

Example 2.6. Max Z = 3x + 2y + z .

Subject to

2x + 4y − 3z ⩽ 25
− x + 2y + 2z ⩽ 35
x , y , z ⩾ 0.

Using the Excel Solver add-in, we solve this problem. Tables 2.10 and 2.11 show the
entry, the solution obtained and the answer for the problem.
From the tables, the optimal values for x = 155, y = 0 and z = 95 and the
maximum value is 560. In this problem, both constraints are of a binding nature,
that is the obtained solution of the decision variables satisfies the constraints and
hence the solution obtained is the optimal basic feasible solution. The solution
report generated by the Excel Solver is given in table 2.12.

Table 2.10. The LPP for example 2.6.

x y z
Coefficients of the
objective function 3 2 1
Optimal values of D
variables RHS
Constraint 1 2 4 −3 0 Leq 25
Constraint 2 −1 2 2 0 Leq 35

Table 2.11. Solution.

x y z
Coefficients of the
objective function 3 2 1
Optimal values of D
variables 155 0 95 560 RHS
Constraint 1 2 4 −3 25 Leq 25
Constraint 2 −1 2 2 35 Leq 35

2-11
 Modern Optimization Methods for Science, Engineering and Technology

Table 2.12. Answer report for example 2.6.

Objective Cell (Max)

Cell Name Original Final Value

Value

$H$3 Optimal values 0 560

of D variables

Variable Cells

Cell Name Original Final Value Integer

Value

$D$3 Optimal values 0 155 Contin

of D variables x
$E$3 Optimal values 0 0 Contin
of D variables y
$F$3 Optimal values 0 95 Contin
of D variables z

Problem-Constraints-Set

Cell Name of the Cell Formula-Type Status/Result Slack

Add Constraint Value—RHS

$G$4 Constraint 1 25 $G$4 <= $I$4 Binding 0

$G$5 Constraint 2 35 $G$5 <= $I$5 Binding 0

2.4 Artificial variable techniques

There are two methods that fall into this category, namely the big-M method/
Charne’s penalty method and the two-phase method. This section deals only with
the two-phase method. As the name suggests, there are two phases to solve the given
LPP, the initial basic feasible solution is found in phase 1; if such an initial basic
feasible solution exists, the optimal basic feasible solution is obtained in phase 2.
Steps involved in the two-phase simplex method:
1. Convert the given problem into standard form by adding slack/surplus
variables to the set of constraints.
2. For the constraints of =⩾ and = type, add an artificial variable, in addition to
the surplus variables (artificial variables play the role of slack variables).
Phase 1: Ignore the original objective function.
3. Set the new objective function so as to minimize the sum of artificial
variables and express the basic variables in terms of non-basic variables
subject to the given constraints and solve as in the simplex method.
4. Consider the final iteration of phase 1 and start with phase 2.

2-12
 Modern Optimization Methods for Science, Engineering and Technology

5. In phase 2 start with the original objective function subject to the constraints
in the final iteration of phase 1, without the artificial variables column and
solve as usual to obtain the optimal basic feasible solution.

Example 2.7. Obtain the optimal basic feasible solution for the LPP using the two-
phase method.
Minimize Z = 4x + y ,
subject to
3x + y = 3
4x + 3y ⩾ 6
x + 2y ⩽ 4
x , y ⩾ 0.

Phase 1
Minimize w = A1 + A2 ,
subject to the constraints
3x + y + A1 = 3
4x + 3y − S1 + A2 = 6
x + 2y + S2 = 4
x , y , S1, S2, A1, A2 ⩾ 0.

The initial table is shown below.

We have to modify the objective function that is expressed by the basic variables
A1, A2 and S2 in terms of non-basic variables x , y, S1.
The new objective function is w = 7x + 4y − S1 − 9. The initial table is presented
in table 2.13.
Solve as usual to obtain the optimal solution of phase 1 and the final iteration
table is presented in table 2.14.

Phase 2
Minimize Z = 4x + y,
subject to the constraints

Table 2.13. Initial table.

Basic variables x y S1 S2 A1 A2 Solutions

W 7 4 0 0 0 0 9
A1 3 1 0 0 1 0 3
A2 4 3 −1 0 0 1 4
S2 1 2 0 1 0 0 4

2-13
 Modern Optimization Methods for Science, Engineering and Technology

Table 2.14. Final optimal table—phase 1.

B variables x y S1 S2 A1 A2 Solution

Z 0 0 1/5 0 1 1 18/5
x 1 0 1/5 0 3/5 1/5 3/5
y 0 1 −3/5 0 4/5 3/5 6/5
S2 0 0 1 1 1 1 1

Table 2.15. Initial table—phase 2.

B variables x y S1 S2 Solution

Z 0 0 1/5 0 18/5
X 1 0 1/5 0 3/5
Y 0 1 −3/5 0 6/5
S2 0 0 1 1 1

1 3
x + 0y + S1 + 0S2 =
5 5
3 3
0x + y − S1 + 0S2 =
5 5
0x + 0y + S1 + S2 = 1
x , y , S1, S2 ⩾ 0.
In the final iteration of phase 1, the basic variables identified are x, y and S2.
Again express the basic variables in terms of non-basic variables to obtain the
new objective function of phase 2. The initial table of phase 2 is shown in table 2.15.
On solving this, we obtain x = 2/5, y = 9/5 and the minimum value of Z is 17/5.

2.5 Duality
Each LPP is associated with another LPP called the dual linear problem, whereas
the original problem is called the primal problem. In this section, we will see the
conversion of primal to dual and dual to primal. If one of the primal or dual
problems has an optimal solution the other also has an optimal solution. One can
obtain the optimal values of a primal problem by solving the dual problem and
vice versa.
Steps involved in writing the dual of the given LPP:
1. Convert all constraints into ⩽ type, if the objective function is maximization;
otherwise convert all constraints into ⩾ type.
2. Convert all the constraints into the equalities type by adding slack/surplus
variables.

2-14
 Modern Optimization Methods for Science, Engineering and Technology

3. If the objective function of the primal is maximization then the objective

function of the dual is minimization and vice versa.
4. The number of dual variables is equal to the number of primal constraints.
5. The number of dual constraints is equal to the number of primal variables.
6. The coefficients of the objective function of the dual are the right-hand side
of the primal constraints.
7. Reverse the inequality of the primal to obtain the dual problem constraints
with dual variables.
8. The coefficient of the dual constraints are the primal variable coefficients
(column).

Note that dual of a dual is a primal.

Example 2.8. Write the dual of the given LPP.

Max Z = 2x + 3y + u ,
subject to
x + 2y − u ⩽ 15
2x + 3y + 15u ⩽ 50
x , y , u ⩾ 0.
The dual of the given problem is
Minimize W = 15r + 50s ,
subject to
r + 2s ⩾ 2
2r + 3s ⩾ 3
− r + 15s ⩾ 1
r , s ⩾ 0.
Here r and s are the dual variables.

2.6 Sensitivity analysis

The examination of the effect of marginal changes in the problem on the optimal
solution is known as sensitivity analysis. That is, how small changes in the
coefficients of the decision variables and/or right-hand side values of the constraints
will affect the optimal solution. Also, one could use sensitivity analysis to find the
range of optimality and a reduced cost for the coefficients of the objective function
by allowing a small change in the objective function coefficients. Similarly a small
change in the right-hand side of the constraint helps us to study how this change
affects the optimal solution. This is called a shadow price.

2-15
 Modern Optimization Methods for Science, Engineering and Technology

Example 2.9. Solve the following LPP and answer the following questions,
Maximize Z = 8x + 9y + 6z ,
subject to
5x + 4y + 2z ⩽ 120
x + 2y + z ⩽ 55
x + 2y + 3z ⩽ 30
x , y , z ⩾ 0.
1. Find the range of optimality of the coefficients of the objective function.
2. Find the shadow prices of all the constraints.
3. If the rhs of the first constraint is changed to 120, what will happen to the
optimal solution?

The obtained optimal solution from solving using Excel is given in table 2.16. The
optimal values of x = 21.67, y = 4.17 and z = 0. The maximum value of Z = 210.83.
The first and third constraints are binding, whereas the second one is not binding
and it indicates that there are 25 units of unused resources available.

The sensitivity report generated by the Excel Solver is given in table 2.17.

Table 2.16. Answer report for example 2.9.

Solver Engine—Excel Solver Options

Objective Cell (Max)

Cell Add Name Original Value Opt/Final Value

$G$3 0 210.83

Variable Cells

Cell Add Name Original Value Final Value Integer

$C$3 x 0 21.67 Contin

$D$3 y 0 4.17 Contin
$E$3 Z 0 0 Contin

Constraints—Set

Cell Name Cell Value Formula Status Slack

$F$4 125 $F$4 <= $H$4 Binding 0

$F$5 30 $F$5 <= $H$5 Not binding 25
$F$6 30 $F$6 <= $H$6 Binding 0

2-16
 Modern Optimization Methods for Science, Engineering and Technology

Table 2.17. Report for the sensitivity analysis.

Problem-Variable Cells

Final Reduced Objective Max. Allowable Max. Allowable

Cell Name Value Cost Coefficient Increase Decrease

$C$3 x 21.67 0 8 2.13 3.5

$D$3 y 4.17 0 9 7.00 1.31
$E$3 z 0 −2.83 6 2.83 1E+30

Constraints Set

Final Shadow Constraint Max. Allowable Max. Allowable

Cell Name Value Price R.H. Side Increase Decrease

$F$4 125 1.17 125 25 65

$F$5 30 0.00 55 1E+30 25
$F$6 30 2.17 30 25 5

The optimality range for C1 is 4.5 ⩽ C1 ⩽ 10.125. The optimality range for C2 is
7.69 ⩽ C2 ⩽ 16. The optimality range for C3 is C1 ⩽ 8.83 and no lower bound.
Within these ranges the optimal solution remains unchanged.
The shadow price for each constraint is given in the table and the range for
optimality is also presented. Any change in the rhs of the first constraint is within the
range, but the optimal value will be decreased by 5 × 1.17 = 5.85. This sensitivity
report helps the decision maker in a better way.

2.7 Network models

Any communication/transport network can be viewed as a graph G = (V, E), where
V is the set of nodes/vertices and E is the set of links/edges.

2.7.1 Shortest path problem

Let G be the given network. A link/edge from the i-node to the j-node is denoted by
(i , j ) ∈ E and Cij is the link cost/maximum capacity the link can hold/distance/
duration between i and j. Let xij be the amount of flow in the link (i , j ). We wish to
find the shortest distance from an s-node to a t-node. The LPP formulation of this
problem is given as:
• The objective function is given by mininmize Z = ∑(i, j ) ∈ E Cijxij , where xij can
assume either 1 if (i , j ) is on the shortest path, otherwise 0.
• The constraints are defined for each node and they must satisfy the
conservation of flow. That is total in-flow = total out-flow.

Example 2.10. Find the shortest path between nodes 1–4 in the network, the link
cost/distance between the nodes are given as: Cost(1,2) = 3; Cost(1,3) = 4; Cost
(2,3) = 2; Cost(2,4) = 6; Cost(3,4) = 4 (figure 2.2).

2-17
 Modern Optimization Methods for Science, Engineering and Technology

6
2
3 2
1 4
4 4

Figure 2.2. Network diagram.

Table 2.18. Optimal solution table.

x1 x1 x1 x1 x2 x2 x2 x2 x3 x3 x3 x3 x4 x4 x4 x4
1 2 3 4 1 2 3 4 1 2 3 4 1 2 3 4
3 4 2 6 4

0 0 1 0 0 0 0 0 0 0 0 1 0 0 0 0 8
E
1 1 1 q 1
E
1 −1 −1 0 q 0
E
1 1 −1 0 q 0
E
1 1 1 q 1

The corresponding LPP is given as:

Objective function is Minimize Z = 3x12 + 4x13 + 2x23 + 6x24 + 4x34.
Subject to the constraints:
For node-1: x12 + x13 = 1

For node-2: x12 = x23 + x24

For node-3: x13 + x23 = x34

For node-4: x24 + x34 = 1

xij ∈ {0, 1}, ∀ (i , j ) ∈ E (G ).

This formulated LPP can be solved in the routine manner. The optimal solution is
given in table 2.18. The shortest path from 1 to 4 is 1–3–4 and the minimal cost is 8.

2.8 Dual simplex method

The other method for solving the LPP is the dual simplex method. In this method,
we start with an optimal solution and move towards feasibility. Note that this
method fails to produce an optimal feasible solution for some problems.

2-18
 Modern Optimization Methods for Science, Engineering and Technology

Steps involved in the dual simplex method:

1. Convert the objective function of the given problem into minimization, if the
objective function is of maximization.
2. All constraints must be of ⩽ type, if not convert.
3. Set up the initial table. Identify the leaving variable by choosing the most
negative value in the solution column of the initial table and the correspond-
ing variable is the leaving variable.
4. Construct the ratios between z-row values and the leaving variable row. Note
that this ratio is constructed only for non-basic variables and the denomi-
nator must be negative.
5. Choose the minimal ratios and the corresponding variable (column head) is
the entering variable.
6. Construct the next table as in the simplex method for the next iteration.
7. Check for optimality, i.e. (a) all values in the z-row must be ⩽ 0 and (b) all
values in the solution column must be non-negative.

Note that if we cannot construct the ratio after identifying the leaving variable, that
is, all values in the leaving variable row are non-negative, then the problem has an
infeasible solution.

2.9 Software packages to solve LPP

LPP is intended for modeling real-world problems and hence one can find two types
of software packages, one type represents the modeling perspective and the other
type is solvers.
The most popular commercially available software packages (solvers) are OSL,
MATLAB, LINDO, CPLEX, LOOQ, EXCEL, etc. The popular ones from the
modeling perspective are GAMS, AMPL, MPL, Operation Research in NCS, etc.
The statistical package R also provides packages to solve LPP.

Further reading
[1] Dantzig G B 1963 Linear Programming and Extensions (Princeton, NJ: Princeton University
Press)
[2] Dantzig G B and Thapa M N 1997 Linear Programming (New York: Springer)
[3] Bazarra M S, Jarvis J J and Sharali H D 2005 Linear Programming and Network Flows
(New York: Wiley Blackwell)
[4] Hiller F and Liberman G J 2005 Introduction to Operations Research 8th edn (New York:
McGraw Hill)
[5] Winston W L and Albright S C 2007 Practical Management Science 3rd edn (Boston, MA:
Cengage South Western)
[6] Winston W L 1996 The teachers’ forum: management science for MBA at Indiana
University INFORMS J. Appl. Anal. 26 105–11
[7] Anderson D R, Sweeny D J, Williams T A, Camm J D and Cochran J J 2016 An Introduction
to Management Science: Quantitative Approaches to Decision Making 14th edn (Boston, MA:
Cengage Learning)

2-19
 Modern Optimization Methods for Science, Engineering and Technology

[8] Murthy K G, Kabadi S N and Chandrasekaran R 2000 Infeasibility analysis for linear
systems: a survey Arab. J. Sci. Eng. 25 3–18
[9] Greenbreg H J 1993 How to analyze the result of linear programs—part 1: preliminaries
INFORMS J. Appl. Anal. 23 56–67
[10] Greenbreg H J 1993 How to analyze the result of linear programs—part 2: price interpretation
INFORMS J. Appl. Anal. 23 97–114
[11] Greenbreg H J 1993 How to analyze the result of linear programs—part 3: infeasibility
diagnosis INFORMS J. Appl. Anal. 23 120–39

2-20
 IOP Publishing

Modern Optimization Methods for Science, Engineering and

Technology
G R Sinha

Chapter 3
Multivariable optimization methods for risk
assessment of the business processes of
manufacturing enterprises
Vladimir Gorbunov

The rapid rate of development of science and technology opens up broad prospects
for the development of innovative businesses. However, opening a new business is
associated with investment costs for the organization and development of this
business. Before you meet these costs, you should make sure that the business idea in
question is realistic. This can be achieved using special programs that allow you to
create a mathematical model of the business and evaluate its possible future
outcomes. However, business is associated with many random sources of data.
The business model must be able to work with random variables as input data. This
chapter discusses the possibility of forming such a model using the example of the
program ‘E-Project’. This program allows one, on the basis of the input source data,
to determine the cash flow of the project during the entire period of its implementa-
tion and business performance indicators, and to build charts that reflect the internal
and external data flows of the project. The program allows you to perform risk
accounting and calculation of business process parameters taking into account these
factors. The project manager has the opportunity to analyze the impact of anti-risk
measures on the model and decide on the feasibility of their implementation. The
decision is made on the basis of the project performance indicators, quantitative
assessment of risk factors and the value of anti-risk costs.

3.1 Introduction
Optimization is the mathematical problem of maximizing or minimizing some
function of one or more variables under constraints. Optimization is the basis of
economic analysis of production processes. Typically, the value to be optimized is

doi:10.1088/978-0-7503-2404-5ch3 3-1 ª IOP Publishing Ltd 2020

 Modern Optimization Methods for Science, Engineering and Technology

related to the performance of the project or facility in question. The optimized

version of the object should be evaluated by a quantitative measure—the criterion of
optimality. The criterion of optimality is a quantitative assessment of the optimized
quality of the object.
On the basis of the chosen criterion of optimality, a function is made of the target
that represents the dependence of the optimality criterion on the parameters that
affect its value. The types of optimality criterion or target criterion functions are
defined by a specific optimization task. Thus, the task optimization is reduced to
finding the extremum of the objective function.
The most general statement of the optimal problem is the expression of the
optimality criterion in the form of an economic indicator (profit, productivity, cost
of production, profitability). The criterion of optimality in the general case should
reflect the most significant aspects of the process, to yield a numeric measure and
easily explain the physical meaning.
When setting specific optimization problems, it is desirable to write the optimality
criterion in the form of an analytical expression.In the case where the random
disturbances are small and their impact on the object can be ignored, the optimality
criterion can be presented as a function of input (X), output (Y) and control (U)
parameters:

L = F (X1, X2, … , Xn, Y1, Y2, … , Yn, U1, U2, … , Un).

Since Y = F(U), it is possible to write for fixed X

L = F (U ).

In this case, any change in the values of the control parameters affects the change in
the value of L, since the control parameters are directly included in the expression of
the optimization criterion and thus change the output parameters of the process,
which depend on the control ones.
If the random perturbations are large enough and they must be taken into
account, then experimental and statistical methods should be used to obtain a model
of the object in the form of a function that is fair only for the studied local area, and
the optimality criterion will take the form:

L = F (X , U ).

In optimization problems there are simple and complex optimization criteria. An

optimality criterion is called simple if you want to define the extremum of the
objective function without setting conditions on any other values. Such criteria are
usually used in solving specific optimization problems (for example, determining
minimum fuel consumption, optimal product processing time, etc). If you know the
functional relationship between the target function and the control variable, you can
directly calculate the value of the target function for a fixed value of the control
variable.

3-2
 Modern Optimization Methods for Science, Engineering and Technology

If the objective function F(x1, x2, …, XP) has continuous partial derivatives in its
arguments, then putting the partial derivatives of F in x equal to zero and solving n
equations together,
(dF / dxi ) = 0, i = 1, 2, … , n ,
provides the value of the objective function.
One-parameter optimization provides a search of the extrema of the functions of
one variable. The limiting condition is the one-modality of the objective function on
the studied interval. The search for the minimum (or maximum) of the objective
function is performed by different methods, which differ in the speed of finding the
desired value. One such method is the dichotomy method. It uses the values of the
objective function when searching for its extremum.
For example, take the function F(x) (figure 3.1). It is necessary to find x ,
delivering the minimum (or maximum) of the function F(x) on the interval [a, b] with
a given accuracy ε , i.e. to find
x = arg min F (x ), x ∈ [a , b ].
At each step of the search process, divide the segment [a, b] in half, x = (a + b)/2—
the coordinate of the middle of the segment [a, b]. Calculate the value of the function
F(x) in the neighborhood ±ε of the computed point x:
F1 = F (x − ε ),
F2 = F (x + ε ).
Compare F1 and F2 and discard one of the halves of the segment [a, b] (figure 3.1).
When searching for a minimum:
• If F1 < F2, then discard the segment [x, b], then b = x.
• Otherwise, drop the segment [a, x], then a = x.

The division of the segment [a, b] continues until its length is less than the specified
accuracy ∣b − a∣ ⩽ ε . The algorithm of this method is presented in figure 3.2.
In the output, x is the coordinate of the point at which the function F(x) has a
minimum (or maximum), FM is the value of the function F(x) at that point.
The optimality criterion is called complex if it is necessary to set the extremum of
the objective function for several changing variables. The procedure for solving the
optimization problem usually includes additional restrictions on these changing

Figure 3.1. Determination of the minimum by dichotomy.

3-3
 Modern Optimization Methods for Science, Engineering and Technology

Figure 3.2. The scheme of the algorithm of the method of dichotomy.

parameters. In this case, the solution of the optimization problem is reduced to the
sequential execution of the following operations: drawing up a mathematical model
of the optimization object; selection of the optimality criterion and determination of
the objective function; the establishment of possible restrictions on the variables; and
the choice of the optimization method, which will find the extreme values of the
required quantities.
The methods of multidimensional optimization are usually distinguished by the
type of information they need in the process:
• Direct search methods (zero-order methods) that need only the values of the
target function.
• Gradient methods (first order methods) which additionally require first order
partial derivatives of the objective function.
• Newtonian methods (second order methods) that use second order partial
derivatives.

3-4
 Modern Optimization Methods for Science, Engineering and Technology

Figure 3.3. IDEF0 business model.

The recurrent scheme of most multivariate optimization methods is based on the

expression Xk + 1 = Xk + λksk. They differ from each other by the way of choosing
the step length λk and the search direction sk. The first is a scalar and the second is a
vector of unit length. Methods of multidimensional optimization are reduced, in one
form or another, to methods of one-dimensional optimization.
Modern production processes consist of a set of interrelated sub-processes, each
of which has its own limitations and optimization conditions. Optimization of such
processes is performed by several performance indicators in compliance with a set of
limiting factors. Special computer programs are used to analyze and optimize such
processes. For the features of multi-dimensional optimization of production business
processes taking into account risk factors, consider the example of using the
program ‘E-Project’.

3.2 A mathematical model of a business process

Risks in manufacturing enterprises are associated with the appearance of random
factors, which ultimately lead to changes in planned results. A financial plan allows
you to consider and take into account the manifestation of potential risks.
The financial plan of the project involves the construction of business models,
which show all the actions performed with the reflection of used material values,
labor costs and the duration of the execution of all intermediate operations. Special
modeling languages, e.g. the unified modeling language, are used to build such
models [1]. This language describes the simple components of the procedure of a
business process and reveals the relationship between the internal sub-processes and
external data streams. Figure 3.3 shows an example view model of the company’s
activities (IDEF0 methodology). This company is planning the development of a
training course to sell through an online shop.

3-5
 Modern Optimization Methods for Science, Engineering and Technology

The main parts of the methodology are given in the diagram. The diagram shows
the functions of the system in geometrical rectangles, and also the existing relations
between the functions and the external environment. The rectangles represent
specific processes, functions, works or tasks that have a purpose and lead to a
marked result. Arrows indicate the interaction processes between them and the
external environment.
The model includes three types of documents (graphical charts, glossary, text),
which refer to each other. Information about the system is displayed in the graphical
diagrams with blocks and arrows, and their connections. The blocks represent the
basic functions of the model elements. These functions can be broken down
(decomposed) into their component parts and presented in the form of more detailed
charts. The decomposition process continues until the subject is described at the level
of detail necessary to achieve the objectives of a specific project. The glossary is
created and maintained by a set of definitions, key words and explanations for each
element of the chart and describes the essence of each element. The text gives an
additional description of the operation of the system.
Often there are cases where some arrows do not make sense to continue
considering in the child diagrams below a certain level in the hierarchy, or vice
versa—individual blocks have no practical value above a certain level. On the other
hand, sometimes you need to get rid of the separate ‘conceptual’ arrows and stop
using them deeper than a certain point.
Standard IDEF0 introduces the concept of tunneling to solve such problems. The
designation ‘tunnel’ in the form of two parentheses around the beginning of the arrow
indicates that this arrow is not inherited from the parent functional block and appears
(from the ‘tunnel’) only on this chart. In turn, the same designation around the end of
the arrow in close proximity to the block receiver refers to the fact that the arrow is a
child of this block diagram, and if displayed will not be considered in the diagram.
The business process model allows financial models to determine the cost
characteristics of processes and their interaction in time. The financial plan of the
project should reflect all costs associated with its preparation, take into account the
cost of manufactured goods or services, and determine income from the sale of
goods or services. The projects differ in the time interval start-up of the business, its
development and completion. During the project the prices of goods, raw materials
and debt capital can change and, in financial terms, these changes must be taken into
account.

3.3 The market and specific risks, the features of their account
A financial plan is an important document containing detailed information about
cash flow for current operations as well as the investment and financing activities of
the enterprise. The manager may obtain information from this plan about:
• The sources of funds and directions of their use.
• The excess cash in the accounts of the enterprise. The extent to which the
enterprise extension is provided at the expense of their own and borrowed funds.
• Any additional loans.

3-6
 Modern Optimization Methods for Science, Engineering and Technology

When making a report on cash flows for previous periods the entrepreneur uses
the actual data available in the accounting records. The preparation of the forecast
for the coming period requires more detailed analysis of the current situation and the
trends of its change. Often entrepreneurs plan the distribution of funds for the best
and worst cases and also for the most real situation.
A small enterprise can take into account the probabilistic nature of the factors
influencing its activity. In this case, the calculation of the indicators for the future
period is not reduced to the calculation of the three variants of development. In
order to obtain correct results, we have to calculate a single process, which
mathematically defines the probability of achieving the possible result of the activity
of the enterprise depending on the identified characteristics of the involved
processes. Such forecasting will enable the entrepreneur to choose the right develop-
ment strategy for the enterprise, ensuring the timely payment of the accepted debt.
Values such as the price of the traded goods or services, the cost of materials or
components, labor costs, etc, are connected through mathematical dependencies
with the performance of the company. The variation of the initial indicators requires
reproducing the calculations. In this regard, there is a need for an automated means
of calculation for established procedures to instantly output when the source data
change.
To help developers speed up the preparation of the business plan and provide the
necessary information for the investor to know the level of quality of the design
documents, the entrepreneur uses specialized programs for financial planning.
Traditionally, such systems are developed versions of the document templates. The
most popular systems provide capabilities similar to the program Ехсеl, and all of their
value for the user lies in a well-chosen list of topics. Filling it out, he or she will obtain
a more or less acceptable financial plan. For example, the most popular system in this
group is the program for Business Plan Pro, with hundreds of thousands of users.
The ability of some of the financial planning programs to account for risk factors
is important because of the acceleration of the rates of economic development in
every industry. An example of such a program is ‘E-Project’ [2, 3]. The program has
the following features:
• Incorporates a probabilistic calculation of the business processes.
• Uses a widespread and reliable software package.
• The ability to navigate freely in the calculation methodology, adapted to
specific user requests by creating their own forms of source data and
calculation algorithms.
• The ability to enter data in the form of arbitrary shapes, and the results of
calculations in the form of required reports. There are good editors for the
formation of forms and reports.
• Performs an advanced analysis of the credit-worthiness of the project, i.e. the
dependence of the results of calculations and changes in loan terms.
• A probabilistic calculation of the financial risk of the project depending on
the probability characteristics of the source data.
• The results of the calculations are in the form of tables and diagrams.
• Has the ability to convert data into HTML format.

3-7
 Modern Optimization Methods for Science, Engineering and Technology

The software package ‘AE-Project’ consists of two related programs: a program

for the calculation of economic indicators, realized in the program Excel, and a
textual blank business plan that is implemented in Word. Using an input form in the
analysis module, one enters all of the source digital data for the project. It then
calculates the results and determines the course of economic processes associated
with the implementation of the project.
Working with a design file starts with the opening title page of the project, where
the author writes the project name, author name, chosen project start date, the
interval calculation of the project (month, quarter) and the duration of the
calculated period. In this form, you select the currency format in which to engage
in financial calculations (RUB, thousand RUB, USD, thousand USD, etc).
After filling in and confirming the passport data, the program displays a main
form, consisting of ten control buttons (figure 3.4).
The first eight control buttons provide data input on the project. The form of
‘Investment costs’ is a list of all activities that must be performed to implement the
project. For each of these events the data are entered on its cost (with options for
average, minimum and maximum values), data at the beginning of the activity and
its duration, the availability of the plants and equipment associated with the event,
and the rate of depreciation for each fixed asset group. The data input forms will be
used to determine the financial flows required to support the production of products
or the provision of services, to chart and plan the schedule of works for project
implementation, and for the computation of the aggregate depreciation at any point
in your project.
The main characteristics of the form ‘Constant expenses’ is the accounting of all
project costs not related to production or services. Such expenses include costs
associated with building maintenance, and payments for utilities, communications,
transportation, security, etc. The data entry form ‘Products’ is associated with the
data describing the income and expenses for each planned type of product or service.
The form ‘Production’ is the total amount spent on marketing research, it also

Figure 3.4. The main menu of the project.

3-8
 Modern Optimization Methods for Science, Engineering and Technology

answers the questions of how much and when the manufactured products listed in
the form ‘Products’ will be implemented. The sales in this form will be submitted as
the number of products, the characteristics of which are presented in the form
‘Products’.
The form ‘Finance’ is used to enter the initial data on the financial activities and
considers such indicators as equity, loans, repayment of loans, interest on loans,
grants and government funding, and the payment of dividends.
The input form ‘Taxes’ specifies the source data for the calculation of tax
planning. The procedure of payment and the amount of taxes depends on the legal
form of the enterprise, and the accepted the forms of accounting and tax reporting.
In this form, the authors of the draft introduce the adopted tax rate for the primary
taxable base indicators: payroll, income, imputed income, profit and property.
The form ‘Staff’ is used to determine the financial flows associated with staff
wages, receiving a set salary. This form indicates the period during which the
employee will work on the project, his or her position and set salary. You can plan
roles without names when you do not know the names of the professionals who will
be involved in the project.
To calculate the impact of risks the software uses the input form ‘Common
project risks’. All risks of the project can be divided into two categories: market and
special. Market risks form due to fluctuations in price parameters, i.e. the instability
of the market. These indicators can change from their average value both upwards
and downwards. Such risks must be characterized by the standard parameters for
random distributions (expectation, variance), which are determined by the entered
data and reviewed by the input forms. The second type (special risks) is associated
with the specific situation in the project, which can occur with some probability and
this will cause appropriate financial changes.
To assess the impact of special risks the following characteristics are considered:
the period of manifestation of the risk factor, the probability of a risk situation R
and the financial cost upon the occurrence of a risk situation Q. For the construction
of the types of risks one of the accepted categories should be allocated: technical,
organizational, financial, environmental, technological, etc. The program includes
measures to reduce the impact of risks. The effect of these events will change the
parameters P and Q on P1 and Q1. These events are recorded in the form of the
investment measures the cost of these activities Q2 and their execution time.
Table 3.1 reflects the approach to risk assessment and events. Boolean variable K
can take the value 0 or 1. When K = 1 the event risks are characterized by the
parameters of Р1Q1, if K = 0 the event fails and the risk is perceived with the
parameters PQ.
The introductory data in the form end with the formation of the financial model
of the project. The results for the generated models can be observed in the output
reports that are in accordance with the selected algorithm and the results are given as
tables (a button on the main menu ‘Reports’) or graphics (click main menu ‘Chart’).
Figures 3.5 and 3.6 show examples of output graphs.
Modern enterprises should take into account in their plans all tangible and
intangible factors that can affect their condition in the future. The result of such

3-9
 Modern Optimization Methods for Science, Engineering and Technology

Table 3.1. Risks and measures for compensation of financial risks.

No. Risk (finance) PQ P1Q1 Event Q2 K

1 Does not increase the return on 10 4 Prepare to present information 20 0

investment. and analytical materials for
the attraction of investments.
2 To continue the high production 10 15 To introduce an automation system 60 0
costs per unit of output. to reduce the cost of production.
3 Will not decrease the average 20 6 To automate the system of 30 1
collection time of accounts collection receivables.
receivable.
4 Will not increase the net income 60 12 Change contract. 30 1
of a separate contract.
5 Reduced sales growth. 50 15 To perform a range of promotion 10 1
to increase sales.
6 Reduced labor productivity. 30 20 To introduce a machine to the 60 0
assembly area.

Figure 3.5. Example of chart ‘The cash flow project’.

3-10
 Modern Optimization Methods for Science, Engineering and Technology

Figure 3.6. Example of chart ‘Cash flow distribution’.

planning is a balanced scorecard. This metric is different from financial planning

because it harmoniously takes into account financial and nonfinancial indicators
that affect the activities of the enterprise. The office of strategic planning is usually
done through risk management. In the formation of a balanced, strategic develop-
ment plan, it is recommended that all risk indicators are split into four groups.
Financial indicators occupy the main place in the system of strategic planning.
Nonfinancial indicators include data for the company’s customers and markets,
characteristics of internal business processes, internal staff development and the
growth potential of the company. These four areas of activity of the enterprise form
the basis of the system of strategic planning. For individual companies other specific
areas of activity can be dealt with, reflecting the peculiarities of their production. All
of the indicator systems for the strategic planning of the united cause-and-effect
relationships are such that changes in either direction indirectly or directly affect the
performance in other directions.
A component of the strategic plan related to customer service shows how the
organization looks in the eyes of customers, i.e. reflects the competitive capabilities
of the company. The component for work with customers is important for the
overall strategy of the organization because it clearly defines the queries with key
customers, choosing a market position and the dynamics of the client groups on
which it focuses. The assessment tool can serve in examining risk factors that
determine the state of the client component of the strategic map. Possible deviations
from the intended plans will be due to the risks of working with external clients. The
results of detection of risk factors for the category ‘Customers’ produced a set of
measures to reduce the impact of these factors.

3-11
 Modern Optimization Methods for Science, Engineering and Technology

A section of the strategic plan relating to the internal business processes necessary
to achieve planned results defines the main innovations. The events section is aimed
at the development of new products and improvement of the competitiveness of the
enterprise. The performance section characterizes the processes contributing to the
achievement of the efficiency of the enterprise, the receipt of planned financial
results, improving internal technology, production process and meeting the chang-
ing demands of buyers. Insufficient attention to the internal processes can lead to
activation associated with the risks of these processes.
The term development implies a structure that the organization must implement
to ensure their development and growth in the long term. To ensure long-term
success and prosperity requires a constant updating of processes taking into account
new technical developments. The successful development of the organization
involves its human resources, information systems and organizational procedures.
To support their activities in the market, the company should invest in staff
development, information technology, systems and procedures. Insufficient atten-
tion to these factors can lead to the emergence of risk factors associated with the
personnel of the company and the growth of the company.

3.4 Measurement of the risk of using the discount rate, expert

assessments and indicators of sensitivity
Using the discount rate is widespread to account for risk through the introduction of
special allowances for the risk-free discount rate. The formula for the discount rate is
T
CFt CF(T +1) 1
NVP = ∑ (1 + * , (3.1)
t=1
+ i )t i−g (1 + i )T

where NPV is the net present value of future cash flows, T is the number of settlement
periods within the planning horizon, CFt is cash flow over the t period, CF(T+1) is the
cash flow terminal of the first period, i is the value of the discount rate and g is the
growth rate of cash flow in the post-planning period (percent per annum) [4].
In the general case, calculating the NPV considers the cash flow generated by the
project lifecycle (during the planning horizon). Sometimes cash flow for the project
in the post-planned period is omitted in the calculation of the NPV to outside
investors (with the aim of increasing the reliability of the calculations). In formula
(3.1) the first term characterizes the cash flow from the project during the planning
horizon, the second does so in the post-planned period. Calculation of the cost cash
flow in the post-planned period is carried out using the Gordon model [5]. The
discount rate represents the average return that an investor received when investing
in the project alternative under consideration.
Calculation (optional) of the discount rate is determined based on:
• Ways of taking into account inflation when calculating cash flow.
• The project participant for which NPV is calculated.
• Information about project components.

3-12
 Modern Optimization Methods for Science, Engineering and Technology

The project over time inevitably changes the factors that determine the value of
the bet. For example, in the start-up phase the business may experience a permanent
decline in risk through risk reduction ‘non-realization of the project’. After the
payback period, the risk of investors associated with a possible ‘return’ of funds is
also reduced to zero. But at the same time, other changes can lead to an increased
exposure to risk factors and this is equivalent to an increase in the discount rate.
Therefore, the discount rate assumption is the basic assumption of the calculation of
equation (3.1)—usually the adoption of constant values of the discount rate for the
entire lifecycle of the constant bet throughout the project lifecycle in the preparation
of preliminary calculations.
The simplified model developed by the magazine Business Valuation Review [7]
for emerging markets may be used to build value models of the company and its
elements. This model is based on the assumption that the yield on government
Eurobonds reflects the risks associated with investing in the share capital of ‘ideal
companies’, i.e. companies with no flaws. The disadvantages of a real company are
equal to the risks specific to the company and the specific business. These flaws are
marked as premium to the discount rate due to these risk factors (see table 3.2).
For practical application of this model for risk management, it is necessary to
increase the risk-free rate in accordance with table 3.2. Each allowance should be
determined taking into account the possibility of its appearance in the project and
the consequences after the considered risk.
The method of expert estimations consists in the possibility of using the
experience of experts in the analysis process of the project and considering the
influence of diverse qualitative factors. The formal peer-review process often comes
down to the following. The management of the project (firm) develops the list of
evaluation criteria in the form of expert (polling) sheets containing the questions.
Corresponding weighting coefficients are assigned for each criterion (at least are
calculated), and are not disclosed to the experts. The answers are compiled for each
criterion, the weight of which is not known to the experts. The experts should have
information about the evaluation project and through examination should be able to
analyze the questions and mark the chosen answer. Next, the completed expert

Table 3.2. Estimates of allowances for ‘deviation of the ideal’ corporate risks.

View corporate risk Interval values

Quality control 0%–8%

Company size 0%–5%
Financial structure 0%–10%
Commodity/territorial diversification 0%–5%
Client diversification 0%–4%
Profits: rules and retrospective predictability 0%–5%
Other risks 0%–5%

3-13
 Modern Optimization Methods for Science, Engineering and Technology

sheets and the output or results of the examination are processed accordingly on the
basis of well-known computer packages for the processing of statistical information.
In practice, risk analysis and decision-making is often not worth the task of
obtaining quantitative characteristics. What is important is comparative analysis,
which experts use to assess the occurrence of risk events on a simplified scale of
gradations. For example, the place each of the events on the graph in the axes of
‘impact’ is ‘probably’. The diagram consists of nine cells, each of which corresponds
to a single set of estimates (figure 3.7). For example, an event characterized by the
estimated ‘low impact, low probability’ should be displayed in the lower left cell of
the chart, and event assessments ‘low impact, high likelihood’ should be displayed in
the lower right cell, etc.
The whole chart is divided into three approximately equal parts. The three cells of
the diagram located at the bottom left are an area of insignificant risk. The three
cells of the diagram in the upper right are an area of significant risk. The remaining
part of the diagram (three cells) is an area of medium risk. Thus, the risk associated
with event A is insignificant, the risk for event B is average and the risk of event C is
substantial. The resulting diagrams, which in accordance with expert assessments
apply to all risk events, are called risk maps. This map shows what risk events can
take place, what the correlation between different types of risks is and how risks
should be given maximum attention (in this example, risk events C). This approach
is widespread in the practice of risk management for companies in the real world.
Risk managers typically use three or five (rarely seven) grades for probability of
exposure and materiality. The described chart is a convenient way to visualize risk.
In practice, there are other ways of visualization, such as using a circular or a
color chart.
Sensitivity analysis and scenario analysis are the sequential steps in quantitative
risk analysis; the latter allows us to avoid some of the shortcomings of sensitivity
analysis. However, the scenario method is the most effective and can be applied
when the number of possible values for NPV of course. However, the expert faces an
unlimited number of different scenarios in the risk analysis of the investment project.
The actual method of assessing the individual risk of the project helps to solve this
problem (simulation); the basis of this method is the probabilistic assessment of the
occurrence of various circumstances. By using specialized software packages for the
calculation of the economic efficiency of projects, the evaluation of the impact of
risks is obtained in the form of output tables and graphs reflecting the impact of risk
factors on the project output.

The probability
Low Medium High
Strong Event C
The Impact Average Event A Event B
Weak

Figure 3.7. The risk map.

3-14
 Modern Optimization Methods for Science, Engineering and Technology

In the sensitivity analysis, analyze the sensitivity of the project to the main
parameters and modify one of the input parameters of the project. By sequentially
changing project parameters, the developer defines input variables that strongly
affect the project result. For these variables, measures are developed to reduce their
impact. The aim of the project sensitivity analysis is to determine the impact of
varying factors on the financial result of the project. The most common method used
for sensitivity analysis is simulation. NPV is used as the integrated indicator of the
financial result of the project. The result of sensitivity analysis can be reduced to
similar conclusions, e.g. the project allows the reduction of the price of sale by 16%,
an increased sales volume by 11% and increased direct costs by 14%.
Figure 3.8 shows a graph of the impact on the net present value, changes in sales
volume, sales price and direct costs.
The sensitivity analysis does not perform the measurement of risk, but only
assesses the influence of various factors on the results of the financial activities of the
enterprise. The analysis allows one to identify the factors that most affect the
amount of profit and then consider activities that control these factors to find ways
to organize activities to neutralize these factors.
Sensitivity analysis (determination of the critical point) can be produced without
the use of special computer programs based on the conditions of break-even
production. Consider this analysis as an example.
In the calculation of the predetermined period from the start of series production
the total costs of production (P) are defined as
P = V*M + F
where F and V are, respectively, fixed and variable costs.
Turnover (O) is defined as the product of the price of the product (C) and the
quantity (M)
O = C * M.

Figure 3.8. Graph of the evaluation of the sensitivity of the project to the sales, the cost of materials and
constant costs.

3-15
 Modern Optimization Methods for Science, Engineering and Technology

The break-even condition describes equality:

P = O or C * M = V * M + F .
Thus, the critical (break-even) volume of production is
M = F /(C − V ).
Let us analyze the break-even conditional production period. For example, the
company’s revenue in the conditional period is planned to be the amount of
14.04 monetary units. It can be assumed that it is planned to implement a
conditional piece of the 14.04 or 14.04 conditional sets of settings for the price of
1 monetary unit. The full current costs of this period are 12.26 currency units,
including fixed costs of 2.38 and variable costs of 9.88.Variable costs per unit:
V = 9.88/14.04 = 0.704.
The break-even threshold obtained from the break-even equation is
1 * x = 0.704 * x + 2.38 where x = 8.04.
Break-even is reached when there is a sales volume of 8.04 units of money.
Critical values for fixed costs are calculated as follows: 14.04 = 0.704 * 14.04 + F,
where F = 4.16. Fixed costs can only increase up to 4.16 and can be 42% higher than
planned.
The critical value of the variable costs is determined from the equations of
momentum and costs, and is calculated as V * 14.04 + 2.38, where V = 0.84. Thus,
variable costs per unit of output may increase to 0.84, i.e. the safety margin is 46% of
that originally planned.
A critical value of the sales price P is determined from the conditions of equality,
profit is zero. The price of conventional products (or the price of conventional sets) is
defined as 1 unit of money. The price level at which profit is zero is defined by the
equation of momentum and costs:
14.04 * Pr = 0.704 * 14.04 + 2.38; Pr = 0.87.
This means that the selling price can only be 13% lower than planned if other
conditions are constant.
The risk-category represents a probability, so in the process of assessing
uncertainty and quantifying risk it is possible to use probabilistic calculations.
The most important indicator for the measure of the financial risk of an enterprise is
the level and probability of losses. This figure has a decisive impact on the level of
profitability of the financial operations of the company. These two indicators are
closely related and constitute a single system of ‘profitability–risk’. The ratio of the
level of risk and return is one of the main underlying concepts of financial risk
management, in accordance with which the level of profitability of financial
operations under other equal conditions is always accompanied by an increase in
the level of their risk and vice versa. In addition, the level of financial risk is a main
indicator in assessing the level of financial security of an enterprise, which character-
izes the degree of protection of its financial activities from external and internal
threats. Therefore, the assessment of the level of risk in the management of financial

3-16
 Modern Optimization Methods for Science, Engineering and Technology

activities of the company is included in the preparation of almost all managerial

decisions.
The level of financial risk characterizes the probability of its occurrence under the
influence of a particular risk factor (or groups of factors) and possible financial loss
upon the occurrence of a risk event. Concrete methodological tools for evaluation of
risk level formed according this definition allow one to solve specific tasks of the
financial management of the enterprise. Consider the key design and performance
indicators.
The level of financial risk describes a common algorithm for estimating this level.
It is represented by the following formula:
Ur = Pl * Fl,
where Ur is the appropriate level of financial risk, Pl is the probability of financial
risk and Fl is the size of possible financial losses when implementing this risk.
In practice this algorithm calculates the size of possible financial losses usually
expressed by an absolute amount and the probability of occurrence of financial risk—
one of the coefficients is the measurement of this probability (the coefficient of
variation, beta coefficient, etc). Accordingly, the level of financial risk when it is
calculated according to the following algorithm will be expressed in an absolute
measure, which greatly simplifies the comparison of project alternatives.
The variance characterizes the degree of variability of the studied indicator (in
this case the expected income from the realization of financial operations) with
respect to its average value. The more fluctuations—the greater the risk. The
variance is calculated according to the following formula:
n
σ2 = ∑(Ri − R¯ )2 * Pi ,
i=1

where σ is the variance, Ri is the specific value of the possible variants of the
2

expected income for the financial transactions, R̄ is the average expected value of
income under financial transactions, Pi is the potential frequency (probability) of
obtaining separate variants of the expected income on financial transactions and n is
the number of observations.
Dispersion does not give a complete picture of deviations ΔX = X − R¯ , which are
more pronounced for the risk evaluation. However, dispersion allows establishing a
link between linear and quadratic deviations using the well-known Chebyshev
inequality.
The probability that a random variable X deviates from its expectation by more
than a given tolerance ε > 0 does not exceed its variance divided by ε2, i.e.
D
P ( ∣X − R ∣ > ε ) ⩽ .
ε2
This shows that a small risk of dispersion deviation corresponds to a small risk,
and according to a linear deviation of point X are likely to be within ε, the
neighborhood of the expected values.

3-17
 Modern Optimization Methods for Science, Engineering and Technology

The root mean square (standard) deviation is one of the most common in
assessing the level of individual financial risk as the variable determines the degree of
absolute variability and is calculated by the following formula:
n
σ= ∑(Ri − R¯ )2 * Pi .
i=1

The standard deviation σ is a dimensional quantity and is specified in the same

units in any measured variable characteristic. The advantage of standard deviation is
that when you reach the observed distribution (e.g. distribution of investment
income) to normal this parameter can be used to determine the boundaries in which,
with a given probability, one should expect the value of a random variable.
The coefficient of variation (CV) lets you define the level of risk if the average
expected income from financial operations differ [6]. The calculation of the
coefficient of variation is carried out according to the following formula:
σ
CV = ± * 100%.
R¯
The coefficient of variation is a dimensionless quantity. With it, you can even
compare the variability of traits, expressed in different units of measurement. The
coefficient of variation varies from 0% to 100%. The greater the ratio, the greater the
variability. For different values of the coefficient of variation are established the
following qualitative assessments: 10% is weak variability, 10%–25% is moderate
variability and over 25% is high variability.
The beta coefficient (β) allows you to evaluate individual or portfolio systematic
financial risk in relation to the risk level of the financial market as a whole. Usually,
this indicator is used to assess the risk of investing in individual securities and is
computed by the formula
K × σi
β= ,
σp
where β is the beta coefficient, K is the degree of correlation between the level of
profitability on individual types of securities (or their portfolio) and the average level
of profitability of the group equity instruments in the market as a whole, σi is the
standard deviation of return on the individual securities (or on their portfolio as a
whole), and σр is the standard deviation of return on the stock market as a whole.
The level of financial risk of individual securities is based on the following values of
beta coefficients: β = 1 is average; β > 1 is high level; and β < 1 is low level.
The risk not only of a particular transaction but also the business of the company
as a whole can be assessed using the probabilistic evaluation method (analyzing the
dynamics of its income) for a certain period of time. The choice of specific
assessment methods is determined by the availability of the necessary information
base and the skill level of management personnel.
A new methodology has been used and developed in the last decade for evaluating
the measures of financial risk with the indicator ‘cost of risk’ or ‘value at risk’

3-18
 Modern Optimization Methods for Science, Engineering and Technology

(VAR). The value at risk is a measure of statistical estimation expressed in monetary

form of the largest possible size of financial losses in the prescribed form of the
probability distribution of the factors influencing the value of the assets (tools) and a
given level of the probability of occurrence of these losses over the estimated period
of time.
From the above definition, it is clear that the methodology of the calculation of
VAR consists of three main elements. One of these elements is to establish the
characteristics of the original factors. The risk manager establishes the form of the
probability distribution of the risk factors affecting the value of the assets (tools) or
their total portfolio. It can be the normal distribution, Laplace distribution, t-test,
etc. One of the qualities of VAR is to set the risk-manager level of probability that
the maximum possible size of financial losses will not exceed the estimated value of
this indicator. The specific value of VAR is chosen by the risk manager based on his/
her risk preferences. In the modern practice of financial risk management this level is
usually in the range of 90%–99%. A visual representation of the formation of VAR
gives the plot shown in figure 3.6.
As can be seen from the graph in figure 3.9, the revenue curve illustrates the
normal probability profit distribution on the financial instrument in a predetermined
billing period of time. The field in this graph between −2σ and +3σ corresponds to
the chosen confidence level (90% of the area under the curve), and between −3σ and
−2σ characterizes the value of possible losses beyond the confidence level (10%). On
the chart, the VAR determined is the amount of −732.6 thousand rubles. This
corresponds to a maximum size of possible financial losses on the financial instru-
ment under the given confidence level and estimated valuation period; the value of
VAR in the diagram separates the value of income beyond the limits of the
confidence interval (10%).

Figure 3.9. Graphical method of determining the value indicator VAR.

3-19
 Modern Optimization Methods for Science, Engineering and Technology

In order to fully describe the risk using the VAR measure, you must first specify the
probability (small enough to consider the event ‘almost’ impossible), or the level of
confidence associated with this probability value. If the probability is set as 5%, it
means a confidence level of 95% (100%–5%) and represents the result in the form
VAR95% (pronounced ‘VAR at the 95% level’). The level of 95% is rather arbitrary,
each individual sets this level based on their relationship to the possible unlikely events,
and an understanding of what is considered an ‘almost’ impossible event. Therefore,
VAR can be used with other levels of confidence, e.g. 90% or 99% (when talking about
VAR90% or VAR99%). In addition, the assessment or the calculation of VAR is in
practice used in the time horizon of the game (financial operations). Therefore, when
speaking about risk, VAR determines what is the minimum financial result that can be
obtained with a certain confidence level during a certain period of time.
Here is an example. The statement ‘evaluation of the VAR of the risk of lower
returns during the next week is minus 2% at a confidence level of 95%’ or briefly ‘a
week VAR95% = −2%’ means that:
• With a probability of 95%, the yield of the planned operation will be at
least −2% for the week.
• With a probability of 95%, the loss for the week will not exceed 2%.
• A weekly loss of over 2% is possible with a probability of 5%.

There is a strong correlation between the two measures of risk—variance and

VAR in the case of normal distribution. Since the normal distribution is completely
determined by two parameters M and σ, any part of this distribution (in particular,
any quantile) is determined by these two parameters. This means that for the normal
probability distribution the relationship between the variance and the VAR at any
confidence level is unambiguous and has the form:
VARi = M [X ] − Z (1 − i ),
where Z (1 − i) is the quantile of order (1 − i) standard normal distribution.
The values of the tabulated quantiles, we present several important special cases:
VAR90% = M [X ] − 1.283 * σ ;
VAR95% = M [X ] − 1.645 * σ ; (3.2)
VAR99% = M [X ] − 2.326 * σ .

These formulas are of practical importance. In the vast majority of cases the
probability distribution of the results of economic games is not known. However, it
is often possible to estimate some characteristics of the unknown distribution, in
particular, the expected outcome and variance. Then you can make the assumption
that unknown to us the distribution is very similar to normal, and we can estimate
VAR using equation (3.2). This assumption is close to the truth for games in the
financial markets, as the prices of many important assets are determined by many
random factors that are often inconsistent and conflicting. Even if the probability
distribution of the results of each of these random factors is not normal, their joint
distribution will tend to normal.

3-20
 Modern Optimization Methods for Science, Engineering and Technology

Analytical models of business processes require a description of the procedures

used and the task in the analytical form of the input process parameters. The
complexity of this approach lies in the analytical conversion of the input
parameters in accordance with the ongoing business processes. By adopting some
limitations on input actions, it is possible to simplify the analytical analysis. For
example, in the E-Project [2] the parameters of the random input signals used only
two characteristics: M—mathematical expectation and D—standard deviation.
Operations with such variables are performed according to the rules of operations
with random variables [3]:
• For any two random variables and the expectation of their sum M (X + Y)
M (X + Y ) = M (X ) + M (Y ). (3.3)
The variance of the sum of two random variables is equal to the sum of their
variances plus twice the correlation time Kxy:
D(X + Y ) = D(X ) + D(Y ) + 2Kxy . (3.4)
• The mathematical expectation of the product of two random variables is
equal to the product of their mathematical expectations plus the correlation
time Kxy:
M (XY ) = M (X ) * M (Y ) + Kxy . (3.5)
• The variance of independent random variables
D(XY ) = D(X ) * D(Y ) + M (X )2 * D(Y ) + M (Y )2 * D(X ). (3.6)

Most of the results of typical business processes are built on the combination of
the operations of addition, subtraction and multiplication that can be defined by
equations (3.3–3.6).
In the E-Project package the normal distribution of the output is accepted. This
assumption is possible in cases when the business process involved a large number of
random variables, and the final result is a complex combination of these input
effects. It is possible to rate the difference between the results of modeling the
business process using the analytical model and the method of simulation. For
example, for changing the law Triang input data (figure 3.10), the distribution of the
output can be represented by a normal distribution (figure 3.11) [7, 8].
From this figure, it is seen that model business processes at various distributions
of the input parameters allow the output process to describe the distribution of the
normal law. This example shows a normal distribution and Welbull distribution.
The standard deviation of all the distributions differs by a negligible amount. An
approximation of these distributions to normal will be more accurate the more
processes will affect the output parameters. The output distribution final results are
built according to the dependences of equations (3.2)–(3.5) for non-symmetric
distribution input data and will have less asymmetry than the input distribution.
Figure 3.12 presents the simulation result of a mixture of input data, distributed
over six different laws. The result of the comparison of the obtained distribution

3-21
 Modern Optimization Methods for Science, Engineering and Technology

Figure 3.10. The distribution of the input parameter ‘Sales’.

Figure 3.11. The distribution of the output parameter of the business process for Triang distribution
input data.

shows that a normal distribution has a matching mathematical expectation and

standard deviation with the result from the simulation values of the distribution of
the output parameter. The simulation results show that the model of the business
projects can be based on analytical relations (3.3–3.6), with simplified analytical
calculations and process simulations.

3-22
 Modern Optimization Methods for Science, Engineering and Technology

Figure 3.12. The distribution of the outputs of the business process while increasing aggregate sales from
various distributions of input data.

The accuracy of the analytic calculation output parameters for business processes
depends on the accuracy of determining the same statistical characteristics of the
input parameters.
An important step of risk management is to reduce the cost of risks and anti-risk
action when you run anti-risk measures to modify the characteristics of risk factors.
Optimization of risks can be reduced to evaluating the cost of anti-risk measures and
changing impact after the occurrence of the risk factors resulting from the
application of anti-risk measures. With a limited budget, you cannot perform all
anti-risk events. We select those that have the greatest effect. We can account for the
feasibility of carrying out certain anti-risk activities based on an integrated assess-
ment of the financial results of the project with different combinations of parameter
K (table 3.1). E-Project uses a special add-in that enables the adopted criteria and
conditions of restrictions to choose the optimal value of parameter K for all the risks
involved.
First, activities are planned with the greatest efficiency [9]. Figure 3.13 shows the
dependence of financial loss from the implementation of the 24 anti-risk measures
and the costs of implementation of these activities.
Various algorithms can be proposed to automate the selection of anti-risk
measures. For example, the set of events fits within the allocated budget or choice
of activities, the effectiveness of which exceeds the threshold. The result of this
optimization is presented in figure 3.14, with the financial losses from the original
risk and the same rate for a project with risks.
The proposed method allows us to assess the impact of risk factors on the
efficiency of the project, to assess the impact on financial performance anti-risk
activities and to choose those which provide the greatest effect according to the
chosen criterion of project evaluation.

3-23
 Modern Optimization Methods for Science, Engineering and Technology

Figure 3.13. Losses from risks and the cost of measures from them.

Figure 3.14. Losses from risk source and risk after special events.

3.5 Conclusion
The business process model based on risk allows you to more accurately plan the
development of a business. It allows the opportunity to develop activities that improve
the final result of the project with minimal cost before the start of the project.

References
[1] Sommerville I 2001 Software Engineering 6th edn (Reading, MA: Addison-Wesley), 693 p
[2] Gorbunov V L 2013 Business Planning with Risk and Efficiency Assessment of Projects:
Scientific-Practical Manual (Moskow: RIOR: Infra-M.) p 248
3-24
 Modern Optimization Methods for Science, Engineering and Technology

[3] Gorbunov V L 2004 A database ‘AE-Project’ certificate of registration No. 2004629261

[4] Mikhailets V B 2002 The discount rate in the evaluation Russian Business Newspaper issue 90
pp 147–50
[5] Gregori A 2003 Strategic Valuation of Companies (Practical guide) (M. Kvinto-Consulting)
[6] Gracheva M V 2001 Risk Analysis of the Investment Project (Moskow: UNINI-DANA)
[7] Nersesian R L 2013 Energy Risk Modeling (New York: Palisade)
[8] Morris J R and Daley J P 2009 Introduction to Financial Models for Management and
Planning Textbook (London: Chapman and Hall)
[9] Gorbunov V L 2014 Accounting balanced scorecard of the enterprise in the system of business
planning Proc. of the Int. Conf. Innovative Approaches to Solving Technical and Economic
Problems (Moscow: MIEE)

3-25
 IOP Publishing

Modern Optimization Methods for Science, Engineering and

Technology
G R Sinha

Chapter 4
Nonlinear optimization methods—overview and
future scope
Somesh Kumar Dewangan, Siddharth Choubey, Jyotiprakash Patra
and Abha Choubey

The technology for combinatorial optimization is changing quickly, and complexity

of the fundamental technology is increasing with size expansion of extent of
problems. The nonlinear programming problem related to the discrete lower bound
farthest point examination problem is treated using methods for calculation where
the need to linearize the yield criteria is avoided. The calculation is an interior-point
strategy and is totally open, as no specific component discretization or yield basis is
required. We classify current programming packages for processing the necessary
nonlinear optimization problems. The packages include interior-point methods,
progressive linear/quadratic programming methods, and increased Lagrangian
methods. For each package the standard methodological components are discussed.
The new methods are connected to the single-class client traffic equilibrium problem,
the multi-class client traffic equilibrium problem under social marginal cost
evaluation, and the stochastic transportation problem. In a constrained arrangement
of computational tests the calculations end up being very efficient. Moreover, a
feasible strategy with a multi-dimensional search for the stochastic transportation
problem is created.
The traffic task problem is a nonlinear model which depicts how every traveler
limits his/her own movement costs to achieve the optimal solution. In the admin-
istration of venture portfolios, the objective is to choose a combination of
speculations in order to enhance returns while limiting risks. For understanding
non-convex or large-scale optimization problems, deterministic methods may not be
suitable for producing globally optimal results in a reasonable time due to the high
complexity of the problems. Heuristic methods are therefore introduced to reduce
the computational time for an optimization problem, however, the result obtained is

doi:10.1088/978-0-7503-2404-5ch4 4-1 ª IOP Publishing Ltd 2020

 Modern Optimization Methods for Science, Engineering and Technology

not necessarily a possible or globally optimal solution. The two types of optimiza-
tion methods have their advantages and disadvantages. Therefore, combining
deterministic and heuristic techniques is suggested for handling large-scale optimi-
zation problems to find a global optimum. This chapter attempts to provide a future
research direction in the area of deterministic and heuristic methods to improve the
computational effectiveness of finding a globally optimum answer for various real-
world application issues. As we introduce new ideas into model design, model
examination, model changes, global restriction handling and hybridization, the
advantages of this new optimization method can be used without expert knowledge.
With increased complex optimization problems, hybrid optimization calculations
and programming language concepts become essential.

4.1 Introduction
This section discusses an overview of optimization, non-linear programming (NLP)
and a few case studies.

4.1.1 Optimization
Optimization can be related to existing or specifically built scientific models. The
idea is that one might want to locate an extremum result of a model by shifting a few
parameters or factors [1]. The typical motivation for discovering suitable parameter
values is used in selection choices in design optimization.
Optimization is the process of achieving the best outcome under given conditions.
In design, development, support, etc, engineers need to make choices. The objective
of every such choice is either to limit costs or to increase advantages [2].
The costs and advantages can normally be communicated as a component of
certain design factors. Thus, optimization is the process of finding the conditions
that can give the maximum or the minimum estimate of value based on the suitable
conditions. It is also commonly established as a hidden guideline in the investigation
of numerous complex choice or allocation problems. Utilizing optimization theory,
one solves a complex choice problem, including the choice of qualities for a number
of interrelated factors, by concentrating on a single target function, designed to
evaluate the execution and measure the nature of the choice. This one function is
maximized (or minimized, depending upon the details) subject to criteria that may
confine one of the choice variable qualities. In this case a reasonable single
component of a problem can be isolated and described by a function, be it a benefit
or risk in a business setting, or speed or distance in a physical problem [3].
No single strategy is available to handle all optimization problems efficiently.
Thus, various methods are used to address different kinds of problems.
Linear and nonlinear optimization methods address the following problem:
finding numerical values for a given set of variables with the feasible goal satisfying
certain criteria, called target functions. The solution with target function achieves its
minimum value among all the combinations of possible variables. An example is the

4-2
 Modern Optimization Methods for Science, Engineering and Technology

dietary regimen problem, for finding a combination of foods which fulfills the health
requirements at minimum expense. In contrast to traditional problems of connected
arithmetic, the majority of which start in materials science, linear and nonlinear
optimization problems for the most part need solutions given by closed formulae,
and must be understood through numerical methods, with calculations performed
on PCs [4].
In numerous numerical programming applications, linear optimization assump-
tions or approximations may allow proper depiction of the problem over the range
of variables being considered. In nonlinear optimization, nonlinear constraints
f (x1, x2 , x3, … , xn)
of the decision variable are used. If the possible solution space is bounded by
nonlinear constraints then the method used to find possible solution is called non-
linear programming (NLP).

Statement of an optimization problem

Current literature suggests that no advances were made until the twentieth century in
the field of optimization, when computing power made the execution of optimiza-
tion strategies conceivable and thus invigorated further research [5].
Significant improvements in the field of numerical strategies for unconstrained
optimization are reported [1, 6–11]. These include the advancement of the simplex
method [12], the principle of optimality [13], and the necessary and sufficient conditions
of optimality [14]. Optimization in its broadest sense can be considered to tackle any
design problem, for example:
• Minimum weight of flying machines.
• Optimal (minimum time) directions for space missions.
• Minimum load of electric networks.
• Optimal production management, resource allocation and scheduling.
• Finding shortest routes.
• Optimum pipeline networks.
• Minimum handling time in production frameworks.
• Optimal control.

An optimization, or a scientific programming problem, can be described as

x = (x1, x2 , x3, … , xn)
which reduces to f(x), subject to the limitations

for k = 1, … , m , gk(x ) ⩽ 0
for k = 1, … , p , lk(x ) = 0.

A design vector can be defined through variable x known as f(x) with objective
function gk (x ).

4-3
 Modern Optimization Methods for Science, Engineering and Technology

4.1.2 NLP
NLP is similar to linear optimization due to the fact that it has a goal, general
requirements and variable constraints. The thing that matters is that a nonlinear
program incorporates in any event one nonlinear function, which could be the target
function or a few or the majority of the constraints.
In nonlinear optimization, models are intrinsically significantly more difficult to
improve. There are some reasons behind this, which are briefly described as follows:
1. In numerical methods for tackling nonlinear optimization we have
restricted data about the problem. This implies that it is difficult to
recognize a local optimum from a global optimum. In such a case, we
have accessible data, we expect that one point x is itself the estimation of
the target function at x. There is sufficient data to determine when you are
near the minimum or maximum, yet there is no opportunity to determine
whether there exists an alternative and better local maximum [10, 11].
2. In a linear optimization system there is a set number of points to search for
the optimum solution in a linear program, by checking optimum points, or
corner points, of the feasible polytope space. In contrast, in a nonlinear
optimization system an optimum solution could be in any place—at an
optimum point, along an edge of the feasible space, or in the interior of the
feasible space [15].
3. In the event that nonlinear constraints are considered, there might be
various diverse feasible regions, regardless of whether you can locate the
optimum inside a specific feasible space.
4. In nonlinear problems, specific initial conditions may produce diverse final
solutions, where there might be various distinctive minima, and beginning
at some other point may produce an alternative final solution point and
target function value.
5. In a linear optimization system one will either discover a point that fulfills
every one of the requirements or it will specifically confirm that no feasible
points exist anywhere. However, in nonlinear procedures there is no such
certainty.
6. In linear optimization methods, the initial segment routine finds a solution
that fulfills every one of the constraints and from there on these are never
broken, however, in a nonlinear optimization strategy, finding a solution
that fulfills the folding and curving conditions is difficult in itself, regardless
of whether a solution is found sooner or later; the uniformity may again be
violated when the calculation attempts to move to another point that has a
superior estimation of the solution [16].
7. In a linear optimization strategy, we have only a few possible results: a
globally optimum solution point, feasible yet unbounded, and infeasible.
8. There is a huge number of complex numerical hypotheses and various
solution calculations. The reason why NLP is so much more difficult
than LP is due to its being dependent on possible nonlinear capabilities.

4-4
 Modern Optimization Methods for Science, Engineering and Technology

There are also considerably more reasons why NLP is difficult, which
depend on increasingly practical considerations.
9. It is difficult to decide the suitable conditions that result in obtaining the
best possible solution.
10. Various calculations converge at certain conditions and provide the
optimum solution.
11. Different conditions used by different users may produce distinctive
solution.

In nonlinear optimization, a few things which were considered difficult in the past
can now be generally achieved effectively:
• Derivatives. The derivative functions are regularly required by solvers (those
involved in attempting optimum solution). There were two methods: numer-
ical estimate by limited differences, or a modeling function which can provide
code to the subsidiary.
• Input groups. At a certain point every solver had its own specific information
organization to depict the model and if we found that one solver was not
successful for the current problem it was a monotonous and mistake-laden
process to recode the model into an alternative configuration and thus it could
be attempted by different solver [17, 18].

4.1.3 Nonlinear optimization problem and models

This section highlights a few models of the non-linear optimization problem [18, 19].

Quadratic optimization models

This model suggests a linear structure

Ax = b ,

which is over-selected (larger number of conditions than variables), and we can

obtain least squares arrangement by dealing with the NLO problem [1]

min Ax − b 2 (quadratic optimiztion problem).

Further,

min∥Αx − b∥2 = (Ax − b)T (Ax − b) = xT AT Ax − 2bTAx + bTb .

The concrete mixing problem can be considered as an example. The important

characteristics for the strength of concrete are its sand and rock structure and these
are used in particular proportion for an optimum design and architecture. For every
type of concrete architects and engineers could determine the ideal combination of
these characteristics.

4-5
 Modern Optimization Methods for Science, Engineering and Technology

Scientific model
This model suggests a plan that operates on different vectors v = (v1, v2, v3, … , vn )
T, where 0 ⩽ vi ⩽ 1.

Portfolio examination model

In this model, we define different variables, arrangement preferences and distinction
of hard and soft elements [20]. Numerically, xi is the degree placed into the major
mean-fluctuation problem:

min { 1
2
x ˆTV (x ): f ˆTx = ʎ , Dx = d , x ⩾ 0 . }
4.2 Convex analysis
Convex analysis is used in cases where the nonlinear optimization problem utilizes
curved objective function [15].

4.2.1 Sets and functions

As per the principles of sets we have two situations, x1 and x2, in IRn [21].

Explanation 4.1. Considering two points x1, x2, IRn we obtain

x = λx1 + (1 − λ)x 2 ,
where the two points x1, x2 result in a convex combination.
In this scenario, C ⊂ IRn is convex if any two points x1, x 2 ∈ C and if every raised
combination of x1, x 2 ∈ C .
In other ways, we can say that the line partition interfacing two sensitive
motivations behind a convex set is contained in the set [22]. A parabola
f (x ) = ax 2 + bx + c with a > 0 can also be defined by the same function.

Explanation 4.2. This type of function f : C → R defined on a convex set C is called

convex if for all x1, x2 ∈ C and 0 ⩽ λ ⩽ 1 one satisfies
f (λx1 + (1 + λ)x 2 ) ⩽ λf (x1) + (1 − λ)f (x 2 ).

Explanation 4.3. The function of f : C → R, where C ⊂ IRn is defined as an n + 1-

dimensional set:
{(x , T ): f (x ) ⩽ T , x ∈ C, T ϵ IR}.

Explanation 4.4. f : C → IR is also denoted as a one group value C becomes the

convex function for all x1, x 2 ∈ C and for 0 < λ < 1 we can write

f (λx1 + (1 + λ)x 2 ) < λf (x1) + (1 − λ)f (x 2 ).

4-6
 Modern Optimization Methods for Science, Engineering and Technology

Explanation 4.5. A function f : C → IR , characterized by the upper raised set is

called concave if the capability profile of the function is curved.

4.2.2 Convex cone

Explanation 4.6. The set C ⊂ IRn is defined as a convex cone for all x ∈ C and 0 ⩽ λ
satisfying λx ∈ C .
A convex cone does not contain any subspace with the exception of the function.

4.2.3 Concave function

Concave functions are essentially the opposite of convex functions.

Explanation 4.7. A function f(x) is called concave if, for every y and z and for 0 ⩽ λ
⩽ 1,
f ∣[λγ + (1 − λ)z ].

4.2.4 Nonlinear optimization: the interior-point approach

This section discusses logarithmic boundary method for dealing with with curved
optimization. We will consider the CO issue in the following equation,
(CO ) min{f ((x )): x ∈ F},

where F means the feasible space, which is given by

F ≔ {x ∈ IRn: gj (x ) ⩽ 0, 1 ⩽ j ⩽ m}.

The constraint function gj:IRn → IR(1 ⩽ j ⩽ m ) and the objective function

f : IRn → IR are convex limits [23].
When managing calculations for solving CO functions f and gj (l ⩽ j ⩽ m )
are taken into consideration and it is also assumed that f(x) is linear, for example
f(x) = −cTxf is defined as c ∈ c ∈ IRn . In case this is not valid, one may present an
extra factor xn+1, an extra requirement f (x ) − xn+1 ⩽ 0f (x ) − xn+1 ⩽ 0 and a limit
xn+1 and then objective function converges as linear.
We can write
⎧ min −cT x
⎪
(CPO)⎨ gj (x ) ⩽ 0, j = 1, … , m .
⎪
⎩ x ∈ IRn

The Lagrange–Wolfe dual of CPO is expressed as

4-7
 Modern Optimization Methods for Science, Engineering and Technology

⎧ max −cT x + ∑m y g (x )
⎪
⎪ m
j=1 j j
⎨∑ j = 1yj ∇gj (x ) = c .
⎪
⎪
⎩ yj ⩾ 0, j = 1, … , m

The primary feasible region is

F 0 :={x ∈ IRn: gj (x ) < 0, j = 1, … , m}.

This will satisfy the IPC if F 0 is nonempty [24].

4.2.4.1 Genetic algorithms

Genetic algorithms (GAs) are based on natural selection, both constrained and
unconstrained types of optimization. Biological evolution is considered as a basis for
GA to work as an optimization method [3].

4.2.4.2 Simulated annealing

Simulated annealing (SA) is an effective and general form of optimization. It is
useful in finding global optima in the presence of large numbers of local optima.
‘Annealing’ refers to an analogy with thermodynamics, specifically with the way that
metals cool and anneal.
The ‘finding that every specie has obtained is exemplified’ [25, 26].

Genetic algorithm procedural steps are:

1. Start with different variables s and q.
2. Set initial value of q as 0.
3. Evaluate the value of function D(q).
4. With the help of incremental operators q ++, take the function from D(q) to
D(q−1).
5. Combine the functions of (D(q)).
6. Return to step 3 and repeat.

4.2.4.3 Scatter search

This strategy joins favored subsets of reference centers to create a fundamental point
by weighted linear combinations, and selects the best members to transform into the
source of the new reference point. After introduction and assessment, the dissipate/
tabu search calculation characterizes the population of arrangement X1……Xpop_size
into a few sets [27].

Procedural steps are:

1. Start with different variable s and q.
2. Set initial value of q as 0.
3. Calculate and classify the value of function D(i).
4. With the help of incremental operators q ++, create the function E(q).
5. Calculate the function E(q).

4-8
 Modern Optimization Methods for Science, Engineering and Technology

6. Take the function from D(q) to D(q−1) and E(q).

7. Calculate the function (D(q)).
8. Return to step 4.

This was derived from true mechanics for discovering least cost answers for immense
upgrade issues [7]. It aggregates up the hillclimbing systems and discards their
standard damage: dependence of the course of action on the starting stage, and
accurately certifications to pass on a perfect plan. This is cultivated by displaying a
probability ρ of affirmation ρ = 1.

The procedure is described as [28]:

1. Start with 0 to q.
2. Define the temperature T.
3. Arbitrarily define the variable Xc .
4. Calculate Xc .
5. Go to step 2.
6. Go to step 3.
7. Create a new variable Xn with Xc . If f (Xc ) < f (Xn ). Then Xc ⤎Xn .
8. Temperature T is assigned to function D(T, q).
9. q is incremented by 1 and repeat the steps from Step 2.

4.2.4.4 Evolutionary strategies

Evolutionary strategies are methods used for parameter optimization problems.
These are advances of development techniques having many members in the field of
optimization problem solving approaches. This can be written as [29]
(μ + λ) − ESs and (μ, λ) − ESs .

Evolutionary procedure uses the following steps [30]:

1. The initial value of q is 0.
2. Evaluate the value of function D(q).
3. Start the loop with condition of endless.
4. With the help of incremental operators q + +, take the function from D(q) to
D(q−1).
5. Combine the function of (D(q)).
6. Go to step 3.

4.2.4.5 Sequential unconstrained minimization technique (SUMT)

Reducing a nonlinear problem to linear constraints is not difficult when the
requirements are nonlinear. The main reason is that it is difficult to move along
the limit of a nonlinearly constrained region, while it is generally easy to move along
the limit of a linearly constrained region. Rosen’s anticipated inclination technique
of nonlinear programming with linear constraints provides a decent methodology
for moving along the limit of linearly constrained regions [6].

4-9
 Modern Optimization Methods for Science, Engineering and Technology

It is advantageous to compose the nonlinear programming problem with non-

linear imbalance and uniformity requirements in the following way. Choose q to
minimize f(q) subject to

gi (q ) ⩾ 0 the value of l is denoted as 1 to l

,
gi (q ) = 0 the value of l is denoted as 1 to m

where point z is such that gi (z ) > 0 i = 1, … , l .

Define the function (called the P function) [9]
i m
1
P(q , r1) = f (q ) + r1∑ + r1−1/2 ∑ gi 2(q ),
g
i=1 i
( q ) i=i+1

where r1 is a positive number.

As a starting point determine z n such that gi (z n ) > 0 i = 1, … , l , proceed from
n
z to a point of z(r1) that approximates the minimum of P(z, r1) in the set of points [15],
i m
1
P(q , r2 ) ≡ f (q ) + r2∑ + r2−1/2 ∑ gi 2(q ),
g
i=1 i
( q ) i=i+1

i.e. r1 > r2 > 0 starting from z(r1) approximates the minimum of P(z, r2).

4.3 Applications of nonlinear optimizations techniques

Here, we will discuss a few applications of non-linear optimization techniques.

4.3.1 LOQO: an interior-point code for NLP

Interior-point systems are a well-known augmentation to nonlinear programming
compared to the ordinary expansion of the simplex methodology [31]. The
methodologies perform very well when interior-point systems are used as proce-
dures for compelled curved nonlinear optimization that minimizes f(x)
subject to b ⩽ q(x ) ⩽ b + s ,
where variables b, h and r take values in m and l, z, and u take values in m.

4.3.2 Digital audio filter

In digital audio filter applications, the presence of noise and other similar signals is
analyzed and their effect is minimized by optimization methods [32]. Hence, the
condition of the tweeters is noteworthy yet the woofer can be put anywhere within
hearing range. For sound systems it is not even essential to have two woofers as the
low repeat signal can be joined into one. The speakers would then have the
alternative to be proposed to work ideally in a smaller range of frequencies [2].
With computerized measurement one has significantly more control over how the
split is accomplished than can be achieved with a basic signal.

4-10
 Modern Optimization Methods for Science, Engineering and Technology

4.4 Future research scope

The best local and global search heuristics structures in this theory are mainly based
on calculations utilizing different physical and numerical parameters. These param-
eters center around a few calculations for randomization and creating neighborhood
arrangements. The basic aspects of actualizing these calculations efficiently and
successfully depend on exploiting huge databases and on maintaining suitable potential
solutions for updating possible accessible moves. Similarly to interior-point techniques
for direct programming, the number of cycles is affected only a little by the problem
estimate. The new techniques are connected to the single-class client traffic balance
problem, the multi-class client traffic harmony problem under socially negligible cost
estimation, and the stochastic transportation problem. In a constrained arrangement of
computational tests the calculations end up being very effective. Moreover, an
achievable direct technique with a multi-dimensional search is created for the stochastic
transportation problem.
In the last three decades, the field of linear programming has grown rapidly due to
the pioneering contributions of the active researchers in the field. However, as a tool
for multi-objective decision analysis, the field of fuzzy target programming is
relatively new in the area of fuzzy multi-objective decision making. Interval
programming is even newer than fuzzy programming, so there remains a huge
open area for new research and development. There is also wide scope for new
research in different application areas for the development of the field of stochastic
programming. The implementation of genetic algorithm methods to various multi-
objective decision making problems is still at its initial stage. There is a vast scope for
future research in this field for the development of a new generation of information
technology applications. The potential which the genetic algorithm techniques offers
over existing techniques is enormous.

References
[1] Klafszky E 1976 Geometric Programming, Seminar Notes 11.1976 (Budapest: Hungarian
Committee for Systems Analysis)
[2] Moore C 1993 Braids in classical gravity Phys. Rev. Lett. 70 3675–9
[3] Chenciner A and Montgomery R 2000 A remarkable periodic solution of the threebody
problem in the case of equal masses Ann. Math 152 881–901
[4] Coleman J O 1998 Systematic mapping of quadratic constraints on embedded fir filters to
linear matrix inequalities Proc. of 1998 Conf. on Information Sciences and Systems
[5] Spergel D N 2000 A new pupil for detecting extrasolar planets, arXiv: astro-ph/0101142
[6] Ho J K 1975 Optimal design of multi-stage structures: a nested decomposition approach
Comput. Struct. 5 249–55
[7] Karmarkar N K 1984 A new polynomial–time algorithm for linear programming
Combinatorica 4 373–95
[8] Anstreicher K M 1990 A standard form variant, and safeguarded linesearch, for the modified
Karmarkar algorithm Math. Program. 47 337–51

4-11
 Modern Optimization Methods for Science, Engineering and Technology

[9] Jarre F 1990 Interior-point methods for classes of convex programs Technical report SOL 90-
16 Systems Optimization Laboratory, Department of Operations Research, Stanford
University, CA
[10] Han C-G, Pardalos P M and Ye Y 1991 On interior-point algorithms for some entropy
optimization problems Working paper Computer Science Department, Pennsylvania State
University, University Park, PA
[11] Kortanek K O and No H 1992 A second order affine scaling algorithm for the geometric
programming dual with logarithmic barrier Optimization 23 501–7
[12] Dantzig G B 1987 Origins of the simplex method Technical report SOL 87–5 https://apps.
dtic.mil/dtic/tr/fulltext/u2/a182708.pdf
[13] Lin Y 2002 Bellman’s principle of optimality and its generalizations General Systems Theory:
A Mathematical Approach (Kluwer), pp 135–61
[14] http://web.math.ku.dk/~moller/undervisning/MASO2010/notes/LKKT.pdf
[15] Vanderbei R J 1999 LOQO user’s manual—version 3.10 Optimiz. Methods Soft. 12 485–514
[16] National Electrical Manufacturers Association (NEMA) 2018 Volt/VAR optimization
improves grid efficiency Technical report http://assets.fiercemarkets.net/public/sites/energy/
reports/voltvarreport.pdf
[17] den Hertog D, Roos C and Terlaky T 1993 The linear complementarity problem, sufficient
matrices and the criss–cross method Linear Algebra Appl. 187 1–14
[18] Jarre F 1994 Interior-point methods via self-concordance or relative Lipschitz condition
(Würzburg: Habilitationsschrift)
[19] Lustig I J, Marsten R E and Shanno D F 1994 Interior point methods for linear
programming: computational state of the art Oper. Res. Soc. Am. J. Comput. 6 1–14
[20] Baxter J and Bartlett P L 2001 Infinite-horizon policy-gradient estimation J. Artif. Intell.
Res. 15 319–50
[21] Coleman J O and Scholnik D P 1999 Design of nonlinear-phase FIR filters with second-
order cone programming Proc. of 1999 Midwest Sym. on Circuits and Systems
[22] Bertsekas D P 1995 Nonlinear Programming (Belmont, MA: Athena Scientific)
[23] Jarre F 1996 Interior-point methods for convex programming ed T Terlaky Interior-Point
Methods for Mathematical Programming (Dordrecht: Kluwer), pp 255–96
[24] Chenciner A, Gerver J, Montgomery R and Simó C 2002 Simple choreographic motions of
N bodies: a preliminary study Geometry, Mechanics, and Dynamics ed P Newton, P Holmes
and A Weinstein (New York: Springer)
[25] Broucke R 2003 New orbits for the n-body problem Proc. of Conf. on New Trends in
Astrodynamics and Applications
[26] Vanderbei R J, Spergel D N and Kasdin N J 2003 Circularly symmetric apodization via star-
shaped masks Astrophys. J. 599 686–94
[27] Vanderbei R J, Spergel D N and Kasdin N J 2003 Spiderweb masks for high contrast
imaging Astrophys. J. 590 593–603
[28] Kasdin N J, Vanderbei R J, Spergel D N and Littman M G 2003 Extrasolar planet finding
via optimal apodized and shaped pupil coronagraphs Astrophys. J. 582 1147–61
[29] Duan Y, Chen X, Houthooft R, Schulman J and Abbeel P 2016 Benchmarking
deep reinforcement learning for continuous control Int. Conf. on Machine Learning (ICML)
pp 1329–38
[30] Sen P K and Lee K 2014 Conservation voltage reduction technique: an application guideline
for smarter grid IEEE Trans. Indus. 52 2122–8

4-12
 Modern Optimization Methods for Science, Engineering and Technology

[31] Lobo M S, Vandenberghe L, Boyd S and Lebret H 1998 Applications of second-order cone
programming Technical report Electrical Engineering Department, Stanford University
http://rutcor.rutgers.edu/~alizadeh/CLASSES/12fallSDP/Papers/socp.pdf
[32] Bendsøe M P, Ben-Tal A and Zowe J 1994 Optimization methods for truss geometry and
topology design Struct. Optimiz. 7 141–59

4-13
 IOP Publishing

Modern Optimization Methods for Science, Engineering and

Technology
G R Sinha

Chapter 5
Implementing the traveling salesman
problem using a modified ant colony
optimization algorithm
Zar Chi Su Su Hlaing, G R Sinha and Myo Khaing

In this chapter, the main modifications induced by ant colony optimization (ACO)
are presented. In fact, the traveling salesman problem (TSP) could easily be solved
by using an improved version of the ant colony optimization method. The same is
attempted in this chapter. There are two main ideas in the proposed algorithm for
the modification of the ant algorithm. The first phase involves defining the candidate
set which is applied to the construction of the ant algorithm. The solution
construction phase includes defining the value of exploitation parameter q0. The
second phase focuses on the variation of pheromone information that is used to
adapt the heuristic parameter automatically throughout the algorithm runs.
Additionally, a local search algorithm is applied to all the solutions that are
produced by ACO and thus the performance of ACO is improved for TSP related
applications.

5.1 ACO and candidate list

ACO is a very popular and much researched area in the metaheuristic domain. It
improves the convergence speed and escaping from local optimal solutions under the
condition of guaranteeing solution quality. The ACO metaheuristics are population
based constructive metaheuristics. At each step of the algorithm, each ant first
searches for an element (the next city for the TSP) to add it to the ant’s solution.
When the size of the problem is increased in any application, then the difficulty
increases and thus ACO is said to have this problem as bottleneck. It is also difficult
when applied to large-sized problems in the solution construction process. As ACO
is a constructive metaheuristic in nature, the ants do consider the whole set of

doi:10.1088/978-0-7503-2404-5ch5 5-1 ª IOP Publishing Ltd 2020

 Modern Optimization Methods for Science, Engineering and Technology

elements which are possible at each and every step, and then the current solution is
added just before choosing one element. It can therefore be seen that in most
algorithms utilizing the ant concept, the evaluation of utilities is done based on
reachable elements, otherwise the techniques employing ACO can perform adversely
due to the large value of the runtime, because if the construction of appropriate sub-
sets is not done appropriately then this problem is likely to occur.
At the solution construction phase, there are possible problems that cause slow
convergence due to the fact that the ants start scanning the set of all possible state
elements (cities) before choosing a particular one and the probability is very small to
visit the same state for most ants. Then, the computational time can be large for each
scanning step of the algorithm. When the TSP size is large scale, such a situation can
occur. This is suggested even though we are aware that the system performance can
be improved as well as the efficiency in a number of real world applications of ACO
and when large TSPs are taken into consideration. So, the major problems of ACO
include slow convergence or speed, long runtimes, long execution times, trapped
local optima, etc.
To consider the above mentioned TSP, there is a limited number of elements
(cities), called the candidate list (CL), which can reduce the scanning set of possible
elements. The CL is a limited set of possible neighborhood elements from the current
element. CLs are created statically or dynamically using available knowledge of the
problems. Using a CL, ACO algorithms can reduce the search space of problems in
the solution construction phase.
The local search procedure employs CL strategies that help to improve the ACO
generated solutions and outcomes. However, the implementation of optimization
has its own constraints in strategies related to 2-Opt and 3-Opt local search
heuristics, because the use of the construction phase in ACO becomes not very
appropriate. These strategies aim to improve the ant colony system (ACS) so that
candidate set based strategies can successfully be applied in the construction phase.
Now, there are two important approaches which are used for solving general
purpose problems in each of the elements with a strong relationship between the
elements. These approaches are the Delaunay graph candidate set approach and
nearest neighbor approach, which provide relationships between elements such as in
the TSP. A static candidate set is used in all such strategies as an a priori candidate
which is derived and does not receive any updates or modifications while being
applied in the process of the ACO metaheuristic. On the other hand, throughout the
search process, dynamic candidate set strategies are required.

5.2 Description of candidate lists

A CL restricts the number of possible choices that need to be considered at each
construction step. In the case of the TSP, the CL is chosen for any city i and contains
a fixed number nn of cities having the smallest distance to city i. These cities are
called nearest neighbors. In order to determine them, all cities are sorted by their
distance to city i in increasing order and the first nn cities are inserted in city i’s CL.
For the tour construction rule (equation (5.2) or, respectively, equation (5.1)) the

5-2
 Modern Optimization Methods for Science, Engineering and Technology

biased search is performed only for cities in the CL. If all cities from this list have
already been visited, one of the remaining cities is chosen.

5.3 Reasons for the tuning parameter

There are many disadvantages in the basic form of ant system (AS) implementation,
a few of them are trapping, premature convergence and stagnation in the behavior of
the system. If all ants start producing similar or equal solutions again and again then
the behavior is said to be stagnated or in stagnation. This demands new ways or
modified and adaptable paths, which is a very difficult task but can result in a
situation of obtaining the best possible local solution. This may arise when a large
amount of pheromone trail on the edges of solutions is seen and some advanced
phases of search process are also observed, which actually happens as a consequence
of improper tuning of ACO parameters such as α, β, ρ.
The ACS has this problem of stagnation which can be observed if the parameters
are not tuned properly. The individual candidates with the highest and lowest values
for q0 affect the construction phase of ACO, which is not very surprising, but the
individuals try to exploit the values. Ranks which are low valued are far from
pheromone information and these ranks with low values affect the algorithm. If the
value goes down too low the search becomes unsupervised and erratic even though
the pheromone trail is fine. On the other hand, a high value of rank improves the
solution and helps in obtaining the local best possible option which is not based on a
probability concept but is obtained on the basis of a high value of rank. As an effect,
short tours are created rather than long tours and therefore the candidates
supporting such exploration are considered as less fit and consequently their
corresponding genes start vanishing. Sometimes, short tours result which usually
support the exploration and the solution becomes better, but the problem of
premature convergence still remains to some extent.
The problem of stagnation and premature convergence can easily be handled and
avoided using a random number q that is generated and compared with pheromone
trails on edges. For high values of trails τ, the random number is smaller and vice
versa. So, the algorithm does not succeed if a suitable value of trail is not chosen,
which means that it should not be too high, otherwise the selection of the best path
and point cannot be made. If the value of the trail is very high, then the algorithm
tends to edges and therefore the updating of these values is required in a dynamic
manner.

5.4 The improved ACO algorithm

At this stage, a modified version of the ACS is presented. The modifications include
a flexible state transition rule, dynamic candidate set strategy and heuristic
parameter updating rule [1–4, 6]. The system introduces a mechanism for escaping
from tapping in a local optimal solution and increases the efficiency of the
algorithm. The complete IACO algorithm is described in detail below.

5-3
 Modern Optimization Methods for Science, Engineering and Technology

Transition probability
Ants choose the next city from the current city using the state transition rule which is
the same as for the ACS. The rules are taken from the ACS of Dorigo and Stutzle. In
this stage, an ant will iteratively select from city to city. Some of the time an ant will
select a city using the random proportional action choice rule as before. Then the ant
chooses only the ‘best’ city based on the visibility and the pheromone trail.
Formally, an ant k positioned on city x chooses a city y to move to according to
⎧
⎪ arg max
y=⎨ { α
u∈Jk (x ) [τ(x,u ) ] . [η(x,u ) ]
β
,} if q ⩽ q0
, (5.1)
⎪
⎩Y , otherwise

where q is a number which has been randomly generated in the range between 0 and
1, q0 is a constant, Jk(x) is the set of cities which are not visited by the kth ant and Y
is a random city according to
⎧ [τ(x,y )]α . [η(x,y )] β
⎪
⎪ , if y ∈ Jk(x )
Pk(x , y ) = ⎨ ∑ [τ(x,u )]α . [η(x,u )] β , (5.2)
⎪
⎪
u∈Jk (x )

⎩ 0, otherwise

where Pk(x, y) is the probability of choosing city y from city x, Jk(x) is the set of
cities which are not unvisited by ant k, τ(x, y) is the pheromone value for moving
from city x to city y, η(x, y) is the local heuristic (distance) for moving from city r to
city s which is 1/d(x, y), α determines the significance of pheromone information and
β determine the significance of heuristic visibility.

Exploration and exploitation

A suitable adaptation is required to be carried out in the algorithm, which is between
the exploration and exploitation of the search space. The experience gained so far is
considered as exploitation, which is gained from visibility and trail amount, whereas
the exploration is counted as the number of unvisited or relatively unexplored or
relatively less explored possible set of elements in search space regions. Exploitation
is the action of taking the path with maximum probability. Exploration is the
process of calculating the probability where the ants decide which path to take based
on this probability. Hence, a path with higher probabilities would have a chance of
being more favorable to choose. So exploitation or exploration is an important
factor of the algorithm. The decision of the exploitation or exploration factor is
defined by a factor q0, the exploitation factor, which has a point value between 0 and
1 (a floating point value).
The tours are constructed in the AS by the ants which differ to a large extent from
the best possible values of tours. In the ACS of [8], the parameters q0, β and ρ are
taken as fixed values or constants whose values do not vary when the construction of
the solution is attempted for the ACS algorithm. In Dorigo et al [8], the value for q0
is chosen as 0.9, which implies that tours will be constructed by ants after exploring

5-4
 Modern Optimization Methods for Science, Engineering and Technology

10% of the total time and exploiting 90% of the time. This further suggests that the
relative significance of exploitation can be determined by q0 in comparison to biased
exploration, on the other hand, the desirability of the edges is determined by the
values of β and ρ.
In the algorithm suggested in this chapter, the next city is chosen by the ants with
more exploration at the beginning and the ants tend to have more exploitation at the
end of the algorithm. The information which is accumulated by exploitation values
is utilized properly. The values of q0 and β are permitted to vary and the appropriate
exploitation can be achieved. A new value for q0 is updated after every iteration of
the suggested scheme, which can be understood through
IterationCounter
q0 = (5.3)
MaxIteration

⎧ 0.1 if q0 < 0.1

⎪
q0 = ⎨ 0.9 if q0 > 0.9 , (5.4)
⎪
⎩ q0 otherwise

where IterationCounter is an important number that indicates the current iteration

cycle, and maximum number of iterations is MaxIteration.
When we increase the value of q0 then we will obtain the exploitation more
frequently and with greater value. The greater value of exploitation enables the ants
to use the edges of tours, which are globally best with a greater frequency. In such
cases, a local optimum may occur, which can be avoided by finding the same tour
iterating again and again, keeping in mind that the global best tour is not improved
after completing a number of iterations. This condition is not advisable because such
exploitation for a best global tour can erase memory and disturb the chosen edges
for the best global tour. Here, we also have to perform dynamic updating of heuristic
parameter so that the global best tour is not affected and proper utilization of
memory is also maintained.

Global pheromone update

In the process of global pheromone updates, the pheromone values are allowed to be
updated by the global best ant once the best tour is detected [10–13]. After the
completion of an iteration, the global pheromone updating rule is applied to obtain
the pheromone update and this is actually an important step in the updating process.
If there is no updating of the pheromone value, each ant will repeatedly find the
same probability on all moves and find the same solution again and again. The
evaporation and depositing of pheromone is only added to the best route found
when the beginning of the trial is detected or observed. The global updating rule is
⎧(1 − ρ)τ (x , y ) + ρ(Lgb)−1, if (x , y ) ∈ global_best_tour
τ (x , y ) = ⎨ , (5.5)
⎩ τ (x , y ), otherwise

5-5
 Modern Optimization Methods for Science, Engineering and Technology

where ρ is the decay parameter of pheromone and Lgb is the length of the global best
tour. As the effect of this rule, the ants’ searching concentrates around the best
known tour and also increases exploitation.

Local pheromone update

The local updating rule is used whenever we find the ant moving from one city to the
next city, and on movement from a city to the next city only the local updates are
applied. At the end of each iteration, while the global pheromone update is applied
to all edges once, the local pheromone update is applied to many. In the case where
local updates are not done, the ants will become non-collaborative in various
iterations. The ants will then only make use of the beginning of the pheromone trail.
The updates by ants on the pheromone trail over the edges are achieved by the
following rule:
τ (x , y ) = (1 − ρ)τ (x , y ) + ρτ0, (5.6)
where ρ is a constant whose value is found in the range (0, 1) and τ0 is the initial
value of the trail level. This initial value for the trail is calculated as
1
τ0 = , (5.7)
n · L nn
where n is the number of cities, and L nn is tour length covered by the nearest
neighbor search.
Using the above strategy, i.e. local updating, there is a decrease in the pheromone
concentration upon the traversal of edges. Subsequently, encouragement is given to
ants choosing other edges so that different solutions are produced. This avoids the
possibility of obtaining the same solutions by the ants during one iteration. If we
look at the original ACS, the ants are found to choose the next city or next element
and, based on the set of possible cities, the ants move to non-visiting cities. However,
this is a time-consuming process as ants, if they do so, will consume more time. If the
set of elements is restricted, then the efficiency of IACO can be improved considering
k nearest neighbors using a dynamic CL.

5.4.1 Dynamic candidate set based on nearest neighbors

The CL uses a strategy to attempt to obtain better performance of the algorithm
used by the ants. The strategies are defined and interpreted nicely in Gambardella
and Dorigo [7, 9] which involves the searching process of ACS applied over a large
amount of data. If a fixed candidate size is used, which is not flexible option, then
various data sizes cannot be addressed, but in order to improve the performance of
the algorithm the system requires the application of different CLs that are dynamic
in nature. A dynamic candidate list (DCL) can take on an appropriate number of
nodes depending on the total number of nodes [14, 15, 17, 18]. The DCL also
determines which sets are to be calculated and calculated again (recalculated)
throughout the search procedure. The static data structure has a limited number of
closed cities of preference which can be visited, as when we increase the number of

5-6
 Modern Optimization Methods for Science, Engineering and Technology

cities with the increase in the distance of cities. While constructing the solution the
next city is chosen by the ants, then the probability of movement from city i to city j
needs to be computed and, based on the probability of transfers, the next city is
chosen. If a city is not found by an ant, then that city is moved out of the total
number of the CLs, but this limits the selection capability of the ant and the list
subsequently becomes smaller. While moving, all different cities which are found by
the ants are added to the list. When the algorithm uses the candidate set strategy
then a few factors become very important:
• The number of edges in the global optimal solution in a set of candidates.
• Restriction of candidates to the scope of choosing or selection.

The improvement results in the performance of the algorithm on application of a

dynamic CL in the process of construction of the solution space. The DCL is based
on the theory of the nearest neighbor approach. The cities which are outside the
DCL are not allowed to be covered by the ants. This results in the following:
• A more limited number of cities than the actual number n.
• A limited value for the size of the candidate.

For instance, computation of the CL is determined and size is calculated for

Oliver30 data through the following procedure:

Procedure: Candidate list selection/determination

1. Initialize an empty node Node and MaxLength.

2. Set DCL = n/4 /*candidate list size*/.
3. if DCL > MaxLength DCL = MaxLength.
4. Get the cities which have not yet been visited.
5. repeat
1. for i = 1 to n.
2. Find the unvisited city j within the nearest neighborhood of k (j ∈ N (k )).
3. Calculate the distance between city j and city r.
4. if distance < distance of previous city j .
5. Node ← city j (move city j to NL).
6. end if
7. end for
8. DCL ← Node.
9. until list of candidates (DCL) is full.

The above procedure proved high speed in the performance of the algorithm
using the DCL strategy, which further helps in obtaining an improved computa-
tional time and improved quality of solution. Experimental studies have shown that
this proposed method results in an improvement of solution quality and a significant
performance gain.

5-7
 Modern Optimization Methods for Science, Engineering and Technology

5.4.2 Heuristic parameter updating

We studied a number of ACO strategies and observed that in the theoretical
background dealing with ACO parameters, α and β have not been attempted much
and therefore the amount of research on this specific domain is limited. These
parameters are very important as they allow ACO for controlling the relative
weights of pheromone trails and also the information of the heuristic search. The use
of parameters in the analysis of the possibility of the use of heuristic information is
pertinent since the construction of probable solutions results in exploitation of
specific knowledge for the problem chosen. The value referred to as the heuristic
value can obtain a prior value or, also in the runtime of all TSP problems, during the
initialization step the heuristic information η is calculated and once it is computed it
should not change throughout the operation of the algorithm. Another parameter β
actually describes the relative influence between the level of pheromone and the
information of the heuristic search, and hence the heuristic desirability η or the
information can be controlled by β. The larger the value of β, the faster convergence
becomes and the speed of the algorithm becomes very high, which subsequently
results in local best solutions in a more efficient manner. If the value of β = 0, then
only the information in the pheromone is taken into consideration, which leads to
stagnation in the algorithm. A stagnation situation in the algorithm means that the
ants will obtain the same solution rather than a sub-optimal solution. Thus, the
value of β needs to be investigated properly so that an appropriate adjustment of this
parameter can obtain the best local solution without any stagnation. High quality
tours also need appropriate heuristic information for the initial search stages.
Initially, the pheromone trail amount was kept as equal for all the edges, but this
did not allow a favorable path for the ants in the construction of the solution. This
situation demands a larger value of parameter β.
Later, when implementing the algorithm, the heuristic parameters may require
certain changes, but not much, due to the fact that the pheromone trails are
gathering enough information as per the requirement and the information may
misdirect the search process. So, we have to be very aware of what value is required
for the heuristic parameter, sometimes a constant value is also desired and
accordingly we have to provide this, particularly in conventional ACO algorithms.
High values of parameters are required in achieving better quality tours and the
influence of the pheromone is also reduced by a considerable extent. Under such
situations, ants can search their paths effectively in the process of construction of
feasible solutions. In the initial process, a small value can help, but performance
improvement always requires high values of parameters [14, 15]. The values of
parameters need to be adaptively changed as per the requirement so that perform-
ance is also satisfactory and no stagnation or other such problems prevail. This
study suggests that dynamic updates are required in the values of heuristic
parameters for ACO towards improved search performance.
Now, when the algorithm begins to be executed then the pheromone information
for all paths is equal to each other and the value of entropy is highest at this point of
time. After every iteration, updates are done on the pheromone and it is also

5-8
 Modern Optimization Methods for Science, Engineering and Technology

enhanced on the path chosen so that the entropy value starts slowly decreasing. If
the entropy is allowed to decrease, then it may also reach zero and that may not lead
to a global best tour and the situation will be termed as premature convergence. In
order to deal with this type of difficulty, which occurs due to behavioral defects,
some complex optimization problems require colony optimization which can handle
the issues related to entropy. The discussion of entropy issues is required for dynamic
updates of heuristic parameters, which can also control the value of entropy.
Shannon’s theory of entropy as information theory is very popular in this area,
also known as Shannon’s information theorem [3]. This was introduced in 1948 and
is usually defined as a measure of uncertainty which is concerned with events having
disorder in a system. The entropy represents that information which is associated
with the probability of the occurrence of the event. Why this concept is relevant is an
important question. It is because of the fact that the ACO algorithm has a path
selection procedure and the selection of the path is also not certain, which means
there is uncertainty. Therefore, we suggest the entropy information to be estimated
in ACO as variation of the pheromone matrix. Each trail in this case would be a
random variable in the pheromone matrix. Such entropy is defined as
r
H (X ) = −∑Pi log Pi , (5.8)
i=1

where pi is the probability of occurrence of each trail in the pheromone matrix. For a
symmetric n city in the TSP, there are n(n − 1)/2 distinct pheromone trails and r = n
(n − 1)/2. It is obvious that the probability of each trail is the same, and H will be the
maximum (Hmax), given as
r r
1 1
Hmax = −∑Pi log Pi = −∑ log = log r . (5.9)
i=1 i=1
r r

We are suggesting using entropy as an index or degree of how much information

is trained by the pheromone trails, and updates in the heuristic parameter are
required accordingly. It is worth noting here that the heuristic parameter β value has
to be considered as an integer value because this can avoid complicated computation
processes. The value of β is also used as a power in equations (5.1) and (5.2). The
suggested system behaves adaptively and the following rules are proposed with
regard to β:
⎧ 5 0.9 < H ′ ⩽ 1
⎪
⎪ 4 0.7 < H ′ ⩽ 0.9
β=⎨ (5.10)
⎪ 3 0.5 < H ′ ⩽ 0.7
⎪
⎩ 2 0 < H ′ ⩽ 0.5

Hmax − Hcurrent
H′ = 1 − , (5.11)
Hmax

5-9
 Modern Optimization Methods for Science, Engineering and Technology

where H ′ is the value of entropy of the pheromone matrix that is currently being
considered.

5.5 Improvement strategy

After the success of the proposed algorithm, there needs to be a lot of effort to
improve the performance of the algorithm. The solution can be improved using a
local search, keeping in mind the fact that we have to look in its neighborhood as
well for better ones. The ants have generated the tours and local search methods are
now responsible for improving the performance. In the case of the TSP, this involves
improving a given tour with small local changes. Having found a tour, the order of
the cities is slightly changed into another valid tour by applying one of the following
methods.

5.5.1 2-Opt local search

The simplest way to change a TSP tour is to exchange two edges. The 2-Opt
algorithm is basically used to eliminate two edges from the tour, and then start
rebuilding the two chosen paths. This is done by a reconnection process of two other
edges in some different way so that we can obtain a new tour. This process is
generally called a 2-Opt move. The way in which reconnection can work is shown in
figure 5.1(a), where we can see [18] that two paths are reconnecting and generating a
new and valid tour. This can be obtained for shorter tours only. The process of
removing and reconnecting tours continues until there remains on further scope of
improvement in Opt-2 and now the resulting tour is 2-optimal.
For example, to realize this, two non-adjacent edges, call them ab and cd, are
chosen (note that ab means that city a is visited before city b on the tour). These
edges are deleted and new edges ac and bd are inserted and hence form a new TSP
tour (see in figure 5.1(b)). Local search in 3-Opt starts working in a similar manner
to 2-Opt, but rather than removing two edges we need to remove three, which means
that the exchange of three edges in the tour is considered to compute the final best
tour. There are two possible ways in which we can reconnect the three paths into a
valid tour.

c a c a

d d
b b

(a) (b)
Figure 5.1. A 2-Opt local search. Edges ab and cd from the graph illustrated in (a) were deleted and the tour
was completed again by inserting edges ac and bd, thus resulting in the valid tour shown in the graph (b).

5-10
 Modern Optimization Methods for Science, Engineering and Technology

5.6 Procedure of IACO

This section briefly describes the proposed algorithm step by step.

Algorithm. Improved ACO (IACO) algorithm for TSP.

1. Initialize
Find pheromone trails τ0 using a nearest neighbor heuristic
for every edge(x, y)
Set τ(x, y) = τ0
end for
2. Calculate the maximum entropy
globalBestTour ← ϕ
globalBestTourLength ← ∞
determine candidate_list strategy
3. Place the n ants randomly on the starting city
for k = 1 to n ants do
Place and sum with the starting city of the kth ant to its tabular list
end for
localBestTour ← ϕ
localBestTourLength ← ∞
Define the value of exploitation parameter q0
4. for k = 1 to n do /*n is the number of ants */
q ← random
repeat
Find the next unvisited node j for kth ant from kth ant’s current node i by using state
transition rule with CL
Append j into the kth ant’s tabu list
Perform local pheromone update
until tour has been completed by ant k
Apply local search (2-Opt or 2.5 Opt) to improve tour
Compute tourLength of kth ant
if tourLength < localBestTourLength then
localBestTourLength ← tourLength
localBestTour ← best tour found
end if
end for
if localBestTourLength < globalBestTourLength then
globalBestTourLength ← localBestTourLength
globalBestTour ← localBestTour
end if
5. for each edge (r, s) belonging to the global best tour
Perform global pheromone update
end for
6. for every pheromone τ(r, s)
Compute value of entropy for current pheromone trails

5-11
 Modern Optimization Methods for Science, Engineering and Technology

end for
Update the heuristic parameter
7. // Check end condition
if (end_condition = true or MaxIteration)
print global best tour
exit
else
go to 3.
End if

5.7 Flow of IACO

The flow diagram of the suggested IACO algorithm is shown in figure 5.2. This
begins from initializing the parameters and terminates by obtaining the best tour
results. The steps and procedures shown in the flow diagram are self-explanatory.

5.8 IACO for solving the TSP

This section explains how the IACO executes, which is applied to solve a
considerable problem for TSP and the candidate set is defined as a set of cities. In
the following example, there are eight cities P to W and two ants (ant A1 and ant A2).
Assume that you place the starting city of ant A1 and ant A2 on city P and Q,
respectively. When an ant k needs to move from current city x to next city y, it first
searches in predefined set of CLs. Then, it also adds its current start city to a short-
term set of lists called the tabu list and applies equation (5.3), where it produces a
value of the q0 parameter and also produces a randomly generated value for
parameter q, and compares the two parameter. When a random parameter q ⩽ q0,
the state transition rule, equation (5.1), is applied to choose the next city. An ant k
exploits the knowledge available about the problem and moves to city (y) which has
the highest product of the amount of pherormone trail on the edge (x, y) and the
shortest distance between the two cities. When q > q0, the random proportional rule,
equation (5.2), is applied in which the ant k uses it to explore new solutions using
the candiate list. Each city has a candidate list (cl); the number of cities listed is the
length of the CL. In this example, assume that the length of the CL, cl = 2, where the
candidate list of city (P) includes cities (Q) and (W ) and they will be explored by ant
A1 before the other cities. It is assumed that ant A1 goes from city (P) to city (Q),
then city (Q) will be added to the tabu list to avoid being visited again by the same
ant. After moving from city (P) to city (Q), ant A1 updates the pheromone on the
edge between the two cities using the local pheromone updating rule, equation (5.6).
For the next step, ant A1 again computes the possibilities of moving from its current
city (Q) to those other cities that are not in its tabu list (R to W ) using the same
equation (5.1) and so on until ant A1 visits all the other cities as shown in figure 5.3.
Ant A2 also proceeds in such a way to find its solution.

5-12
Modern Optimization Methods for Science, Engineering and Technology

Start

Initialize parameter

Calculate max entropy of pheromone matrix

Candidate list strategy

Place ant in a randomly chosen city

Determine probabilistically as to which

city to visit nextwith candidate list

Move next city

Local pheromone update

More cities yes

to visit
Repeat m ants
no

Local search (2-Opt, 2.5 Opt) apply to

improve tour

Global pheromone update to the best tour

found by ants

Calculate entropy value of current pheromone

matrix

Heuristic parameter updating using entropy

value

no Maximum
Iteration

yes

Best Tour Result

End

Figure 5.2. Flow of improved ACO algorithm.

5-13
 Modern Optimization Methods for Science, Engineering and Technology

Q R Q R
S S
P P
T T

W U W U
V V
(a) (b)

Q R Q R
S S
P P
T T

W U W U
V V
(c) (d)

Q (c) R Q R
(d)
S S
P P
T T

W U W U
V V
(e) (f)

Q R Q R
S S
P P
T T

W U W U
V V
(g) (h)
Figure 5.3. IACO for a simple TSP.

The tour length generated by the ant will be calculated by adding the length of the
edge between each two cities from the tour. The process will be accomplished by
each ant and at the end of the iteration there will be five tours generated by five ants.
The local search improvement is applied to improve the tour consturcted by the ants.

5-14
 Modern Optimization Methods for Science, Engineering and Technology

The shortest of these tours will be selected as the best tour and the edges that form
this tour will be updated using the global update formula in equation (5.5). Then, we
calculate the current entropy by analyzing the pheromone information to update the
heuristic parameter β. Then, the ants are again placed randomly for a second
iteration and redefine the exploitation parameter q0. The algorithm goes on until the
maximum number of iterations has finally found the global best tour.

5.9 Implementing the IACO algorithm

This section describes in detail the steps for implementing an improved ACO
algorithm for the TSP. The basic considerations for the implentation of the IACO
algorithm are similar to the ACS and the necessary changes for implementation are
presented. The problem can be represented as a construction graph used by the
construction procedure, and then it is only necessary to:
1. Define the pheromone trail variables to the construction graph.
2. Define the number of aritificial ants to be used for constructing solutions
according to equation (5.1).
3. A randomized version of the construction procedure.

This section describes an implementation in a pseudo-code description.

Procedure: IACO for TSP

begin
InitializeData
Calculate MaxEntropy
while (not termination) do
ConstuctSolutions
ApplyLocalSearch
UpdateStatistics
UpdatePheromoneTrails
ComputeCurrentEntropy
UpdateHeuristicParameter
end-while
end

In data initialization: (i) read the TSP instance; (ii) calculate the distance matrix of
the read TSP instance; (iii) define and compute the the CLs for all cities; (iv) the ants
randomly place the their starting cities; (v) the algorithm’s parameters must be
initialized; and (vi) some variables that keep track of statistical information, such as
number of iterations, or best solution found (best tour length), and best tour, need to
be included.

5-15
 Modern Optimization Methods for Science, Engineering and Technology

Procedure: InitializeData
begin
ReadTSPInstance
ComputeDistances
ComputeCandidateLists
InitializeAnts
InitializeParameters
InitializeStatistics
end

The following two constuction steps are repeated until the tour is finished by all the
ants. When exploiting the procedure the ConstructExploitDecisionRule needs to be
adapted. If not, the procedure ConstructExploreDecisionRule needs to be computed.
In ConstructExploitDecisionRule, a major change that can be seen is that while
choosing the next city, we need to find the city which is not visited from the entire list
of candidates from the current city. Another change which can be seen is the necessity
for dealing with the situations where all cities have been covered in the CL by any ant
k. Under this condition and taking the changes into account, the variable node
maintains its initial value as −1 and the city beyond the CL is chosen. We are required
to choose the maximum product of the pheromone value and heuristic information
[τij]α[ηij]β for moving to the next city. In ConstructExploreDecisionRule there are two
changes as in the ConstructExploitDecisionRule. The exploring procedure helps in
choosing the next city which is unvisited as per the acton choice rule in equation (5.2).

Procedure: ConstructSolutions

begin
curNode ← startNode
q0 ← IterCounter/Iterations
for k = 1 to m ants do
repeat
q ← random number
if(q < q0) then
newNode ← ConstructExploitDecisionRule(k, curNode)
else newNode ← ConstructExploreDecisionRule(k, curNode)
end if
Add newNode to ant k’s tour
LocalUpdatingRule(curNode, newNode)
curNode ← newNode
until ant k completes tour
end for
end

5-16
 Modern Optimization Methods for Science, Engineering and Technology

Procedure: ConstructExploitDecisionRule(k, curNode)

begin
sum_probability ← 0.0 // CandidateListConstructionRule
node ← −1
for j = 1 to DCL do
if kth ant’s node j is not visited in candidate list then
selection_probability ← value of transition probability
node ← j /* city with maximal ταηβ*/
end if
end for
if (node == −1) then // city outside candidate list
for j = 1 to n do
if kth ant’s node j is not visited outside the candidate list then
selection_probability ← value of transition probability
node ← j /* city with maximal ταηβ*/
end if
end for
end if
return node
end

Procedure: ConstructExploreDecisionRule(k, curNode)

begin
sum_probability ← 0.0 // CandidateListConstructionRule
node ← −1
for j = 1 to DCL do
if kth ant’s next node j is not visited in candidate list then
partial_product ← pheromone_value*exp(1/distance,β) /* city with ταηβ*/
sum_probability ← sum_probability + partial_ product
end if
end for
for j = 1 to DCL do
if kth ant’s next node j is not visited in candidate list then
selection_probability ← partial_ product / sum_probability
rno ← random number
if selection_probability >= rno then
node ← j
break
end if
end if
end for
if node = −1 then // city outside candidate list
for j = 1 to n do
if kth ant’s next node j is not visited outside the candidate list then

5-17
 Modern Optimization Methods for Science, Engineering and Technology

partial_product ← pheromone_value*exp(1/distance, β) /*city with ταηβ*/

sum_probability ← sum_probability + partial_ product
end if
end for
for j = 1 to n do
if kth ant’s next node j is not visited outside the candidate list then
selection_probability ← partial_product / sum_probability
rno ← random number
if selection_probability >= rno then
node ← j
break
end if
end if
end for
return node
end

It is evident that the use of CL and computation time are very important factors
for the ants in the construction of solutions, in particular if lower values of these
factors are expected. It can be seen also that the values are considerably reduced
when ants are choosing the cities.
The next step is about local pheromone updating which acts as a trigger after the
ants have moved to the next city.

Procedure: LocalUpdatingRule(curNode, newNode)

begin
value ← (1 − ρ)*pheromone[curNode][newNode] + ρ* pheromone[curNode][newNode]
pheromone[curNode][newNode] ← value
pheromone[newNode][ curNode] ← value
end

Once the solutions are constructed, the generated tours may be improved by a
local search procedure (for example 2-Opt or 2.5-Opt). The next step in an iteration
of the algorithm is the pheromone update (global pheromone updating). This is
implemented by the procedure of UpdatePheromoneTrails, which makes up two
pheromone update phases: pheromone evaporation and pheromone deposit.
Pheromone evaporation decreases the value of the pheromone trails on only the
best path by a constant pheromone decay factor ρ. The pheromone deposit adds
pheromone to the edges belonging to tours constructed by the ant’s best path length.
ComputeCurrentEntropy computes the pheromone information of the current
pheromone matrix to be used in the next step. UpdateHeuristicParameter

5-18
 Modern Optimization Methods for Science, Engineering and Technology

dynamically updates the heuristic parameter based on the entropy value of current
pheromone information.

Procedure: UpdatePheromoneTrails

begin
for i = 1 to n − 1 do
tau ← 1/ bestLength
evaporation ← (1 − ρ) * pheromone[bestPath[i]][bestPath[i + 1]]
deposition ← ρ * tau;
pheromone[bestPath[i]][bestPath[i + 1]] ← evaporation + deposition
pheromone[bestPath[i + 1]bestPath[i]] ← pheromone[bestPath[i]][bestPath[i + 1]]
end for
end

Procedure: ComputeCurrentEntropy

begin
sum ← 0.0
current ← 0.0
max_entropy ← Math.log(noOfNodes*(noOfNodes-1)/2)
for i = 2 to n do
for j = 1 to i do
sum ← sum + pheromone[i][j]
end for
end for
for i = 2 to n do
for j = 1 to i do
current ← current + (−((pheromone[i][j]/sum)*
(Math.log(pheromone[i][j])/sum)))
end for
end for
current ← 1−((max_entropy-current)/max_entropy)
end

5.10 Experiment and performance evaluation

This section reports on the experiments performed to test the proposed system and
compares the results with the ACS ant algorithm. An experimental study is
presented which focuses on comparing the convergence speed of the proposed
system with that of ACS and also the analysis results. To obtain meaningful
performance comparison results, a relative error (degree of approximation) calcu-
lation is conducted. The measurement of the performance of the experiments is
described and the tour length results are also presented.

5-19
 Modern Optimization Methods for Science, Engineering and Technology

The TSP is a classical optimization problem which has been used for the
evaluation of the performance of ant algorithms. This chapter presents experiments
on TSPs comparing the performance of ant colony optimization and the modified
ant colony optimization approach with CL strategy based on entropy. The
implementations use the ACS version of the ant algorithm. The proposed method
includes a candidate set which is applied to the construction phase and one extension
to the update phase of the algorithm’s heuristic parameter based on entropy.

5.10.1 Evaluation criteria

Numerous studies in the field of TSPs and applications using ACO systems suggest
the metrics used by the method in the implementation, such as standard deviation,
best tour iteration and best possible tour. The best tour is interpreted as the
optimum tour with the least computation time and the standard deviation
indicates how much deviation is reported in the solution after each iteration.
The degree of approximation, also referred to as relative error, is also computed,
which determines how much deviation there is in solutions from the optimal
solution. Among several methods reported in the literature, the TSP has found a
good position as a traditional method to address various combinatorial problems
of optimization.

5.10.2 Path evaluation model

The goal of the path evaluation model is to find an improvement space for each path
in an optimal solution. Given a random sample (X1, X2, …, Xn) from an n-
dimensional random variable, the sample mean is
n
∑xi (5.12)
i=1
X = ,
n

where X means the tour length, and n is number of the tour length. The sample
variance σ2 is the second sample central moment and is defined by
n
∑(Xi − X )2 (5.13)
i=1
σ2 = ,
n

where σ is the standard deviation. The algorithm for optimization is better, with
more robustness and better stability, if the value of the standard deviation of a
performance criterion over a number of simulation runs is small. Another similar
factor σ2 is known as variance in the evaluation of performance criterion. The
smaller the value of σ2, the smaller is the range of performance values fitting into
stability like a swarm.

5-20
 Modern Optimization Methods for Science, Engineering and Technology

5.10.3 Evaluation of solution quality

In optimization search problems, the algorithm generates some solution but the
quality of the solution needs to be assessed carefully. Suitable measures are required
for the assessment and there are four popular measures for evaluating the quality of
the solution: approximation, accuracy, completeness and belief.
a. Approximation. The approximation ε describes the size of the deviation of s
from s*. s* means the optimal solution value and s is the value of the feasible
solution:
s − s⁎
ɛ= . (5.14)
s⁎
Note that 0 ⩽ ε < ∞. If the value of approximation is equal to 0, it is an
optimal solution.
b. Accuracy. The accuracy, also called the degree of accuracy, α, indicates how
close the feasible solution s is to the optimal solution s* and is defined as
min{s⁎ , s}
α= . (5.15)
max{s⁎ , s}
Note that 0 ⩽ α ⩽ 1. The optimal solution has an accuracy equal to 1.
c. Completeness. This is the measure of retrieving the complete solution. If the
search ceases, then the solution may not be complete. Adequate resources are
to be used so that no premature convergence takes place. The value of this
measure is between 0 and 1, and for a value of 1, the solution is said to be
complete.
d. Belief. The degree of belief indicates heuristically or statistically the belief
that the solution obtained is the one desired. This is useful in searching in a
probabilistic environment. The degree of belief is a value between 0 and 1. A
totally confident solution has a degree of belief equal to 1.

In this system, we are interested in metaheuristic solutions; therefore, complete-

ness and belief are irrelevant. Approximation and accuracy are duals because one
can be determined from the other. In this system, the approximation is taken
because it is commonly used in approximation searches and also presents the degree
of accuracy of both algorithms’ results.

5.11 TSPLIB and experimental results

The experiments in this approach use problems exclusively from TSPLIB, a widely
known library of TSPs [16]. TSPLIB is an online library, developed by the
University of Heidelberg in Germany, that contains several samples of TSPs and
similar related problems in a list of different files. The optimal solutions of the
problems are recorded within the library, which is a convenient feature of TSPLIB.
Due to this being a prominent source for TSPs, many other research papers explore
the TSP instances defined within TSPLIB. TSPLIB is thus used to benchmark

5-21
 Modern Optimization Methods for Science, Engineering and Technology

solutions. The benchmark instances are given with varying complexity and diffi-
culty. Only strongly connected instances of TSP were chosen from the TSPLIB, and
through symmetric Euclidian TSPs were analyzed, referenced as EUC_2D, meaning
that distances (or weights) between nodes are expressed as an Euclidean 2D
coordinate system. The coordinates are decimal numbers (or doubles).

5.11.1 Experiment 1 (analysis of tour length results)

First, the proposed algorithm was tested to see how it performed; the eil51 case was
chosen as the test case. The results were shown over 20 trials with 20 iterations per
trial (run) for eil51. The parameters used in this test case are described in table 5.1.
The results show how large the deviation from the optimal distance is when
running the proposed system several times, both with and without optimization
including local search optimization.
As seen in figure 5.4, the proposed algorithm obtains an average tour length of
426.6 when using optimization with local search, which is 0.6 higher than the optimal
of 426. The deviation of the proposed algorithm’s result (degree of approximation) is

Table 5.1. General parameters used in the test case.

Parameters Value

α 1
β 2–5
ρ 0.1
No. of ants 10
No. of iterations 20

IACO: Results for 20 runs on eil51.tsp

428

427.5

427
Tour length

TourLength
426.5 Optimal
Average
426

425.5

425
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
Number of runs
Figure 5.4. Results for the eil51 instance using local search (2-Opt).

5-22
 Modern Optimization Methods for Science, Engineering and Technology

Figure 5.5. The best-so-far solution of the tour length of the eil51 instance.

Figure 5.6. The tour best result of each iteration for the eil51 instance.

0.14% from the optimal distance. The proposed algorithm’s results vary from 426
(0% deviation) to 428 (0.47% deviation) and considering half of the running results
are converged to the optimal and average length of the runs are much closed to the
optimal. Figures 5.5–5.7 show the analysis results for best tour-so-far solution, tour
best and standard deviation of tour length, respectively.

5-23
 Modern Optimization Methods for Science, Engineering and Technology

Figure 5.7. Standard deviation of tour length for the eil51 instance.

IACO: Results for 20 runs on eil51.tsp

455
450
445
Tour length

440
435
TourLength
430
Optimal
425
Average
420
415
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
Number of runs
Figure 5.8. Results on the eil51 instance without using local search.

In figure 5.8, when the proposed algorithm does not use the local search
optimization to optimize the tour, the results change dramatically. The test case
was changing the number of iterations to 100 and the remaining parameter values
were kept as in table 5.1. The best solution found is at 429 for 20 trials. It is 0.7%
deviated from the optimal solution 426. The results vary between 429 and 453, with
an average tour length of 440.7 (0.7%, 6.34% and 3.45% deviation, respectively).

5-24
 Modern Optimization Methods for Science, Engineering and Technology

ACS: Results for 20 runs on eil51.tsp

430

429
Tour Length

428

427
TourLength
426 Optimal
Average
425
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
Number of runs
Figure 5.9. Results on the eil51 instance with local search (2-Opt).

Table 5.2. Comparison of the final best solution and convergence number between IACO and DSMACS.

TSP Best length of Best length of Convergence number Convergence number of

problem IACO DSMACS of IACO DSMACS

eil51 426 426 4 5

berlin52 7542 7542 3 4
st70 675 N/A 3 N/A

The analysis results of the ACS are also described in figure 5.9. The ACS
algorithm runs the eil51 test case with the parameter value of β = 5 and the other
parameter values as in table 5.1. ACS obtains an average tour length of 428.25 when
using optimization with local search, which is 2.25 higher (making it a deviation
(degree of approximation) of 0.53%) than the optimal of 426. The ACS algorithm
results vary from 427 (0.24% deviation) to 430 (0.94% deviation) and more than half
of the runs are below the average length of the runs.

5.11.2 Experiment 2 (comparison of convergence speed)

In the second experiment, the proposed approach is compared with DSMACS [5] in
terms of convergence speed. In order to compare the proposed system with
DSMACS, some instances of TSP that are the same as those used in DSMACS
are chosen. A comparison of the final best solution and the convergence number
between the proposed algorithm (improved ACO) and DSMACS are shown in
table 5.2 (the results of DSMACS are directly taken from [5]). It should be noted
that the performance of the proposed IACO algorithm is excellent. It not only
discovers the global optimal solutions for the following TSP instances but also has
very quick convergence speed. The speed of convergence is also shown graphically as
figure 5.10 for the proposed algorithm.

5-25
 Modern Optimization Methods for Science, Engineering and Technology

Convergence result of tour length

432
Best sofar: Tour Length
431
430
429
428
427
426
425
424
423
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
Number of Iterations

Improved ACO

(a) eil51
Convergence result of tour length
7800
Best sofar: Tour Length

7750
7700
7650
7600
7550
7500
7450
7400
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
Number of Iterations

Improved ACO

(b) berlin52
Convergence result of tour length
684
Best sofar: Tour Length

682
680
678
676
674
672
670
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
Number of Iterations
Improved ACO
(c) st70

Figure 5.10. Convergence speed of tour length for (a) eil51, (b) berlin52 and (c) st70 TSPs.

5-26
 Modern Optimization Methods for Science, Engineering and Technology

5.12 Comparison experiment

For the purpose of demonstrating the efficiency of the improved ACO algorithm
proposed in this system, we have constructed a simulation and applied it to problems
from the TSPLIB library: oliver30, eil51, eil76, eil101, berlin52, st70, rat99,
kroA100, lin105 and pr144. In this study the proposed algorithm results are
compared to the results of the ACS algorithm in the aspects of algorithm
convergence and experiment results. The ACS algorithm combines with local search.
In tables 5.3–5.12, the first column is the best distance from the beginning of the
trial (run), as compared to the benchmark distance. The second column shows the

Table 5.3. Comparison results of oliver30.

oliver30 (20 runs,

20 iterations/run) Best solution Average solution Relative error Standard deviation

β=2 420 420 0% 0

β=3 420 420 0% 0
ACO β=4 420 420 0% 0
β=5 420 420 0% 0
IACO algorithm 420 420 0% 0
Optimum 420

Table 5.4. Comparison results of eil51.

eil51 (20 runs,

20 iterations/run) Best solution Average solution Relative error Standard deviation

β=2 427 429.05 0.23% 1.62

β=3 427 429.3 0.23% 1.63
ACO β=4 427 429.05 0.23% 1.63
β=5 427 428.21 0.23% 0.92
IACO algorithm 426 426.6 0% 0.68
Optimum 426

Table 5.5. Comparison results of eil76.

eil76 (20 runs,

20 iterations/run) Best solution Average solution Relative error Standard deviation

β=2 541 546.8 0.56% 3.53

β=3 538 546.1 0% 3.46
ACO β=4 541 545.75 0.56% 4.27
β=5 538 543.95 0% 2.14
IACO algorithm 538 539.9 0% 1.33
Optimum 538

5-27
 Modern Optimization Methods for Science, Engineering and Technology

Table 5.6. Comparison results of eil101.

eil101 (20 runs,

20 iterations/run) Best solution Average solution Relative error Standard deviation

β=2 637 647.35 1.27% 3.47

β=3 636 644.9 1.11% 3.95
ACO β=4 638 644.45 1.43% 3.66
β=5 639 643.05 1.59% 3.10
IACO algorithm 629 635.55 0% 3.03
Optimum 629

Table 5.7. Comparison results of berlin52.

berlin52 (20 runs,

20 iterations/run) Best solution Average solution Relative error Standard deviation

β=2 7542 7561.85 0% 53.11

β=3 7542 7581.25 0% 54.95
ACO β=4 7542 7599.65 0% 66.56
β=5 7542 7556.2 0% 36.51
IACO algorithm 7542 7542 0% 0
Optimum 7542

Table 5.8. Comparison results of kroA100.

kroA100 (20 runs,

20 iterations/run) Best solution Average solution Relative error Standard deviation

β=2 21 282 21 420.85 0% 90.91

β=3 21 282 21 383.75 0% 82.06
ACO β=4 21 292 21 375.75 0% 59.75
β=5 21 282 21 353.15 0% 63.03
IACO algorithm 21 282 21 297.95 0% 24.31
Optimum 21 282

average distance of the trials. The third column shows the relative errors, also called
degree of approximation ((best solution-optimum)/optimum) which is how much the
deviation of the optimal distance from the best solution, and the fourth column is the
standard deviation.
In all cases, the proposed algorithm shows better performance than the ACS
algorithm. The experiment shows that the improved ant colony algorithm proposed

5-28
 Modern Optimization Methods for Science, Engineering and Technology

Table 5.9. Comparison results of lin105.

lin105 (20 runs, 20 iterations/

run) Best solution Average solution Relative error Standard deviation

β=2 14 397 14 439.35 0% 41.71

β=3 14 397 14 416.3 0% 32.51
ACO β=4 14 397 14 417.85 0% 47.83
β=5 14 397 14 404.65 0% 33.24
IACO algorithm 14 397 14 391.35 0% 16.72
Optimum 14 397

Table 5.10. Comparison results of pr144.

pr144 (20 runs,

20 iterations/run) Best solution Average solution Relative error Standard deviation

β=2 58 537 58 583.5 0% 35.46

β=3 58 537 58 592 0% 51.48
ACO β=4 58 537 58 589.65 0% 54.33
β=5 58 537 58 571.4 0% 27.03
IACO algorithm 58 537 58 537 0% 0
Optimum 58 537

Table 5.11. Comparison results of rat99.

rat99 (20 runs,

20 iterations/run) Best solution Average solution Relative error Standard deviation

β=2 1217 1233.2 0.5% 8.02

β=3 1219 1234.85 0.66% 10.07
ACO β=4 1215 1232.45 0.33% 9.33
β=5 1213 1228.5 0.17% 10.14
IACO algorithm 1211 1216.4 0% 4.87
Optimum 1211

in this approach achieved better results for TSPs, and its efficiency of solutions is
better than that of ant colony algorithm.
Looking at the above tables the deviation (relative error) from the optimal
solutions can be seen, and the differences in the deviations between the IACO
algorithm and the ACS are compared. Moreover, the average distance of the IACO

5-29
 Modern Optimization Methods for Science, Engineering and Technology

Table 5.12. Comparison results of st70.

st70 (20 runs,

20 iterations/run) Best solution Average solution Relative error Standard deviation

β=2 676 680.7 0.15% 2.83

β=3 675 680.2 0% 3.17
ACO β=4 675 680.5 0% 2.55
β=5 675 680 0% 2.55
IACO algorithm 675 676.8 0% 1.91
Optimum 675

Comparison of Tour Length Result

423
Best so far: Tour Length

422

421

420

419
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
Number of Iterations
Improved ACO ACS

Figure 5.11. Comparison of the tour length result of oliver30.

and ACS are more obvious. Figures 5.11–5.20 present the comparison of con-
vergence results of the tour length to obtain the best results for oliver30, eil51,eil76,
eil101, berlin52, kroA100, lin105, pr144, rat99 and st70, respectively. The tested TSP
datasets were executed for 20 trials, with the same number of iterations for each
dataset. The best values obtained from a trial are given below for each dataset as in
the figures. These figures show the best solution found since the start of the algorithm
run, or the iteration number at which the best solution was found. The distance which
is illustrated in the graph is the best distance found so far in the iterations.
In figure 5.21 the performance and optimal solution can be seen for IACO and
ACS with the index 100 (also called degree of accuracy). The IACO algorithm
obtains better results than the ACS and all of the results are strikingly close to the
optimal solution.
The analysis results for the larger TSP instances can be seen in table 5.13. These
problems were tested with 20 iterations for 30 trials. The number of ants used for all
these test cases is 10. The parameters are as follows α = 1, β is a dynamic value of the
algorithm and the pheromone decay parameter ρ is 0.1. However, β = 5 is used for
the ACS and other parameters are used as the proposed algorithm. Both of these

5-30
 Modern Optimization Methods for Science, Engineering and Technology

Best so far: Tour Length

Comparison of Tour Length Result
436
434
432
430
428
426
424
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
Number of Iterations
Improved ACO ACS

Figure 5.12. Comparison of the tour length result of eil51.

Comparison of Tour Length Result

555
Best so far: Tour Length

550
545
540
535
530
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
Number of Iterations
Improved ACO ACS

Figure 5.13. Comparison of the tour length result of eil76.

Comparison of Tour Length Result

655
Best so far: Tour Length

650
645
640
635
630
625
620
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
Number of Iterations
Improved ACO ACS

Figure 5.14. Comparison of the tour length result of eil101.

5-31
 Modern Optimization Methods for Science, Engineering and Technology

Comparison of Tour Length Result

Best so far: Tour Length 7950
7850
7750
7650
7550
7450
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
Number of Iterations
Improved ACO ACS

Figure 5.15. Comparison of the tour length result of berlin52.

Comparison of Tour Length Result

22200
Best so far: Tour Length

22100
22000
21900
21800
21700
21600
21500
21400
21300
21200
21100
21000
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
Number of Iterations
Improved ACO ACS

Figure 5.16. Comparison of the tour length result of kroA100.

Comparison of Tour Length Result

14800
Best so far: Tour Length

14750
14700
14650
14600
14550
14500
14450
14400
14350
14300
14250
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
Number of Iterations
Improved ACO ACS

Figure 5.17. Comparison of the tour length result of lin105.

5-32
 Modern Optimization Methods for Science, Engineering and Technology

Comparison of Tour Length Result

Best so far: Tour Length 58700
58680
58660
58640
58620
58600
58580
58560
58540
58520
58500
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
Number of Iterations
Improved ACO ACS

Figure 5.18. Comparison of the tour length result of pr144.

Comparison of Tour Length Result

1290
Best so far: Tour Length

1280
1270
1260
1250
1240
1230
1220
1210
1200
1190
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
Number of Iterations
Improved ACO ACS

Figure 5.19. Comparison of the tour length result of rat99.

Comparison of Tour Length Result

710
Best so far: Tour Length

705
700
695
690
685
680
675
670
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
Number of Iterations
Improved ACO ACS

Figure 5.20. Comparison of the tour length result of st70.

5-33
 Modern Optimization Methods for Science, Engineering and Technology

Accuracy results
100.5
100
Accuracy %

99.5
99
98.5 ACS
IACO
98
Optimal
97.5

TSP instances
Figure 5.21. The accuracy of IACO and ACO performance compared to the optimal solution.

algorithms tested all the problems with local search optimization (2-Opt). The
results were shown as the best length, average tour length and relative errors of the
algorithms. Most results of the IACO algorithm achieved the optimal solutions.
Therefore, the results of the IACO algorithm are more satisfactory than the results
of the ACS algorithm. Furthermore, the results of both algorithms are compared to
the optimal solution with the degree of accuracy. In figure 5.22 it can be seen that
most of the proposed algorithm results achieve the optimal solution and some results
are less deviated from optimum. The log graphs are also presented.

5.13 Analysis on varying number of ants

The section presents the analysis results when the number of ants is varying. The
analysis on the number of ants consists of two parts: analysis of the number of ants
starting at different cities versus the same city, and analysis of an increasing number
of ants versus an increasing number of iterations.

5.13.1 Analysis of ants starting at different cities versus the same city
If the TSP finds a good solution with a good tour length, then the algorithm is very
effective when the ants start from different cities. Generally, cities are similar or
identical in implementations of the TSP, and while maturity of the algorithm is
obtained and at the beginning of the tour, ants can freely select any appropriate city
for the next movement. So, ants may also use the cities with short distance and when
they reach the end of the tour almost all cities might have been covered. Once the
cities with small edges or distance are covered, the ants will also be required to reach
the next cities with some bigger edges. The method of selection for a starting city is
random in this work and the search space can therefore take multidirectional space.
The possibility of obtaining better results along with handling short and long edges
is greater.

5-34
 Table 5.13. Tour length results and relative errors (deviation) on several TSP instances.

IACO ACS

TSP Optimum Best length Average tour Relative errors Best length Average tour Relative errors
problems (1) (2) length (deviation) ((2)−(1))/(1) (3) length (deviation) ((3)−(1))/(1)

ch130 6110 6123 6175.3 0.21% 6144 6200.6 0.56%

ch150 6528 6528 6573.9 0% 6548 6600.83 0.31%
d198 15 780 15 815 15 894.03 0.22% 15 900 15 994.77 0.76%
kroB100 22 141 22 141 22 183.33 0% 22 146 22 278.5 0.02%
kroC100 20 749 20 749 20 789.6 0% 20 753 20 906.1 0.02%

5-35
kroD100 21 294 21 294 21 379.6 0% 21 309 21 584.7 0.07%
kroE100 22 068 22 068 22 135.77 0% 22 116 22 284.67 0.22%
kroA150 26 524 26 524 26 806.5 0% 26 820 27 189.47 1.08%
pr76 108 159 108 159 108 303.5 0% 108 304 108 723.4 0.13%
pr124 59 030 59 030 59 110.67 0% 59 076 59 228.23 0.08%
pr152 73 682 73 682 73 772.73 0% 73 818 74 243.67 0.18%
pr226 80 369 80 377 80 628.16 0.01% 80 524 80 959.67 0.19%
rat195 2323 2339 2356.17 0.69% 2352 2379.27 1.25%
Modern Optimization Methods for Science, Engineering and Technology
 Modern Optimization Methods for Science, Engineering and Technology

Accuracy results
100.50

100.00

99.50
Accuracy %

99.00
ACS
98.50
IACO
98.00 Optimal

97.50

TSP instances

Figure 5.22. Accuracy of IACO and ACO performance compared to the optimal solution.

Table 5.14. The effect of the same starting city and random starting city.

Same starting city Random starting city

TSP Optimal Best length Average Relative Best length Average Relative
problem (1) (2) length error((2)−(1))/(1) (3) length error((3)−(1))/(1)

eil51 426 430 441.6 0.94% 428 438.5 0.47%

The effects when all ants start from same city randomly can be seen in table 5.14.
The eil51 (51-city) problem has been tested and accordingly we found the best
possible path using 100 iterations in 20 trials. For this data, the best result was
obtained when ants started from same city randomly. In the case of the best optimal
solution, the distance is greater than the optimal distance by 2.0. The results for
distances greater than 4.0 are also highlighted.

5.13.2 Analysis on an increasing number of ants versus number of iterations

One way to analyze the solution of tour length is using more iterations or more ants.
This case was tested using the eil51 TSP instance without using local search with 20
trials. As seen in the figure, the results seem vary random between the two extreme
values found in figure 5.23 (428 and 455) and in figure 5.24 (436 and 466). In
figures 5.23 and 5.25, the former figure shows the result of testing with 10 ants in
each iteration and the latter figure presents the testing result of using 100 ants in each
iteration. It can also be seen that the tour length value is better when varying the
number of iterations than the number of ants.

5-36
 Modern Optimization Methods for Science, Engineering and Technology

Tour length on number of ants 10 & iterations 100

460

450
Tour length

440

430
TourLength
420 Average

410
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
Number of runs
Figure 5.23. Tour length result for the eil51 instance with 10 ants and 100 iterations.

Tour length result found on eil51

500
490
Best so far: tour length

480
470
460
450
440
430
420
410
400
1 10 19 28 37 46 55 64 73 82 91 100
Number of Iterations
Improved ACO

Figure 5.24. Tour length distance achieved for eil51 using 10 ants and 100 iterations.

The algorithm was run for 100 iterations, where each iteration used 10 ants and
the best solution was found after 60 iterations. The best value was quite close to the
optimum value (figure 5.26).
Summarized results for the eil51 TSP instance:

Best distance achieved: 428

Optimum solution: 426
Iteration which achieved best distance: 65

5-37
 Modern Optimization Methods for Science, Engineering and Technology

Tour length on number of ants 100 & iterations 10

470

460
Tour length

450

440 TourLength

430 Average

420
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
Number of runs
Figure 5.25. Tour length result on the eil51 instance with 100 ants and 10 iterations.

Tour length result found on eil51

490
480
Best so far: tour length

470
460
450
440
430
420
1 2 3 4 5 6 7 8 9 10
Number of Iterations
Improved ACO

Figure 5.26. Tour length distance achieved for eil51 using 100 ants and 10 iterations.

The algorithm was run for 10 iterations, where each iteration used 100 ants and
the best solution was found at 10 iterations. The best value was not far from the
optimum value.
Summarized results for the eil51 TSP instance:

Best distance achieved: 436

Optimum solution: 426
Iteration which achieved best distance: 10

5-38
 Modern Optimization Methods for Science, Engineering and Technology

Figure 5.27. Tour length results for an increasing number of ants.

Figure 5.28. Tour length results for an increasing number of iterations.

It can also be seen that the solution result is better when increasing the number of
iterations than the number of ants.
Another alternative way to analyze the varying number of ants and iterations is
also presented in the following figures, which suggest that the chance of obtaining a
good result would be higher. Figure 5.27 shows the analysis result for an increasing
number of ants throughout the iterations (for example, the first iteration uses 1 ant,
the second iterations use 21 ants, 41 ants, 61 ants and so on and the last iteration uses
181 ants). This test iterates 10 times. Figure 5.28 shows the result of increasing the
number of iterations to 200 iterations. The individual iteration uses 10 ants. The
average length of the increasing number of ants is 545.96 and that of the increasing
number of iterations is 457.89. It can be seen that increasing the number of iterations
obtain better quality results than increasing the amount of ants.

5-39
 Modern Optimization Methods for Science, Engineering and Technology

5.14 IACO comparison results

We discuss here the analysis of results obtained for the suggested algorithm, IACO,
both using CL and without it. Three TSP instances were tested, i.e. oliver30, eil51
and berlin52, which are highlighted in table 5.15 in four different columns. The
distance was found for 500 iterations among 10 trials in the first column, in
comparison to the optimal distance, which is considered as a benchmark distance.
The average length is reported in the second column and the best number of
iterations is in the third column. The average time is taken as 10 trials in the fourth
column where 500 iterations are there in each run or trial. Local search optimization
is used and convergence was found to be satisfactory (table 5.15).

Table 5.15. IACO comparison results.

IACO with candidate list IACO without candidate list

TSP Best Average Best Average Best Average Best Average

problem Optimal length length iteration time length length iteration time

oliver30 420 420 420.5 69 5.66 420 423.3 356 7.30

eil51 426 426 430.8 382 8.61 431 444.3 475 12.87
berlin52 7542 7542 7635.2 235 9.33 7542 7678.6 302 13.86

Tour Length Result on oliver30

500

480
Best so far: Tour Length

460

440

420

400

380
1 51 101 151 201 251 301 351 401 451 501
Number of Iterations
with canidate list without candidat list

Figure 5.29. Comparison of the tour length result of IACO with the candidate list and without the candidate
list for oliver30.

5-40
 Modern Optimization Methods for Science, Engineering and Technology

Tour Length Result on eil51

520

500
Best so far: Tour Length

480

460

440

420

400

380
1 51 101 151 201 251 301 351 401 451 501
Number of Iterations
with candidate list without candidate list

Figure 5.30. Comparison of the tour length result of IACO with the candidate list and without the candidate
list for eil51.

Figures 5.29–5.31 show the results with CL and without CL, respectively. We
have achieved considerable improvements in our searching ability, with the number
of iterations reduced and less time taken.

5.15 Conclusions
Research studies suggest that ACO is a modern and emerging research field in the
area of optimization and is used for a number of engineering problems, in particular
in the age of artificial life and operations research. One of the most important
applications is swarm intelligence, used in applications related to AI. The simu-
lations using ant colonies and exploring their behavior foraging for the food resulted
in the birth of ACO whose principal element is pheromone information. This
chapter has discussed the basics of ACO, establishing a metaheuristic structure
which provides a number of options for implementation in the design of algorithms.
Various alternative approaches are explained briefly, highlighting successful algo-
rithms such as MMAS and ACS. ACO algorithms are an emerging field and are
considered as state-of-the-art algorithms in solving combinatorial problems of
optimization.
One major contribution of the chapter is suggesting IACO as an improved ant
colony optimization algorithm, that is designed for addressing combinatorial
optimization problems such as TSPs. The search space, dynamic updates, compu-
tation time and best possible tours have been addressed well with a substantial
amount of results and discussions. The heuristic parameter and selection of the
values have been discussed in terms of how the different values affect the

5-41
 Modern Optimization Methods for Science, Engineering and Technology

Tour Length Result on berlin52

9200
9000
8800
8600
Best so far: Tour Length

8400
8200
8000
7800
7600
7400
7200
7000
6800
1 51 101 151 201 251 301 351 401 451 501
Number of Iterations

with candidate list without candidate list

Figure 5.31. Comparison of the tour length result of IACO with the candidate list and without the candidate
list for berlin52.

performance of the system. The suggested approach proved to produce much better
results compared to a single approach which is used to solve TSPs. Optimal results
are obtained and some solutions to the problems were found to be optimal. An
interesting finding regarding the time taken was made that suggests the time required
in finding the solution can be optimized if flexible and dynamic updates are properly
applied.

References
[1] Colorni A, Dorigo M and Maniezzo V 1991 Distributed optimization by ant colonies Proc.
of Ecal91—Eurpoean Conf. on Artificial Life (Paris, Amsterdam: Elsevier) pp 134–42
[2] Colorni A, Dorigo M and Maniezzo V 1992 An investigation of some properties of an ant
algorithm Proc. of the Parallel Problem Solving from Nature Conf. (PPSN 92) (Brussels,
Amsterdam: Elsevier) pp 509–20
[3] Shannon C E 1948 A mathematical theory of communication Bell Syst. Tech. J. 27 379–423
[4] Pintea C-M and Dumitrescu D 2005 Improving ant system using a local updating rule Proc.
of the Seventh Int. Symp. and Numeric Algorithms for Scientific Computing (SYNASC’05)
(Piscataway, NJ: IEEE)
[5] Wang C-X, Cui D-W, Zhang Y-K and Wang Z-R 2006 A novel ant colony system based on
Delauney triangulation and self-adaptive mutation for TSP Int. J. Inform. Technol. 12 89–99

5-42
 Modern Optimization Methods for Science, Engineering and Technology

[6] Hung K S, Su S F and Lee S J 2007 Improving ant colony optimization for solving traveling
salesman problem J. Adv. Comput. Intell. Intell. Inform. 11 433–42
[7] Gambardella L M and Dorigo M 1995 Ant-Q: a reinforcement learning approach to the
traveling salesman problem Proc. of the Twelfth Int. Conf. on Machine Learning (San
Francisco, CA: Morgan Kaufmann) pp 252–60
[8] Dorigo M, Maniezzo V and Colorni A 1996 The ant system: optimization by a colony of
cooperating agents IEEE Trans. Syst. Man Cyber. B 26 29–41
[9] Dorigo M and Gambardella L M 1997 Ant colony system: a cooperative learning approach
to the traveling salesman problem IEEE Trans. Evolut. Comput. 1 1–24
[10] Dorigo M and Gamgardella L M 1997 Ant colonies for the traveling salesman problem
BioSystems 43 73–81
[11] Dorigo M, Birattari M and Stuzle T 2006 Ant colony optimization—artificial ants as a
computational intelligence technique IEEE Comput. Intell. Mag. 1 28–39
[12] Dorigo M and Stützle T 2004 Ant colony optimization (Cambridge, MA: MIT Press)
[13] Dorigo M and Stützle T 2002 The ant colony optimization metaheuristic: algorithms,
applications, and advances Handbook of Metaheuristics ed F Glover and G Kochenberger
(Amsterdam: Kluwer)
[14] Randall M and Montgomery J 2002 Candidate set strategies for ant colony optimisation Ant
Algorithms, 3rd International Workshop, ANTS 2002, Proceedings ed M Dorigo, G di Caro
and M Sampels (Lecture Notes in Computer Science (including subseries Lecture Notes in
Artificial Intelligence and Lecture Notes in Bioinformatics) vol 2463) (London: Springer),
pp 243–49
[15] Stutzle T and Hoos H H 1997 MAX–MIN ant system and local search for the traveling
salesman problem Proc. of the 1997 IEEE Int. Conf. on Evolutionary Computation
(ICEC’97) (Piscataway, NJ: IEEE), pp 309–14
[16] TSPLIB 2005 TSPLIB: Library of Sample Instances for the TSP University of Heilderberg,
Department of Computer Science http://iwr.uniheidelberg.de/groups/comopt/software/
TSPLIB95/tsp/
[17] Hlaing Z C S S and Khine M A 2011 An ant colony optimization for solving traveling
salesman problem Int. Conf. on Information Communication and Management (ICICM),
IPCSIT vol 16 (Singapore: IACSIT)
[18] Hlaing Z C S S and Khine M A 2011 Solving traveling salesman problem by using improved
ant colony optimization algorithm Int. J. Inform. Educ. Technol. 1 404–49

5-43
 IOP Publishing

Modern Optimization Methods for Science, Engineering and

Technology
G R Sinha

Chapter 6
Application of a particle swarm
optimization technique in a motor imagery
classification problem
Rahul Kumar and Mridu Sahu

In many real-world problems one needs to find the best solution from all feasible
solutions to solve any particular problem. Finding the best among all solutions is a
basic theme of optimization. Minimization of the time and space to solve real
problems is still a challenging task in many areas such as the biomedical, behavior
and prediction sciences, etc. Optimization establishes relationships among problem
objectives, predefined constraints on these and the targeted variables. Optimization
techniques have been widely employed by researchers to improve the performance of
computers in many cognitive detections. This chapter presents the application of
particle swarm optimization (PSO) in motor imagery (MI). The dynamic state of the
brain is simulated with the action of different body parts in MI. The brain–computer
interface (BCI) provides a conduit between the brain and computer, and it performs
classification of MI action, which is very helpful for those people who are paralyzed
due to high-level spinal cord injury and are not able to perform any muscular
activities. The quality of signals and their performance on classifiers are still crucial
challenges in MI classification. For classification various steps are required, such as
preprocessing, feature extraction, selection, etc. The presented chapter includes
various feature extraction techniques and selections for the classification of MI. For
the recording of brain activity electroencephalography (EEG) is normally used. This
is the non-invasive mode of recording brain signals. The brain signal from the
recording is noisy and complex data, and several preprocessing steps are involved to
improve the quality of signals. The transformation of the signal from one domain to
another is called signal transformation, which has been done in the current chapter
through wavelet transformation; this also helps in extracting features from these

doi:10.1088/978-0-7503-2404-5ch6 6-1 ª IOP Publishing Ltd 2020

 Modern Optimization Methods for Science, Engineering and Technology

signals. EEG signal nonlinearity is a serious problem in finding a solution for the
detection of many diseases and disorders. Fuzzy logic is a powerful tool for handling
nonlinearity. This chapter focuses on the classification of EEG signal using an
adaptive neuro-fuzzy inference system (ANFIS). The discrete wavelet transform
(DWT) has been used here for useful feature extraction.
Also, in this chapter, PSO has been applied for optimizing the network parameters
of ANFIS. In standard ANFIS all parameters are updated using a gradient descent
algorithm. The problem present in gradient descent is that when the search space is
large, the complexity of gradient computations is high. The PSO method is inspired by
the biological nature of bird flocking and fish schooling, which is applied here to tune
the parameters of the network. Normally PSO finds an iterative candidate solution for
the specified problem. It is a metaheuristic, which provides a set of rules that is high
level and independent of the problem. PSO can optimize the functional module, tune
the neuro and fuzzy systems as well as modify the rule set generated from various
systems. In the presented study, PSO minimizes the mean square error (MSE) by
tuning the parameters of membership functions of the fuzzy inference system (FIS)
and increases the performance for classification of right hand and foot movement
detections. PSO possesses a central advantage on the implementation side as it is easy
and requires very few parameters. Two models have been proposed: one is based on
PSO and the other is gradient descent for comparative analysis. The results confirm
that the PSO based model gives better accuracy than the others.

6.1 Introduction
Optimization plays an important role in our day-to-day lives. In many disciplines,
such as the scientific, social, economics and engineering fields, optimization is used
to find a desirable solution to problems by adjusting their parameters. The main aim
of optimization is finding the best solution among all feasible solutions that are
available to us. A feasible solution satisfies all constraints in optimization problems.
Currently the problems in optimization are multi-objective and multidisciplinary.
To solve that kind of complex problem, not only is gradient descent based
optimization used, but also the evolutionary algorithms such as genetic algorithms,
and particle swarm optimization (PSO) are employed. Over the years, many
methods have been developed for optimization. The PSO method is a popular
method of optimization which is inspired by the social behavior of birds and fish.
PSO is used in many fields, such as the medical field, engineering, social economics,
etc. In medicine, biomedics is one of the important fields which focuses on the study
of the brain and its behavior (how the brain works in different environments). In this
field, the design of a stable and reliable brain–computer interface (BCI), which
provides the basic connection between a brain and computer, is a major challenge.
To design a BCI, various problems need to be overcome. To solve these problems
PSO can be used. Channel selection is an important step in recording brain signals
and also a major challenge in recording brain signals. PSO is widely used for
selecting the minimum number of channels. During the recording of signals, a lot of
noise and artifacts are generated that affect the performance of the model. To

6-2
 Modern Optimization Methods for Science, Engineering and Technology

remove those artifacts and noise, PSO based filters are used. Feature extraction
methods are used to obtain the best possible characteristics from brain signal for
dimensionality reduction. Feature selection is a method which is applied after
extraction to filter out irrelevant and redundant features. PSO based feature
selection is a widely used technique to reduce the dimensionality.
Human motor activity is described by notable adaptability; a human can perform
a number of undertakings, for example, walk forward and backward, run, dance,
shuffle and produce a wide range of activities. We appear to have the option to
produce a practically limitless stream of movements to achieve objectives in the
surrounding environment [1]. The motor system is based on brain activity. However,
due to spinal cord injury and other neurological diseases the motor ability can be
damaged [2]. People may be paralyzed and they are not able to perform any
muscular activity. To restore the damaged motor function, a concept called MI,
which is fundamentally a mental task in which patients perform the motor task in
their mind without performing any physical task. This concept is used with a brain–
computer interface (BCI), which provides a connection between the brain and
external environment (figure 6.1) [3]. An MI based BCI performs the classification of
MI action such as movement of the hands and feet and other motor related action. It
provides one-way communication for a patient who is suffering from a motor
disability. BCI is categorized into two types: invasive and non-invasive. In invasive
BCI, electrodes are placed in the head using cranial surgery in order to compute
brain activity. This method provides a higher signal-to-noise ratio (SNR), but its
downsides are a higher infection rate and other side effects. To counter the negative
effects of the above method, non-invasive BCI was introduced, but it has a lower
SNR. This calls for the development of more precise, non-invasive BCI in order to
accurately measure brain activity.
In the BCI, people produce a different brain activity pattern by performing the
MI action and the pattern is identified and converted into a command and control

Figure 6.1. Brain–computer interface system.

6-3
 Modern Optimization Methods for Science, Engineering and Technology

signal by the system. Identification depends on the classification algorithms. The

performance of a BCI depends on the features and classification algorithm. To select
the most suitable classifier for designing the BCI, it is important to understand which
types of features we have to use. Various features are extracted in order to design the
BCI, such as band power (BP) [4], power spectral density value [5, 6], amplitude
values of EEG signals, autoregressive (AR) and adaptive autoregressive parameters
[7], time-frequency features and inverse model-based features [8].
In the BCI classifier there are two basic problems: first, the curse of dimension-
ality and, second, the bias–variance trade-off [3]. Solving these problems is the main
challenges for classification algorithms. This chapter focuses on the classification of
EEG signal using ANFIS and DWT, used here for the extraction of useful features.
Also, in this chapter, PSO has been applied for optimizing the network
parameters of ANFIS. In standard ANFIS all parameters are updated using a
gradient descent algorithm. The problem present in gradient descent is that when the
search space is large, the complexity of gradient computation is high. The PSO
method is inspired by the biological nature of bird flocking and fish schooling, which
is applied here to tune the parameter of the network. Normally PSO finds an
iterative candidate solution for a specified problem. It is a metaheuristic, which
provides a set of rules that is high level and independent from the problem. PSO is
also able to optimize the functional module, tune the neuro and fuzzy systems, as
well as modify the rule set generated from various systems.

6.1.1 Literature review

Previously, a lot of work has been done on MI and BCI application. Channel
selection for the classification of brain signals is crucial in designing a reliable and
stable BCI. PSO plays an important role in channel selection for recording signals,
filtering the signal with optimum parameters, and PSO is also used by some
researchers to find an optimum number of features, which is more relevant and
efficient in classification. A number of classification algorithms are used for EEG
signals, such as support vector machines (SVMs), neural networks (NNs), Bayes’
classifier, K-nearest neighbor classifier and linear discriminate analysis (LDA), etc.
To improve the performance and efficiency of a classifier, one needs to find the
optimal parameter for which the algorithm gives the best performance. Optimization
methods such as the PSO method, genetic algorithms and others are used for finding
these parameters.
BCI systems are generally not used as a real-time applications because of the low
accuracy of pattern recognition algorithms. In particular when the channel number
is very large, and with the increment of the channel. A poor arrangement of the
channels with high computation complexity influences the accuracy of the classifier.
To solve these problems different approaches have been proposed.
Wei and Wang [9] observed that when the number of channels is high, it is very
tedious to perform analysis on multi-channel EEG signals. It requires the prepara-
tion of a suitable recording, and a tedious calculation which will take a lot of
computation time. To solve this problem the authors proposed binary multi-

6-4
 Modern Optimization Methods for Science, Engineering and Technology

objective particle swarm optimization (BMOPSO) for optimizing the number of

channels. The principle objective of this work was to improve the performance of a
BCI by reducing the number of channels by considering the problem of channel
selection as a search problem. For the experiment they used a dataset which consists
of the EEG signals of five healthy subjects. They included three different MI tasks:
right hand, left hand and foot moment. The method successfully reduced the number
of channels without affecting the accuracy. BMOPSO was not only used for
minimizing the number of channels, but also used to maximize the mutual
information. The authors found that, through reducing the number of channels,
preparation of the recording was easier and the classification algorithm took less
time to compute. The proposed method also has some disadvantages, such as a slow
learning convergent velocity and, due to the adopting capability of the gradient
method, converging to local minima cannot be avoided.
Hasan and Gan [10] describe the effective multi-objective PSO technique to
resolve the channel selection problem for BCI. The authors describe two methods for
channel selection: multi-objective particle swarm optimization and sequential
floating forward search. In this paper the authors used dataset1 of BCI
Competition IV. The dataset consisted of EEG signals for seven subjects, each
subject having three MI tasks (right hand, left hand and foot). From each subject
AR features were calculated and applied as input into an LDA. If the dimension of
features was more than 20, then principal components analysis was applied for
dimension reduction. In the end, the authors concluded that MOPSO selected a
smaller number of channels. However, there are some channels that are selected by
both methods as the most important channel for classification.
Lv and Liu [11] observed that when a common special pattern algorithm is used
for extracting features from multi-channels, and if the number of channels is large,
then the common spatial pattern (CSP) algorithm will address an overfitting
problem that is inconvenient for clinical operation. On the other hand, most of
the channels contain noisy and redundant data that affect the performance of the
classifier. To solve that problem the CSP filters’ discrimination and channel number
were integrated in a single unit. They then used BPSO for selecting the group of
channels. In the experiment the authors used the BCI2003 dataset IV and BCI2005
dataset I. The authors found that maximum accuracy was achieved with
9–14 channels.
Kumar et al [12] used a combination of PSO and rough set theory to calculate a
minimum number of related features from large dimension extracted features. In
order to perform classification on multiclass MI they used a neighborhood rough set
classifier. The authors followed four phases of preprocessing and classification. In
the first phase the BCI-MI dataset was processed, which was gathered from
computerized devices. For this work, the MI datasets were taken from BCI
Competition IV Dataset IIa. In the second phase, raw EEG signals were processed
in order to improve the quality of signals; for this purpose filter and feature
extraction methods were used. In the third step, feature selection was used to reduce
the dimensionality of features from the set feature extracted from the wavelet
transform method. In this step, the combination of PSO rough set theory was used to

6-5
 Modern Optimization Methods for Science, Engineering and Technology

select the dominating features. The fourth step used k-fold cross-validation, in which
the dataset is divided into two parts, one for training other data for testing and then
performing the classification algorithm. The result demonstrates that the proposed
technique improves precision more than other comparablee classification algo-
rithms. By using this method the authors achieved significant improvements in terms
of sensitivity and positive predictive value. Using the proposed method, the MI task
classification accuracy reached up to 80.9848%.
Hsu [13] proposed an automatic artifact elimination method where feature
selection is performed using a quantum-behaved PSO algorithm from a set of
extracted features. Optimal features are selected through QPSO and these selected
features are used as input into the SVM. The author proposed automatic artifact
elimination to remove the EOG artifacts and improve the classification accuracy. At
the same time a number of features such as spectral power, asymmetry ratio, MFFV
and other features are calculated and then combined together. After that QPSO is
used to select the features which enhance the classification accuracy.
Xu et al [14] used PSO based CSP method for feature extraction. Two parameters,
frequency band and time interval, affect the performance of CSP. So, using PSO the
authors found the optimal values for the parameters which improve the discrim-
inative ability of CSP.
Filters are used to improve the quality of signals by removing the noise and
irrelevant data. The performance of a filter depends on parameters such as the
frequency band time interval and others. To improve the performance such
parameters need to be optimized. For that, several researchers used PSO to optimize
the parameters.
Ahirwal et al [15] studies an adaptive noise canceler for improvement of EEG
signal filtering. Different versions of PSO were used to design the adaptive noise
canceler and perform a comparative analysis on different parameters for a varied
range of particles and inertia weights.

6.1.2 Motivation and requirements

BCIs provide communication for those people whose motor cells are damaged and
are not able to perform any muscular activity. By using a BCI they can communicate
with the outside environment and control devices such as wheelchairs, artificial hands
and feet, etc. Designing a BCI is not an easy task, as there are several challenges such
as the selection of channels for recording, the dimensionality and complexity of
signals, and the performance of the classifier which is used for recognition of imagery
action or brain pattern. The aims of this chapter are as follows:
1. To reduce the dimensionality of the data by applying the feature extraction
method.
2. To design the ANFIS model for classification which is a combination of NN
and fuzzy logic and is capable of handling nonlinearity of data.
3. To improve the performance of ANFIS; the parameters of ANFIS are tuned
using PSO.

6-6
 Modern Optimization Methods for Science, Engineering and Technology

The remaining sections of this chapter are as follows: in section 6.2 the
fundamentals of PSO are briefly explained and in section 6.3 the proposed model
is described. Some results of experiments are presented in section 6.4 and we provide
some remarks and conclusions in the last section.

6.2 Particle swarm optimization

PSO was designed by Kennedy and Eberhard in 1995 [16] inspired by the communal
deportment of bird flocking and fish schooling. PSO is a stochastic optimization
method that computes the global best solution using a number of search points
called particles. In a PSO system, there is a group of particles, and all particles are
moving around a multidimensional search space, and each particle shares its
location with other particles and they update their position. Through communica-
tion, they reach the best position [17]. The direction of each particle is defined on the
basis of a set of particles neighboring the particle and previous experience. The
group of particles is organized according to some sort of communication structure or
topology, such as the star topology, mesh-topology and wheel topology. There are
different versions of PSO and each version is a modified version of a previous one
with a high convergence property. The PSO was mainly created for real-valued
problems, but most problems are based on discrete values, where the domain of a
variable is finite. It includes problems such as integer programming, scheduling and
routing. In 1997, Kennedy and Eberhart developed a discrete binary version of PSO
for solving discrete optimization problems [18], where the location of each particle is
represented by a binary number (0 or 1) and velocity is defined as the probability of
change in the state of the particle. This algorithm has a better convergence rate
compared to other versions of PSO. In PSO, most of the time it is trapped in the
local minima. To solve this problem [19] presented a new version of PSO called
guaranteed convergence PSO which includes extra particles that search the region
around the current global best to find the final global best. In [20] the authors
proposed a new algorithm personal best position PSO where velocity is updated with
a new equation. The modification has been carried out through vanishing the global
best term within the velocity update equation of the standard PSO. Many PSO
versions are not able to solve multiple objective problems because of the way in
which they exchange information in order to find the global solution. To solve that
problem niche PSOs were introduced to solve complex problems with multi-
objective functions; one version of PSO is explained in [21].
PSO looks like an evolutionary algorithm, and it has emerged that PSO is
practically identical to the genetic algorithm [22, 23]. A considerable amount of
exploration has demonstrated the viability of PSO in improving the persistent and
discrete advancement issues. There are different applications of PSO in different
fields such as NN training, electric power systems, telecommunications, control,
data mining, design, combinatorial optimization, signal processing, and many
others. However, when the objective function has a large dimension then most
often PSO fails to find an optimal solution. A fall in local minima is the main issue in
PSO and to handle this issue several modifications were proposed in the essential

6-7
 Modern Optimization Methods for Science, Engineering and Technology

particle driven equation [24, 25], to keep the particle velocity or accelerate them
using an adaptation technique or randomized method. These proposed modifica-
tions work well and have the ability to avoid the fall of local minima. However, PSO
still does not guarantee finding a global solution with high dimension search space
advantages.

6.2.1 The mathematical model of PSO

PSO is computational, effective and easier to implement compared to other
mathematical algorithms and evolutionary algorithms [26]. PSO initializes with a
population of a random solutions called particles. Each particle has an individual
velocity. All the particles are moving around the search space with a velocity that is
dynamically adjusted as per previous behavior. Every particle has a tendency to
move toward the best solution in the search space. In PSO once problems are
defined, particles are placed in the search space and each one evaluates the fitness
value for a given function at its current location. After that particles update their
current locations and best according to information shared by all other particles.
The next iteration takes place when all particles have moved. With each iteration
particles reach close to the optimal solution. In PSO the solution is associated with
two vectors, position (Xi ) and velocity ( Vi ). In D dimensionsl search space Xi = [xi1,
xi2 , xi 3, xi 4 , …, xiD ] and Vi = [vi1, vi2 , vi 3, vi 4 , …, viD ] are the vectors which are
associated with each particle i. X1 show the location of the first particle. It is
necessary to define the movement direction and speed to update the position vector.
Vector velocity (Vi ) can be used for doing the same thing:

Vit+1 = w Vit + c1r1 ( Pit − Xit ) + c2r2 ( G t − Xit ) (6.1)

Xi t+1 = Xi t + Vi t +1, (6.2)

where
c1 and c2 are positive constraints,
r1 and r2 are two randomly generated numbers with a range [0, 1],
w is inertia weight,
Pi t is the best position of particle i achieved based on its own experience,
G t is the best particle position among all,
t is the iteration index.

The position vector represents the location of a particle in the search space and
the velocity shows in which direction and at what speed (inertia of movement) the
particles move. In each iteration, these two vectors are updated with the above two
equations. The new velocity is calculated by the multiplication of the current
velocity and a variable called inertia w, and has c1 and c2 positive constraints for the
last two components in the velocity vector. The first term is called the inertia weight
as it maintains the current velocity and moment direction. The second component is
the cognitive component, it is also known as an individual component because it
considers the particle best position and current position. The third component is a

6-8
 Modern Optimization Methods for Science, Engineering and Technology

social component; here the particle calculates the distance between its current
position and the best position for the whole swarm.
The cognitive and social component have a great impact on the moment of a
particle and this impact can be changed by tuning the coefficient. The inertia weight
tunes the exploration and exploitation and is normally decreased from 0.9 to 0.4 or
0.2. In real-word problem when searching for global optima, variation in the inertia
parameter will balance the exploration and exploitation. To demonstrate the impact
of the parameter in PSO, let us consider a simple objective function called a sphere
(equation (6.3)). Calculate the average objective value for a different range of inertia
weights ([0.9 0.2], [0.9 0.3], [0.9 0.5] [0.5 0.3]) with c1 = c2 = 2. It was observed that
the function quickly converges for adaptive inertia weight in the range [0.9 0.2].
Figure 6.2 shows the convergence of the function for various ranges of inertia
weight,
f (x ) = ∑(x 2 ). (6.3)

Let us also demonstrate the effect of c1 and c2 on the convergence of functions.

We consider w = [0.9 0.2] for c1 = 0, c2 = 2 and c1 = 2, c2 = 0. It was observed that the
functions quickly converge toward an optimal solution for c1 = 0 and c2 = 2 and
when c1 = 2 and c2 = 0 the particle only searches one part of the search space.
Because there is no information exchange between the particles, they are unable to
find the global best solution. Figure 6.3 shows convergence of a function for various
values of c1 and c2 .

Figure 6.2. Impact of inertia weight in PSO.

6-9
 Modern Optimization Methods for Science, Engineering and Technology

Figure 6.3. Impact of c1 and c2 parameter.

6.2.2 Constraint-based optimization

The constraint is a limit applied to the particle, which prevents us from going forever
in a particular direction. The constraint applies to multiple variables. Each search
space has a boundary and we can use the constraint to simulate this. The set of
constraints indicates the upper and lower bound of all parameters. In this way, we
can define the boundary for the search space. If we apply the limit to the variable it
significantly reduces the size of the search space. The constraint divides the search
space into two parts: feasible and infeasible. In figure 6.4, all shaded areas are
feasible. An infeasible solution indicates that if we find a solution that violates the
constraint then we do not consider it as a desirable solution and it should be avoided.
When a large number of constraints divides the space into too many parts, the
infeasible region is dominant over the feasible region. This means the solutions for
optimization algorithms are more likely to be in a feasible region. To solve this
problem the optimization algorithm should find an isolated region and identify
promising ones. The constraint optimization problem is represented as the following
nonlinear problem.
Minimize
f (x1, x2 , x3, x4, … , xn−1, xn), (6.4)

6-10
 Modern Optimization Methods for Science, Engineering and Technology

Figure 6.4. Feasible and infeasible regions.

subject to
gi (x1, x2 , x3, x4 … , xn−1, xn) ⩾ 0 , i = 1, 2, 3, … , m (6.5)

hi (xi , x2 , x3, x4, … , xn−1, xn) ⩾ 0, i = 1, 2, 3, … , p (6.6)

lbi ⩽ xi ⩽ ubi , i = 1, 2, 3, … , n . (6.7)

To handle the constraints, penalty functions are used. These convert the objective to
a constraint, meaning that constraints are handled inside the objective function
without any modification in the algorithm [27]. The penalty function is represented
as follows.
Minimize
f (x ) + σp(x ), (6.8)
where

p(x ) = { 0 if x ∈ S
+ ∞ otherwise
. } (6.9)

The penalty function returns a 0 if solution x is feasible, moreover in this case there
is no penalty. In the case that x is an infeasible solution, the penalty function
penalizes it by returning a greater objective value than the actual.

6.3 Proposed method

In the proposed method we perform classification of EEG signal using PSO based
ANFIS. Before classification, signals are first acquired from the brain using a 10–20
electrode placement method on the scalp and after that segmentation is performed
according to the given trial, and then, using DWT, approximate and detailed
coefficients are calculated which are used as features for classification. Figure 6.5
shows the proposed method.

6-11
 Modern Optimization Methods for Science, Engineering and Technology

Figure 6.5. Proposed work.

Here 18 features were extracted as detailed and approximate coefficients,

these features were applied as input into the ANFIS. The method is briefly
explained below.

6.3.1 Materials and methods

In this section we describe the methods and materials used in the proposed work.
The complete details are provided below.

6.3.1.1 Dataset
The dataset was acquired from BNCI Horizon-2020 [28]. The dataset related to two-
class MI, where signals are recorded for the right hand and foot movement. As per
the Graz-BCI training paradigm, a single session is carried out for recording,
training and feedback. Sessions have a total of eight runs and of the eight runs five
are used for training and three are used for testing. Of a total of 20 trials for a set, per
class 50 trials are for training and 30 are for testing. In this recording, process
participants had the task of performing sustained (5 s) kinesthetic MI as instructed
by a cue. The EEG signal is recorded using an Ag/Agcl electrode, with 512 Hz
sampling frequency using a bio-signal amplifier. Electrodes are positioned as per the
10–20 standard. Dataset processing is shown in figure 6.6.

6.3.1.2 Feature extraction

A feature is a distinctive measurement, transform extracted from a set of patterns.
Features are used to differentiate each class level with the goal of dimensionality
reduction, which reduces the computational complexity. In feature extraction we
remove redundant and irrelevant data and improve the quality of signals. EEG
signals are random in nature and have a lot of artifacts and noise so, instead of
applying these signals as input into a classifier, first it is important to improve the
quality of the signals. For this, first signals are filtered by both the time domain as
well as spatial domain and then a feature extraction method is deployed. For a better
representation of EEG signals, there are various feature extraction methods such as
fast Fourier transform (FFT), wavelet transform, principle component analysis

6-12
 Modern Optimization Methods for Science, Engineering and Technology

Figure 6.6. Description of dataset processing.

(PCA), independent component analysis (ICA), auto-regression (AR), etc. Among

all these methods WT is a widely used feature extraction method because it provides
both time and frequency analysis. WT represents EEG signals more accurately. In
our work we used a DWT feature extraction method, which is described in
detail below.
DWT is a widely used method, with a multiresolution analysis capability in which
we perform analysis on both time and frequency [29]. It facilitates the multi-
resolution analysis of EEG signals. The signals are split into two parts using a high
pass filter and a low pass filter simultaneously. These filters satisfy some condition
called the admissibility condition. The approximate and detailed coefficient is
computed at each level and the approximate coefficient is again split into two parts,
the approximate and detailed coefficient. Repeat this process until a specific level is
reached. Wavelet coefficients are an efficient representation of EEG signals [30]. In
the context of computation time and complexity of signals, DWT is better compared
to other methods. Figure 6.7 represents the nth level of DWT decomposition.
The dilation function ∅j,k (n) relies on the low pass filter and the wavelet function
ψj ,k (n) relies on the high pass filter which is denoted as

∅j ,k (n ) = 2 j /2h(2 j n − k ) (6.10)

ψj ,k(n ) = 2 j /2g(2 j n − k ), (6.11)

where n = 0, 1, 2, 3, …, N − 1, j = 0, 1, 2, 3, …, J − 1, k = 0, 1, 2, 3, …, 2j − 1, and
J = log2M . M represents the length of the EEG signal. Approximate coefficient Ai
and detailed coefficient Di are calculated using the following equations:
1
Ai = ∑n x(n) × ∅j,k (n) (6.12)
M

1
Di = ∑n x(n) × ψj, k(n). (6.13)
M

6-13
 Modern Optimization Methods for Science, Engineering and Technology

Figure 6.7. DWT decomposition of n levels.

6.3.2 Classification
ANFIS is one of the neuro-fuzzy models. Both NN and fuzzy logic are individual
systems. When increasing the complexity of the model it is difficult to calculate the
membership function and fuzzy rules encourage the development of another
approach that has the tendencies of both systems. This approach is called the
adaptive neuro-fuzzy system, which gives the advantages of both NN and fuzzy
logic. The advantage of fuzzy logic is that fuzzy rules describes real-world problems
in a better way. The second thing is interpretability; this means it is easy to explain
how every single output value of the fuzzy system is generated. The problem with the
fuzzy system is that it needs expert knowledge to create rules and it takes a long time

6-14
 Modern Optimization Methods for Science, Engineering and Technology

to tune the parameter (membership function). Both problems arise simply because it
is not possible to tune the fuzzy system. However, for NN, it trains well but it is
remarkably difficult to use earlier learning about the considered system. To over-
come the disadvantages of these two systems, some researchers combine both the
NN and fuzzy logic systems. The hybrid system ANFIS was proposed by Jang [31].

6.3.2.1 Structure of ANFIS

ANFIS is the standard method for transforming human knowledge or experience
into the rule base and database of the FIS [31]. ANFIS extracts features from a
dataset and, according to a given error criterion, adjusts the system parameters. For
training, a back-propagation gradient descent algorithm in combination with the
least squares method is used. FIS is a basic part of ANFIS, which employs fuzzy
rules for modeling the qualitative aspects of human knowledge. FIS consists of five
functional blocks (figure 6.8).
ANFIS first used the Sugeno model for creating FIS, and for better under-
standing we have to consider the following if–then rules:
Rule 1: if (x is A1) and ( y is B1) then ( f1 = p1 x + q1y + r1)

Rule 2: if (x is A2 ) and ( y is B2 ) then ( f2 = p2 x + q2y + r2 ),

where x , y and Ai , Bi represent the given input and antecedent, respectively. The fi
includes f1 and f2 in terms of pi , qi and ri whose values are found through the training

Figure 6.8. Functional block of FIS [31].

6-15
 Modern Optimization Methods for Science, Engineering and Technology

Figure 6.9. Structure of ANFIS [31].

process. The structure for two rules is shown in figure 6.9. Circles represent fixed
nodes while squares represent adaptive nodes. Five layers of ANFIS are described.
In the first layer, it receives incoming crisp input and determines rules for each
crisp input. All the node are adaptive nodes. This layer is also called the fuzzification
layer. The output of this layer is given as

Oi1 = μ Ai (x ) i = 1, 2, (6.14)

where μ Ai (x ) is any membership function that accepts input x and provides a

membership value, μ Ai (x ) for the Gaussian membership, the function is given by

⎧ ⎛ x − c ⎞2 ⎫
⎪ ⎪

μ Ai (x ) = exp⎨ −⎜ i
⎟ ⎬, (6.15)
⎩ ⎝ 2ai ⎠ ⎭
⎪ ⎪

where ai and ci represent the parameters of the membership function.

Nodes are fixed in the second layer which multiplies the incoming signals and
sends the product out, given as
wi = μ Ai (x ) · μBi (y ) i = 1, 2. (6.16)

This output is known as the firing strength of the rule.

Nodes in the third layer are fixed as well. Input coming from the second layer is
normalized in this layer. We call the output of this layer the normalized firing
strength and it is computed as

6-16
 Modern Optimization Methods for Science, Engineering and Technology

wi
Oi3 = wi = i = 1, 2. (6.17)
w1 + w2
The nodes are adaptive in the fourth layer and their output is the product of the
input from the third layer and first-order polynomial as
Oi4 = wl ( pi x + qi y + ri ) i = 1, 2. (6.18)

The fifth layer consists of a single fixed node which aggregates all incoming signals.
Hence, the overall output is
2 2
(∑i = 1wi fi )
Oi5 = ∑wi fi i = 1, 2. (6.19)
i=1
w1 + w2

6.3.2.2 Learning method

The process of learning in ANFIS is very similar to a NN, where the back-
propagation algorithm is used to learn; learning means adjusting the weight of
synapses between neurons according to the error criteria. In ANFIS there are two
types of parameters: antecedent and consequence. These parameters play an
important role in the derivation of rules. On the basis of rules, inputs are mapped
into the output. In the rules the ‘IF’ part consists of the antecedent parameter and
the ‘THEN’ part contains the consequence parameter. The ANFIS learning
algorithm is used in order to tune the set of parameters.
In ANFIS, a blend of back-propagation and least squares estimation is used for
learning. The algorithm is used to learn both the separate parameters. Back-
propagation is used to learn the antecedent (input) parameter and least squares is
used for the consequence parameter. The learning procedure consist of two passes:
the forward pass and backward pass. In the forward pass the least squares method
optimizes the consequent parameter with the predetermined antecedent parameter.
After finding the optimal consequent parameter, the backward pass begins imme-
diately. In the backward pass the gradient decent method is used to tune the
antecedent parameter. The final output is evaluated by using the consequent
parameter. The output is changed according to the antecedent parameter by means
of the standard back-propagation algorithm. The output is calculated in terms of
MSE and root MSE. Error is calculated in the following way:
1 n
MSE = ∑ ( y − ok )2
n k=1 k
(6.20)

1 n
RMSE = ∑ k=1( yk − ok )2 . (6.21)
n

6.3.2.3 ANFIS-PSO
Neuro-FIS is optimized by adapting the membership function parameter and
consequent parameter, so objective functions are minimized. The back-propagation

6-17
 Modern Optimization Methods for Science, Engineering and Technology

algorithm is a widely used adaptation method of FIS that is used to recursively solve
the optimization problem. The difficulty with this algorithm is trapping in local
minima. To overcome this problem, evolutionary algorithms such as the genetic
algorithm and PSO have been used [32, 33]. In this proposed method we perform
two-class MI classification based on ANFIS-PSO. Here PSO is used to tune the
antecedent and consequent parameters of ANFIS and examine the performance
accuracy. The total number of the antecedent parameter is the sum of all the
parameters in each membership function. The antecedent parameter (ai , ci ) is related
to the Gaussian membership function shown in equation (6.15). The consequent
parameters are output parameters ( pi , qi and ri ) that are shown in equation (6.18).
The mathematical model of ANFIS is quite similar to the NN where the weights are
updated according to the error criteria. In ANFIS the antecedent parameters are
calculated which are associated with the membership function. In the rules shown in
equation (6.1) the ‘IF’ part contains the membership function parameters and the
‘THEN’ part contains the linear output variable. These parameters are evaluated as
weights; the performance of ANFIS depends on the structural parameter and
training related parameter. In this chapter, an ANFIS system is trained by evolu-
tionary algorithms (EAs) such as PSO in order to optimize the classification error.
The whole process of this method is shown in figure 6.10. As is evident, the
optimization of ANFIS is based on four steps:
1. Initialize the ANFIS system parameters.
2. Evaluate the output of the initial ANFIS and calculate the MSE by
comparing the actual output value and target value.

Figure 6.10. Flow diagram of ANFIS-PSO.

6-18
 Modern Optimization Methods for Science, Engineering and Technology

3. The output is not too good so in order to minimize the MSE a learning
process should be performed.
4. Using PSO, membership functions are optimized in order to minimize the error.

Steps 2 and 3 are repeated until a satisfactory result is obtained.

Moreover, if the number of parameters is large, then the computation times for
these algorithms are also high. In this chapter, we used PSO algorithms to tune the
membership function. The procedure is shown in figure 6.10.

Algorithm 1. Adopted research methodology

1. Data loading: Load the EEG data acquired from BNCI Horizon-2020.
2. Generating basic FIS: Generate the initial fuzzy inference system.
3. Training ANFIS: Tune the parameters of the Gaussian membership function and optimize the
MSE value using PSO.
4. Classification: Apply ANFIS-PSO on the training and testing data.
5. Performance evaluation: Calculate the MSE value for classification on the training and
testing data.
END.

6.2.3.4 Optimization problem

ANFIS has two main parameters: antecedent and consequent. To define the problem,
we first need to identify these parameters. As mentioned above, the antecedent
parameters (ai , ci ) and consequent parameters ( pi , qi and ri ) play an important role in
the optimization of ANFIS. For each membership function, all the values of parameters
are calculated and saved as vectors and applied for training for optimized values. In a
vector the number of the variable represents the dimension of the search agent.
In the next step we define the objective function. Objective functions are the
fitness of a problem; if we obtain an acceptable value of the objective function
training should stop, otherwise we repeat the process. In ANFIS the performance is
measured terms of the MSE value. For every sample, MSE values are calculated. To
evaluate the overall performance, calculate the average of MSE for all training
samples. The average MSE value is calculated according to the equation
1 n
MSE = ∑ ( y − ok )2 ,
n k=1 k
where yk is the targeted output and ok is the actual output predicted by the ANFIS.
The length of the training sample is denoted by n and k denotes the kth training
sample. So the optimization problem can be defined as
Minimize: f (x ) = MSE. (6.22)

So the above function is considered as an objective function and we minimize its

value using PSO.

6-19
 Modern Optimization Methods for Science, Engineering and Technology

Table 6.1. Average of the approximate and detailed coefficients for channel C3 of subject 1.

Dataset D1 D2 D3 D4 D5 A5

S01 4.0387 4.2069 3.2889 4.9283 7.3535 76.1838

S02 0.5686 1.0407 2.1799 4.8142 6.9992 84.3975
S03 2.3672 3.1341 4.4636 9.0005 13.4167 67.6180

Table 6.2. Description of the parameters used in PSO.

Description Considered value/size

Population size 25
Inertia weight 1
Personal learning coefficient 1
Global learning coefficient 2

Table 6.3. Performance of ANFIS.

Performance of ANFIS

Training Testing

Subject MSE RMSE MSE RMSE

S01 1.4724e−08 0.000 121 34 8.5311 2.9208

S02 8.3516e−07 0.000 913 87 2.9647 1.7278
S03 3.4101e−09 5.8396e−05 4.1703 2.0421

6.4 Results
The acquired EEG signals are segmented according to a given trial in the dataset.
We use DWT for the decomposition of EEG into five level wavelet coefficients. The
average approximate and the detailed coefficients for channel C3 subject 1 are
shown in table 6.1.
The extracted features are used as input for ANFIS and ANFIS-PSO. For
ANFIS-PSO different parameters are used during the training. The description of all
the parameters is provided in table 6.2. In order to perform the classification, the
data are divided into two parts, one for training and the other for testing; here 80%
of the data are used for training and 20% of the data are used for testing.
The predicted RMSE and MSE values for three subjects are shown in tables 6.3
and 6.4, figures 6.11 and 6.12 shown the training and testing performance of subject
1 for ANFIS and figures 6.13 and 6.14 shown the training and testing performance
of subject 1 for ANFIS-PSO.

6-20
 Modern Optimization Methods for Science, Engineering and Technology

Table 6.4. Performance of ANFIS-PSO.

Performance of ANFIS-PSO

Training Testing

Subject MSE RMSE MSE RMSE

S01 0.222 39 0.471 58 0.349 15 0.590 89

S02 0.205 23 0.453 02 0.317 52 0.563 02
S03 0.225 36 0.474 72 0.283 06 0.532 03

Figure 6.11. Training performance of ANFIS.

Figure 6.12. Testing performance of ANFIS.

6-21
 Modern Optimization Methods for Science, Engineering and Technology

Figure 6.13. Training performance of ANFIS-PSO.

Figure 6.14. Testing performance of ANFIS-PSO.

6.5 Conclusion
In this chapter classification of two-class MI action of the right hand and foot
moment is performed using an ANFIS and ANFIS-PSO. Before classification, some
preprocessing methods such as segmentation and feature extraction are also
experimented. To extract features the DWT method was used, so in this way we
calculate six features for each channel (C3, Cz, C4) and a total of 18 features for
each subject. These feature are applied as input for the classifier. It is found that PSO
performed well for tuning the parameters of ANFIS with increased accuracy.

6-22
 Modern Optimization Methods for Science, Engineering and Technology

References
[1] Mulder T 2007 Motor imagery and action observation: cognitive tools for rehabilitation J.
Neural Transm. 114 1265–78
[2] Kübler A, Kotchoubey B, Kaiser J, Wolpaw J R and Birbaumer N 2001 Brain–computer
communication: unlocking the locked Psychol. Bull. 127 358
[3] Lotte F, Congedo M, Lécuyer A, Lamarche F and Arnaldi B 2007 A review of classification
algorithms for EEG-based brain–computer interfaces J. Neural Eng. 4 R1
[4] Pfurtscheller G, Neuper C, Flotzinger D and Pregenzer M 1997 EEG-based discrimination
between imagination of right and left hand movement Electroencephalogr. Clin. Neurophysiol.
103 642–51
[5] Chiappa S and Bengio S 2004 HMM and IOHMM modeling of EEG rhythms for
asynchronous BCI systems European Symposium on Artificial Neural Networks ESANN
[6] Millan J R and Mouriño J 2003 Asynchronous BCI and local neural classifiers: an overview
of the adaptive brain interface project IEEE Trans. Neural Syst. Rehabil. Eng. 11 159–61
[7] Penny W D, Roberts S J, Curran E A and Stokes M J 2000 EEG-based communication: a
pattern recognition approach IEEE Trans. Rehabil. Eng. 8 214–5
[8] Qin L, Ding L and He B 2004 Motor imagery classification by means of source analysis for
brain–computer interface applications J. Neural Eng. 1 135
[9] Wei Q and Wang Y 2011 Binary multi-objective particle swarm optimization for channel
selection in motor imagery based brain–computer interfaces 2011 4th International
Conference on Biomedical Engineering and Informatics (BMEI) vol 2 (Piscataway, NJ:
IEEE), pp 667–70
[10] Hasan B A S, Jan J Q and Zhang Q 2010 Multi-objective evolutionary methods for channel
selection in brain computer interface: some preliminary experimental results IEEE Congress
on Evolutionary Computation (CEC) pp 1–6
[11] Lv J and Liu M 2008 Common spatial pattern and particle swarm optimization for channel
selection in BCI 3rd International Conference on Innovative Computing Information and
Control (Piscataway, NJ: IEEE), p 457
[12] Kumar S U and Hannah Inbarani H 2017 PSO-based feature selection and neighborhood
rough set-based classification for BCI multiclass motor imagery task Neural Comput. Appl.
28 239–58
[13] Hsu W-Y 2013 Application of quantum-behaved particle swarm optimization to motor
imagery EEG classification Int. J. Neural Syst. 23 1350026
[14] Xu P, Liu T, Zhang R, Zhang Y and Yao D 2014 Using particle swarm to select frequency
band and time interval for feature extraction of EEG based BCI Biomed. Sig. Process.
Control 10 289–95
[15] Ahirwal M K, Kumar A and Singh G K 2012 Analysis and testing of PSO variants through
application in EEG/ERP adaptive filtering approach Biomed. Eng. Lett. 2 186–97
[16] Eberhart R and Kennedy J 1995 A new optimizer using particle swarm theory MHS’95.
Proceedings of the Sixth International Symposium on Micro Machine and Human Science
(Piscataway, NJ: IEEE), pp 39–43
[17] Omrana M, Engelbrechta A and Salmanb A 2005 Particle swarm optimization method for
image clustering Int. J. Pattern Recogn. Artif. Intell. 19 297–322
[18] Kennedy J and Eberhart R C 1997 A discrete binary version of the particle swarm algorithm
1997 IEEE International Conference on Systems, Man, and Cybernetics. Computational
Cybernetics and Simulation vol 5 (Piscataway, NJ: IEEE), pp 4104–8

6-23
 Modern Optimization Methods for Science, Engineering and Technology

[19] Bergh F V D and Engelbrecht A P 2006 A study of particle swarm optimization particle
trajectories Info. Sci. 176 937–71
[20] Singh N and Singh S B 2012 Personal best position particle swarm optimization J. Appl.
Comput. Sci. Math. 12 69–76
[21] Kumar A, Singh B K and Patro B 2016 Particle swarm optimization: a study of variants and
their applications Int. J. Comput. Appl. 135 24–30
[22] Poli R, Kennedy J and Blackwell T 2007 Particle swarm optimization Swarm Intell. 1 33–57
[23] Angline P 1998 Evolutionary optimization versus particle swarm optimization: philosophy
and performance difference Int. Conf. Evolution. Program. 1447 601–10
[24] Parsopoulos K and Vrahatis M 2001 Particle swarm optimizer in noisy and continuously
changing environments Artificial Intelligence and Soft Computing ed M H Hamza
(Anaheim, CA: IASTED/ACTA Press) pp 289–94
[25] Hendtlass T 2005 WoSP: a multi-optima particle swarm algorithm 2005 IEEE Congress on
Evolutionary Computation vol 1 (Piscataway, NJ: IEEE), pp 727–34
[26] Silva A, Neves A and Costa E 2002 Chasing the swarm: a predator–prey approach to
function optimisation Proc. of the MENDEL2002––8th Int. Conf. on Soft Computing Brno,
Czech Republic
[27] Parsopoulos K E and Vrahatis M N 2002 Particle swarm optimization method for
constrained optimization problems Front. Artif. Intell. Appl. 76 214–20
[28] Steyrl D, Scherer R, Förstner O and Müller-Putz G R 2014 Motor imagery brain–computer
interfaces: random forests vs regularized LDA-non-linear beats linear Proc. of the 6th Int.
Brain-Computer Interface Conf. pp 241–4
[29] Ren W, Han M, Wang J, Wang D and Li T 2016 Efficient feature extraction framework for
EEG signals classification 2016 Seventh International Conference on Intelligent Control and
Information Processing (ICICIP) (Piscataway, NJ: IEEE), pp 167–72
[30] Jahankhani P, Kodogiannis V and Revett K 2006 EEG signal classification using wavelet
feature extraction and neural networks IEEE John Vincent Atanasoff 2006 International
Symposium on Modern Computing (JVA’06) (Piscataway, NJ: IEEE), pp 120–4
[31] Jang J-S R 1993 ANFIS: adaptive-network-based fuzzy inference system IEEE Trans. Syst.
Man Cybernet. 23 665–85
[32] Catalão J P d S, Pousinho H M I and Mendes V M F 2011 Hybrid wavelet-PSO-ANFIS
approach for short-term electricity prices forecasting IEEE Trans. Power Syst. 26 137–44
[33] Ghomsheh V S, Shoorehdeli M A and Teshnehlab M 2007 Training ANFIS structure with
modified PSO algorithm Mediterranean Conference on Control and Automation (Piscataway,
NJ: IEEE), pp 1–6

6-24
 IOP Publishing

Modern Optimization Methods for Science, Engineering and

Technology
G R Sinha

Chapter 7
Multi-criterion and topology optimization using
Lie symmetries for differential equations
Sailesh Kumar Gupta

Lie addressed the unification problem of all the apparently different integration
methods for differential equations and investigated some elegant and simple
algebraic structures, commonly known as continuous transformation groups, which
hold the key to the problem. This connection with differential equations naturally
led to the study of the topological manifolds associated with group structure.
Interestingly, the manifolds needed for solving different aspects related to differ-
ential equation systems can be thought to be locally some open subsets defined in an
n-dimensional Euclidean space where one can choose the local coordinates freely.
The key concept turns out to be that of symmetry or the infinitesimal generator
associated with the group. Once the generators associated with a system of equations
become available, many applications become immediate. For ordinary differential
equations (ODEs), the symmetries and their generators always lead to the possibility
of going a step forward in the integration of the equation, and if one has a
sufficiently large number of symmetries, the complete integration of the differential
equation is guaranteed in most cases by the method of quadrature alone. For most
partial differential equations (PDEs), one cannot write down the general solution
but has to rely on various ansatz such as the similarity solutions, traveling waves,
separable solutions and so on. These ansatz methods usually lead to the problem of
solving some ODE generated in the process. These methods involve nothing more
than looking for solutions that are invariant under a particular group of symmetries
associated with the equations. Further, it is well documented in the associated
literature that any linear combination of the infinitesimal generators associated with
a given system of equations also leads to some group invariant solutions for the same
system of equations. Thus, it is easy to understand that we are left with a situation
for dealing with infinite group invariant solutions for the system. Hence, we have to

doi:10.1088/978-0-7503-2404-5ch7 7-1 ª IOP Publishing Ltd 2020

 Modern Optimization Methods for Science, Engineering and Technology

devise a mechanism to classify all the inequivalent solutions only and the corre-
sponding problem leads to the optimization problem for the group invariant
solutions.

7.1 Introduction
Accurate formulations of different physical problems can be achieved with the help
of nonlinear ordinary or partial differential equations (ODEs and PDEs). A vast
literature exists on all of the solution methods, such as analytical, numerical or
approximate, but the concept of Lie symmetry groups provides a complete analytical
understanding of all these apparently different solution methods. Lie [1] began the
unification of all these different methods and founded the concept of continuous
transformation groups or Lie groups [2–4]. He found that the mathematically
challenging nonlinear conditions generated under the action of a continuous trans-
formation group, to preserve the invariance of a differential equation system, could
be systematically replaced by some simple linear conditions associated with the
generators of the group. The connection of Lie groups with the differential equations
naturally led to the study of the topological manifolds [5–7] associated with the group
structure. These manifolds can be thought, at least locally, to be some open subsets of
an n-dimensional Euclidean space in which the coordinates can be chosen freely, at
least locally. The idea of the infinitesimal generator of the group plays an important
role in the process. The infinitesimal generators are like a vector field of a given
manifold and their flows coincide with the one-parameter groups that they generate.
Thus, the starting point of the study of a differential equation using the group
techniques basically requires familiarity with two aspects. First, the idea of flow in a
vector field and, second, the infinitesimal criterion of invariance for the system under
the action of group transformations. The prolongation formula of vector fields
becomes the main tool in calculating the symmetry groups of a system [2, 3], and
requires the introduction of spaces which also includes the derivatives of different
dependent variables along with the independent and dependent variables as coor-
dinates of the manifold. The space of all these variables is called the jet space. Once
the symmetry groups are calculated for a system, many applications become
immediate. For the ODEs, the existence of Lie symmetries always leads to the
possibility of going a step forward in the integration of the equation, and if one has a
sufficiently large number of symmetries, then the complete integration of the
differential equation is guaranteed in most cases by the method of quadrature alone.
Thus, it becomes very important to ascertain the possibility of maximum reduction in
order of the equation in this case. When it comes to the PDEs, usually one obtains
several arbitrary functions which are not very useful in practical applications. For
most PDEs, one cannot write down the general solution but has to rely on various
ansatz such as the similarity solutions, traveling waves, separable solutions and so on.
These ansatz methods usually lead to solving some ODEs generated in the process.
These methods involve nothing more than looking for solutions that are invariant
under a particular symmetry group of transformations for the system. Further, it is a
well known fact that for a given system of differential equations possessing

7-2
 Modern Optimization Methods for Science, Engineering and Technology

symmetries, one can have a system of infinite group invariant solutions. This is due to
the fact that any linear combination of the symmetry generators of the system again
leads to a symmetry associated with the same system. Hence, we have to devise a
mechanism to classify all the in-equivalent solutions only, and the problem leads to
the optimization of the associated group invariant solutions for the system. The
classified solutions give the optimal group invariant solutions [2, 3].
Thus, in this chapter we will try to address the different issues discussed so far
with examples. We start with a brief introduction of the fundamentals of topological
manifolds and then establish the connection between the groups and differential
equations [3]. We discuss the methods to calculate the group invariant solutions for
the differential equations and classify them to establish the corresponding optimized
solutions. Finally, concluding remarks are given with further reading suggestions.

7.2 Fundamentals of topological manifolds

A Lie group should have the properties of a topological group, that is, one must be
able to assign a topology under which the group action and its inverses are both
continuous functions. A methodological discussion to explore the analytical
structures associated with the Lie groups requires familiarity with the concept of
an analytic manifold. Lie’s works lead to the concept of defining a Lie group in
terms of analytic manifolds whose objects are the differential equations themselves.

7.2.1 Analytic manifolds

The theoretical studies of connected analytic manifolds can be systematically carried
out by treating them as some smooth surfaces in a Euclidean space Rn . Each point in
the space is assigned coordinates with the property of possessing continuous
mapping with its inverse, which is also continuous. Analytically, if X ⊂ Rn is
some open subset of manifold Z and Ψ: X → Y is any diffeomorphism such that
Y ⊂ Rn , then Ψ is a C ∞ map with C ∞ inverses. Thus, for different elements on X,
some equivalent objects on Y will always exist. The coordinate charts Θρ : Xρ → Yρ
assign the structure of a topology to the manifold Z, such that each open subset
W ⊂ Yρ ⊂ Rn , Θ−ρ 1(W ) is an open subset of Z. Further, if α ≠ α˜ are points in Z, then
one can have open sets X containing α and X̃ , containing α̃ such that X ∩ X˜ = Φ
(null set) and further, if the overlap function Θσ ∘ Θ−ρ 1: Θρ(Xρ ∩ Xσ ) → Θσ (Xρ ∩ Xσ ) is
smooth, then Z is an analytic manifold.

7.2.1.1 Coordinate transformations and maps

If coordinate charts are used to formulate different elements on a manifold, then it
must be ensured that the definition is coordinate independent. The problem
reduces to that of studying objects under a coordinate transformation. In actual
situations, it is found that the studies can be carried out more easily in local
coordinates. If one is able to find some special coordinate charts in which the
objects under study take some simple form then many complicated calculations can be
considerably simplified. A local coordinate map Θρ : Xρ → Yρ with a diffeomorphism

7-3
 Modern Optimization Methods for Science, Engineering and Technology

Ψ: Yρ → Y˜ρ , of Rn is called a coordinate transformation. For the manifolds Z and M

with the property of smoothness, then this property is retained by the functional map
L: Z → M if, for Θρ : Xρ → Yρ ⊂ Rm on Z and Θ̃σ : X˜σ → Y˜σ ⊂ Rn on M , then the
composite map defined by Θ̃σ ∘ L ∘ Θ−ρ 1: Rm → Rn is also smooth. If a manifold
possess a subset inheriting all its properties, then the subset is a submanifold.

7.2.2 Lie groups and vector fields

7.2.2.1 Group
A group can be defined as a set S endowed with the following properties:
• Group composition law: For any two elements a, b ∈ S taken in a definite
sequence we obtain a product element c ∈ S such that
a · b = c ∈ S.

• Associativity: The product of any definite sequence of three elements is

unambiguous, i.e. if a, b, c ∈ S , then the following holds:
a · (b · c ) = (a · b ) · c .

• Existence of identity: There is a special element e ∈ S , such that

e · k = k = k · e, k ∈ S.

• Inverses: For each k ∈ S , there exists a unique inverse k −1, such that
k · k −1 = k −1 · k = e.

7.2.2.2 Lie group

An s-parameter Lie group can be defined as a topological group S such that S is also
a smooth manifold endowed with the following maps:
• Θ: S × S → S , Θ(a, b ) = a · b, a, b ∈ S .
• i: S → S , i (a ) = a−1, a ∈ S .

7.2.2.3 Local transformation group

Each of the transformations of the manifold Z is associated with a Lie group S, if for
each k ∈ S we can find an associated map from Z to itself. Analytically, if S is a
group for the local transformations on Z endowed with ϒ ⊂ S , such that
{e} × Z ⊂ ϒ ⊂ S × Z , it is associated with a smooth map Π: ϒ → Z such that:
• If a, b ∈ S , and z ∈ Z , then

a · (b · z ) = (a · b ) · z . (7.1)

• For all the elements z ∈ Z ,

e · z = z. (7.2)

7-4
 Modern Optimization Methods for Science, Engineering and Technology

• For each a ∈ S and z ∈ Z , if the product a · z is defined, then the following

hold
a − 1 · (a · z ) = z .

7.2.2.4 Vector fields

In a given m-dimensional manifold Z we can define a tangent space to all the points
z ∈ Z . The objects defining the space will be the collection of all the tangent vectors
at z belonging to different curves passing through the point. It is denoted by the
m-dimensional vector space TZ∣z and the elements defining the basis are
{ ∂
∂z1
,…,
∂
∂z m }. The corresponding tangent bundle of manifold Z is given by
TZ =
TZ∣z . ∪
z⏟
∈Z
If A ∈ Z , then A∣z ∈ TZ∣z defines the corresponding tangent vector at points when
z ∈ Z and A∣z is also assumed to vary smoothly. If written in terms of (z1, … , z m ),
i.e. the local coordinates, then we can write the vector field expression
∂ ∂ ∂
A∣z = ξ1(z )
∂z1
+ ⋯ + ξ m (z ) m = ∑ξi(z ) ∂ .
∂z zi
The coefficient functions ξ(z ) have smoothly varying properties. The hydrodynamic
steady fluid flow can be taken as a good example. In this case, the points z ∈ Z can
be associated with the vector A∣z , which physically gives the velocity of the fluid
through the point z. For the vector field A, the curve z = ϕ(ε ) gives a smooth
parametrized integral curve. The related parametrized tangent vector and the
corresponding vector field A coincide at a given point and are given by
dϕ
= A∣ϕ(ε ) .
dε
The points z = ϕ(ϵ ) = (ϕ1(ϵ ), … , ϕ m(ϵ )) in local coordinates, satisfy the ODEs

dz k
= ξ k(z ), k = 1, … , m . (7.3)
dε
Now it is guaranteed from the theory of differential equations that for smoothly
varying ξ k (z ) we can have a unique solution to the initial value problem of equation
(7.3) with the initial condition
ϕ(0) = z0. (7.4)
Again, if we are given smoothly varying functions h(z ) with domain z ∈ Z , then h · A
becomes a smooth vector field, with (h · A)∣z = h(z )A∣z and h · A = ∑h(z )ξ i (z ) ∂ i .
∂z

7.2.2.5 Flow of vector fields and infinitesimal generators

Given a smoothly varying curve C belonging to the manifold Z, parametrized
by ϕ(ϵ ): P → Z , where P is a sub-interval in R , we obtain m smooth functions

7-5
 Modern Optimization Methods for Science, Engineering and Technology

ϕ(ϵ ) = (ϕ1(ϵ ), … , ϕ m(ϵ )). The corresponding tangent vector to the manifold Z at
1 m
each point z = ϕ(ϵ ) of C is given by the derivative ddϕε = ( ddϕε , … , ddϕε ). Now, for the
field A ∈ Z , the curve C through the point z ∈ Z is given by α (ϵ, z ). We call it the
flow of A and it has the properties
α(β, α(ϵ , z )) = α(β + ϵ , z ), z ∈ Z, (7.5)
for all β, ϵ ∈ R with
α(0, z ) = z (7.6)
and
d
α(ϵ , z ) = A∣α(ε,z ) (7.7)
dϵ
for all ϵ where defined. Comparing equations (7.5) and (7.6) with equations (7.1) and
(7.2), we find the similarity between the group action of the Lie group R on Z and
the respective flow of A. The transformations defined above are popularly called
one-parameter transformation groups and A is is called the infinitesimal generator.
Now, making use of Taylor’s theorem, we can write
α(ϵ , z ) = z + ϵξ(z ) + O(ϵ 2 ),
where (ξ1(z ), … , ξ m(z )) are the coefficients of A.
The curves C for the vector field A are the orbits of the one-parameter group
action. Conversely, if a one-parameter transformation group, α (ε, z ), acts on Z then
the corresponding generator is obtained by calculating (7.7) at ε = 0 and is given by
d
A∣z = α(ε , z ).
dε ε=0

In usual practice the flow of A is referred to as exponentiation and can be expressed as

exp(ϵA)z ≡ α(ϵ , z ).

This notation allows us to rewrite the properties of the flow of A listed above for all
z ∈ Z as
d
exp[(β + ϵ )A]z = exp(β A)exp(ϵ A)z , exp(0A)z = z , [exp(ϵ A)z ] = A∣exp(ε A)z .
dϵ

7.2.2.6 Action of vector flows on functions

Let A ∈ Z be a vector field and h: Z → R be a mapping. Now the flow generated by
A changes h and the infinitesimal change in the local coordinates can be calculated
using Taylor’s theorem as
h(exp(ϵ A)z ) = h(z ) + ϵ A(h)(z ) + O(ϵ 2 ).

7-6
 Modern Optimization Methods for Science, Engineering and Technology

The process of differentiation can be continued and if substituted into the Taylor
series assuming convergence, we obtain the Lie series for the action of the flow on h
and this can be written as
∞
ϵk k
h(exp(ϵ A)z ) = ∑ A (h)(z ).
k=0
k!

7.3. Differential equations, groups and the jet space

We consider a differential equations system Λ with the set of equations given by
Eν(z , v(n)) = 0, ν = 1, … , l
involving p independent variables z = (z1, …, z p ), and q dependent variables
v = (v1, …, vq ) with v(n) signifying the nth order derivatives of v with respect to z.
Let the respective solutions of the system be given by v = h(z ). Let
Z = (z1, …, z p ) = R p be the space representing the independent variables, and
V = (v1, …, vq ) = R q represent the space of the dependent variables. A symmetry
group is a local group of transformations S acting on X ⊂ Z × V , which is some
open subset, such that S transforms solutions of E to its own solutions. If the graph
of the function v = h(z ) is given by
Γh = {(z , h(z )): z ∈ Ξ} ⊂ Z × V ,
then the transformation of the graph Γh by g can be written as
g . Γh = {(z˜ , v˜ ) = g . (z , v): (z , v) ∈ Γh}.
Now, for a proper analytical treatment we need to reduce the domain of Ξ and h, so
that the transform g. Γh = Γh˜ gives the transformed function v˜ = h˜ (z˜ ) for all
elements g near the identity. Thus, we can say that the function h̃ is the transform
of h by g and can be written as h˜ = g. h.
Now the main task for studying systems like Eν is to establish a workable criterion
to find out the symmetry group for a given system of differential equations with the
help of some group transformations. This criterion will be the infinitesimal trans-
formations requiring an appropriate geometrical setting. This idea leads to the
visualization of a prolonged space of the space Z × V , so that the final space includes
the various partial derivatives occurring in the system in addition to the independent
and dependent variables. Now, let the κth order derivative of h(z ) be represented by
∂ κh(z )
∂J h(z ) = ,
∂z j1∂z j2 … ∂z jκ
where J = (j1 , j2 , …, jκ ) is the set of integers with 1 ⩽ jκ ⩽ p indicating the
derivatives that are being considered, not necessarily in a given order. Further, let
the space V (n) = V × V1 × ⋯ × Vn include the derivatives of the functions v = h(z ).
The maximum order of the derivative will be n. The resulting total space Z × V (n) is
called the nth jet space. Generally, we are interested only in the system Λ defined on

7-7
 Modern Optimization Methods for Science, Engineering and Technology

X ⊂ Z × V , where the nth jet space has the definition, X (n) ≡ X × V1 × ⋯ × Vn .

Now, it is straightforward to generalize that if v = h(z ) is the graph in X, then the
nth prolongation pr (n)h(z ) has to be the graph in the nth jet space X (n). Now, to make
matters clear we see an example. If, p = 2, q = 1 with v = h(i , j ), we obtain the
second prolongation
⎛ ∂h ∂h ∂ 2h ∂ 2h ∂ 2h ⎞
v(2) = pr (2)h(i , j ) = (v ; vi , vj ; vii , vij , vjj ) = ⎜h ; , ; , , ⎟
⎝ ∂i ∂j ∂i 2 ∂i ∂j ∂j 2 ⎠
all evaluated at (i , j ). Thus, treating E as the map E : Z × V (n) → Rl , such that the
mapping leads to the l-dimensional Euclidean space from the space Z × V (n), then
we can write
ΛE = {(z , v(n)): E (z , v(n)) = 0} ⊂ Z × V (n).
Hence we can say v = h(z ) is a solution if
Γ (hn) ≡ {(z , pr (n)h(z ))} ⊂ ΛE = {E (z , v(n)) = 0}.

7.3.1 Prolongation of group action and vector fields

The action of S on the nth jet space X (n) leads to pr (S ), which is the nth prolongation of
S. Thus, its action on the function v = h(z ) should lead to the transformed function
v˜ = h˜ (z˜ ). This symmetry condition needs to be connected to the geometric condition
relating to that of the corresponding sub-variety ΛE , so that it becomes invariant of
pr (n) S . Now keeping in line with the definition of a group transformation, the
prolongation of the corresponding infinitesimal generators can be written as
d
pr (n) A∣(z,v (n) ) = pr (n)[exp(ϵ A)(z , v(n))]
dε ε=0

for any (z, v(n) ) ∈ X (n). Now, any vector field in X is given by
p q
∂ ∂
A= ∑ξi(z , v) ∂ + ∑ ϕα(z , v) ∂ .
i=1
zi α=1
vα

Hence we can say that ξ (i )(z, v ) and ϕα(z, v ) determines the coefficient of the nth
prolongation of A and can be written as
p q
∂ ∂
pr (n)A = ∑ξi(z , v) + ∑∑ϕαJ (z , v) ∂ .
i=1
∂z i α=1 J
vJα

7.3.2 Total derivatives of vector fields and general prolongation formula

It is clear from the discussions so far that we need explicit expressions for the
prolonged vector field, pr (n) A . Once this is achieved then we can make a connection
between the infinitesimal criterion of invariance of system Λ and the corresponding

7-8
 Modern Optimization Methods for Science, Engineering and Technology

symmetry groups S attached to it. Now, the group transformations gε = exp(ϵ A) for
the vector fields A can be written as
(z˜ , v˜ ) = gε(z , v) = (Θε(z ), v)
with the components Θi (z ) satisfying the equation

d Θiε(z )
= ξi (z ).
dz ε=0

Now differentiating the expression with respect to the variable ϵ at ϵ = 0, we obtain

the infinitesimal generator pr (1) gε . For the vector field A we obtain the first
prolongation
p p
∂ ∂
pr A = ∑ξ (z ) i +
(1) i
∑ϕ j (z , v(1)) ∂ , (7.8)
i=1
∂z j=1
vj

where
p
d ∂Θ−κ ε
ϕ j (z , v(1)) = ∑ (Θε(z ))vκ ,
dε ˜j
ε=0 κ= 1 ∂z

with the assumption gε−1 = g−ε in the domain of definition. As the functions are
assumed to be smooth, the order of the differentiation can be interchanged and,
simplifying, we can write
p
∂Θ κ
ϕ j (z , v , vz ) ≡ ϕ j (z , v(1)) = −∑ vκ ,
κ= 1
∂z j
which provides the expression for pr (1) A in (7.8). Now the above expression can be
written in terms of the total derivative of ϕ with respect to z as
∂ϕ ∂ϕ
ϕ j (z , v(1)) = Dj ϕ(z , v) = j
+ vj .
∂z ∂v

7.3.2.1 Definition
Given H (z, v(n) ), the expression
q
∂H ∂H
Di H = + ∑∑vJα/i ∂ , (7.9)
∂z i α= 1 J
vJα
∂v α k +1 α
where for J = (j1 , …, jκ ) the expression vJα/i = Ji = i∂ j v j represents the ith
∂z ∂z ∂z ⋯ ∂z κ
total derivative of H . In (7.9) the summation domain of J goes up to order n.

7.3.2.2 Theorem 7.1

Let
p q
∂ ∂
A= ∑ξi(z , v) ∂ + ∑ ϕα(z , v) ∂
i=1
zi α=1
vα

7-9
 Modern Optimization Methods for Science, Engineering and Technology

be a vector field in Z, so that the corresponding nth prolongation

q
∂
pr (n)A = A + ∑∑ϕαJ (z , v) ∂
α= 1 J = 1
vJα

becomes a vector field in the jet space Z (n), with 1 ⩽ jκ ⩽ p, 1 ⩽ κ ⩽ n , then the
coefficient functions ϕαJ of pr (n) A are given by

⎛ p ⎞ p
ϕαJ (z , v(n)) = DJ ⎜⎜ϕα − ∑ξiviα⎟⎟ + ∑ξivJα/i , (7.10)
⎝ i=1 ⎠ i=1

with viα = ∂v α /∂z i etc.

Proof. The theorem will be proved in general by the very simple and effective
method of the induction principle. We note that the second jet space X (1+1) belongs
to a subspace of the first jet space (X (1) )(1). To be precise, we note that the first-order
derivative of vJα gives the second order derivative of the same. Similarly, the space
X (n+1) can be treated as a subspace of the nth jet space because its first jet space is
given by (X (n) )(1). Thus, we can say that the vector fields on the manifold X (n−1) are
given by pr (n−1) A , which can then be prolonged to (X (n−1) )(1) by the use of first-order
prolongation formula. The resulting vector field is now restricted to Z (n) and its
subspaces which in turn will determine the expression for pr (n) A . In the process the
new nth order coordinates in (X (n−1) )(1) are given by vJα/κ = ∂vJα /∂z κ where
J = (j1 , …, jn−1 ), 1 ⩽ κ ⩽ p and 1 ⩽ α ⩽ q . At this point we use the definition of
the total derivative so that the coefficient of ∂/∂vJα/κ , in (pr (n−1) A)(1), becomes
p
ϕαJ /κ = DκϕαJ − ∑DκξivJα/i . (7.11)
i=1

Now if we are able to prove that (7.10) solves (7.11) in closed form then the proof is
complete. Using induction, we find
⎧
⎪
⎛ p ⎞ p p ⎫
⎪
ϕαJ /κ = Dκ⎨ D J ⎜⎜ϕα − ∑ξiviα⎟⎟ + ∑ξivJα/i − ∑DκξJα/i ⎬
⎪ ⎪
⎩ ⎝ i=1 ⎠ i=1 i=1 ⎭
⎛ p ⎞ p p
= DkDJ ⎜⎜ϕα − ∑ξiviα⎟⎟ + ∑(DκξivJα/i + ξivJα/iκ ) − ∑DκξivJα/i
⎝ i=1 ⎠ i=1 i=1
⎛ p ⎞ p
⎜ i α⎟
= DκDJ ⎜ϕα − ∑ξ vi ⎟ + ∑ξivJα/iκ ,
⎝ i=1 ⎠ i=1
∂ 2v Jα
with vJα/iκ = . Thus, the form of ϕαJ /κ is like (7.10), which completes the proof. □
∂z i ∂z κ

7-10
 Modern Optimization Methods for Science, Engineering and Technology

7.3.3 Criterion of maximal rank and infinitesimal invariance for differential equations
The given system Λ of equations Eν will possess the Jacobian matrix
JE (z, v(n) ) = (∂Eν /∂z i , ∂Eν /∂vJα ) and if E (z, v(n) ) = 0, then the maximal rank con-
ditions state that the matrix is of maximum rank with rank l .

7.3.3.1 Theorem 7.2

Let every infinitesimal generator A ∈ S on manifold X satisfy the equation
pr (n)A[Eν(z , v(n))] = 0, ν = 1, … , l whenever E (z , v(n)) = 0, (7.12)
then S will also be the symmetry group for the system of equations Λ defined in X, if
the system Λ of equations Eν is of maximal rank.

Proof. Using theorem 7.1 and the condition of maximal rank for the associated
Jacobian matrix, the proof is immediate [3].

7.3.4 Differential equations and symmetry groups

Theorem 7.2, along with formula (7.12) gives an effective analytical tool to calculate
the symmetry group S of any system Λ . The key idea is to start with an arbitrary
infinitesimal generator A. The coefficients ξ and ϕ have a functional dependence on z
and v. The corresponding prolonged vector field pr (n) A will have the transformed
coefficients ϕαJ (z, v ). These transformed coefficients involve the partial derivatives of
ξ i and ϕα . Hence, one can say that the infinitesimal criterion of invariance (7.12)
involves z, v, the z derivative of v, ξ i , ϕα and their higher order partial derivatives
with respect to z and v. The natural constraints of the system allow us to eliminate
any dependencies among the derivatives of v . The remaining unconstrained partial
derivatives of v can then be equated to zero. The resulting PDE system is a set
involving the functions ξ i and ϕα . These equations form the symmetry determining
equations for the system Λ . The corresponding general solution generates the
infinitesimal symmetries for the system Λ . Once the symmetries are available we
can use their algebraic properties for different applications. Using exponentiation of
the given vector fields, the general explicit expressions for the symmetry groups can
be calculated. We will illustrate the procedure with the following example.

7.3.4.1 The Korteweg–de Vries equation (KdV)

This equation [3] has two (p = 2) independent variables z ≡ (z, t ) and one (q = 1)
dependent variable v and is given by
E (z , v(3)) ≡ vt + vvz + vzzz = 0, (7.13)
where vt = ∂v /∂z, vzzz = ∂ 3v /∂z 3 etc. Using theorem 7.2, we can write any vector field
as
∂ ∂ ∂
A = ξ (z , t , v ) + η(z , t , v) + ϕ(z , t , v) .
∂z ∂t ∂v

7-11
 Modern Optimization Methods for Science, Engineering and Technology

This generates a one-parameter infinitesimal group of (7.13) if

pr (3)A[E (z , v(3))] = 0 when, E = 0.
Simple calculations give
ϕt + ϕ zzz + vϕ z + vz ϕ = 0 (7.14)

whenever v is a solution of (7.13). In (7.14) the different expressions for ϕ represent

the coefficients of prolongations of A with ϕ zzz representing the coefficients of ∂/∂vzzz
in pr (3) A , etc. Now the coefficient ϕ zzz is given by

ϕ zzz = Dz3ϕ − vz Dz3ξ − vt Dt3η − 3vzz Dz2ξ − 3vzt Dz2η − 3vzzz Dzξ − 3vzzt Dxη .

Similar expressions can be obtained for other coefficients as well. Once this is done
one needs to put all these in (7.14) while making sure to replace vt by −(vvz + vzzz ) in
all the expressions before simplifying. The working rule to analyze these equations is
to start first from the solution of the coefficients of highest order derivatives equated
to zero. Proceeding in this manner we see the coefficient of vzzt is Dzη. Thus, the
solution of Dzη = 0 gives the solution that η depends on t only. Next the coefficient
of vzz2 gives ξv = 0, which gives the result ηt = 3ξz from the coefficient of vzzz . Thus,
we obtain the result ξ = 13 ηt z + σ (t ). Next the coefficient of vzz is zero thus making ϕ
linear in v and v a function of t only. Next vz gives the equation
−ξt − v(ϕv − ηt ) + v(ϕv − ξz ) + ϕ = 0,

and the terms without any derivatives of u give

ϕt + ϕzzz + vϕz = 0.

The final general solution gives

ξ = c1 + c3t + c4z ,
η = c2 + 3c4t ,
ϕ = c3 − 2c4v ,

where c are arbitrary constants and hence we obtain the four-dimensional vector
fields for the KdV given by
∂
A1 =
∂z
∂
A2 =
∂t
∂ ∂
A3 = t +
∂z ∂v
∂ ∂ ∂
A 4 = 3t + z − 2v .
∂t ∂z ∂v

7-12
 Modern Optimization Methods for Science, Engineering and Technology

7.3.5 Differential invariants and the group invariant solutions

7.3.5.1 Definition
For the action of S on X ⊂ Z × V , the nth order differential invariant of S is given
by a smooth function Δ: X (n) → R . If pr (n)g · (z, v (n) ) is defined then the expression
for Δ for all g ∈ S is given by:
Δ(pr (n)g · (z , v (n))) = Δ(z , v (n)), (z , v (n)) ∈ Z (n).

7.3.5.2 Group invariant solutions and steps for calculations

A solution v = h(z ) of the system Λ is also the S-invariant solution, if the invariance
is maintained for all the elements in S. Precisely, for each g ∈ S , the functions h(z )
and g. h are the same on their common domains and a solution v = h(z ) has a
locally S-invariant subset given by the graph Th ≡ {(z, h(z ))} ⊂ X . Now we give the
algorithmic steps for finding such a solution with s-dimensional orbits of the group.
• Using the basic prolongation formula and the criterion of invariance, find all
the vector fields A for the given system.
• The system Λ is reduced to a transformed system depending on the new p − s
independent variables and choosing s = p − 1, find all the s-dimensional
subgroups of S.
• For a chosen symmetry group S, the set of all functionally independent
invariants are constructed and are divided into two categories of new
independent yi = yi (z, v ) and new dependent variables u j = v j (z, v ),
respectively.
• Next we solve for the p − s of the original independent variables z and denote
the solution by z̃ . Similarly, all the v are also solved in variables y, u and the
remaining variables are sorted out as zr . Further, the z-derivatives of any S-
invariant dependent variable v are transformed in terms of new variables and
their derivatives as well as zr , using the chain rule. All of them are then
substituted in the original equation E (z, v(n) ) = 0 so that we find the reduced
system E /S (y, v(n) ) = 0, which will be independent of zr . The solutions
u = f (y ) can be reverted back to the original variables and gives the S-
invariant solutions v = h(z ) for the original system.
• Finally, we repeat all the steps for all the different elements of the symmetry
group S. The result is a complete set of group invariant solutions for the
original system Λ .

Now we give an example of group invariant solution for the KdV equation. We
consider the group of scaling symmetries (z, t , v ) → (εz, ε3t , ε−2v ), generated by the
generator A4 of the KdV equation. We can find functions F (z, t , v ) for any of its
generators Ai, such that they satisfy the equations Aj F ≡ 0. This equation can be
solved for each of the generators by the method of characteristics and can be written as
dz dt du
= = . (7.15)
ξ (z , t , v ) η(z , t , v ) ϕ(z , t , v)

7-13
 Modern Optimization Methods for Science, Engineering and Technology

The solutions give the invariants for the system corresponding to the generator
under consideration. The invariants are then treated as new variables for equation
(7.13). Using the procedure enumerated before the global invariants of A4 in the
1 2
upper half space (t > 0) for the KdV are given by y = t − 3 z and u = t 3 v. Now
treating u as the function u(y ) we obtain the reduced equation
1 2
u yyy + uu y − yu y − u = 0.
3 3
This equation gives solutions in terms of the second Painlevé transcendent [3] given
by equation
1 3 1
wyy = w + yw + κ
3 3
with u = wy − 61 w 2 and κ being a constant. Thus, the final similarity solution can be
written in terms of the second Painlevé transcend for the KdV.

7.4 Classification of the group invariant solutions and optimal

solutions
For each of the s-parameter subgroups H ⊂ S , of a differential equation system in
p > s independent variables, we will find group invariant solutions. It is a well
established fact [3] that there are always an infinite subgroups associated with a
given differential equation system and hence they have infinite group invariant
solutions. Thus, it becomes very important to classify the solutions so that only the
inequivalent solutions are sorted out for the system, thus giving the optimal group
invariant solutions. The process is fairly algorithmic and will be made clear using the
cylindrical Korteweg–de Vries (cKdV) equation as an example.

7.4.1 Adjoint representation for the cKdV and optimization of the group generators
The cylindrical Korteweg–de Vries (cKdV) equation originates in different physical
situations in non-planar cylindrical geometry and is given by
v
vt + + vvz + vzzz = 0. (7.16)
2t
The cKdV has the four-dimensional Lie point symmetry generators [8, 9] given by
∂
A1 =
∂z
∂ ∂ ∂
A2 = 3t +z − 2v
∂t ∂z ∂v
1 ∂ 1 ∂
A3 = 2t 2 + t− 2
∂z ∂v
1 ∂ 3 ∂ ∂
A4 = 2zt 2
∂z ∂t
1
( 1
+ 4t 2 + zt − 2 − 4vt 2
∂v
. )
7-14
 Modern Optimization Methods for Science, Engineering and Technology

The first step to obtaining optimal group invariant solutions of equation (7.16)
following the procedure given in [3] is to find the adjoint representation of the
generators. Now for any n-dimensional symmetry algebra S generated by the vector
fields {A1, A2, …, An}, we can have equivalent elements
⎧
⎪
n n ⎫
⎪
⎨
⎪
A= ∑ai Ai , w= ∑bi Ai⎬ ∈ S,
⎪
⎩ i=1 i=1 ⎭
if any one of the following conditions is satisfied:
• For g ∈ G we have a transformation Adg (w) = A , where Adg is the adjoint of
g and is written as Adg (w) = g −1wg .
• There is a constant ϑ , such that A = ϑw .

The use of the Lie series provides an effective tool to find the adjoint system of any
symmetry group [3] and can be constructed using the relation
ϵ2 ϵ3
Adg(exp(ϵ A))w = w − ϵ[A , w] + [A , [A , w]] − [A , [A , [⋯ + ]]]w, A ,
2! 3!
where [A, w] = Aw − wA , is the commutator between the two vectors. The tables 7.1
and 7.2 give the commutator table and the adjoint system for the cKdV, respectively.
The (i , j )th entry in table 7.1 gives [Ai , Aj ] and in the table 7.2 the (i , j )th entry
indicates Adg (exp(ϵ Ai))Aj .

7.4.1.1 Theorem 7.3

Let S be the full vector field group for the cKdV equation with the basis
{A1, A2, … , A 4} and let A ∈ S be a subgroup. If v = h(z ) is an A-invariant solution
to the cKdV, then the transformation h˜ (z ) = gh(z ) gives a à -invariant solution,
where A ≡ Adg (A) = g Ag −1 for all g ∈ S .

Proof. We take any vector Ai ∈ S . Let ψ = {z, t , v} denote the set of all variables
for the cKdV equation and T (ϵ ) denote a symmetry mapping generated by vector Ai
such that
T (ε ): ψ ↦ exp(ε A i)ψ = ψ˜ .
We further define the action of T ≡ T (ε ) on any smooth function F (ψ ) by
TF (ψ ) = F (Tψ ) = F (ψ˜ ). (7.17)
Now using (7.17) and assuming the existence of T −1 we can write
˜ F (ψ ) = T AF (ψ ) = T AT −1F (ψ˜ ).
A (7.18)

Using the inverse transform theorem we can write F (ψ˜ ) = h(ψ ) and F (ψ ) = h(ψ˜ )
and this is always true for group invariant solutions generated by Lie symmetry
groups [2, 3] and hence we can write, from (7.18),

7-15
 Modern Optimization Methods for Science, Engineering and Technology

Table 7.1. Commutator table for cKdV.

A1 A2 A3 A4

A1 0 A1 0 A3
A3 3
A2 −A1 0 A
2 4
2
A3 0 −
A3 0 0
2
A4 −A3 3
− 2 A4 0 0

Table 7.2. Adjoint action for cKdV.

Adg A1 A2 A3 A4

A1 A1 A2 − ϵ A1 A3 A 4 − ϵ A3
ε 3
A2 e ε A1 A2 e− 2 A3 e− 2 ε A 4
A3 A1 ε
A2 + 2 A3 A3 A4
A4 A1 + ϵ A3
3
A2 + 2 ϵ A 4 A3 A4

˜ h(ψ˜ ) = T AT −1h(ψ ).
A (7.19)

Now an invariant is a function h(z, t , v ) for the cKdV if it satisfies the equation

A ih ≡ 0. (7.20)

Hence using (7.19) and (7.20) we can say that if f = h(z ) is an invariant solution for
A then f˜ = h(z˜ ) will be an invariant solution for à , when
˜ = T AT −1 = exp(ϵ A i)A exp( −ϵ A i) ≡ Ad Ai(A).
A □

7.4.1.2 Optimal system of subalgebras for the cKdV

The concept of optimal systems of subalgebras was first used in [2]. Starting from the
generators of the cKdV group S, any general generator A ∈ V , where V is the space
of all the generators, can be written as
A = c1A1 + c2 A2 + c3A3 + c4 A 4.
The coefficients ci (i = 1, 2, 3, 4) are arbitrary constants. It is obvious that the
transformation has infinite subalgebras corresponding to which there are infinite
group invariant solutions. To tackle this problem one looks for subalgebras that are
similar, meaning they can be connected by some transformation of the group
thereby also connecting the group invariant solutions generated by each member of
a similar class. Thus, it will be sufficient to take any one member of the similar class

7-16
 Modern Optimization Methods for Science, Engineering and Technology

as a representative of the class. The set of all inequivalent classes gives the optimized
subalgebra for the system.
Now following the procedure in [3] and using the tables 7.1 and 7.2 we will find
the optimal system of subalgebras and their corresponding representatives for
equation (7.16). The main idea is to simplify as many coefficients of the vector
(ci), as much as possible, by application of the adjoint map given in table 7.2. For the
cKdV, the function ζ (A) = c2 is found to remain an invariant under the action of the
full adjoint (Adg ) group and can be expressed as
ζ(Adg(A)) = ζ(A), A ∈ V, g ∈ S.
Thus, at first we must start by considering different values of c2 . We suppose c2 ≠ 0
and set c2 = 1. The transformed vector is then acted on by Adg (exp( − 23 c4 A 4)) so that
the coefficients of A 4 vanish and we obtain a new vector
⎛ ⎛ 2 ⎞⎞
A′ = Adg⎜exp⎜ − c4A4 ⎟⎟A = c1′A1 + c3′A3 + A2
⎝ ⎝ 3 ⎠⎠

for some other constants c1′, c3′ depending on the original constants c1, c3 and c4 . The
new vector is acted by Adg (exp( −2c3′A3)) so that it is transformed to A″ = c1″A1 + A2 ,
which can again be transformed by Adg (exp(c1″A1)) so that the coefficients of A1
vanish and the new vector A‴ is equivalent to A2. Thus, we see that under the adjoint
action the vector A with c2 ≠ 0 generates subalgebras that are equivalent to the
subalgebra spanned by A2 . The rest have c2 = 0 and are of the form
A = c1A1 + c3A3 + c4 A 4.
Next we suppose c4 ≠ 0 and put c4 = 1. The vector is then transformed by
Adg (exp(c3A1)) so that the coefficients of A3 vanish and we obtain the new
transformed vector
A′ = c1′A1 + A 4.
No further simplification is possible for this vector and therefore the coefficient c1′
can be chosen only as 0 or ±1. Hence, any one-dimensional vector with c2 = 0 and
c4 ≠ 0 is spanned by either
A 4, A 4 + A1 or A 4 − A1.
All the remaining cases can be solved similarly [9] and we obtain an optimized
system of subalgebras for the cKdV to be spanned by the vectors
A1, A2, A3, A 4, A 4 + A1, A 4 − A1.
Again we note that (x , t , v ) → ( −x , t , −v ) gives the discrete symmetry associated
with the cKdV which maps A 4 − A1 to A 4 + A1, and thus reduces the number of
inequivalent subalgebras for the cKdV equation to five, i.e.
A1, A2, A3, A 4, A 4 + A1, (7.21)

7-17
 Modern Optimization Methods for Science, Engineering and Technology

which gives the optimal system of subalgebras for the cylindrical Korteweg–de Vries
(cKdV) equation.

7.4.2 Calculation of the optimal group invariant solutions for the cKdV
Having found the optimized system of one-dimensional subalgebras for the cKdV,
we are now in a position to calculate the corresponding optimized system of group
invariant solutions in this case. However, it must be noted that each solution will
have a singularity at t = 0. Now we proceed with each of the generators in the
optimal subalgebra (7.21):
i. A1 = ∂∂z
The invariants for the generator are given by y = t , u = v and the
corresponding reduced equation is
du u
+ = 0.
dy 2y
Simple integration gives the solution u(y ) = a / y , where a is a constant.
Thus, in terms of original variables the solution becomes
1
v = at − 2 . (7.22)

ii. A2 = 3t ∂∂t + z ∂∂z − 2v ∂∂v .

For this generator ξ(z, t , v ) = z , η(z, t , v ) = 3t , ϕ(z, t , v ) = −2v. The
1 2
invariants for the generator are given by y = zt − 3 and u = t 3 v. The reduced
equation for (7.16) is given by
d 3u du ⎛⎜ y⎞ u
+ u − ⎟ − = 0. (7.23)
dy 3
dy ⎝ 3⎠ 6
The solutions of equation (7.23) transform to the group invariant solution
2
( ) 1
v = t − 3 u zt − 3 . (7.24)

iii. A3 = 2t 2 ∂∂z + t − 2 ∂∂v .

1 1

The invariants in this case are y = t and u = z /2t − v. The corresponding

reduced equation is given by
du u
+ =0
dy y
with solutions u = b /y, where b is an arbitrary constant. The final solution in
terms of v is
z b
v= − . (7.25)
2t t

7-18
 Modern Optimization Methods for Science, Engineering and Technology

1
iv. For v4 , the invariants are y = zt − 2 and u = vt − z/2. The reduced equation
after two integrations gives
⎛ du ⎞2 u3
⎜ ⎟ =− + mu + n ,
⎝ dy ⎠ 3
where m and n are arbitrary constants. It has solution u = −12℘(y ), where
the Weierstrass elliptic functions (℘) [4] satisfy the equation
⎛ d ℘ ⎞2
⎜ ⎟ = 4℘ 3 − g℘ − g1, (7.26)
⎝ dy ⎠

where g and g1 are the invariants for ℘(y ). Now, if a, b and c are the roots of
equation (7.26) then we can have the following situations:
• When a < b < c then the solution for t > 0 gives the cnoidal wave
solution [4]

z + 2δ 2δ ⎛ δ ⎞
−1
v= + 2 dn 2⎜ zt 2 , s⎟ , (7.27)
2t s t ⎝ 6s 2 ⎠

where dn(y, s ) is the Jacobian elliptic function with amplitude

1
β = (c − a )/2 and the modulus s = [(c − b )/(c − a )] 2 .
• If a = b < c , then for the boundary conditions u(y ) → a , i.e. du
dy
,
d 2u
→ 0 as y → ∞ we obtain the famous soliton-like solution [4],
dy 2

z + 2a 1 ⎛ c − a −1⎞
v= + (c − a )sech 2⎜ zt 2 ⎟ . (7.28)
2t t ⎝ 12 ⎠

• If a = b = c , then we have the solution with some constant ν ,

z 12
v= − . (7.29)
2t −1
t(zt 2 − ν )2

1 1 1
v. For A 4 + A1, the invariants are y = zt − 2 + 1/4t and u = vt − yt 2 /2 − t − 2 /8.
The resulting reduced equation can be integrated to give the first Painlevé
transcendent,
d 2u u2 y
2
+ − = σ,
dy 2 4
where σ is a constant. The corresponding solution in terms of v are given by,
1 3
zt − 2 1 t− 2 1 ⎛ 1 1⎞
v= + + + u⎜zt − 2 + ⎟ . (7.30)
2 8t 8 t ⎝ 4t ⎠

7-19
 Modern Optimization Methods for Science, Engineering and Technology

The solutions for v between equations (7.22)–(7.30) are the required

optimized system of solutions. By definition, under suitable group action,
any of the group invariant solutions lead to some other group invariant
solutions for the same equation and an example for the cKdV can be found
in [9].

7.5 Concluding remarks

Historically, the transformational properties of some group actions on a given space
were found to have the properties of Lie groups. Cartan [5] was the first to use
coordinate charts in building a concrete foundation of the concept of a manifold.
The general analytical methods using the concepts of manifolds appeared in the
work of Palais [10], in which group invariant solutions are also discussed.
Subsequently, the concepts of quotient manifolds were also introduced by Olver
[3]. Although Lie had knowledge of the adjoint representation of Lie groups,
Ovsiannikov [2] used it to classify the group invariant solutions and gave the concept
of optimized solutions for differential equations. The technique was used further by
Patera et al [11], who applied it to some very important symmetry groups in physics.
Further generalization techniques were proposed by Ovsiannikov [2] and the duo
Bluman and Cole [12], and also by Ames [13], and are popularly known as the
nonclassical methods. A direct method for PDEs was introduced by Clarkson and
Kruskal [14]. The analytical connections of all these methods were proved by Levi
and Winternitz [15], while Nucci and Clarkson [16] showed that it is not quite as
general as the nonclassical approach. Galaktionov [17] has provided a further
promising generalization called nonlinear separation.

References
[1] Lie S 1888 Theorie der Transformationsgruppen (Leipzig: Teubner)
[2] Ovsiannikov L V 1982 Group Analysis of Differential Equations (New York: Academic)
[3] Olver P J 1993 Applications of Lie Groups to Differential Equations (New York: Springer)
[4] Ibragimov N H 1994 Lie Group Analysis of Differential Equations vol 1 (Boca Raton, FL:
CRC Press)
[5] Cartan E 1930 La Théorie des Groupes Finis et Continus et l, Analysis Situs Mém. Sci. Math.
42 (Paris: Gauthier-Villars)
[6] Chevalley C 1946 Theory of Lie groups I (Princeton, NJ: Princeton University Press)
[7] Warner F W 1971 Foundations of Differentiable Manifolds and Lie Groups (Berlin: Springer)
[8] Zakharov N S and Korobeinikov V P 1980 Group analysis of the generalised Korteweg–de
Vries–Burger’s equation J. Appl. Math. Mech. 44 668
[9] Gupta S K and Ghosh S K 2017 Classification of optimal group-invariant solutions:
cylindrical Korteweg–de Vries equation J. Optim. Theory Appl. 173 763–9
[10] Palais R S 1957 A global formulation of the Lie theory of transformation groups Memoirs of
the American Mathematical Society Number 22 (Providence, RI: American Mathematical
Society)
[11] Patera J, Winternitz P and Zassenhaus H 1975 Continuous subgroups of the fundamental
groups of physics. I. General method and the Poincaré group J. Math. Phys. 16 1597–614

7-20
 Modern Optimization Methods for Science, Engineering and Technology

[12] Bluman G W and Cole P D 1969 The general similarity solution of the heat equation
J. Math. Mech. 18 1025–42
[13] Ames W F 1965 Nonlinear Partial Differential Equations in Engineering (New York:
Academic)
[14] Clarkson P and Kruskal Z 1989 New similarity reductions of the Boussinesq equation
J. Math. Phys. 30 2201–13
[15] Levi D and Wintemitz P 1989 Nonclassical symmetry reduction: example of the Boussinesq
equation J. Phys. A 22 2915–24
[16] Nucci Z C and Clarkson P 1992 The nonclassical method is more general than the direct
method for symmetry reductions. An example of the Fitzhugh–Nagumo equation Phys. Lett.
A 164 49–56
[17] Galaktionov V A 1990 On new exact blow-up solutions for nonlinear heat conduction
equations with source and applications Diff. Int. Eq. 3 863–74

7-21
 IOP Publishing

Modern Optimization Methods for Science, Engineering and

Technology
G R Sinha

Chapter 8
Learning classifier system
Kapil Kumar Nagwanshi

The learning classifier system (LCS) is a technique which utilizes the power of
genetic algorithms and machine learning to make decisions based on specific rules.
The general components of LCSs include rule based and decision systems, learning
algorithms and rule discovery systems. The current chapter aims to discuss different
learning classifier systems for optimization. To understand each of these classifier
systems some datasets are required. MATLAB® has been utilized to simulate the
behavior and track the use of each of these classifiers. Examples of such learning
systems include tree, K-nearest-neighbor, support vector machine discriminant,
Bayes’ and ensemble classifiers. The degree of each of these classifiers has also been
chosen to describe the performance. The principal component analysis may used to
reduce the learning dimension to obtain a faster result. Multicore CPU and GPU
play a significant role in speeding up the system’s performance. Note, classifiers can
be used with every classification problem. This chapter also helps readers to choose
the best classifier for their work. Some other tools are also discussed at the end of the
chapter. The performance of the system depends on the dataset used. Case studies
suggest the type of result, such as a scatter plot, confusion matrix, parallel plot and
ROC curves. Later in this chapter, Python is introduced for developing the learning
classification. This chapter also aims to describe cloud-based products such as
BigML® and Microsoft® AzureML® for the optimization of classification problems.
The chapter ends with concluding remarks including benchmarking.

8.1 Introduction
The Internet is full of data. If you want to know some details about any particular
topic, it will be able to suggest anything from a few items to a very long list. From
a downloaded search, concluding a result can be a straightforward task in a few
cases, but for all other cases it will be a complex problem. There are a significant

doi:10.1088/978-0-7503-2404-5ch8 8-1 ª IOP Publishing Ltd 2020

 Modern Optimization Methods for Science, Engineering and Technology

number of patterns available to determine a result. These patterns can be obtained

from a different dataset either created by an individual person or available on the
Internet for free or a fee. Learning classifier systems (LCSs) are effective in
recognizing such patterns to solve a variety of problems from difficult to
elementary questions.
Holland [1] introduced the concept of the LCS. Later, in 1988, a guest editorial
was written by Goldberg and Holland [2] which described the utilization of genetic
algorithms (GAs) and machine learning. In the same editorial, they also explained
the GAs and classifier systems with a significant set of references. According to
Urbanowicz and Moore [3], ‘A learning classifier system (LCS) is defined as a model
of rule based machine learning technique based on evolutionary computing that
consolidates a discovery component, for example, genetic algorithm (GA) with a
learning component such as supervised, unsupervised or reinforcement learning’.
Similarly, the classifiers are defined by mapping the input stimuli to the output
action through a population of condition–action–payoff rules. LCSs are suitable for
classification as well as optimization of any problem domain that is solved or
approximated by a distributed set of local approximations [4]. This chapter starts
with the background of LCS, followed by a MATLAB demonstration of a different
set of classifiers using a graduate admissions dataset under a Creative Common
licence [5].

8.2 Background
In the year 2000 Holland et al [6] presented the answer to what a learning classifier
system was, according to the best researchers of the time. Holland divided the
classifier algorithm into three parts: the first part addresses parallelism and
coordination, the second part gives the credit assignment and the third part describes
the rule discovery. Then he described how classifier systems deal with these issues,
followed by rule implementation and the description of future directions. Booker
described the importance of classification mechanisms in parallelism as well as
standardization, that allows construction of a block technique for data processing
through the use of differentiation in conflict management.
Riolo [6] found that the main features of such a classifier system should be as
follows: (i) A communications message board. (ii) A representation of understanding
based on rules. (iii) A contest to activate guidelines, biased by inputs, past
performance and anticipated outcome projections. (iv) Parallel firing of rules, with
endogenously emerging consistency and coordination of operation as a rapidly
growing state created and preserved by the bid handling dynamics. Only at the
effector interface is explicit conflict resolution exclusively implemented. (v)
Temporal difference techniques of some sort are used to allocate credit, e.g. the
traditional algorithms of the bucket-brigade, some-sharing system, Q-learning
algorithms, etc. Note that various types of lines of credit may be assigned
simultaneously, e.g. traditional power is anticipated to payoff depending on past
performance, some rule of payoff accuracy or consistency [7], or a certain measure of
the capacity to forecast a subsequent state [8, 9]. (vi) Heuristics suitable for the

8-2
 Modern Optimization Methods for Science, Engineering and Technology

modeling scheme create the finding of the rule. Examples include activated bonding
to detect the surprise-triggered forecast of asynchronous causal links [10–12], or
traditional GAs with mutation and recombination. Holmes et al [13] adapted LCS
for the design, implementation and evaluation of EpiCS, for knowledge discovery in
epidemiological monitoring of a national child automobile passenger protection
program. The paper was concluded with significant detail of regression analytics.
Urbanowicz and Moore [3] explain that if the problem is of a high degree of
complexity, then LCS can solve the problem efficiently. The LCS has be applied in
biology, computer science, artificial intelligence, machine learning, evolutionary
biology, evolutionary computation, reinforcement learning, supervised learning,
evolutionary algorithms and GAs to build the solution domain as a learning
classifier system. Zang et al [14] utilized XCS with memory conditions (absent in
standard XCS) to solve ‘maze environments and the perceptual aliasing problem’
with robust output.

8.3 Classification learner tools

Several tools are available to describe the learning classifier system. In this chapter,
the author will describe the MATLAB® Classification Learner App, BigML®, and
Microsoft® AzureML®. In this text, the author has used these tools wherever it is
applicable to affect the output of the classifier system.
The readers need to have a licensed or demo version of MATLAB to run their
experiments. It is recommended that the latest version of MATLAB is installed.
Currently, readers can try the version MATLAB 2019a. If the reader’s organ-
ization has already purchased MATLAB and other cloud-based tools, it is
straightforward to start working on it. The BigML and AzureML tools have
been used in order to understand the solution graphically. BigML and AzureML
are both cloud-based services one can use by merely creating a trial account on the
latest websites.

8.3.1 MATLAB®: classification learner app

The Classification Learner app of MATLAB® trains models to classify data
through a diverse set of classifiers to examine supervised machine learning. It will
start with data exploration, then selection of features, followed by specific validation
schemes if any, and then the models are trained to evaluate outcomes. Automated
training could be undertaken to search for the most elegant form of classification,
including decision trees, discriminating assessment, support vector machines
(SVMs), logistic regression, nearest neighbors, naive Bayes and classification of an
ensemble. Learning can be accomplished by providing a piece of input information
with a number of observations and associated responses, such as labels or categories.
Training is used to train a template to produce fresh information responses. The
trained model can also be used in workspace programming with the export of
produced MATLAB code [15–17].

8-3
 Modern Optimization Methods for Science, Engineering and Technology

8.3.2 BigML®
BigML® is an advanced tool for machine learning, which contributes a collection of
robustly engineered algorithms established to elucidate real-life problems. The
algorithms based on supervised learning have been utilized to solve classification,
regression and time series forecasting while unsupervised learning has been used to
provide cluster analysis, anomaly detection, topic modeling and principal compo-
nents analysis by implementing a single, standardized framework. In this chapter,
the author has chosen BigML for its applicability for understanding the character-
istic features of different classifiers. BigML provides a broad set of modeling
scenarios to analyze and compare, which makes analysis easy. One can directly
open an evaluation account to understand BigML through the webpage https://
bigml.com [18–20].

8.3.3 Microsoft® AzureML®

The Microsoft Azure Machine Learning Studio is a collaborative, drag-and-drop
tool you can use to build, test and deploy predictive analytics solutions on data of
interest. It provides a facility to publish models as web services that can efficiently be
utilized by custom apps or BI tools such as Excel. In most solutions, one needs to
provide the dataset as an input. Next, one is expected to split the dataset into the
training dataset and testing dataset, followed by choosing one or more algorithm to
train using the train model block. Then the output of the train model block will be
validated by the score model block. Finally, the production of the score model is
evaluated by the evaluate model block to obtain the output in the form of
performance parameters such as ROC, confusion matrix, etc. Readers can obtain
a subscription through the AzureML website https://studio.azureml.net [21].

8.4 Sample dataset

Let us describe each classifier type using the graduate admissions dataset as shown in
table 8.1 [5]. In this chapter the author is not going to justify the predictions of
different models. This chapter focuses on the learning process and its steps to obtain
the output. The dataset consists of 500 rows and nine columns, namely, Serial No.,
GRE Score, TOEFL Score, University Rating, Sum of Product (SOP), LOR,
CGPA, Research and Chance of Admit. Of these nine columns, Serial Number has
no role in prediction, so this column is deleted from the dataset. The column with the
heading Chance of Admit will act as the response class and the remaining columns
act as predictors.
Figure 8.1(a) shows the graph of 61 response classes of the original dataset and
figure 8.1(b) shows eight response classes after rounding-off the values in the Chance
of Admit column from the original dataset. This alteration of data has been done in
order to understand the classification learner problem.

8-4
 Modern Optimization Methods for Science, Engineering and Technology

Table 8.1. Sample data from the graduate admission dataset.

Serial GRE TOEFL University Chance

No. score score rating SOP LOR CGPA Research of admit

1 337 118 4 4.5 4.5 9.65 1 0.92

2 324 107 4 4 4.5 8.87 1 0.76
3 316 104 3 3 3.5 8 1 0.72
4 322 110 3 3.5 2.5 8.67 1 0.8
5 314 103 2 2 3 8.21 0 0.65
6 330 115 5 4.5 3 9.34 1 0.9
7 321 109 3 3 4 8.2 1 0.75
8 308 101 2 3 4 7.9 0 0.68
9 302 102 1 2 1.5 8 0 0.5
10 323 108 3 3.5 3 8.6 0 0.45
11 325 106 3 3.5 4 8.4 1 0.52
12 327 111 4 4 4.5 9 1 0.84
13 328 112 4 4 4.5 9.1 1 0.78
14 307 109 3 4 3 8 1 0.62
15 311 104 3 3.5 2 8.2 1 0.61

Figure 8.1. Graduate admission dataset with (a) 61 classes and (b) 8 classes.

8.4.1 Splitting the dataset

The admission dataset can be split manually or one can run the following MATLAB
code to split the dataset for training into the Training dataset, and for testing into the
Testing dataset to evaluate the algorithms:

8-5
 Modern Optimization Methods for Science, Engineering and Technology

admissiontable = readtable('Admission_Predict.csv');
save('admissiontable1');
load admissiontable;
[m,n] = size(admissiontable); P = 0.70;
idx = randperm(m);
Training = admissiontable(idx(1:round(P*m)),:);
save('Training');
Testing = admissiontable(idx(round(P*m)+1:end),:);
save('Testing');

After running the above code, the admissions table is split into the Training
table with 350 rows and the Testing table with 150 rows. On the other hand,
AzureML provides a Dataset Split module which by default splits the dataset into a
50:50 ratio of training:testing. The subscriber can change this value by selecting this
module.

8.5 Learning classifier algorithms

This section deals with learning classifier algorithms in detail with the help of
different tools. The following algorithms are discussed here:
• Logistic regression classifiers.
• Decision tree classifiers.
• Discriminant analysis classifiers.
• Support vector machines classifiers.
• Nearest neighbor classifiers.
• Ensemble classifiers.

Before going to test each of the above stated models it is required to load the
Admission prediction training dataset using the following MATLAB command (see
section 8.4.1 on splitting dataset for the sample dataset):

» load Training;

After running the above command the variable names were modified as GREScore,
TOEFLScore, UniversityRating, SOP, LOR, CGPA, Research and
ChanceOfAdmit to make them valid MATLAB identifiers. The original names are
saved in the VariableDescriptions property. Now the stated table (Dataset) is
available for training of the classification learners. Subsequently, we can now start
the classification learner app and start the new session from the workspace with the
admissiontable dataset available in the workspace. The app will automatically
select the predictors and response (by default the last column). One can select any

8-6
 Modern Optimization Methods for Science, Engineering and Technology

number of predictors from the available list. After selecting predictors and
responses, it is necessary to validate the dataset by pressing the Start Session
button. By default, the cross-validation is five-fold to protect against overfitting of
the data.
The Classification Learner starts with the scatter plotting of the data. After this
step, it is necessary to select the classifier(s) to see the classification performance. A
scatter plot is useful to check the predictability of the response. The plot has the
option to choose different x and y parameters. Figure 8.2(a)–(f) show the scatter plot
to examine seven variables for predicting the response by choosing different options
on the different x and y parameters under Predictors to reflect the distribution of
response variable ChanceOfAdmit in different colors.
Note: In MATLAB the classification learner app does not support the regression
classifier, and for this purpose there is a separate app called the regression learner.
The regression learner is not able to give prediction in the form of a confusion matrix
and ROC directly, so to see this the output of AzureML has been used.

Figure 8.2. Showing the response in color coded format with respect to GRE versus other predictors.

8-7
 Modern Optimization Methods for Science, Engineering and Technology

8.5.1 Logistic regression classifiers

This section will explain the concept of the logistic regression classifier with the help
of AzureML as well as MATLAB. The logistic function shown in equation (8.1) is
also recognized as the sigmoid function that statisticians use to express the
characteristics of population growth in ecology, propagating and maximizing the
environmental carrying potential. It is an S-shaped curve that can take any real-
valued amount and trace it to a value between 0 and 1, but never at those boundaries
exactly:
1
f (t ) = . (8.1)
1 + e −t
Figure 8.3 shows the logistic regression flow using AzureML. Here the author has
selected a multiclass logistic regression classifier and a two-class logistic regression
classifier. The admissions dataset is divided into training and testing datasets. The
training model takes two inputs, the left-hand side uses the algorithm, and the right-
hand side uses the training dataset. The output of the Train Model block is provided
as the left-hand side input of the score model, and the right-hand output of the split
data module is taken as the right-hand input of the score model block. The Evaluate
model block can take two sources of information for comparison of the algorithm.
In this case, because the nature of the algorithm is varied, i.e. one is multiclass and
the other is two-class, so the current author took two different Evaluate model
blocks.
The output of the Evaluate model for multiclass logistic regression is shown as a
confusion matrix (see section 8.6.1 for more details on the confusion matrix) shown
in figures 8.1 and 8.7, and for a two-class regression model it is shown in figure 8.4 as

Figure 8.3. Multiclass and two-class logistic regression experimentation in AzureML.

8-8
 Modern Optimization Methods for Science, Engineering and Technology

Figure 8.4. ROC two-class logistic regression experimentation in AzureML.

Table 8.2. The confusion matrix and results obtained from AzureML. (ALL: average log loss.)

Predicted class

Class 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 ALL Precision Recall

0.3 0 0 0 2 1 0 0 0 2.55 0 0
0.4 0 0 0 0 2 0 0 0 2.87 0 0
0.5 0 0 0 4 4 1 0 0 1.88 0 0
0.6 0 0 0 3 12 0 0 0 1.33 0.33 0.2
0.7 0 0 0 0 21 6 1 0 1.13 0.47 0.75
0.8 0 0 0 0 5 10 3 0 1.14 0.45 0.56
0.9 0 0 0 0 0 5 16 0 0.84 0.67 0.76
1 0 0 0 0 0 0 4 0 2.59 0 0

ROC (see section 8.6.2 on the receiver operating characteristic). Further, the output
can be taken as the .csv file shown in table 8.2 to download the results of a different
set of algorithms.
It is assumed that the reader is aware of the logistic function, therefore, the author
does not intend to show the logistic function plot, although readers can go to https://
www.geeksforgeeks.org/understanding-logistic-regression/ for more details. One can
also assume that the target variable is categorical. Based on the number of classes,
logistic regression can be categorized as follows: (i) Binomial logistic regression is
used if there are two possible target classes 0 or 1, for example admitted versus not

8-9
 Modern Optimization Methods for Science, Engineering and Technology

admitted. (ii) Multinomial logistic regression is used when the target variable can
have three or more possible classes which are not ordered, i.e. the types have no
quantitative significance, for example IPL cricket teams in India can be ‘Chennai Super
King’ versus ‘Rajasthan Royals’ versus ‘Kings Eleven’. (iii) Ordinal logistic regression
is used if target classes are ordered, for example a test score can be described as ‘very
poor’, ‘poor’, ‘good’ or ‘very good’. For this purpose, each category can be given a
score, such as 0, 1, 2 and 3.

Algorithm 1. Logistic regression classifier

1. Initialize the weights at t = 0 to w(0).

2. repeat t = 0 : n.
1 N yn xn
3. Compute the gradient: ∇Ein = − ∑ .
N n=1 1 + e ynwT (t )xn
4. Update the weights: w(t + 1) = w(t) −η∇Ein.
5. until Stopping Condition is false.
6. return final weights w.

Consider wi to be the weights or coefficients, xi to be the input. h represents the

hypothesis set that a selected classifier brings and θ is a sigmoid or logistic function.
Then the regression function s , shown in figure 8.5, can be given by
d
s= ∑i=0 wixi . (8.2)

Now we come to linear classification. For this, the characteristic function for h
has binomial values and the shape of the curve is like ˩.̄ Second, for linear regression
the characteristic function for h has binomial values and the shape of the curve is
like _/ .̄ Finally, for logistic regression the characteristic function for h has
continuous values and the shape of the curve is like ∫ .
Logistic regression is presented as an algorithm. One is a classification technique
based on supervised learning; therefore, it helps us to converge those uncertain
posteriors with a differentiable decision function. For logistic regression, w(0) can be

Figure 8.5. Logistic regression function for four inputs with inputs x0 = b and w0 = 1.

8-10
 Modern Optimization Methods for Science, Engineering and Technology

0, it makes sense to start at η = 0.5 because it is the most uncertain state. Again,
maybe it is not a logistical regression issue as other concepts such as neural
networks, as the issue of logistical optimization is smoothly convex. Generally
speaking, we use a blend of requirements as a stop condition. Let us use the
ubiquitous threshold (Δw < threshold ), for example. The stochastic gradient descent
is from time to time a stronger strategy. It is extremely effective and often produces
excellent outcomes. The conjugate gradient is the best of all the derivative-based
techniques. The concept is to optimize the second order without effectively
calculating the Hessian.
In MATLAB based experimentation on the same dataset, figure 8.6 exhibits
different graphical results for the LR classifier. The responses have been plotted for
predictor versus residual, predicted response versus residual, true response versus
predicted response, etc. The confusion matrix obtained from the Evaluate model as a
.csv file, shown in table 8.2, also gives output in the form of average log loss,
precision and recall. In this way, one can use a logistic regression classifier to obtain
and validate the applicability of the algorithms used (figure 8.7).

Figure 8.6. Showing the responses in logistic regression classifiers.

8-11
 Modern Optimization Methods for Science, Engineering and Technology

Figure 8.7. Confusion matrix multiclass logistic regression experimentation in AzureML.

8.5.2 Decision tree classifiers

The classification tree algorithm or decision tree algorithm refers to the family of
algorithms based on supervised learning that can be used to solve regression and
classification issues. With a faster training time, the decision tree belongs to a non-
parametric white box class of machine learning methods which does not rely on
probability distribution hypotheses to express ongoing decision-making logic based
on the SOP or disjunctive normal form prototype, and is not readily accessible in the
black box class of algorithms along with the artificial neural networks. The time
complexity of decision trees depends on the amount of records and the amount of
features in the dataset. Decision trees can accurately handle high-dimensional
information. The purpose of this method is to produce a training model for
predicting answers by learning the rules of decision making deduced from the
training dataset. Compared to imminent classification algorithms, the tree algorithm
is straightforward and attempts to determine the solution using representation of the
tree. The tree’s internal nodes constitute a trait and a class tag for the leaf node.
Figure 8.8(a) gives sample code snippet and corresponding tree diagram for the tree

8-12
 Modern Optimization Methods for Science, Engineering and Technology

classification algorithm in figure 8.8(b)–(c). Algorithm 2 gives the pseudocode for the
tree classifier. Initially, it will begin by organizing the most fitting trait of the dataset
at the root of the tree, accompanied by splitting the training set into subsets such that
each subset comprises data with the same attribute value. This operation operates on
each subset indefinitely until leaf nodes are found in all tree branches.

Algorithm 2. Tree classifier

1. Arrange the fittest attribute using attribute selection measures (ASM), such as Gini index,
information gain or gain ratio to split the records of the dataset at the root of the tree.
2. Make that attribute a decision node and divide the training set into subsets.
3. Every subset comprises data with the same value for an attribute.
4. Repeat step 1 and step 2 on each subset until it finds leaf nodes in all the branches of the tree.

All the internal nodes in a decision tree act like decision nodes which contain
features or attributes, while leaf nodes give the response or outcome through
branches which act as decision rules (see figure 8.8(a) and (b)). A topmost decision
node is known as the root of the tree.
The measurement of of the attribute in the attribute selection measure choice or
splitting rules is a method for selecting the splitting criteria that best compute the
partition. It enables us to manage breakpoints on a specified node for tuples. By
defining the specified dataset, ASM offers a class for each function and chooses the
highest resulting function as a dividing variable. The information gain, gain ratio
and Gini index are some common choice criteria [22, 23].
Information gain is a contraction in entropy that measures the distinction
between pre-split entropy and the dataset’s median post-split entropy based on the
function numbers given. Suppose that Pi is the probability that an arbitrary tuple
exists in D ∈ Ci . The median quantity of data needed to differentiate a tuple class
tag in D is Info(D ) given by equation (8.3), in equation (8.4) InfoA (D ) is the
anticipated data required to classify a D tuple depending on the characteristic
∣D ∣
partitioning A, j characterizes the weight of the jth partition and the attribute A
∣D∣
with the peak information gain, and v is the amount of discrete values in trait A.
Equation (8.5) can determine Gain(A). The dividing attribute at node N is selected:
m
InfoA(D ) = −∑ pi log2pi (8.3)
i =1

v ∣Dj ∣
InfoA(D ) = ∑ j=1 × Info(Dj ) (8.4)
∣D∣

Gain(A) = Info(D ) − InfoA(D ). (8.5)

According to Navlani [24], Gain(A) is appropriate for an attribute with many
implications. But for a uniquely identified attribute, such as the AADHAR number

8-13
 Modern Optimization Methods for Science, Engineering and Technology

Figure 8.8. Classification tree algorithm: (a) sample rules; (b) sample tree diagram in BigML; and (c) sample
tree diagram in MATLAB (pruning level 9 of 13).

8-14
 Modern Optimization Methods for Science, Engineering and Technology

of any student, it has zero Info(D ). Due to the tidy partitioning, it maximizes the
benefit of data Gain(A) and creates unusable partitioning. The gain ratio
GainRatio(A) provided in equation (8.7) can be used to resolve this problem, it
handles the bias issue by using divided data to normalize the information gain
SplitInfoA(D) as shown in equation (8.6). For trait A, the maximum value of the gain
ratio GainRatio(A) is chosen as a splitting attribute:
v ∣Dj ∣ ∣Dj ∣
SplitInfoA(D ) = −∑ × log2 (8.6)
j =1 ∣D∣ ∣D∣

Gain(A)
GainRatio(A) = . (8.7)
SplitInfoA(D )

The Gini index provides another way of splitting the decision tree. For a given
partition of data D into D1 and D2 , the Gini index of D can be estimated by
∣D1∣ ∣D 2 ∣
GiniA(D ) = Gini(D1) + Gini(D2 ), (8.8)
∣D∣ ∣D∣
with
m
Gini(D ) = 1 − ∑i=1pi2 . (8.9)

If a continuously valued attribute is in the exercise, then select a couple of

neighboring values as a prospective split point with a lower Gini index being taken
as the split point; on the other side, if it has a discrete valued attribute, a subset
giving a minimum Gini index has been chosen as a dividing attribute.

8.5.3 Discriminant analysis classifiers

Discriminant analysis is a classification technique based on Bayes’ theorem to
estimate the probability estimates that distinct classes produce data based on
distinctive Gaussian distributions of predictors in each of the response classes.
Linear discriminant analysis aims to map the higher dimensional space character-
istics into a smaller dimensional space. If the classes are well separated and if the
classes are linearly separable, then other classification techniques such as logistic
regression work well, but if the case is complicated with no linearly separated classes,
in this case linear discriminant analysis works well. It gives a method to produce
classifiers based on labeled training data for a predictive model of group member-
ship to apply on new cases. Unlike some other classifiers which work on a
two-dimensional plane, the LDA can work well with higher dimensionality and
hyperplanes. The discriminant analysis further produces an equation that can be
used to classify new examples [25]. Because of space limitations, the author
recommends reading an article written by Tharwat et al [26]. It is an excellent
tutorial with a graphical representation of the LDA algorithm. The five general steps
for performing a linear discriminant analysis are presented in algorithm 3.

8-15
 Modern Optimization Methods for Science, Engineering and Technology

Algorithm 3. Discriminant analysis classifier

1. Estimate the d-dimensional mean vectors for the distinct classes from the dataset.
2. Compute the scatter matrices (in-between-class and within-class scatter matrix).
3. Compute the eigenvectors (e1, e2, …, ed) and corresponding eigenvalues (λ1, λ2, …, λd) for the
scatter matrices.
4. Sort the eigenvectors by decreasing eigenvalues and choose k eigenvectors with the largest
eigenvalue to form a d × k-dimensional matrix W (every column represents an eigenvector).
5. Use this d × k eigenvector matrix to transform the samples onto the new subspace. This can be
summarized by the matrix multiplication: Y[n, k] = X[n, d] × W[d, k] (where X is a n × d-
dimensional matrix representing the n samples, and y are the transformed nk-dimensional
samples in the new subspace).

Readers can also use the RSS feed by Raschka to go through each of the above
steps in detail using Python [27]. LDA is a classification technique that is easy and
efficient. Many extensions and variants to the method are available because it is easy
and well known. Some popular extensions lead to: (i) regularized discriminant
analysis (RDA), which introduces regularization in variance estimation which is
effectively covariance, moderating the impact of various factors on LDA; (ii) flexible
discriminant analysis (FDA), where nonlinear input components such as splines are
used; and (iii) quadratic discriminant analysis (QDA), in which, when there are
numerous input factors, each category utilizes its own variance or covariance
estimate.
The initial design was termed as discriminant analysis of the linear discriminant
or Fisher’s discriminant analysis. Multiple discriminant analysis has been linked to
the multiclass variant. All of these are now called linear discriminant analysis [28].

8.5.4 Support vector machine classifiers

A support vector machine (SVM) is technically assigned by a separate hyperplane as
a discriminative classifier. In other words, considering the labeling of training data
(supervised learning), the classifier produces an ideal hyperplane categorizing fresh
instances. This hyperplane is a row separating a plane into two sections in two-
dimensional spaces in which each category lies on either hand. In principle, the SVM
algorithm is applied using a kernel. Hyperplane learning in linear SVM is done by
transforming the problem using some linear algebra outside the scope of this
introduction to SVM. A strong understanding is that, instead of the results
themselves, the linear SVM can be interpreted using the inner product of any two
variables. The inner product around two vectors is the sum of each pair of input
values multiplying. For example, the inner product of the vectors [3, 4] and [6, 7] is
4 × 7+3 × 6 or 46. The equation for building a prediction for a new input using the
dot-product flanked by the input (x ) and for each support vector (xi ) can be given by

f (x ) = B 0 + ∑(ai × (x, xi )). (8.10)

8-16
 Modern Optimization Methods for Science, Engineering and Technology

Equation (8.10) encompasses determining the inner products of a new input

vector (x ) with all support vectors in the training data. The coefficients B0 and ai
must be projected from the training data by the learning algorithm. In linear kernel
SVM the dot-product is known as the kernel which is given by
K (x , xi ) = ∑(x × xi ). (8.11)

The kernel describes the resemblance or measure of range between fresh information
and vectors of assistance. The dot-product is the degree of resemblance used for linear
SVM or a linear kernel since the distance is a linear input combination. Other kernels,
such as polynomial kernels and radial kernels, can be used to transform the input space
into higher dimensions. This is called the trick of the kernel or kernel trick. It is suitable to
use more complex kernels as it allows lines to differentiate the classes that are bent or
even more complex. This in turn can lead to more accurate classifiers. Further, the
SVM classifiers can be divided into the following. (i) Polynomial kernel SVM, where
we can use a polynomial kernel instead of the dot-product, for instance
K (x , xi ) = 1 + ∑(x × xi )d , and where the degree of polynomial to the learning
procedure must be defined by the side. When d = 1 this will contribute to the linear
kernel. The polynomial kernel enables input space for curved rows. (ii) Radial kernel
SVM, which is more complicated. For example, one can see the value of
2
K (x , xi ) = e (−γ × ∑(xxi )) with γ as some learning parameter. A worthy default value
for γ is 0.1, and γ lies between 0 < γ < 1. The radial kernel is quite global and can
generate complicated areas inside the feature space, such as closed polygons in 2D space.
An optimization method must be used to solve the SVM model. A quantitative
optimization method could be used to search for the hyperplane coefficients. This is
inefficient and is not the strategy used in SVM applications, such as LIBSVM, that are
commonly used. You could use stochastic gradient descent if you implement the
algorithm as an workout. Specialized optimization processes exist that re-formulate the
issue of optimization to be a issue of quadratic programming. The sequential minimum
optimization approach is the most common technique for SVM planning. It divides the
issue into sub-problems that can be analytically rather than numerically resolved.

8.5.5 Nearest neighbor classifiers

The nearest neighbor principle continually achieves strong efficiency amongst the
many techniques for statistical pattern recognition without any hypothesis regarding
the distribution of training examples. It includes a number of favorable as well as
adverse instances. A fresh sample is categorized by calculating the distance from the
closest situation, which then determines the ranking of the sample. The KNN
classifier expands this concept by taking the k-points closest to the center and giving
these the symbol K. Tiny and odd links (typically 1, 3 or 5) are frequently selected.
No matter how long k measurements help decrease the impacts of loud points in the
training dataset, cross-validation is used for the selection of k. By following
algorithm 4 m, we can introduce a KNN template.

8-17
 Modern Optimization Methods for Science, Engineering and Technology

Algorithm 4. Nearest neighbor classifier

1. Load the data State Initialize the value of k.
2. for getting the predicted class, iterate from 1 to the total number of training data points do.
3. Calculate the distance between the test data and each row of training data.
Here we will use Euclidean distance as our distance metric since it is the most popular method.
The other metrics that can be used are Chebyshev, cosine, etc.
4. Sort the calculated distances in ascending order based on distance values.
5. Get the top k rows from the sorted array.
6. Get the most frequent class of these rows.
7. return the predicted class.

Many methods are available to improve the efficiency and velocity of the next
classification. One strategy for this issue is the pre-sorting of the practice sets (such as
kd-trees or cells of Voronoi). Another alternative is to select a subset of trained
information so as to approximate the Bayes’ ranking in accordance with the 1-NN
law (through the subgroup). As k can now be restricted to one and redundant data
points taken from the training set, this can produce substantive velocity gains. As k
can now be restricted to 1 and redundant data points are deleted from the practice
pack, important increases in velocity can be achieved. These methods of information
alteration can also enhance efficiency with the removal of misclassification points.

8.5.6 Ensemble classifiers

Ensemble methods are meta algorithms that mix various machine learning techni-
ques in a single predictor model to reduce (bagging) variability or prejudice
(boosting), or to enhance (stacking) projections. Sequential ensembles, for example
AdaBoost, are techniques in which primary learners are sequentially produced.
Sequential approaches’ fundamental motive is to take advantage of the dependency
among base learners. By weighing earlier wrong instances with a greater weight,
general efficiency can be improved. (ii) Parallel ensemble techniques such as the
random forest are produced in linear by the basic learners. The fundamental motive
of parallel techniques is to use the autonomy of grassroots students because errors
can be dramatically decreased by average.
The majority of ensemble techniques use a single basic learning algorithm to
make homogeneous, uniform and basic students. Some techniques also use hetero-
geneous learners which lead to heterogeneous sets. The basic learners have to be
precisely as feasible and as varied as necessary in order to make ensemble processes
more precise than any of their individual participants.

8.6 Performance
The dataset has been converted to obtain the visualization of results in a simple
form, and this is the reason why the results are not up to the mark. Figure 8.9 shows
the prediction response in color coded format with respect to GRE versus
TOEFLScore, where the legends mean: ∙ = correct data and × = incorrect data.

8-18
 Modern Optimization Methods for Science, Engineering and Technology

Figure 8.9. Showing the prediction response in color coded format with respect to GRE versus TOEFLScore
(• = correct data and × = incorrect data).

In this figure, readers can compare the visuals of the decision tree, SVM, NN,
discriminant analysis and ensemble with the original data.
Similarly, figure 8.10 shows the prediction response in color coded format of two
classes with respect to GRE versus TOEFLScore. This figure exhibits the results of
scatter plots for the SVM classifier only. In this set of plots, the author has recorded

8-19
 Modern Optimization Methods for Science, Engineering and Technology

Figure 8.10. Showing the prediction response in color coded format of two classes with respect to GRE versus
TOEFLScore (• = correct data and × = incorrect data).

predictions for two–two classes (0.3–1, 0.4–0.9, 0.5–0.8 and 0.6–0.7) so that the
readers may be able to see the responses and count the correct and incorrect
responses as well. The readers can plot the plots for all the available classifiers. All
the tree tools discussed support different styles of scatter plots.

8.6.1 Confusion matrix

For measuring the performance of learning classifier systems, several measures are
used that depend on four major parameters. A confusion matrix (see table 8.3)
consists of the four results from the binary classification and is a two by two table.1
1. True positive (TP): true is identified as true (correct identification).
2. True negative (TN): false is identified as false (correct identification).
3. False positive (FP): true is identified as false (wrong identification).
4. False negative (FN): false is identified as true (wrong identification).

1
The definitions used in this chapter have been taken from the website: https://classeval.wordpress.com/
introduction/basic-evaluation-measures/.

8-20
 Table 8.3. Confusion matrix and performance measurements for different classifiers.

Predicted footprint match

Positive Negative Total number of incorrect FP + FN

EER =
predictions = PP + FN P+N

Positive True positive (TP) False positive (FP) Total number of TP

PPV =
matches = TP + FP TP + FP
Observed
FP
footprint FDR =
TP + FP
match
FDR = 1 − PPV

8-21
Negative False negative (FN) True negative (TN) Total number of non- TN
NPV =
matches = TN + FN TN + FN

All with footprint All footprint non-match Total number of correct TP + TN

ACC =
match P = TP + FN N = FP + TN predictions = TP + TN P+N
ACC = 1 − EER
TP TN
TPR = P
TNR = N
Informedness = TPR + TNR − l Markedness = PPV + NPV − l
FN FP
FNR = P
FPR = N
Modern Optimization Methods for Science, Engineering and Technology

FPR = 1 − TNR
 Modern Optimization Methods for Science, Engineering and Technology

Therefore, the number of correct identification cases represented by P and the

number of false identification cases represented by N are given by
P = TP + FN (8.12)

N = FP + TN. (8.13)
Figure 8.11 gives the confusion matrix for (a) decision tree, (b) SVM, (c) NN, (d)
discriminant analysis, (e) ensemble and (f) logistic regression. This confusion matrix
is plotted for the number of observations or frequency of occurrence. Diagonal
responses have been shown in green, i.e. if the true class and predicted classes are the
same, and red observations show the true classes that were confused with predicted

Figure 8.11. Confusion matrix for a number of observations.

8-22
 Modern Optimization Methods for Science, Engineering and Technology

classes. The confusion matrix for the number of observations for the logistic
regression algorithm has been taken from the AzureML tool.

Definition 12. Sensitivity, recall or true positive rate ‘is calculated as the number of
correct positive predictions divided by the total number of positives. ... The best
sensitivity is 1.0, whereas the worst is 0.0’:

TP TP
TPR = = . (8.14)
P (TP + FN)

Definition 2. Specificity or true negative rate (TNR) ‘is calculated as the number of
correct negative predictions divided by the total number of negatives. ... The best
specificity is 1.0, whereas the worst is 0.0’:

TN TN
TNR = = . (8.15)
N (FP + TN)

Definition 3. Miss rate or false negative rate (FNR) is the ratio between false
negative and false negative + true positive predictions:

FN
FNR = . (8.16)
(FN + TP)

Definition 4. Precision or positive predictive value (PPV) ‘is calculated as the

number of correct positive predictions divided by the total number of positive
predictions. ... The best precision is 1.0, whereas the worst is 0.0’:

TP
PPV = . (8.17)
(TP + FP)

Definition 5. Negative predictive value (NPV) is measured as the number of correct

negative predictions divided by the total number of of negative predictions. The best
NPV is 1.0, the worst is 0.0:

TN
NPV = . (8.18)
(TN + FN)

Definition 6. False discovery rate (FDR) ‘is calculated as the number of false
positive predictions divided by the total false positive and true positive predictions.
The best FDR is 0.0, while the worst is 1.0. It can also be calculated as (1 − PPV)’:

2
Definitions are taken from https://classeval.wordpress.com/introduction/basic-evaluation-measures/.

8-23
 Modern Optimization Methods for Science, Engineering and Technology

FP
FDR = = 1 − PPV. (8.19)
(FP + TP)

Definition 7. Fall-out or false positive rate (FPR) ‘is calculated as the number of
incorrect positive predictions divided by the total number of negatives. The best false
positive rate is 0.0 whereas the worst is 1.0’:

FP FP
FPR = = = 1 − TNR. (8.20)
N (FP + TN)

Definition 8. Informedness is the distance from (i.e. measured perpendicular to) the
random line joining coordinate (0,0) and coordinate (1,1) in the ROC. You can then
use markedness as the second metric to identify your classification system’s
general value:

Informedness = Sensitivity + Specificity − 1 (8.21)

Markedness = Precision + NPV − 1. (8.22)

Definition 9. Error rate (ERR) ‘is calculated as the number of all incorrect
predictions divided by the total number of the dataset. The best error rate is 0.0,
whereas the worst is 1.0’:

FP + FN FP + FN
EER = = . (8.23)
TP + TN + FP + FN P+N

Definition 10. Accuracy (ACC) ‘is calculated as the number of all correct predic-
tions divided by the total number of the dataset. The best accuracy is 1.0, whereas
the worst is 0.0. It can also be calculated by 1 – ERR’:

(TP + TN)
ACC = = 1 − EER. (8.24)
(P + N)

Figure 8.12 shows the confusion matrix for PPV versus FDR, and figure 8.13
shows the confusion matrix for TPR versus FNR. We only explain the confusion
matrix for four algorithms and the remaining two algorithms are left as an exercise
for the reader. The matrix in figure 8.12 determines the horizontal characteristics of

8-24
 Modern Optimization Methods for Science, Engineering and Technology

Figure 8.12. Confusion matrix for PPV versus FDR.

the predicted classes. The green result is known as the positive predictive value (see
definition 5), while the false discovery rate (see definition 6) is represented by the red
shaded results. The matrix in figure 8.13 determines the vertical characteristics of
true classes. The green shaded result is known as the TPR (see definition 1), while
FNRs (see definition 3) are in red.

8.6.2 Receiver operating characteristic

The receiver operating characteristic curves obtained in this experiment provide the
trade-off between sensitivity (TPR, see definition 5) and fall-out (FPR, see definition
49) or specificity (TNR, see definition 5) as FPR = 1 − TNR (see equation (8.20))
for all possible cut-offs for a test or its combinations. Also, the area under the ROC
curve (AUC) presents an idea about the advantage of employing the test(s), which is
a measure of the usefulness of an experiment. The best cut-off has the highest true
positive rate together with the lowest false positive rate leading to a greater area.
Figure 8.14 displays the receiver operating characteristics for the decision tree,
SVM, NN, and discriminant analysis classifiers for positive class 0.3. One can easily
see the different values of ROC and AUC. The right-hand panel is helpful in the
creation of the other plots as well.

8-25
Modern Optimization Methods for Science, Engineering and Technology

Figure 8.13. Confusion matrix for TPR versus FNR.

Figure 8.14. ROC for positive class 0.3.

8-26
 Modern Optimization Methods for Science, Engineering and Technology

8.6.3 Parallel plot

Figure 8.15 demonstrates the parallel characteristics plot for the decision tree, SVM,
NN, discriminant analysis classifiers and data only. ‘A parallel coordinate plot
displays as a line or chart each column in the data table. A line point reflects each of

Figure 8.15. Parallel plots.

8-27
 Modern Optimization Methods for Science, Engineering and Technology

a row’s attributes. This makes parallel coordinate plots comparable in appearance to

row graphs, but there is a significantly distinct manner in which information is
transformed into a plot. The values are always standardized in a parallel coordinate
plot. This implies that the smallest value in the respective column is set at 0 percent
for each stage along the x-axis and the largest value is placed at 100 percent along
the y-axis in that row. The size of the different columns is completely separate, so
comparing the height of the curve in one column with the height of the curve in
another column is meaningless’ [29, 30].

8.7 Conclusion
This chapter covered an introduction and background to LCS. The focus of this
chapter is to understand the LCS practically, therefore, MATLAB, BigML and
AzureML have been discussed to understand how the results from learning
classifiers can be obtained. The dataset and its processing, such as splitting into
training and testing datasets has been covered in detail with MATLAB code as well
as AzureML procedure. Six learning classifiers, namely logistic regression classifiers,
decision tree classifiers, discriminant analysis classifiers, support vector machines
classifiers, nearest neighbor classifiers and ensemble classifiers have been discussed.
The logistic regression classifier is included in MATLAB as an independent classifier
app, so the author has used AzureML to address various performance matrices and
also MATLAB. The remaining five classifiers have been tested using MATLAB’s
Classification Learner App.
In the latter half of this chapter, different visualizations of performance have been
shown, which include scatter plots, prediction plots, tree diagrams, code snippets,
confusion matrices, parallel plots and receiver operating characteristics. Table 8.4 at

Table 8.4. Performance comparison of LCS.

Prediction Training
Classifier Accuracy speed (ops) time (sec) Remarks

Decision trees 52.8 46 000 0.89 #Split 20, Gini’s diversity index
Discriminant 58.2 29 000 0.69 Linear, full covariance exponential
analysis kernel, MSE = 0.0048
Logistic 0.069 RMSE 12 000 15.41 MAE = 0.0522, R squared = 0.80,
regression basis function is constant
Support vector 60.8 5200 1.93 Linear kernel, box constraint
machines level = l, multiclass method,
one-vs-one
Nearest 57 17 000 2.64 N = 10, Euclidian distance, square
neighbor inverse distance weight
Ensemble 59 2400 3.96 Subspace ensemble dimension 4;
descriminant learning,
30 learners

8-28
 Modern Optimization Methods for Science, Engineering and Technology

the end of this chapter has been provided on the basis of experimentation carried out
by the author for the readers to justify their choices of algorithms. The LCS system
of MATLAB is a potent tool for researchers to obtain results quickly, and hence the
comparison of different classifiers is meaningful in the chapter. BigML is a purely
automated tool with almost negligible user intervention. AzureML is a robust tool
with a high degree of flexibility.

Acknowledgments
The author is grateful to his colleague Dr S P Dubey from Rungta College of
Engineering and Technology, Bhilai, India, who provided insight and expertise that
greatly assisted in the research, although they may not agree with all of the
interpretations/conclusions of this chapter. He is immensely grateful to his mentor
Dr G R Sinha, who always motivates him for research and development, and
academic excellence.

References
[1] Holland J 1976 Adaptation Progress in Theoretical Biology ed R Rosen and F M Snell
(New York: Academic)
[2] Goldberg D E and Holland J H 1988 Genetic algorithms and machine learning Mach. Learn.
3 95–9
[3] Urbanowicz R J and Moore J H 2009 Learning classifier systems: a complete introduction,
review, and roadmap J. Artif. Evol. Appl. 2009 736398
[4] Butz M V 2015 Learning classifier systems Springer Handbook of Computational Intelligence
(Berlin: Springer) pp 961–81
[5] Acharya M S, Armaan A and Antony A S 2019 A Comparison of Regression Models for
Prediction of Graduate Admissions ICCIDS 2019: IEEE Int. Conf. on Computational
Intelligence in Data Science Chennai, India,
[6] Holland J H et al 2000 What Is a Learning Classifier System? Learning Classifier Systems
(Berlin: Springer)
[7] Wilson S W 1995 Classifier fitness based on accuracy Evol. Comput. 3 149–75
[8] Riolo R L 1991 Lookahead planning and latent learning in a classifier system Proceedings of
the First International Conference on Simulation of Adaptive Behavior on From Animals to
Animats (Cambridge, MA: MIT Press), pp 316–26
[9] Stolzmann W 1998 Anticipatory classifier systems Genet. Program. 98 58–64
[10] Holland J H 1983 Escaping brittleness Proc. Second Int. Workshop on Machine Learning
[11] Robertson G G and Riolo R L 1988 A tale of two classifier systems Mach. Learn. 3 139–59
[12] Holland J H, Holyoak K J, Nisbett R E and Thagard P R 1989 Induction: Processes of
Inference, Learning, and Discovery (Cambridge, MA: MIT Press)
[13] Holmes J H, Durbin D R and Winston F K 2000 The learning classifier system: an
evolutionary computation approach to knowledge discovery in epidemiologic surveillance
Artif. Intell. Med. 19 53–74
[14] Zang Z, Li D and Wang J 2015 Learning classifier systems with memory condition to solve
non-Markov problems Soft Comput 19 1679–99

8-29
 Modern Optimization Methods for Science, Engineering and Technology

[15] Zebin T, Scully P J and Ozanyan K B 2017 Inertial sensor based modelling of human activity
classes: feature extraction and multi-sensor data fusion using machine learning algorithms
eHealth 360° (Cham: Springer), pp 306–14
[16] Noor N Q M, Sjarif N N A, Azmi N H F M, Daud S M and Kamardin K 2017 Hardware
Trojan identification using machine learning-based classification J. Telecom. Electron.
Comput. Eng. (JTEC) 9 23–7
[17] Maleki M, Manshouri N and Kayikçioğlu T 2017 Application of PLSR with a comparison
of MATLAB classification learner app in using BCI 2017 25th Signal Processing and
Communications Applications Conf. (SIU)
[18] Nagwanshi K K and Dubey S 2018 Statistical feature analysis of human footprint for
personal identification using BigML and IBM Watson analytics Arab. J. Sci. Eng. 43
2703–12
[19] Kessel M, Ruppel P and Gschwandtner F 2010 BIGML: a location model with individual
waypoint graphs for indoor location-based services PIK-Praxis Informationsverar.
Kommunikation 33 261–7
[20] Zainudin Z and Shamsuddin S M 2016 Predictive analytics in Malaysian dengue data from
2010 until 2015 using BigML Int. J. Adv. Soft. Comput. Appl. 8 18–30
[21] Chappell D 2015 New White Paper: Introducing Azure Machine Learning A Guide for
Technical Professionals, Microsoft Corporation http://davidchappellopinari.blogspot.com/
2015/08/new-whitepaper-introducing-azure.html
[22] Mántaras R L D 1991 A distance-based attribute selection measure for decision tree
induction Mach. Learn. 6 81–92
[23] Liu W Z and White A P 1994 The importance of attribute selection measures in decision tree
induction Mach. Learn. 15 25–41
[24] Navlani A 2018 Decision Tree Classification in Python https://www.datacamp.com/com-
munity/tutorials/decision-tree-classification-python
[25] Hallinan J S 2012 Data mining for microbiologists Systems Biology of Bacteria ed C
Harwood and A Wipat (Methods in Microbiology vol 39) (Amsterdam: Elsevier) pp 27–79
[26] Tharwat A, Gaber T, Ibrahim A and Hassanien A E 2017 Linear discriminant analysis: a
detailed tutorial AI Commun. 30 169–90
[27] Raschka S 2014 Linear Discriminant Analysis—Bit by Bit https://sebastianraschka.com/
Articles/2014_python_lda.html
[28] Sawla S 2018 Linear Discriminant Analysis https://medium.com/@srishtisawla/linear-
discriminant-analysis-d38decf48105
[29] Li M, Zhen L and Yao X 2017 How to read many-objective solution sets in parallel
coordinates IEEE Comput. Intell. Mag. 12 88–100
[30] Hussain A and Vatrapu R 2014 Social data analytics tool (SoDaTo) Advancing the Impact of
Design Science: Moving from Theory to Practice ed M C Tremblay, VanderMeer D,
Rothenberger M, Gupta A and Yoon V (Cham: Springer)

8-30
 IOP Publishing

Modern Optimization Methods for Science, Engineering and

Technology
G R Sinha

Chapter 9
A case study on the implementation of six sigma
tools for process improvement
Bonya Mukherjee, Rajesh Chamorshikar and Subrahmanian Ramani

A blast furnace (BF) is one of the primary processes for the production of crude steel
in an ore based integrated steel plant (ISP). In a BF, iron ore is reduced to pure
molten iron by supplying hot air blasts through the tuyers at the bottom of the
furnace, just above the hearth. A redox reaction takes place due to heat and mass
transfer and, along with the molten iron and slag that are tapped from the bottom of
the furnace, BF gas is emitted from the top of the BF as a by-product of the smelting
process. The raw BF gas carries a dust load of 25–40 g N−1 m−3, which is removed in
a gas cleaning plant (GCP) by spraying with water at high pressure through nozzles
located at various levels of a gas scrubber. The water removes the dust from the BF
gas and exits from the bottom of the scrubber in the form of slurry, is cooled in a
cooling tower and goes to radial settling tanks (RSTs) for removal of dust and is
then recycled back to the scrubber for cleaning of BF gas. The clean BF gas is then
fed into the gas network for consumption as fuel gas in various units of the plant.
Since April 2012, the ISP in question was facing the problem of a high concentration
of suspended solids in the clean water supply of the GCPs of BFs. The average
suspended solids concentration was 277 ppm against the specification of 100 ppm,
which was adversely affecting the functioning of the GCPs, leading to insufficient
removal of dust from BF gas. There were a number of factors that could have
impacted the concentration of total suspended solids (TSS) in the supply water to
GCPs. These factors were: insufficient removal of dust from BF gas in the dust
catchers of BFs; a high concentration of TSS in the fresh make-up water supply
itself; and insufficient and inefficient settling of dust and suspended solids in
the RSTs.
To narrow down the probable root cause/causes within the shortest possible time,
the authors decided to take up the above problem as a six sigma project. Based on

doi:10.1088/978-0-7503-2404-5ch9 9-1 ª IOP Publishing Ltd 2020

 Modern Optimization Methods for Science, Engineering and Technology

the past year’s performance data for the BF dust catchers, GCPs and RSTs, various
six sigma tools and techniques were used to find and filter out the critical inputs that
influence the output to a large extent. Thus, it was possible to quickly carry out a
root cause analysis (RCA) and it was found that the problem was mainly due to
certain operational lacunae in the RSTs and necessary remedial actions were taken.
The TSS levels in the supply water were effectively brought down to less than 100
ppm, and have been maintained at those levels since April 2013. The financial
benefits have been to the tune of Rs. 1 Crore per annum, as validated by the finance
department. Other benefits have included the reduced choking of nozzles/throttle
assembly and U-seals, as well as gas lines, reduced frequency of unscheduled/
emergency breakdowns/shutdowns in GCPs, and better removal of dust from BF
gas, leading to reduced specific fuel consumption and better furnace productivity at
the BF gas consumer end. The non-tangible benefits have been a renewed motivation
in the workforce, due to the elimination of non-value adding activities such as
attending to repetitive unscheduled repairs and sudden breakdowns.

9.1 Introduction
9.1.1 Generation and cleaning of BF gas
BF is one of the primary processes for the production of crude steel in an ore based
ISP. In a BF, counter current heat and mass exchange takes place for smelting of iron
ore using coal and coke to produce hot metal or molten iron in the liquid state. The
solid raw materials such as iron ore lumps, sinter and coke, and flux materials such as
limestone are charged from the top and hot air blast is supplied through the tuyers at
the bottom of the furnace, just above the hearth. A redox reaction takes place due to
heat and mass transfer and along with the molten iron and slag that are tapped from
the bottom of the furnace, BF gas is emitted from the top of the BF as a by-product of
the smelting process. This gas has a calorific value of 780–800 kcal N−1 m−3 and is
utilized as one of the primary fuel gases in an ISP. The raw BF gas carries a dust load
of 25–40 g N−1 m−3. This dust content has to be removed or minimized before the raw
BF gas can be used as a fuel in the various furnaces and stoves of an ISP.

9.1.1.1 BF gas cleaning process

Each BF is equipped with a dry dust catcher in which more than 50% of the dust
load is separated from the gas, and a GCP which is either dry type or wet type.
The functions of the GCPs of BFs are:
a. To reduce the temperature of BF gas.
b. To remove dust content from BF gas using a wet scrubber and atomizer.
c. To maintain the top pressure of BF.

The primary function of the GCPs of BF is to remove the suspended dust particles
from the raw BF gas, before it is supplied to the consumer as fuel. The raw BF gas
emerging from the BFs has a dust load of around 25–40 g N−1 m−3. More than 50%
of the dust is removed in the dry dust catcher. The raw BF gas containing the

9-2
 Modern Optimization Methods for Science, Engineering and Technology

remaining 10–15 g N−1 m−3 of dust then goes to the scrubber where the gas flows
from bottom to top and water is sprayed on it from nozzles at four levels at a rate of
800–1000 m3 h−1. After the scrubber, the raw BF gas passes through a set of parallel
venturi atomizers where again water is sprayed on the gas, to remove the remaining
dust. The dust content of clean BF gas is less than 5 mg N−1 m−3. The clean BF gas
is then fed into the gas network for consumption as fuel gas in various units of
the plant.

9.1.1.2 Problems encountered in cleaning of BF gas

The slime water containing the dust removed from the BF gas collects in a common
slime channel from where it is distributed to radial settling tanks (RSTs) for settling
of the suspended solids and the clarified overflow water is supplied (pumped) back to
the scrubbers of the GCPs, after cooling in cooling towers and adding fresh make-up
water to compensate for evaporation and other process losses. In this particular case,
the slime water was distributed to five RSTs, to RSTs 1 and 2 through one route and
to RSTs 3, 4 and 5 through an alternative route from a common slime water
channel. As per standard operating practices (SOPs), the TSS levels of supply water
to GCP scrubbers, after passing through the RSTs and the addition of fresh make-up
water, should be less than 100 ppm, on average. But, for the duration of a year, it
was observed that the TSS of the supply water to the GCPs was as high as 280–300
ppm, which was affecting the functioning of the GCPs in the following ways:
a. Choking of the nozzles in the scrubbers
b. Erosion of nozzle, throttle assembly and atomizer shell plate surfaces
c. Insufficient removal of dust content from BF gas, thus resulting in insuffi-
cient cleaning of BF gas

The authors decided to use a six sigma (DMAIC) process to bring about process
improvement by reducing the concentration of TSS in the supply water, which was
adversely affecting the operation and equipment health of the GCPs of the BFs.
Various six sigma tools and techniques were used to find and filter out the critical
inputs that influenced the output to a large extent.

9.2 Problem overview

Since April 2012, the ISP in question had been facing the problem of a high
concentration of suspended solids in the supply water of BF GCPs 1–6. The average
suspended solids concentration was 277 ppm against the specification of 100 ppm.
(See figure 9.1)
The TSS in supply water started increasing from the month of April 2012 and was
drastically high in the months of May 2012 and June 2012. The objective of the six
sigma project was to reduce the TSS in the water supplied for gas cleaning from an
average 277 ppm to 130 ppm by May 2013, on the basis of past performance data.
The project was started in December 2012 and the performance improvement
started from April 2013.

9-3
 Modern Optimization Methods for Science, Engineering and Technology

Time Series Plot of TSS Average Supplied

700 678.000

600

500
TSS Average Supplied

400 380.000

321.250
300
222.800 210.000
197.500
200 162.500
145.000 133.750
92.333
100 68.600 58.000 63.000

0
12 01 02 03 04 05 06 07 08 09 10 11 12
011 012 012 012 012 012 012 012 012 012 012 012 012
2 2 2 2 2 2 2 2 2 2 2 2 2
Months

Figure 9.1. Monthly TSS (in ppm) in the supply water of GCPs 1–6.

9.3 Project phase summaries

9.3.1 Definition
The definition phase is the beginning phase of the six sigma improvement process. In this
phase, the project team identifies and defines the problem, creates a map of the process
and outlines the focus of the project for themselves and the customers of the process. The
project team measures both the process baseline and the process entitlement.
Process baseline: The current performance of a process over a timeline before any
change is made the input variable(s). This baseline data can also be used as a
yardstick for measuring future efficiency in the process.
Process entitlement: The historical best performance of a process that a project team
is ‘entitled’ to achieve. It is also based on design data or technical specifications of
the process/equipment. The entitlement analysis also helps to decide whether the
project should address process improvement (DMAIC) or process redesign (design
for six sigma).
On analyzing one year’s data for the TSS of the supply water to GCPs 1–6 from
December 2011 to December 2012 in comparison to historical data, the average or
baseline TSS was found to be 277 ppm, while the entitlement was 68 ppm, based on
standard norms. The P-value of 0.021 (see figure 9.2) also shows that the probability
distribution was not normal for the average TSS of the supply water from April 2012.

9-4
 Modern Optimization Methods for Science, Engineering and Technology

Summary for TSS Average Supplied_Not OK

Anderson-Darling NormalityTest
A-Squared 0.82
P-Value 0.021
Mean 272.31
StDev 172.47
Variance 29746.49
Skewness 1.92045
Kurtosis 3.92183
N 9

Minimum 133.75
1st Quartile 153.75
Median 210.00
3rd Quartile 350.63
100 200 300 400 500 600 700
Maximum 678.00
95% Confidence Interval for Mean
139.74 404.88
95% Confidence Interval for Median
148.99 366.61

95% Confidence Intervals 95% Confidence Interval for StDev

116.50 330.42
Mean

Median

100 150 200 250 300 350 400

Figure 9.2. Graphical summary for the period April 2012–December 2012 before the start of the project.

However, a graphical summary (see figure 9.3) of the TSS data for the four
months before April 2012, i.e. December 2011 to March 2012 shows that the average
TSS was not only below 100 ppm, but also the distribution was normal, i.e. the
process capability was within the limits.
In a hypothesis test, a P-value establishes the significance of the results.
Hypothesis tests are used to establish the validity of a claim made about a dataset.
This claim is known as the null hypothesis.
The P-value is a number between 0 and 1 and can be interpreted as such:
a. A P-value of less than 0.05: strong evidence against the null hypothesis and
hence the null hypothesis can be rejected.
b. A large P-value of more than 0.05: weak evidence against the null
hypothesis, so the null hypothesis cannot be rejected outright and can be
further investigated statistically for validity.
c. P-values very close to 0.05 can go either way.

9.3.2 Measurement
In the measurement phase of a lean six sigma improvement process, the current
performance of the process, the magnitude of the problem and the probable factors
affecting the problem are measured.

9-5
 Modern Optimization Methods for Science, Engineering and Technology

Summary for TSS Average Supplied_OK

Anderson-Darling Normality Test
A-Squared 0.37
P-Value 0.221

Mean 70.483
StDev 15.197
Variance 230.934
Skewness 1.53124
Kurtosis 2.46109
N 4

Minimum 58.000
1st Quartile 59.250
Median 65.800
3rd Quartile 86.400
60 65 70 75 80 85 90
Maximum 92.333
95% Confidence Interval for Mean
46.302 94.664
95% Confidence Interval for Median
58.000 92.333

95% Confidence Intervals 95% Confidence Interval for StDev

8.609 56.661
Mean

Median

40 50 60 70 80 90 100

Figure 9.3. Graphical summary for the period December 2011–March 2012.

Various six sigma tools were used in the measurement phase, namely the process
flow diagram, time series plot, regression and correlation analysis, I/O worksheet,
and C and E matrix, to identify the causes of high TSS in the supply water.
Regression analysis (fitted line plot): Fitted line plots were drawn between the
measured TSS of the return water from GCPs 1–6 to RSTs 1 and 2 and the return
water to RSTs 3, 4 and 5 to check whether the TSS of the return water had any
bearing on the TSS of the supply water. A close correlation was found between the
TSS of the return water before RSTs and supply water after RSTs, as is evident from
figures 9.4 and 9.5.
A fitted line plot is a statistical technique for regression analysis (linear) to find the
best fit line for a set of data points. This is used when experimental data or historical
process data are plotted and the data points are scattered across the plot area.
The formula for fitted line plot or regression line is y = mx + b where m is the
slope and b is the y-intercept.
R2 or the coefficient of determination is a statistical measure of how close the data
are to the fitted line and how much the variation in dependent variables can be
explained by the variation in the independent variable.
An R2 value of 50% and above usually indicates a better correlation between the
dependent and independent variables. A low R2 value usually fails to indicate any

9-6
 Modern Optimization Methods for Science, Engineering and Technology

Fitted Line Plot

TSS Average Supplied_Not Ok = - 244.2 + 0.2776 TSS-Returned avg 1&2_Not Ok
700 S 135.074
R-Sq 46.3%
TSS Average Supplied_Not Ok

R-Sq(adj) 38.7%
600

500

400

300

200

100

0
1000 1200 1400 1600 1800 2000 2200 2400 2600
TSS-Returned avg 1&2_Not Ok

Figure 9.4. Fitted line plot between the average TSS of supply water and the average TSS of return water to
RSTs 1 and 2; with an R2 value of 46.3%.

Fitted Line Plot

TSS Average Supplied_Not OK = - 358.6 + 0.3386 TSS-Returned avg 3,4&5_Not OK
700 S 122.437
R-Sq 55.9%
TSS Average Supplied_Not OK

R-Sq(adj) 49.6%
600

500

400

300

200

100

0
1200 1400 1600 1800 2000 2200 2400 2600
TSS-Returned avg 3,4&5_Not OK

Figure 9.5. Fitted line plot between the average TSS of supply water and the average TSS of return water to
RSTs 3, 4 and 5; with an R2 value of 55.9%.

9-7
 Modern Optimization Methods for Science, Engineering and Technology

straightforward correlation between a single dependent variable with an independent

variable.
The R2 value, however, is not always straightforward and, hence, statisticians
have to rely on adjusted R2 values to establish any kind of correlation between
dependent and independent variables.
After a close correlation was established between the TSS levels of supply water
and the TSS levels of return water from the scrubbers, it was deemed prudent to
examine other factors that might be contributing to the increased TSS of the return
water from the scrubbers of the GCPs.
The raw BF gas passes through a dry dust catcher, where more than 50% of the
dust is removed before it enters the scrubbers. The quantity of dust removed can be
measured from the average number of dump cars that are filled up with dust
removed from the dust catchers periodically. It was felt that inefficient operation of
the dust catchers might cause insufficient removal of dust from the BF gas, which
might have in turn resulted in a higher dust load in the scrubbers, thereby adversely
impacting the TSS of both return water and supply water.
Therefore, fitted line plots were drawn between the average TSS of supply water
and the average number of dump cars (see figure 9.6).
Also, separate fitted line plots were drawn between:
1. The average TSS of return water to RSTs 1 and 2 and the average number of
dump cars (see figure 9.7)
2. The average TSS of return water to RSTs 3, 4 and 5 and the average number
of dump cars (see figure 9.8)

Fitted Line Plot

TSS Average Supplied_Not Ok = - 304.9 + 98.9 AvgNoDumpCars

700 S 173.168
R-Sq 11.8%
R-Sq(adj) 0.0%
TSS Average Supplied_Not Ok

600

500

400

300

200

100
5.0 5.5 6.0 6.5 7.0
AvgNoDumpCars

Figure 9.6. Fitted line plot between average TSS of supply water and the average number of dump cars from
the dust catchers; with an R2 value of 11.8%.

9-8
 Modern Optimization Methods for Science, Engineering and Technology

Fitted Line Plot

TSS-Returned avg 1&2_Not Ok = 1353 + 87.0 AvgNoDumpCars
2600 S 448.629
R-Sq 1.5%
2400
TSS-Returned avg 1&2_Not Ok

R-Sq(adj) 0.0%

2200

2000

1800

1600

1400

1200

1000
5.0 5.5 6.0 6.5 7.0
AvgNoDumpCars

Figure 9.7. Fitted line plot between the average TSS of return water to RSTs 1 and 2 and the average number
of dump cars from dust catchers; with an R2 value of 1.5%.

Fitted Line Plot

TSS-Returned avg 3,4&5_Not Ok = 1529 + 57.2 AvgNoDumpCars

2600 S 405.481
R-Sq 0.8%
TSS-Returned avg 3,4&5_Not Ok

R-Sq(adj) 0.0%
2400

2200

2000

1800

1600

1400

1200
5.0 5.5 6.0 6.5 7.0
AvgNoDumpCars

Figure 9.8. Fitted line plot between the average TSS of return water to RSTs 3, 4 and 5 and the average
number of dump cars from dust catchers; with an R2 value of 0.8%.

9-9
 Modern Optimization Methods for Science, Engineering and Technology

All the three fitted line plots (see figures 9.6, 9.7 and 9.8) show that there is little or
no correlation between the average number of dump cars filled from the dust
catchers and the average TSS of the return water as well as supply water. Thus, the
dust catchers were eliminated from the scope of this project, and only the closed loop
circuit of water supply to the scrubbers of GCPs 1–6 was taken into consideration

Figure 9.9. Process flow diagram of the closed loop treatment circuit for GCP 1–6 supply water for the
removal of total suspended solids.

9-10
 Modern Optimization Methods for Science, Engineering and Technology

for identifying critical independent variables that could significantly affect the
dependent variable, namely the TSS of supply water, in the case of our project.
Process flow diagram (PFD): Through PFD, we created a visual representation of
the closed loop process flow of the supply of water to the GCP scrubbers, exit of dust
laden return water to the common slime channel, its distribution to different RSTs,
settling and removal of the dust, addition of make-up water to the overflow water
and pumping back of the supply water to the scrubbers. The PFD helped us to
identify loops, hidden factories and non-value adding activities in the process (see
figure 9.9).
A process is a combination of sequential or closed loop activities (or sub-
processes) that culminate in an outcome in the form of a product or service. All
activities in a processes can be categorized, in six sigma parlance, into the broad
categories of:
1. Value added: This activity in the process adds form, function and value to the
end product or service and is of value to the customer.
2. Non-value added: This activity in the process does not add any form, function
or value to the end product or service and can be eliminated.
3. Non-value added but necessary: This activity by itself may not add any value
but is necessary in the final sequence of activities in order to add value to the
final product, or may be deemed as value added for another concurrent,
dependent or symbiotic sub-process that leads to the final product or service.

A hidden factory refers to the activities or sub-processes that reduce the quality or
efficiency of an operation or business process, and may result in waste or bad work
as they remain hidden under the day-to-day operations.
The hidden factories can be in the form of:
a. Production loss.
b. Compromised quality.
c. Reduced availability of equipment.
d. Delayed delivery.

Identifying the non-value added activities and the hidden factories can help to address
the operational issues that can adversely affect the quality of the process outcome.
Mapping of the process flow diagram of the closed loop treatment circuit for the
supply water of BF GCPs 1–6 for removal of TSS helped to identify two hidden
factories that were adversely affecting the TSS levels in the supply water.
The two existing hidden factories identified were:
a. The jamming of a valve plate in the common slime channel that was causing
uneven distribution of water into the RSTs, thereby increasing the water load
in RSTs 1 and 2 beyond their capacities, and reducing the water load in
RSTs 3, 4 and 5. This was in turn adversely affecting the TSS settling
capability of the RSTs 1 and 2.
b. Further, it was also found that RST 1 was not working properly, thereby
again impacting the removal of TSS from the return slime water.

9-11
 Modern Optimization Methods for Science, Engineering and Technology

Table 9.1. Input–output worksheet.

Input–output (I/O) worksheet: An I/O worksheet was constructed to provide details

of the inputs and outputs identified in the PFD. Through a spreadsheet in the form
of an I/O worksheet we listed steps from PFD and identified process steps, process
input/output, specification and VA or NVA. The output of this I/O worksheet was
directly fed into the cause and effect matrix and FMEA. (See table 9.1 for the I/O
worksheet.)
Fishbone diagram: The fishbone diagram was made after a brainstorming session
with the extended team members. Individuals from the water management depart-
ment up to the supervisory grades were also included. (See figure 9.10.)
Cause and effect matrix: The team members prepared the cause and effect matrix by
taking inputs from all process steps from the I/O worksheet as well as from the
fishbone diagram. It was found that inputs from man (personnel) and machine were
the most frequent and critical. With these as the starting point, we proceeded further
to carry out the failure mode and effects analysis (FMEA) based on both the I/O
sheet and fishbone diagram. (See table 9.2 for the cause and effect matrix.)

9-12
 Modern Optimization Methods for Science, Engineering and Technology

Fish Bone Diagram-High TSS in Supply Water for GCPs 1-6

Measurements Material Personnel

Improper throttling of sluice

valve
Improper throttling of flow
control valves
No stirring of electrolyte

High/low ratio of water &

polyelectrolyte
Electrolyte composition not
High/low ratio of water & meeting specs
polyelectrolyte Electrolyte composition
Less valve opening
not meeting specs
High TSS in
supply water
Fresh water supply pipe line
Less valve opening
rupture
Fresh water pump failure
Low head of polyelectrolyte
slurry Pipe line
jammed/ruptured
Improper throttling of flow slurry pump not working
control valves
Pump discharge not same as
No stirring of electrolyte capacity of slurry pump
Rotating arm of RSTs damaged

Improper throttling of Mechanical jamming of control

sluice valve valves
Jamming of sluice valve

Environment Methods Machines

Figure 9.10. Fishbone diagram to identify possible causes for high TSS in the supply water.

Table 9.2. Cause and effect matrix.

9-13
 Modern Optimization Methods for Science, Engineering and Technology

Failure modes and effects analysis: The input for FMEA was taken from the detailed
map, cause and effects matrix and fishbone diagram. The important processes and
their outputs were tabled under the column process function. These outputs were
analyzed for their modes of failure and their effects on the primary metric. The risk
priority numbers (RPNs) were calculated, according to their severity, occurrences
and detection. The processes with the highest RPNs were taken for further analysis.
(See table 9.3 for FMEA.)

Table 9.3. Failure modes and effects analysis.

9-14
 Modern Optimization Methods for Science, Engineering and Technology

Table 9.4. Why-why for improper throttling of valve.

Improper throttling of valve

Question Answer

What is your final action? Proper regulation of throttling of

valve
After regulation of throttling, is the distribution of water Yes
proper?
Why Answer Action
Why was throttling not Lack of SOP for throttling. SOP to be made.
proper? Lack of knowledge of the Train the operator.
operator. Prepare a training manual for
operators.

9.3.3 Analyze and improvement

Why-why analysis: Why-why analysis of each process with the highest RPNs was
carried out to identify the root cause and eliminate or reduce it to the minimum extent
possible. The why-why analysis sheets of each cause is as shown in tables 9.4–9.9.
Based on the inputs from FMEA and why-why analysis, certain process improve-
ments in the form of quick wins were carried out immediately. These were:
1. Elimination of the hidden factory: The broken valve plate obstructing the
homogeneous flow of return water in the common slime channel was
removed, thereby facilitating optimum distribution of return water to all
the RSTs.
2. Cleaning of the common slime water channel: This also facilitated the
optimum distribution of return water to all the RSTs
3. Increasing the ratio of polyelectrolyte to water in the polyelectrolyte
emulsion: Ratio of polyelectrolyte to water was increased from 100 g/
3000 l/ 8 h to 250 g/6000 l/ 8 h in RSTs 3–5 and 100 g/3000 l/ 8 h to
400 g/ 10 000 l/ 8 h in RSTs 1–2, to improve the sedimentation rate of the
suspended solids.
4. Lubrication and throttling of distribution valves: Lubrication of valves was
carried out to ensure full throttling of the valves in each distribution channel
to each RST so that the full flow of the return water to each RST could be
ensured. This also resulted in correct distribution of return water to
each RST.

After bringing about the above mentioned improvements, data were collected for the
TSS of return water to RSTs 1 and 2 and RSTs 3, 4 and 5 as well as the TSS of the
supply water at the pump house outlet one day per week, over a period of
two months. On the basis of the new data on the TSS of the supply water, capability
analysis of the process was carried out, as shown in figure 9.11. With Cp and Pp of
more than 1, the process capability has been improved.

9-15
 Modern Optimization Methods for Science, Engineering and Technology

Table 9.5. Why-why for the incorrect ratio of water and polyelectrolyte.

Incorrect ratio of water and polyelectrolyte

Question Answer

What is your final action? Ratio of polyelectrolyte and

water increased.
After increasing the ratio, is the settling of dust particles okay? Yes
Why Answer Action
Why was the ratio of It was felt that the ratio of Regulated increase in
polyelectrolyte to water polyelectrolyte dose to water polyelectrolyte dosing
increased? was low and inadequate. ratio.
Why was the ratio of The exact ratio of Ratio of polyelectrolyte and
polyelectrolyte to water low and polyelectrolyte to water was water told to the operator.
inadequate? not known to the operator.
Higher polyelectrolyte dosing Recommended installation of
leads to higher settling of higher capacity slurry
dust which in turn leads to pumps and cleaning of
choking of the RSTs. slurry pipeline.
Why was the exact ratio of No SOP existing for SOP to be prepared.
polyelectrolyte to water not polyelectrolyte dosing.
known to the operator?

Table 9.6. Why-why for mechanical jamming of valve.

Mechanical jamming of valve

Question Answer

What is your final action? Removal of jamming.

After removal of jamming, is valve ok? Yes
Why Answer Action
Why did you remove jamming? The valve was jammed and Lubricating the valve.
not operating.
Why was the valve jammed? Less frequent operation. Frequent operation of the
valve is needed.
No lubrication/less Frequent lubrication of the
lubrication. valve is required.
Why was valve not operated No standard schedule was Make operation schedule.
frequently? made for the operating
valve.
Why was no schedule made for valve The importance of frequent Impart information to
operation? valve operation was not working group about the
known to the working frequent operation of
group. valve.
Why was valve not lubricated No schedule was made for Prepare the valve
frequently? lubricating the valve. lubrication schedule.

9-16
 Modern Optimization Methods for Science, Engineering and Technology

Table 9.7. Why-why for the deposits of mud and slabs in the slime channel.

Deposits of mud and slabs in slime channel

Question Answer

What is your final action? Cleaning of the slime

channel.
After cleaning, is the distribution of water proper? Yes
Why Answer Action
Why was the slime channel filled Because of slabs falling into Removal of slabs from the
with mud and slabs? the slime channel. slime channel.
Because of a broken valve Broken valve piece
piece stuck in the channel. removed.
Why did slabs fall into the slime Slabs were old and broken. Replacement of slabs done.
channel?
Broken slabs were not Replacement of slabs done.
replaced.
Why were broken slabs not Not much importance given Periodic housekeeping to be
replaced? to housekeeping. done.
Why was a broken valve piece A valve was broken during Prevent breaking of valves
stuck in the channel? operation and was not during operation.
removed.
Why was the valve broken during The valve was not maintained Proper maintenance of the
operation and not removed? properly valve to be done.
Removal of broken valve Attention to be paid to
piece not considered removal of broken parts
necessary. in future.
Why was the valve not NO SOP for valve Prepare a maintenance
maintained properly? maintenance. schedule and
maintenance checklist.
Why was the removal of the The housekeeping was not Prepare housekeeping
broken valve piece not proper. schedule.
considered important?

Box plot analysis was also carried out for three sets of TSS data:
a. Average TSS of supply water for the period December 2011 to March 2012,
when the TSS value was meeting minimum requirements: TSS average
supplied—ok.
b. Average TSS of supply water for the period April 2012 to December 2012,
when TSS values were not meeting minimum requirements: TSS average
supplied—not ok.
c. Average TSS of the supply water for the period April 2013 to May 2013,
after carrying out the six sigma improvements: TSS average supplied—
improved.

9-17
 Modern Optimization Methods for Science, Engineering and Technology

Table 9.8. Why-why for no stirring of electrolyte emulsion.

No stirring of electrolyte emulsion

Question Answer

What is your final action? Stirring of electrolyte with

compressed air.

After stirring, is the settling of dust particles okay? Yes.

Why Answer Action
Why was stirring not done? A lack of facilities for supply of Air supply provided and
compressed air. stirring being done.
Less importance given to stirring. Operator asked to monitor
stirring of polyelectrolyte
with compressed air.
Why was less importance Operator unaware of the Importance of stirring to
given to stirring? operational importance of achieve homogeneous
stirring. polyelectrolyte solution
communicated to operator.
Why was the operator Stirring not included in the SOP Include stirring in SOP.
unaware of the for polyelectrolyte dosing.
operational importance
of stirring?

Table 9.9. Why-why for insufficient pump discharge.

Pump discharge not sufficient

Question Answer

What is your final action? Check existing pump

discharge/install higher
capacity pump.
After checking existing pump discharge, is it okay? Yes.
Why Answer Action
Why did you check the Settled slurry in the RSTs is not being Check discharge of slurry
pump capacity? removed completely, affecting the pumps.
efficiency of the RSTs.
Why is the settled slurry Existing pump is old. Recommend installation of
not being removed higher capacity slurry
completely? pumps in RSTs.
Slurry pipeline may be choked. Pipeline condition checked.
Why is slurry pipeline Not cleaned regularly. Slurry pipeline to be cleaned.
choked?
Why is slurry pipeline NO SOP for cleaning of slurry pipeline. Prepare SOP for cleaning of
not cleaned regularly? slurry pipeline.

9-18
 Modern Optimization Methods for Science, Engineering and Technology

Process Capability of TSS Average Supplied- Apr-May'13

LSL USL
Process Data Within
LSL 0.00000 Overall
Target *
USL 100.00000 Potential (Within) Capability
Sample Mean 70.00000 Cp 1.10
Sample N 7 CPL 1.54
StDev (Within) 15.20000 CPU 0.66
Cpk 0.66
StDev (Overall) 13.13851
CCpk 1.10
Overall Capability
Pp 1.27
PPL 1.78
PPU 0.76
Ppk 0.76
Cpm *

0 15 30 45 60 75 90 105
Observed Performance Exp. Within Performance Exp. Overall Performance
PPM < LSL 0.00 PPM < LSL 2.06 PPM < LSL 0.05
PPM > USL 0.00 PPM > USL 24208.83 PPM > USL 11204.44
PPM Total 0.00 PPM Total 24210.89 PPM Total 11204.49

Figure 9.11. Capability analysis of TSS average supplied for the period April–May 2013—post improvement.

Box plot of TSS Average Supp_Ok, TSS Average Supp_NotOk, TSS Average Supp_Improved
700

600

500

400
Data

300

200

100

0
TSS Average Supplied_OK TSS Average Supplied_NotOK TSS Average Supplied_Improved

Figure 9.12. Box plot analysis of the TSS of the supply water: historical, skewed and improved.

The result is as depicted in figure 9.12.

The probability plot of the improved TSS of the supplied water (April–May 2013)
compared to the historical data (December 11–March 2012) also shows that the
process variations have been minimized and brought within the specification limit, at
95% CI (see figure 9.13). The process variations have been minimized and brought
within the specification limit, at 95% CI.

9-19
 Modern Optimization Methods for Science, Engineering and Technology

Probability Plot of TSS Average Supp_Ok, TSS Average Supp-Improved

Normal-95% CI
99
Variable
TSS Average Supplied_OK
TSS Average Supplied_Improved
95
Mean StDev N AD P
90 70.48 15.20 4 0.369 0.221
67.77 12.60 7 0.478 0.156
80
70
Percent

60
50
40
30
20

10

1
0 20 40 60 80 100 120 140
Data

Figure 9.13. Probability plot of TSS average supplied: historical versus improved.

9.3.4 Control
Various control measures were adopted and practiced for maintaining the processes
and the same activities are evaluated for reducing breakdowns. The control plan is
shown in table 9.10.

9.4 Conclusion
9.4.1 Financial benefits
The financial benefits forecast, on account of savings in terms of reduced breakdown
and reduced replacement of spares, as validated by the finance department, was
0.992 crores/annum and the actual savings from March 2013 to May 2013 were
Rs. 24.8 lakhs.

9.4.2 Non-financial benefits

The non-financial benefits have been in the form of reduced suspended solids (TSS)
of less than 100 ppm consistently since April 2013. Other consequent non-financial
benefits have been as follows (the domino effect):
1. Reduced choking of nozzles/throttle assemblies, U-seals, and drain pots of u-
seals as well as gas lines.
2. Visible reduction in the wearing out/erosion of scrubber shell plates.
3. Fewer unscheduled/emergency breakdowns/shutdowns in GCPs 1–6, due to
BF gas leakages.
4. Renewed motivation in the workforce, due to the elimination of non-value
adding activities towards correcting repetitive unscheduled/sudden breakdowns.

9-20
 Modern Optimization Methods for Science, Engineering and Technology

Table 9.10. Control plan for improved TSS in supply water.

5. The lifespan of the throttle assemblies (septum valves), nozzles, gate valves
and overflow throttles of the scrubber and atomizers will increase.
6. The BF gas leakage problem will be reduced. The safety of GCP operators as
well as the area surrounding GCPs will be ensured.
7. Better removal of dust from BF gas, thereby causing less choking of valves
and burners at the consumer end, leading to reduced specific fuel consump-
tion and better furnace productivity.

9-21
 IOP Publishing

Modern Optimization Methods for Science, Engineering and

Technology
G R Sinha

Chapter 10
Performance evaluation and measures
K A Venkatesh and N Pushkala

Measuring performance is an art and identifying the right tool to evaluate the
performance is another Herculean task. Philosophically, anything and everything is
perception. Evaluating the performance in terms of the perception of the promoters,
the internal customers/employees or the end users are different tasks. This goal of
this chapter is to introduce a tool which could address them all. In this chapter, we
introduce an elegant non-parametric tool, called data envelopment analysis (DEA),
in which the choice of inputs and outputs is based on the needs of the evaluator.
Among the many other tools, the growth of DEA literature is incomparable. DEA is
an extension of linear programming problems, in which one has to solve a number of
linear programming problems in order to obtain relative efficiency scores. Initially
this chapter introduces the various paradigms in performance measurement for
decision making and ends with a discussion of the tool R in addressing DEA.

10.1 Performance measurement models

For success in any environment, the most important aspect is sustainability.
Efficiency is the only parameter which makes an individual or organization
sustainable. Many methods and models are available in the literature and in
practice. Most of them are parametric ones, that is, mundane ones. These methods
and models may not be suitable from the perspective of the observer. Needlessly, the
observer may not be part of the system; sometimes the observer may be a one-time
user of the system. Are there any scientific ways to measure the efficiency? This
chapter will address this question and discuss various models and methods to
measure the performance. Generally these models are used to assess the intended
system to make a decision in order to attain the mission/goal. Optimization
techniques are the most common tools used to make the best choice among the
available options in order to make appropriate decisions. The decision making

doi:10.1088/978-0-7503-2404-5ch10 10-1 ª IOP Publishing Ltd 2020

 Modern Optimization Methods for Science, Engineering and Technology

process can be achieved by deploying any of the following three approaches, or

hybrid approaches: descriptive analysis, prescriptive analysis and normative
analysis.
Further, the decision making process can be classified based on single or multiple
criteria. Multiple-criterion decision making problems (MCDM) can be classified as
multiple attribute decision making problems (MADM) and multiple objective
decision making problems (MODM). MADM evolved from the Bernoulli utility
theory. More precisely, multiple attribute utility theory (MAUT) and out-ranking
methods (such as ELECTRE and PROMETHEE) are the two subcategories of
MADM. Fuzzy MADM has been introduced to deal with situations of uncertainty.
The objective of a multi-objective decision making problem is to attain an
optimal solution for the given set of (contradictory) objectives subject to the set of
constraints. Hence, to associate an optimization technique with MODM is the
natural choice. The foremost aim in these models is to convert the given multiple
objectives into a weighted single objective function, otherwise pareto solutions must
be obtained. Evolutionary algorithms, genetic algorithms, particle swarm optimi-
zation and vector optimization are the popular models of MCDM.
Single criteria decision making problems are modeled as linear programming
problems to obtain the desired optimal solution. In this chapter, we present DEA as
a tool to study performance measurement. In addition to DEA, this chapter deals
with the analytical hierarchy process (AHP) and fuzzy AHP.

10.1.1 Fuzzy sets

This section deals with the fundamentals of fuzzy sets and operations on the fuzzy sets.

Definitions
1. Let X be a set. The fuzzy set A′ is given by A′ = {(a , μA(x ))∣x ∈ X }, where
μA ′(x ):X → [0, 1] is the membership function.
2. The support of the fuzzy set A′ is given by supp(A′) = {x ∈ X ∣μA ′(x ) > 0}.
3. The α − cut of a fuzzy set A′ is defined as A′(α ) = {x ∈ X ∣μA ′(x ) ⩾ α},
∀ αϵ[0, 1].
4. The height of ′ is given by h(′) = supx∈X μA(x ). If h(A′) = 1, then the
fuzzy set is said to be normal, otherwise it is called subnormal.

Operations
The fundamental arithmetic operations addition, subtraction, multiplication and
division (/) are obtained based on the extension principle and α − cut arithmetic.

Extension principle
Let a′, b′ be any two fuzzy numbers and c be a particular event. The fundamental
operations are defined as:
1. μa ′+b ′(c ) = suppx,y∈X {mini(a′(x ), b′(y )∣x + y = c}.
2. μa ′−b ′(c ) = suppx,y∈X {mini(a′(x ), b′(y )∣x − y = c}.

10-2
 Modern Optimization Methods for Science, Engineering and Technology

3. μa ′ Xb ′(c ) = suppx,y∈X {mini(a′(x ), b′(y )∣xXy = c}.

4. μa ′/b ′(c ) = suppx,y∈X {mini(a′(x ), b′(y )∣x /y = c}.

α − cut arithmetic
Let r, s, t be the infimum, the mode and the superimum of the two fuzzy numbers
a′ = [a r , a s , at ] and b′ = [br , b s , bt ], respectively. Using α − cut , one could define
the fundamental operations as:
1. a′(α ) + b′(α ) = [a r (α ) + br (α ), at(α ) + bt(α )].
2. a′(α ) − b′(α ) = [a r (α ) − b s (α ), at(α ) − br (α )].
3. a′(α )Xb′(α ) = [A, B ]
where A = mini{a r , br , a s , b s , at , bt},
B = mini{a r , br , a s , b s , at , bt}.
a ′ (α ) ⎡ ⎤
4. b ′ (α ) = [a r(α ), at(α )]⎢⎣ t1 , br1(α ) ⎥⎦ .
b (α )

10.2 AHP and fuzzy AHP

AHP is one of the MADM methods [1]. The various steps involved in AHP are:
1. Identify the nature of the problem.
2. Construct the hierarchy at all possible levels.
3. Construct pairwise judgement matrix at all levels.
4. Synthesization.
5. Consistency test.
6. Comparisons at all levels.
7. Develop overall rankings to identify the best alternatives.

Let a1 a2, …, an be the attributes and the relative importance is given by w1, w2, …,
wn and the pairwise comparison is given in the matrix form, called an association
matrix
⎡ a11 ⋯ a1n ⎤
A = ⎢ ⋮ ⋱ ⋮ ⎥,
⎢⎣ an1 ⋯ ann ⎥⎦

where aij = 1/aji and aij · ajk = aik . Generally the ratio wi /wj is unknown. The
problem of AHP is to determine aij = wi /wj . Let W be the weight matrix given as
⎡ w1/ w1 ⋯ w1/ wn ⎤
A = ⎢⎢ ⋮ ⋱ ⋮ ⎥⎥ ,
⎣wn / w1 ⋯ wn / wn ⎦
where w = [w1, w2, .., wn ]T is a column vector. The system can be written as
(W − nI )w = 0. One could employ an eigenvalue method to solve the system
(W − nI )w = 0, then obtain the comparative weights by finding the largest
eigenvalue λ of A in magnitude. For the consistency test we introduce two more
indices called the consistency index (CI) and consistency ratio (CR). The index CI is
given by CI = (λ − n )/(n − 1) and CR is given by CR = CI/random CI.

10-3
 Modern Optimization Methods for Science, Engineering and Technology

10.2.1 Fuzzy AHP

To capture the vague information from a real world problem, one has to deploy
fuzzy numbers/linguistic variables. Triangular fuzzy numbers were incorporated in
AHP by van Laaroven and Pedrycz in a pairwise comparison matrix. Subsequently
many deployed various fuzzy numbers with AHP and the theory and models of
FAHP have increased.
The geometric mean method is one of the popular methods to address FHAP.
The pairwise comparisons are represented by linguistic variables using fuzzy
numbers; let us define the fuzzy pairwise comparison matrix A = [aij′] of order n,
where n is the number of attributes with aij ⊙ aji = 1; and aij = wwi .
j
The weights can be computed by geometric mean method as wi′ = si ⊙
1
(s1 ⊕ s2…⊕sn )−1, where si = (a i1 + a i2 + …+ain ) n .
One of the limitations of this method is that it does not consider whether the sum of the
weights is equal to unity. To overcome this situation, one could deploy optimization
techniques to obtain the weights. The major disadvantage of the geometric mean method
is the increasing length of the irrational fuzzy interval, due to the multiplication of fuzzy
numbers. A linear programming method to address the FAHP is presented.

10.2.2 Linear programming method

Given with a fuzzy positive reciprocal matrix A = [aij′] of order n, the pairwise
judgement can be described in the fuzzy interval [Ilij , Fuij ] as
wi
Ilij ⩽ ⩽ Fuij .
wj
For the given level of α − cut,
wi
Ilij (α ) ⩽ ⩽ Fuij (α ), ∀ α ∈ [0, 1]. (10.1)
wj

The inequality (10.1) can be written as

Ilij (α )wj + wi ⩽ 0, and

wi − Fuij (α )wj ⩽ 0.
The above two inequalities can be written in matrix notion as Sw ⩽ 0, where S is the
matrix of order 2m × n.
The linear membership function is used to measure the consistency test towards
the interval judgement
⎧ Sw
⎪1 − t if St w ⩽ Tt
μt (Stw) = ⎨ Tt ,
⎪ 0,
⎩ otherwise
where Tt is the parameter of tolerance.

10-4
 Modern Optimization Methods for Science, Engineering and Technology

The optimal weights are

μD (w) = max w{μ1[mini1 ⩽ t ⩽ m[ μ1(S1, w), μ2 (S2, w), … , μm (Sm, w)]
∣w1 + w2 + … + wn = 1]}.
The linear programming model corresponding to FAHP is given by
Z = maximize θ .
Subject to the constraints
Stw
θ⩽1−
Tt

w1 + w2 + …+wn = 1, w1 > 0, 1 ⩽ i ⩽ n & 1 ⩽ t ⩽ 2m .

Here, θ is the degree of satisfaction.

10.3 Performance measurement in the production approach

Performance evaluation is the most fundamental thing to benchmark. A systematic
comparison of the performance of the system against others is popularly known as
benchmarking. By comparing one’s own performance with other firms, one can help
to identify the position of the firm and opportunities to learn toward improvement.
In general, efficiency is defined as the ratio between the output and the input. This
section introduces necessary definitions for performance evaluation.

Definitions
1. For the given input i , the system produces an output o then production
possibility set is given by T = {(i , o)∣i can produce o}. That is, for the
amount of input i, the system/firm produces the amount of output o. Note
that the inputs may be resources such as raw material and human resources,
and outputs may be the services, number of transactions and so on.
2. Farrell input efficiency tells us the amount of reduction in input i that is able
to produce the same amount of output o. It can be defined as
Fie = min{k∣ki can produce o}, where k ⩾ 0.
3. Farrell out efficiency tells us the amount of increase in the output o, for the
same amount of input i , and is given by Foe = max{l∣i can produce lo}, where
l ⩾ 0.
4. Let (i1, o1) and (i2, o2 ) be the members of the production possibility set
(i2, o2 ) dominates (i1, o1) iff i2 ⩽ i1 and o2 ⩾ o1 and (i1, o1) ≠ (i2, o2 ).
5. A (i , o) is efficient in the production possibility set iff ∃ no member
(il , ol ) ∈ T dominates (i , o).

Let us assume that we are interested in studying the performance of an organization

among its r peers and denote the firm of interest as the tth firm. The tth firm produced
m outputs ot = (ot1, ot2 , … , otm ) using n inputs it = (it1, it2, … , itn ), the production
possibility set for the tth firm is given by Tt = {(ot , it ) ∈ nX m∣it can produce ot }.

10-5
 Modern Optimization Methods for Science, Engineering and Technology

Using Farrell efficieny scores, the ranking of organizations is determined. There are
other methods to find efficiency scores based on distance, called the Shepard distance.
In the literature there are methods based on the reduction of input and expansion
output simultaneously to obtain the efficiency scores. The efficiency obtained based
on Farrell is called technical efficiency (TE). If we replace the input vector with a cost
vector and the efficiency computation is carried out, the resulting computation is
called cost efficiency (CE). Cost efficiency is choosing the right candidate with
minimal cost. The other efficiency is allocative efficiency (AE) and it can be
computed as AE = CE/TE. In the same way one could define revenue efficiency
and profit efficiency.

10.3.1 Free disposability hull

Consider the production possibility set T = (i , o)∣i can produce o}, where with the
given inputs i , o outputs can be produced. It is also true that one could produce the
same output o with more than i inputs. This idea is called free disposability of input
and it can be translated as whenever (i , o) is in T and i′ ⩾ i then (i′, o) is in T .
In the same way, free disposability of output can be defined and it is given as:
whenever (i , o) is in T and o′ ⩽ o then (i , o′) is in T . The combination of free dis-
posability of input and free disposability of output along with the production set of
feasibility for the observed set is called the free disposability hull (FDH).
Convexity: The production possibility set T is said to be convex if the line segment
joining any two points in T lies entirely inside T. Let (i , o) and (i1, o1) be any two
members in T and λ (0 < λ < 1) be a scalar such that λ(i , o) + (1 − λ )(i1, o1)ϵT .
Returns to the scale: Let T = {(i , o)∣i can produce o} be the production possibility
set and be a scalar. If for any possible combination (i , o) can be scaled down or
scaled up for the given scalar λ, that is if (i , o) ∈ T then λ(i , o) ∈ T .
Based the value of λ, various returns to scale are classified as:
1. Constant returns to the scale (CRS), if λ ⩾ 0.
2. Decreasing returns to the scale (DRS), if 1 ⩾ λ ⩾ 0.
3. Increasing returns to the scale (IRS), if λ ⩾ 1.

Based on the need, one could construct the production possibility set in terms of
production function or cost function with CRS, IRS, DRS, FDH and VRS.

10.4 Data envelopment analysis

DEA is one of the non-parametric methods for studying the performance of any firm
and was introduced by Charnes, Cooper and Rhodes (CCR) in 1978, and in 1984 it
was refined by Banker, Charnes and Cooper (BCC). This is an extended application
of the linear programming model. The firms/entities whose performance is to be
studied are called decision making units (DMU). The major advantage of the DEA
model is that it can handle multiple inputs and multiple outputs. DEA models
assume that all considered DMUs are efficient and able to compute a single score of
efficiency for all DMUs, unlike in ratio analysis.

10-6
 Modern Optimization Methods for Science, Engineering and Technology

Based on the requirements, DEA can be input oriented or output oriented

with any of the returns to scale. The various returns to scale are as given in
figures 10.1–10.4.
This chapter only deals with CRS and VRS and is input oriented as well as output
oriented.

10.4.1 CCR model

Now we will present the most basic model, popularly known as CCR DEA. The
underlying principle of this model is to create a piecewise linear convex frontier in
such way that the frontier envelops the assumed input and output data, as per the set
objective function [2]. From the generated linear piecewise frontiers, efficiency scores
are computed by a set of linear programs. The decision making unit (DMU) is the
firm, which has the capability of transforming the inputs into outputs. Suppose we
have n firms, each producing n distinct outputs from m distinct inputs. We have to
compute the relative efficiency score of all considered DMUs. The scenario is
modeled as a fractional linear program with the objective of maximizing the ratio of
the sum of weighted outputs to the sum of weighted inputs subject to the constraints
of non-decreasing weights. As it is a maximization of the ratio, the relative efficiency
score can be at most 1. That is, to find the optimal weights so that the efficiency score
is maximized.

Figure 10.1. CRS.

10-7
Modern Optimization Methods for Science, Engineering and Technology

Figure 10.2. VRS.

Figure 10.3. FDH.

10-8
 Modern Optimization Methods for Science, Engineering and Technology

Figure 10.4. IRS.

The corresponding mathematical problem can be written as

u1y1,0 + u2y2,0 + …. +u nyn,0
Max u, vz = (10.2)
v1x1,0 + v2x2,0 + …+vmxm,0
subject to the constraints
u1y1, j + u2y2, j + …. +u nyn, j
⩽1
v1x1, j + v2x2, j + …+vmxm, j

ui ⩾ 0 (1 ⩽ i ⩽ n ) and vj (1 ⩽ j ⩽ m),
where yik denotes the output k produced by DMU i , xij denotes the input j used by
DMU i , and ui and vj are weights supplied to outputs and inputs, respectively. The
objective of this fractional linear programming problem (10.2) is to obtain the
weights ui , vj which could maximize the ratio DMU0; the DMU0 is being evaluated.
The model (10.2) can be converted to a multiplier model using CCR trans-
formation and is
Maxu,vz = u1y1, j + u2y2, j + …. +u nyn,j (10.3)

subject to the constraints

v1x1,j + v2x2,j + …+vmxm,j = 1

10-9
 Modern Optimization Methods for Science, Engineering and Technology

u1y1,j + u2y2,j + …. +u nyn,j ⩽v1x1,j + v2x2,j + …+vmxm,j

u i ⩾ 0(1 ⩽ i ⩽ n ) and vj (1 ⩽ j ⩽ m).

The relative efficiency scores of all DMUs can be obtained by solving model (10.3) n
(the number of DMUs) times. In optimization problems, we are accustomed to
talking about the dual of the model. Here, the dual of this model is called the
envelopment form. The model (10.2) is just the envelopment model and assumes
constant returns to scale (CRS), but imposing one more condition, that the sum of
lambdas is equal to 1, paves the way for the BCC model with a variable scale to
returns (VRS). All DMUs are operating in an optimal environment, and then one
should consider the CRS. In the majority of reports, this CRS assumption is violated
due to improper competition, and the variety of prevailing regulations and
restrictions. This led to a new measure of modified technical efficiency called
super-efficiency (SE). Hence the estimated efficiency scores based on VRS are at
least equal to the efficiency scores based on CRS.
Another important issue in the DEA approach needs to be discussed: the
orientation, input orientation or output orientation. While measuring the technical
efficiency, the efficiency measure improves the efficiency by the reduction in the
input quantities, whereas the output quantities are unaltered. The technical
efficiency for the cost frontier concides with this approach. The output orientation
is just the reversal of the input orientation that improves the efficiency scores by
proportional reduction of the output quantities, without altering the input quanti-
ties; this approach works well for the revenue frontier. The choice of orientation has
a very minor influence on the efficiency scores, even though there is a strong
correlation between both orientations [3]. The profitability falls under technical
efficiency and hence it can be measured with both orientations. Importantly, the
profit efficiency does not need to fall between 0 and 1 (sometimes, the profit could be
more than 100% and at times it could be negative as well). Whether the profit is
maximum or normalized is identified by the parameter return on expenditures. If the
return of expenditure is minimal or negative, it indicates the inefficiency of the firm.
Both input and output orientation models identify identical frontiers, more precisely
both orientation models fetch the same set of efficient DMUs.

Example 10.1. Multiple input and multiple output.

Consider a chain of restaurants; the promoter of the group wishes to measure the
performance of every branch. He/she decides the input and output to measure, given
in table 10.1.

Let us translate this problem into LPP.

For DMU1
MaxZ = 9u1 + 4u2 + 16u3,
subject to

10-10
 Modern Optimization Methods for Science, Engineering and Technology

Table 10.1. Input and output.

DMU Input 1 Input 2 Output 1 Output 2 Output 3

DMU1 5 14 9 4 16
DMU2 8 15 5 7 10
DMU3 7 12 4 9 13

5v1 + 14v2 = 1

9u1 + 4u2 + 16u3 − 5v1 − 14v2 ⩽ 0

5u1 + 7u2 + 10u3 − 8v1 − 15v2 ⩽ 0

4u1 + 7u2 + 13u3 − 7v1 − 12v2 ⩽ 0

u1, u2, u3, v1, v2 ⩾ 0.

For DMU2
Max, Z = 5u1 + 7u2 + 10u3
subject to
8u1 + 15u2 = 1

5u1 + 7u2 + 10u3 − 8v1 − 15v2 ⩽ 0

9u1 + 4u2 + 16u3 − 5v1 − 14v2 ⩽ 0

4u1 + 7u2 + 13u3 − 7v1 − 12v2 ⩽ 0

u1, u2, u3, v1, v2 ⩾ 0.

For DMU3,
Max Z = 4v1 + 7v2 + 13v3,
subject to
7u1 + 12u2 = 1

4u1 + 7u2 + 13u3 − 7v1 − 12v2 ⩽ 0

9u1 + 4u2 + 16u3 − 5v1 − 14v2 ⩽ 0

5u1 + 7u2 + 10u3 − 8v1 − 15v2 ⩽ 0

u1, u2, u3, v1, v2 ⩾ 0.

One could use any tool to solve the above linear programming and find the efficiency
scores of DMU1, DMU2 and DMU3.

10-11
 Modern Optimization Methods for Science, Engineering and Technology

CCR efficiency: A DMU is efficient in the context of CCR if z = 1, that is there

exists at least one optimal solution for non-negative u and v, otherwise the DMU is
inefficient.
Duality plays a vital role in linear programming problems. We have k DMUs
under consideration, which use m inputs to produce n outputs. The dual of the
model (10.3) can be expressed for a real number z * and a non-negative vector
λ = (λ1, λ2 , … , λk ) as
z∗ = mini z . (10.4)
Subject to
k
∑xijλj ⩽ zxi0. 1 ⩽ i ⩽ n
j=1

k
∑yrj λj ⩽ yr0 . 1⩽r⩽m
j=1

λj ⩾ 0, ∀ j .
The dual model (10.4) is feasible if z ∗ = 1, λ j∗ = 1 and λ 0 = 0, for j ≠ 0. This
process has to be repeated as many times as the number of DMUs.
Note that the optimal solutions occur on the boundary points and not all
boundary points are efficient. In such a scenario, we employ the modified model
(10.5) as
n s
Max∑ si− + ∑sr+, (10.5)
i=1 r=1

subject to
k
∑xijλj + si− = zxi0, 1⩽i⩽n
j=1

k
∑yrj λj − sr+ = yr0 , 1⩽r⩽m
j=1

λj ⩾ 0, 1 ⩽ j ⩽ k

si− ⩾ 0, 1 ⩽ i ⩽ m

sr+ ⩾ 0, 1 ⩽ i ⩽ n ,
where sr+ and si− are slack variables.
DEA efficiency: Any DMU is efficient if and only if z ∗ = 1 and all slack variables
are zero.

10-12
 Modern Optimization Methods for Science, Engineering and Technology

Now we will solve example 10.1 using input oriented and output oriented
methods and present the efficiency scores for three DMUs. The efficiency scores
of DMU1, DMU2 and DMU3 are, 1, 0.7733 and 1, respectively. Clearly DMU2 is
inefficient.
The models given in (10.2) to (10.5) are CCR—i.e. input oriented. In the same
manner, one can define the output oriented models.

10.4.2 BCC model

In the BCC model convexity is assumed. The BCC model was introduced by Banker,
Charnes and Cooper to evaluate the efficiency of the DMU0. The convexity
assumption is the additional feature of BCC compared to CCR.
The corresponding LPP is given as
z∗ = mini z , (10.6)
subject to
k
∑xijλj ⩽ zxi0, 1⩽i⩽n
j=1

k
∑yrj λj ⩽ yr0 , 1⩽r⩽m
j=1

k
∑ λj = 1
j=1

λj ⩾ 0, ∀ j .
BCC efficient: Any DMU is BCC efficient if z = 1, that is there exists one optimal
solution for non-negative u and v, otherwise the DMU is inefficient.
As we did in the CCR model, in the case of an inefficient boundary point we
deploy the following model:
n s
Max∑si− + ∑sr+, (10.7)
i=1 r=1

subject to
k
∑xijλj + si− = zxi0, 1⩽i⩽n
j=1

k
∑yrj λj − sr+ = yr0 , 1⩽r⩽m
j=1

10-13
 Modern Optimization Methods for Science, Engineering and Technology

k
∑ λj = 1
j=1

λj ⩾ 0, 1 ⩽ j ⩽ k

si− ⩾ 0, 1 ⩽ i ⩽ m

sr+ ⩾ 0, 1 ⩽ i ⩽ n ,
where sr+and si− are slack variables.
The BCC output oriented model is given as
w∗ = max w, (10.8)
subject to
k
∑xijλj ⩽ xi0, 1⩽i⩽n
j=1

k
∑yrj λj ⩾ wyr0, 1⩽r⩽m
j=1

k
∑ λj = 1
j=1

λj ⩾ 0, ∀ j .

Example 10.2. We are interested in studying the performance of 27 public sectors

banks in India for the year 2017. The inputs are cash-in-hand (CH), money at call
and short notice (MCS), advances (ADV) and investments in India (II) and outputs
are interest income (IIc) and business per employee (BPE). The input and output are
given in table 10.2.
Of these 27 banks, only 12 are fully efficient. The largest bank in India, SBI, is
also inefficient in the year 2016, whereas under VRS, only four banks are inefficient.
The computed efficiency scores based on CRS-input oriented, VRS-input oriented
and the scale efficiency are given in table 10.3.

10.4.3 Other models

A variety of DEA models is available. To name a few, we have the slack based
measure DEA, network DEA, stochastic DEA, fuzzy DEA and so on. In some
cases, one may not have either inputs or outputs, in such cases, use unity as inputs/
outputs and adopt the chosen model to obtain the efficiency scores [4–7].

10-14
 Modern Optimization Methods for Science, Engineering and Technology

Table 10.2. Inputs and outputs of DMUs.

Banks CH MCS II ADV IIC BPE

Allahabad Bank 6137.051 46 378.76 556 579.193 1560 695.2 188 849.46 148.5
Andhra Bank 8847.228 0 537 209.13 416 888.62 176 346.75 166.5
Bank of Baroda 37 564.451 353 920.31 1108 851.704 2289 771.6 440 612.77 168
Bank of India 26 310.955 217 519.64 1136 107.609 3538 093.2 417 964.7 179.6
Bank of 8189.764 0 362 308.724 299 714.37 130 529.86 180.2
Maharashtra
Bharatiya Mahila 56.178 1350 4788.136 5.841 1558.457 32.9
Bank Ltd
Canara Bank 20 118.673 66 393.138 1413 138.326 2879 455.5 440 221.35 144.462
Central Bank of 16 584.535 12 000 888 200.49 768 034.42 258 878.97 119.478
India
Corporation Bank 12 213.76 6000 632 800.927 805 599.41 194 112.37 187.9
Dena Bank 5975.075 0 352 262.182 390 862.62 106 457.34 146.2
Idbi Bank Limited 15 856.281 14 437.12 989 994.323 1983 065.7 280 431.02 251.8
Indian Bank 5371.302 62.942 505 146.666 297 525.23 162 437.84 153.1
Indian Overseas 18 253.872 57 568.519 757 189.041 758 587.35 235 172.95 124.1
Bank
Oriental Bank of 6947.381 1656.375 656 578.381 403 696.29 200 587.1 168.873
Commerce
Punjab and Sind 2507.299 8000 276 450.374 149 244.64 87 443.409 162
Bank
Punjab National 22 238.461 4690 1525 444.026 3357 959.2 474 243.5 135.9
Bank
Syndicate Bank 8945.607 112 963.74 678 464.959 813 299.55 231 977.8 146.1
Uco Bank 5838.033 45 147.936 818 859.854 593 404.56 185 609.74 126.8
Union Bank of 11 248.14 3619.76 877 728.066 3937 164 321 988.01 155.1
India
United Bank of 5588.093 21 000 447 233.834 115 839.1 99 366.709 123.7
India
Vijaya Bank 5545.97 301.612 418 424.895 172 765.8 120 835.79 145.7
State Bank of Bik 5025.53 0 247 823.746 502 362.9 95 924.728 121.5
and Jai
State Bank of 5191.821 7006.436 380 075.95 694 043.33 141 872.07 147.1
Hyderabad
State Bank of 150 809.19 124 570.23 4376 605.631 9719 560.1 1636 853.1 141.1
India
State Bank of 3882.42 0 201 239.554 393 592.22 71 277.658 118.3
Mysore
State Bank of 2693.868 0 309 170.205 302 991.46 104 570.99 128.8
Patiala
State Bank of 5879.217 11 496.507 360 618.29 407 926.73 96 088.79 117.6
Travancore

10-15
 Modern Optimization Methods for Science, Engineering and Technology

Table 10.3. Various efficiency scores of DMUs.

Banks VRS EFF CRS EFF SCALE EFF

Allahabad Bank 1 0.956 817 0.956 817

Andhra Bank 1 1 1
Bank of Baroda (BOB) 1 1 1
Bank of India (BOI) 1 0.945 135 0.945 135
Bank of Maharashtra 1 1 1
Bharatiya Mahila Bank Ltd 1 1 1
Canara Bank (CB) 1 0.831 885 0.831 885
Central Bank of India (CBI) 1 0.822 517 0.822 517
Corporation Bank (CRB) 0.932 689 0.835 85 0.896 172
Dena Bank (DB) 0.860 194 0.858 416 0.997 934
Idbi Bank Limited 1 0.749 013 0.749 013
Indian Bank (IB) 1 1 1
Indian Overseas Bank (IOB) 0.935 99 0.849 889 0.908 011
Oriental Bank of Commerce 1 0.945 505 0.945 505
Punjab and Sind Bank (PSB) 1 1 1
Punjab National Bank (PNB) 1 0.847 056 0.847 056
Syndicate Bank 1 0.957 495 0.957 495
Uco Bank 1 0.836 404 0.836 404
Union Bank of India (UNBI) 1 1 1
United Bank of India (UBI) 1 0.946 651 0.946 651
Vijaya Bank 1 1 1
State Bank of Bik and Jai 1 1 1
State Bank of Hyderabad (SBH) 1 1 1
State Bank of India (SBI) 1 0.963 954 0.963 954
State Bank of Mysore (SBM) 1 1 1
State Bank of Patiala (SBP) 1 1 1
State Bank of Travancore (SBT) 0.745 526 0.743 562 0.997 366

10.5 R as a tool for DEA

R is an open source tool for statistical computing. Because of its capabilities, it is a
popular and powerful tool. This section will introduce R for solving DEA problems
in an elementary manner. R can be downloaded from https://www.r-project.org/.
Store your input and output as a .csv file and type the following in R-prompt:
First install the Benchmarking package then type: require (Benchmarking):
1. To call your file to R
Dea_data <- read.csv(file.choose())
2. To verify the data and number of DMUs
head(Dea_data) // to show the first six rows of your data
dim(Dea_data) // to show the number of DMUs

10-16
 Modern Optimization Methods for Science, Engineering and Technology

3. Arrange your input as well as output in matrix form

xImat<-cbind(Dea_data$col_name1, Dea_data$col_name2, …, Dea_data
$col_namek)
yOmat<-cbind(Dea_data$col_name1, Dea_data$col_name2, …, Dea_data
$col_namel)
4. To compute the efficiency scores
CRS and input oriented
deaCrsIn<- dea(xImat, yOmat, RTS=‘crs’,ORIENTATION = ‘IN’)
eff(deaCrsIn)
pears(deaCrsIn)// list the pears
deaCrsIn$sx and deaCrsIn$sy // list the slacks in inputx, outputy
5. To get the output in csv file
write.csv(eff(deaCrsIn), file=‘eff_scores.csv’)

References
[1] Venkatesh K A, Pushkala N and Mahamayi J 2017 IFA - Off-Balance Sheet Items and
Performance Evaluation of Public and Private sector banks in India: A DEA approach
CFA Institute https://arx.cfa/post/IFA-Off-Balance-Sheet-Items-and-Performance-Evaluation-
of-Public-and-Private-sector-banks-in-India-A-DEA-approach-4968.html
[2] Cooper W, Seiford L M and Tone K 2007 Data Envelopment Analysis: A Comprehensive Text
with Models, Applications, References and DEA-Solver Software (Berlin: Springer)
[3] Coelli T and Perelman S 1999 A comparison of parametric and non-parametric distance
functions: with application to European railways Eur. J. Operation. Res. 117 326–39
[4] Behr A 2015 Production and Efficient Analysis with R (Berlin: Springer)
[5] Ray S C DEA, Theory and Techniques for Economics and Operations Research (Cambridge:
Cambridge University Press)
[6] AliEmrouznijad M T Performance Management with Fuzzy Data Envelopment Analysis
(Berlin: Springer)
[7] Bogetoft P and Otto L 2011 Benchmarking with DEA, SFA and R (Berlin: Springer)

10-17
 IOP Publishing

Modern Optimization Methods for Science, Engineering and

Technology
G R Sinha

Chapter 11
Evolutionary techniques in the design of
PID controllers
Santosh Desai, Rajendra Prasad and Shankru Guggari

Currently, the world is witnessing great changes in almost all areas, including the
field of system engineering, with steeply increasing system complexity resulting in
outsized systems. Generally these systems are best described by a large number of
differential or difference equations to form a mathematical model and ease the
process of analysis, simulation, design and control. However, the said task is not as
easy as it seems to be. It is very grueling, sometimes not feasible and also proves to
be a costly affair because of what may be called ‘the curse of dimensionality’. Hence
model order reduction has been borne out of the necessity to provide simplified
models, thereby addressing the effects of higher dimensional models and allowing us
to cope with the emergence of present day technology.
Today, various methods are being proposed by authors [1–9] and are available in
the literature. However, choosing the right method is one aspect which needs to be
considered. The area of application may also have an effect in addition to satisfying
the design requirements. Finally, this results in the need for approximations (low
order). Hence there is a need for a suitable technique to combat/improve the
drawbacks of the prevailing methods by developing/proposing new reduction
techniques satisfying the need of the hour. The proposed methods are not only
limited to continuous time systems, but can also be applied to systems presented in
discrete time. The same is justified by applying the proposed technique on several
numerical models. The outcomes are compared to other well-known methods when
subjected to specific test inputs. However, these simulations were performed to
obtain the behavior of the system in open loop configuration, but in most practical
cases some sort of controller always exists to control the system behavior. The design
of such a controller becomes a crucial task, in particular when the plant size is very
large. In that case, the size of the controller also increases, thereby resulting in

doi:10.1088/978-0-7503-2404-5ch11 11-1 ª IOP Publishing Ltd 2020

 Modern Optimization Methods for Science, Engineering and Technology

complicated and costly design. Apart from this, more computational time, and
difficulties in analysis, simulation and understanding of the system arise. Hence,
there is a need for a suitable lower order controller which can be derived from a
higher order controller, preserving the crucial dynamics of the original.
Furthermore, the derived reduced controller will be able to control the original
higher order system satisfactorily. This has resulted in the application of order
reduction methods to address controller reduction problems. In this chapter, the
design of a PID controller is achieved with the aid of evolutionary techniques such as
PSO and BBBC. Further, these techniques are extended to fine-tune the controller
parameters, namely λ (integral order) and μ (derivative order) in the fractional order
proportional integral derivative (FOPID) controller structure.

11.1 PID controller

As per the available literature, authors have proposed different methods [1–9]. But,
the choice of method also plays a major role. The suitability of the method chosen
depends on meeting the design requirements without sacrificing accuracy. This
results in approximations of lower order. The PID controller Gc(s) design connected
in series with an uncontrolled plant Gp(s) is of prime importance in this chapter. The
behavior of the overall closed loop system designed has to be stable enough to drive
under unity feedback conditions. Also, the responses of the proposed system and the
reference model should match each other in the time domain. Figure 11.1 indicates
two approaches [10–12] that are being considered here for the design of the
controller, namely plant/process reduction and controller reduction.
The process reduction method comprises reducing the original plant Gp(s) to
Rp(s). Then a suitable controller Rc(s) is designed and placed in series with Rp(s), as
shown in figure 11.2. Further, the closed loop response of RCL(s) is obtained
assuming unity feedback. The block diagram in figure 11.2 depicts the original,
reduced controller configurations using the so called reference model M(s). This
method is also referred to as the direct approach.
In the second (controller reduction) approach, the (higher order) controller Gc(s)
is designed on the basis of the higher order original uncontrolled plant Gp(s).
Then, the transfer function of the closed loop configuration GCL(s) is reduced

High Order Plant High Order

Gp(s) Controller
High order controller
Gc(s)

Plant Controller
Reduction Reduction

Reduced Order Reduced Order

PlantRp(s) Low Order Controller Controller Rc(s)
Design

Figure 11.1. Controller design approaches.

11-2
 Modern Optimization Methods for Science, Engineering and Technology

+
Gc(s) Gp(s)
-
R(s) E(s) u(s) C1(s)

R(s)
+ E(s) u(s) C2(s)
Rc(s) Rp(s)
-

R(s) M(s) C(s)

Figure 11.2. Original and reduced controller configuration with reference model.

appropriately with unity feedback and compared to RCL(s). The procedure carried
out in the controller reduction approach is also referred to as an indirect approach.
Here, the reduction process is carried out at the last stages and the possibility of
error propagation is ruled out. In the direct approach the issue of error creeps in
since the reduction is carried out during the initial stage of design.
In the present study, both direct and indirect design approaches have been
considered. Another optimization technique based on big bang big crunch (BBBC)
theory [13] is being used to tune the parameters of the PID controller for the said
purpose. This approach has yielded better results when compared to the HNA
assisted approach. In addition to this PSO, another evolutionary method, is also used
for optimizing the design parameters. The evolutionary technique under discussion
helps in searching for and selecting the best possible set of parameters which satisfies
the design requirements. When the design task is completed, the closed loop
responses are then compared with the reference model M(s). The reference model
M(s), also called the specification model or standard model, is the desired closed loop
system. To conclude, M(s) meets all the desired performance specifications and acts
as the basis for comparison.

11.1.1 Design procedure

Figure 11.1 indicates the two approaches of the controller design [10]. The first step
involves the design of the higher order controller as per the plant followed by
reducing to a lesser order. The closed loop behavior of the controller (higher order)
with the plant and the controller (lower order) with the plant are compared with that
of the reference model. Approximate model matching in the Pade sense is used to
obtain the controller parameters. The controllers (full order and reduced order) are
then compared in terms of their performance, as shown in figure 11.2.

11-3
 Modern Optimization Methods for Science, Engineering and Technology

11.1.1.1 Plant reduction and controller design (direct method)

The following steps are followed for design purposes based on the Pade sense.
Step 1. A reference model M(s) is obtained based on time/frequency/complex
domain specifications of the plant Gp(s). The controlled system response
(with unity feedback) in a closed loop is similar to M(s).
Considering the given original plant Gp(s), reference model M(s) represented by
ao + a1s + … + ams m
Gp(s ) = ; m<n (11.1)
bo + b1s + b2s 2 + … + s n

go + g1s + … + gus u
M (s ) = ; u < v. (11.2)
ho + h1s + h2s 2 + … + h vs v

Step 2. Obtain the equivalent open loop specification model from the reference
model M(s) using

˜ (s ) = M (s )
M . (11.3)
1 − M (s )
Step 3. Consider the structure of the controller given by

po + p1 s + … + pk s k
Gc(s ) = ; k < j. (11.4)
qo + q1s + q2s 2 + … + qj s j

Step 4. Compare the response of the system with the reference model to obtain the
unknown controller parameters by
Gc(s )Gp(s ) = M ˜ (s )
or
˜ (s ) ∞ (11.5)
M
Gc(s ) =
Gp(s )
= ∑eis i ,
i=0

where eis = coefficients of power series expansion (s = 0).

Equate (11.4) and (11.5) in the Pade sense to determine unknown control
parameters pi and qi:
p0 = q0e0
p1 = q0e1 + q1e0
p2 = q0e2 + q1e1 + q2e0
⋮⋮⋮
pi = q0ei + q1ei −1 + … + q1e0 (11.6)
0 = q0ei +1 + q1e1 + … + qi +1e0
⋮⋮⋮
0 = q0ei +j + q1ei +j −1 + … + qj ei .

The controller with the desired structure is obtained by solving (11.6).

11-4
 Modern Optimization Methods for Science, Engineering and Technology

Step 5. Obtain the closed loop transfer function using

Gc(s )Gp(s )
GCL(s ) = . (11.7)
1 + Gc(s )Gp(s )

Step 6. RP(s) is obtained from GP(s) using the reduction method. Steps 4 and 5 are
repeated and the transfer function (closed loop) for RCL(s) (reduced order
model) is given by
R c(s )Rp(s )
R CL(s ) = . (11.8)
1 + R c(s )Rp(s )

11.1.1.2 Indirect approach: controller design and reduction

Here, the original plant Gp(s) is considered and a high order controller Gc(s) is
designed. The closed loop transfer function GCL(s) is determined considering
feedback as unity. Finally, GCL(s) is reduced to obtain the reduced closed loop
transfer function RCL(s).

11.1.2 Method 1: PID controller design using PSO

PSO, as a global optimization algorithm, has gained popularity in both academia
and industry. Its simplicity and intuitiveness, ability to handle both discrete and
continuous variables, quite easy implementation, and exploration of the majority of
the problem space are its advantages. As a result, it is being used to solve
optimization problems related to engineering systems. Another main attraction of
PSO is that it works well for any dimensional problem. Hence, it is used to find the
optimum for functions. These functions include single, multi-objectives and may be
nonlinear or linear. Although, the problem exists in local minima, it is uncompli-
cated to code and understand its most basic form. Conceptually, PSO and the
genetic algorithm (GA) are similar and are easier to implement. To conclude, PSO is
comparable to and may be better than GA [14, 15].
The PSO algorithm was originally introduced by Kennedy and Eberhart in 1995
[16]. A swarm can be formally defined as a group of mobile agents that communicate
with each other directly or indirectly [17]. Since its inception, many problems in
various engineering fields have benefited because of its simple computation and
information sharing within the algorithm, as it derives its internal communications
from the social behavior of individuals (particles). A multi-dimensional optimization
problem [18] is represented by each individual (particle). However, the appropriate-
ness of the solution is based on the function which is related to an optimization
problem.
The process of PSO begins by initializing the population of particles, randomly
positioned across the search range with an initial random velocity having values not
greater than a certain percentage of the search space in each direction. Each particle
(candidate solution) is conveyed as a location within the problem search space. This
is established by reviving its velocity at consistent interims towards both the

11-5
 Modern Optimization Methods for Science, Engineering and Technology

positions pbest [19] (personal best) and gbest (global best) found by the entire
swarm. pbest and gbest are iteratively updated for each particle, until a better or
more dominant solution (in terms of fitness) is found. This process continues, until
the maximum iterations are reached or specified criteria are met. The equation
governing the movement of the particles of the swarm is
vi = vi + c1r1(pi − xi ) + c2r2(pg − xi ), (11.9)

where
vi is the velocity vector of the ith particle,
xi is the position vector of the ith particle,
pi is the n-dimensional personal best of the ith particle found from initialization,
pg is the n-dimensional global best of the swarm found from initialization,
c1 is the cognitive acceleration coefficient and
c2 is the social acceleration coefficient

and the position is updated using

xi = xi + vi . (11.10)
The classical version of the PSO algorithm defined by velocity update equations
(11.9) and (11.10) inherits a weakness that can be nullified by inertia weight w or
constriction coefficient χ. The method of introducing inertia weight was first
introduced by Shi and Eberhart [20]. The modified velocity equation is given by
vi = w*vi + c1r1(pi − xi ) + c2r2(pg − xi ). (11.11)

According to Eberhart and Shi [21], the optimal strategy is to initially set w to 0.9
and reduce it linearly to 0.4, allowing initial exploration followed by acceleration
toward an improved global optimum.
The problems in velocity update equations (11.9) and (11.10) were addressed by
Clerc [15] by introducing constriction coefficient χ so as to result in
vi = χ [vi + c1r1(pi − xi ) + c2r2(pg − xi )], (11.12)

where χ is computed as
2
χ= , (11.13)
∣2 − ϕ − ϕ(ϕ − 4) ∣
where ϕ = c1 + c2, ϕ > 4.
The velocity equation forms the main component of the PSO algorithm and is
unlikely to become stuck in the local minima [22]. The basic PSO algorithm is as
follows:
Step 1. [Start.] Initialize the position, velocity and the personal best of each particle
in the swarm at random.
Step 2. [Evaluate fitness value.] For each iteration, the particles will move progres-
sively into the solution space. The action involves updating the particle velocity
and its movement, and assessing the fitness function for the current position.

11-6
 Modern Optimization Methods for Science, Engineering and Technology

Step 3. [Compare fitness function.] The fitness function of the current position and
gbest are compared. Revisit the above stages for the whole particles.
Step 4. [Maximum iteration.] Continue until the iteration limit or until the termination
criteria are reached. Stop, and return the best solution gbest, otherwise update
w and go to the next iteration.
Step 5. [Loop.] Go to step 2, evaluate fitness value.

The flowchart as shown in figure 11.3 indicates the process flow of PSO. Table 11.1
gives the typical parameters used for PSO in the present study.

Start

Specify the parameters for PSO

Initialize swarm

Calculate velocities and new postion

Evaluate swarm and update each

particle

No Evaluate fitness Yes

function Stop
good ?

Figure 11.3. PSO optimization process.

Table 11.1. Typical parameters used for PSO.

Parameters Value

Swarm size 20
Maximum generations 100
c1, c2 2, 2
wstart, wend 0.9, 0.4

11-7
 Modern Optimization Methods for Science, Engineering and Technology

11.1.2.1 Illustrative examples

Numerical examples are presented to illustrate both indirect and direct methods of
PID controller design. The design parameters are optimized using PSO and
solved in detail for the first example. Later, the results are compared to other
methods.

11.1.2.1.1 Direct method

Example 11.1. A regulator problem is considered and its transfer function and
reference model are represented as [23]

s 5 + 8s 4 + 20s 3 + 16s 2 + 3s + 2
Gp(s ) =
2s + 36.6s 5 + 204.8s 4 + 419s 3 + 311.8s 2 + 67.2s + 4
6

0.023s + 0.0121
M (s ) = 2 .
s + 0.21s + 0.0121

Step 1. Consider M(s) and determine the equivalent open loop transfer function
using (11.3):
3 2
˜ (s ) = 0.023s + 0.01693s + 0.002819s + 0.0001464 .
M
s 4 + 0.397s 3 + 0.05137s 2 + 0.002263s

Step 2. Let the desired controller be according to (11.5) and this is given by
M˜ (s ) 1
Gc(s ) = = (0.064707 + 0.767859s + 0.801795s 2 − 4.681159s 3 + …).
Gp(s ) s

Step 3. Take the PID controller structure as

K2 K s + K2 + K3s 2
Gc(s ) = K1 + + K3s = 1 .
s s

Step 4. Gc(s) is compared to the power series expansion to obtain the parameters K1,
K2 and K3. This results in the PID controller
0.064707 + 0.767859s + 0.801795s 2
Gc(s ) = .
s

Step 5. Obtain GCL(s) using (11.7)

GCL(s )
0.8228s 7 + 7.349s 6 + 22.66s 5 + 29.02s 4 + 16.03s 3 + 4.981s 2 + 1.728s + 0.1294
= .
1.823s 7 + 25.65s 6 + 125.1s 5 + 238.5s 4 + 172s 3 + 38.58s 2 + 3.728s + 0.1294

11-8
 Modern Optimization Methods for Science, Engineering and Technology

Step 6. Apply the PSO technique to attain a reduced order model (second) from
GP(s). This is achieved by minimizing the ISE between the Gp(s) and Rp(s)
using the formula
0
I= ∫∞ [y(t ) − yr (t )]2 dt , (11.14)

where y(t) = unit step response of Gp(s) and yr(t) = unit step response of Rp(s).
Thus
0.02555s + 0.01036
R ppso(s ) = 2
.
s + 0.4756s + 0.01036

Step 7. Similarly using GA and HNA, the reduced system will be

0.01414s + 0.009369
R pga(s ) = 2
s + 0.1436s + 0.009369
0.0113s + 0.0736
R phna(s ) = 2 .
s + 0.1436s + 0.009369

Step 8. Now, following step 4 in 11.1.1.1, the controller structure will be

M˜ (s )
R cga(s ) =
R pga(s )
1
= (0.06471 + 0.714036s + 1.89926s 2 + …).
s

Step 9. Considering the PID controller structure as

K2
R c(s ) = K1 + + K3s
s
or
K1s + K2 + K3s 2
R c (s ) = .
s

Step 10. The values of K1, K2 and K3 are obtained with the comparison of power
series coefficients:
0.06471 + 0.714036s + 1.89926s 2
R cga(s ) = .
s

Step 11. Obtain the transfer function of RCLPSO(s), RCLGA(s) RCLHNA(s) as per
step 6 in 11.1.1.1

11-9
 Modern Optimization Methods for Science, Engineering and Technology

1.2

0.8
Amplitude

0.6 M (s)
GCL (s)
RCLPSO (s)
0.4

0.2

0
0 10 20 30 40 50 60 70 80 90 100
Time (sec)

Figure 11.4. Comparison of step responses.

Table 11.2. Qualitative comparison of original and reduced order plants in terms of time domain specifications
for example 11.1.

System Rise time tr (s) Peak overshoot % Mp Settling time ts (s)

M(s) 28.1 0.003 83 46.2

GCL(s) 22.3 0 45.1
RCLPSO(s) 28.8 0 43.6
RCLGA(s) 28 0 45.3
RCLHNA(s) 30 0.95 42.1

0.04853s 3 + 0.03861s 2 + 0.00933s + 0.0006703

R CLPSO(s ) =
1.049s 3 + 0.2142s 2 + 0.01969s + 0.0006703
0.01165s 3 + 0.07664s 2 + 0.005604s + 0.004762
R CLGA(s ) =
1.012s 3 + 0.417s 2 + 0.0792s + 0.004762
0.05177s 3 + 0.0445s 2 + 0.007102s + 0.0006152
R CLHNA(s ) = .
1.052s 3 + 0.1885s 2 + 0.0167s + 0.0006152
The unit step responses obtained are compared in figure 11.4. All the responses
are found to be comparable. Table 11.2 shows the qualitative comparison of the
original and reduced order plants in terms of time domain specifications.

11-10
 Modern Optimization Methods for Science, Engineering and Technology

11.1.2.1.2 Indirect method

Example 11.2. Consider a stable original practical system [24] along with a reference
model represented by

248.05s 4 + 1483.3s 3 + 91931s 2 + 468730s + 634950

Gp(s ) =
s 6 + 26.24s 5 + 1363.1s 4 + 26803s 3 + 326900s 2 + 859170s + 528050
4
M (s ) = 2 .
s + 4s + 4

Step 1. Considering M(s), the equivalent transfer function in an open loop

configuration is obtained using (11.3)

˜ (s ) = 4s 2 + 16s + 16
M .
s 4 + 8s 3 + 20s 2 + 16s
Step 2. According to (11.5), the controller transfer function is given as

M˜ (s ) = Gc(s )Gp(s )
or
M ˜ (s )
Gc(s ) =
Gp(s )
1
= (0.8316 + 0.5313s − 0.2841s 2 + 0.1159s 3 + …).
s
Step 3. Choose the controller structure as
k (1 + k1s )
Gc(s ) = .
s(1 + k2s )

Step 4. Comparing the power series expansion coefficients with controller structure
results in
K = 0.8316, K1 = 1.1735, K2 = 0.5347
and Gc(s) for the original plant will be
0.976s + 0.8316
Gc(s ) = .
0.5347s 2 + s

Step 5. Then GCL(s) will be

GCL(s ) =
242.5s 5 + 1656s 4 + 9.11 × 10 4s 3 + 5.347 × 105s 2
+ 1.011 × 106s + 5.28 × 105
.
0.5363s 8 + 15.07s 7 + 757.3s 6 + 1.598 × 10 4s 5 + 2.038 × 105s 4 + 8.788 × 105s 3
+ 1.677 × 106s 2 + 1.539 × 106s + 5.28 × 105

11-11
 Modern Optimization Methods for Science, Engineering and Technology

0.8
Amplitude

0.6 M(s)
GCL(s)
RCLPSO(s)
0.4

0.2

0
0 1 2 3 4 5 6 7 8 9
Time (sec)

Figure 11.5. Comparison of step responses.

Table 11.3. Qualitative comparison of original and reduced order plants in terms of transient response
parameters for example 11.2.

System Rise time tr (s) Peak overshoot % Mp Settling time ts (s)

M(s) 1.68 0 2.92

GCL(s) 1.76 0.239 2.81
RCLPSO(s) 2.57 0 4.57
RCLGA(s) 1.82 0.158 2.79
RCLHNA(s) 2.57 0.02 4.6

Step 6. The third order reduced model using PSO is obtained by reducing the
transfer function (closed loop) by minimizing (11.14):
0.8622s 2 + 2.05s + 0.9609
R CLPSO(s ) =
s 3 + 3.258s 2 + 3.172s + 0.9609
0.4844s 2 + 2.393s + 1.674
R CLGA(s ) = 3
s + 3.233s 2 + 4.045s + 1.674
0.9633s 2 + 3.88s + 1014
R CLHNA(s ) = .
1.176s 3 + 1.404s 2 + 1190s + 1013
The comparison of the step responses of M(s), GCL(s) and RCLPSO(s) is depicted
in figure 11.5. These responses are found to be comparable in terms of transient
response specifications with the responses from other methods in table 11.3. It
can be concluded that the result obtained by PSO is comparable.

11-12
 Modern Optimization Methods for Science, Engineering and Technology

11.1.3 Method 2: PID controller design using BBBC

Techniques based on the theory of evolution of the universe, such as BBBC, are
found to be useful for optimization [13]. In the present section, the BBBC based
method is utilized to optimize the values of the PID controller parameters. It is seen
that this method is better/comparable to the conventional methods in the following
solved examples.

11.1.3.1 Illustrative examples

An example is being considered to illustrate that BBBC outperforms PSO, GA and
HNA methods. The same is verified by two methods (indirect and direct) of PID
controller design by comparing in terms of time domain specifications.

11.1.3.1.1 Direct method

Example 11.3. Consider the regulator problem in example 11.1 in section 11.1.2.1.1

s 5 + 8s 4 + 20s 3 + 16s 2 + 3s + 2
GP(s ) =
2s + 36.6s 5 + 204.8s 4 + 419s 3 + 311.8s 2 + 67.2s + 4
6

0.023s + 0.0121
M (s ) = 2 .
s + 0.21s + 0.0121
Step 1. Following steps 1 to 5 of example 11.1 in section 11.1.2.1.1, we obtain
GCL(s )
0.8228s 7 + 7.349s 6 + 22.66s 5 + 29.02s 4 + 16.03s 3 + 4.981s 2 + 1.728s + 0.1294
= .
1.823s 7 + 25.65s 6 + 125.1s 5 + 238.5s 4 + 172s 3 + 38.58s 2 + 3.728s + 0.1294

Step 2. The second order model is obtained after the original system is reduced using
the proposed BBBC by minimizing (11.14) to obtain
0.0233s + 0.01176
RPBBBCOA(s ) = .
s 2 + 0.2035s + 0.01176

Step 3. Similarly, using PSO, GA and HNA, the reduced transfer function is
0.02555s + 0.01036
R ppso(s ) = 2
s + 0.4756s + 0.01036
0.01414s + 0.009369
R pga(s ) = 2
s + 0.1436s + 0.009369
0.0113s + 0.0736
R phna(s ) = 2 .
s + 0.1436s + 0.009369

11-13
 Modern Optimization Methods for Science, Engineering and Technology

Step 4. Now, the reduced controller will be

M˜ (s )
R CBBBC(s ) =
RPBBBC(s )
1
= (0.06191 + 0.76252s + 0.5764s 2 − …).
s

Step 5. Take the PID controller structure as

K2
R c(s ) = K1 + + K3s
s
or
K1s + K2 + K3s 2
R c (s ) = .
s

Step 6. K1, K2 and K3 of the controller are obtained using power series coefficients
0.06191 + 0.7625s + 0.5764s 2
R CBBBC(s ) = .
s

Step 7. Second order model is obtained by reducing Gp(s) and GC(s) (PSO, HNA,
BBBC, GA) in closed loop results as
0.01343s 3 + 0.02455s 2 + 0.01041s + 0.000728
R CLBBBC(s ) =
1.013s 3 + 0.2281s 2 + 0.02217s + 0.000728
0.04853s 3 + 0.03861s 2 + 0.00933s + 0.0006703
R CLPSO(s ) =
1.049s 3 + 0.2142s 2 + 0.01969s + 0.0006703
0.01165s 3 + 0.07664s 2 + 0.005604s + 0.004762
R CLGA(s ) =
1.012s 3 + 0.417s 2 + 0.0792s + 0.004762
0.05177s 3 + 0.0445s 2 + 0.007102s + 0.0006152
R CLHNA(s ) = .
1.052s 3 + 0.1885s 2 + 0.0167s + 0.0006152
The response of M(s), GCL(s) and RCLBBBC(s) for a given step input is plotted in
figure 11.6. Further, the outcomes are compared with other methods as listed in
table 11.4.

11.1.3.1.2 Indirect method

Example 11.4. A sixth order stable practical system (example 11.2 in section
11.1.2.1.2) is considered:
248.05s 4 + 1483.3s 3 + 91931s 2 + 468730s + 634950
Gp(s ) =
s 6 + 26.24s 5 + 1363.1s 4 + 26803s 3 + 326900s 2 + 859170s + 528050
4
M (s ) = 2 .
s + 4s + 4

11-14
 Modern Optimization Methods for Science, Engineering and Technology

0.8
Amplitude

M(s)
GCL(s)
0.6 RCLBBBC(s)

0.4

0.2

0
0 10 20 30 40 50 60 70 80 90 100
Time (sec)

Figure 11.6. Comparison of step responses.

Table 11.4. Qualitative comparison of original and reduced order plants in terms of transient response
parameters for example 11.3.

System Rise time tr (s) Peak overshoot % Mp Settling time ts (s)

M(s) 28.1 0.003 83 46.2

GCL(s) 22.3 0 45.1
RCLBBBC(s) 28.2 0 47.5
RCLPSO(s) 28.8 0 43.6
RCLGA(s) 28 0 45.3
RCLHNA(s) 30 0.95 42.1

Step 1. Following the same steps 1 to 5 of example 11.2 in section 11.1.2.1.2, we

obtain the closed loop transfer function GCL(s) as
GCL(s ) =
242.5s 5 + 1656s 4 + 9.11 × 10 4s 3 + 5.347 × 105s 2
+1.011 × 106s + 5.28 × 105
.
0.5363s 8 + 15.07s 7 + 757.3s 6 + 1.598 × 10 4s 5 + 2.038 × 105s 4
+ 8.788 × 105s 3
+1.677 × 106s 2 + 1.539 × 106s + 5.28 × 105

11-15
 Modern Optimization Methods for Science, Engineering and Technology

1.2

0.8
Amplitude

0.6
M(s)
GCL(s)
0.4 RCL3BBBC
RCL2BBBC

0.2

0
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5
Time (sec)

Figure 11.7. Comparison of step responses.

Step 2. GCL(s) is reduced to third and second order using BBBC by minimizing
(11.14) and is given by

−0.04996s 2 + 7.323s + 25.83

R CL3BBBC(s ) =
s 3 + 14.13s 2 + 33.04s + 25.83
0.2917s + 2.797
R CL2BBBC(s ) = 2 .
s + 3.083s + 2.797

Step 3. The reduced third order systems obtained by PSO, GA and HNA are
0.8622s 2 + 2.05s + 0.9609
R CL3PSO(s ) =
s 3 + 3.258s 2 + 3.172s + 0.9609
0.4844s 2 + 2.393s + 1.674
R CL3GA(s ) = 3
s + 3.233s 2 + 4.045s + 1.674
0.9633s 2 + 3.88s + 1014
R CL3HNA(s ) = .
1.176s 3 + 1.404s 2 + 1190s + 1013
The comparison of the step responses of M(s), GCL(s), RCL3BBBC(s) and RCL2BBBC(s)
is depicted in figure 11.7. It is seen that GCL(s), RCL3BBBC(s) and RCL2BBBC(s) almost
overlap each other and are competitive. The same can be concluded by examining
table 11.5, which compares the responses from other methods in terms of transient
domain specifications. It can be concluded that the result obtained by BBBC is
outstanding.

11-16
 Modern Optimization Methods for Science, Engineering and Technology

Table 11.5. Qualitative comparison of original and reduced order plants in terms of transient response
parameters.

System Rise time tr (s) Peak overshoot % Mp Settling time ts (s)

M(s) 1.68 0 2.92

GCL(s) 1.76 0.239 2.81
RCL3BBBC(s) 1.76 0.213 2.81
RCL2BBBC(s) 1.76 0.0583 2.84
RCLPSO(s) 2.57 0 4.57
RCLGA(s) 1.82 0.158 2.79
RCLHNA(s) 2.57 0.02 4.6

11.2 FOPID controller

PID controllers have become quite popular because of their various advantages [25].
However, it is observed that there is scope for better tuning of PID controllers [26].
In the recent past, different tuning methods have been proposed and this is still a
challenging task [27]. Here, a FOPID controller is being proposed, providing some
sort of solution to the problems under consideration. Both FOPID and PID
controllers are quite similar. The difference lies in the introduction of additional
flexibility in design. This is accomplished due to the presence of derivative λ and
integral μ in fractional order and results in better gain variations and phase
characteristics. To conclude, fractional order controllers have proved to be quite
powerful in robust control design.
The FOPID controller proposed [28] is applicable to linear systems [29] and this
can be termed as a drawback. However, due to various tuning methods [29, 30],
FOPID controllers are being used widely. In this regard, the use of evolutionary
techniques such as PSO [31] and GA [32, 33] in tuning has proved to be fruitful. As a
result, FOPID controllers based on evolutionary techniques have gained importance
when compared to conventional methods.
The PIλDμ, also known as the FOPID controller, was dealt with in both the time
[28] and frequency [30] domains. The FOPID controller generally takes the form
KDs λ+μ + KPs λ + KI
G FOPID(s ) = , (11.15)
sλ
where λ, μ are real numbers (positive), and Kp, KI, KD represent the proportional,
integral and differential gains, respectively.
The FOPID controller provides flexibility in the design of the PID control as
shown by the shaded portion in figure 11.8. The values of λ and μ define the four
control coordinates, i.e. P, PI, PD and PID of the classical PID. This results in
generalizing the conventional PID controller (integer order) and expands it from
point to plane.

11-17
 Modern Optimization Methods for Science, Engineering and Technology

Figure 11.8. Fractional order PID and the classical PID (integer and fractional order).

11.2.1 Statement of the problem

The original nth order LTI-SISO system is taken into account represented by the
transfer function
n
∑bj s j−1
j=1
Gp(s ) = m , (11.16)
∑ais i −1

i=1

where bj, ai are scalar constants, n is the order of the numerator and m is the order of
the denominator polynomial. The FOPID controller GFOPID(s) in the form of
(11.15) is the objective of this proposal. Later, the response of Gp(s) connected in
series with GFOPID(s) (with unity feedback) injected with unit step input is obtained.
The closed loop response meets the specified time domain parameters, i.e. MP, TS,
Tr, steady state error (SSE) and integral square error (ISE).

11.2.2 BBBC aided tuning of FOPID controller parameters

The introduction of evolutionary based tuning methods has changed the way
problems are being solved. It is noted that these methods are successful in almost
all areas and hence have become popular amongst researchers. BBBC [34–37] is one
such evolutionary technique being used for optimizing design parameters [13]. Here,
BBBC is used to optimize the values of the FOPID controller for a given
uncontrolled original plant GP(s). This is accomplished by tuning the areas of
search space through the interaction of individuals and has proven to be fruitful. The
response of the system obtained in closed loop (unity feedback) is stable, satisfying
time domain specifications.

11.2.3 Illustrative examples

Example 11.5. A third order original system [32] represented by Gp(s) is considered.
It is desired to design a suitable FOPID controller to provide better MP, TS, SSE and
ISE when injected with a unit step signal:

11-18
 Modern Optimization Methods for Science, Engineering and Technology

40
Gp(s ) = 3 2
.
2s + 10s + 82s + 10

Step 1. Assume the structure of the FOPID controller as

KDs λ+μ + KPs λ + KI
G FOPID(s ) = .
sλ

Step 2. Minimize the fitness function E using BBBC:

∞

E= ∫ [y(t ) − yr (t )]2 dt , (11.17)

0

where y(t) = the unit step response of the higher order model and yr(t) = unit step
response of the reduced order model at time t, and the peak overshoot is given by
−σd π
Mp = exp( ωd ) 100, (11.18)

where, σd ± jωd represents the dominant pole position.

Step 3. This results in
5.3896s 0.9+1.1 + 27.0406s 0.9 + 19.6417
G FOPIDBBBC(s ) = .
s 0.9

Step 4. The transfer function in the open loop configuration is given by

G OL(s ) = GP(s )G FOPIDBBBC(s )
⎛ 40 ⎞⎛ 5.3896s λ+μ + 27.0406s λ + 19.6417 ⎞
=⎜ 3 ⎟⎜ ⎟.
⎝ 2s + 10s 2 + 82s + 10 ⎠⎝ sλ ⎠

Step 5. The transfer function in the closed loop configuration is given by

Gp(s )G FOPIDBBBC(s )
GCLBBBC(s ) =
1 + Gp(s )G FOPIDBBBC(s )
215.584s 2 + 1081.624s 0.9 + 785.668
= .
2s 3.9 + 10s 2.9 + 215.584s 2 + 82s1.9 + 1091.624s 0.9 + 785.668

Step 6. The FOPID controller tuned with the aid of GA [32] is

0.36015s λ+μ + 6s λ + 12.24
G FOPIDGA(s ) = .
sλ

Step 7. The transfer function in the closed loop configuration (for λ = 0.9, μ = 1.1) is
given by

11-19
 Modern Optimization Methods for Science, Engineering and Technology

14.40s 2 + 240s 0.9 + 489.6

GCLGA(s ) = 3.9
.
2s + 10s + 14.406s 2 + 82s1.9 + 250s 0.9 + 489.6
2.9

Step 8. GCLBBBCRED(s) is obtained with the aid of BBBC

4.858s 2 + 3.25 × 10−5s + 0.07416
GCLBBBCRED(s ) = .
s 3 + 4.859s 2 + 0.0153s + 0.07416
The unit step responses of GP(s), GCLBBBC(s), GCLBBBCRED(s), GCLGA(s) [32] (λ = 0.9,
μ = 1.1) are compared, as shown in figure 11.9. It is noted that the method proposed
with the aid of BBBC yielded improved results compared to GA. Figure 11.10
indicates the unit step response of the integer PID controller. Further, the fractional
order PID controller (λ = 0.9, μ = 1.1) exhibits better results. The responses of
GCLBBBC(s) (various values of λ, μ) when subjected to a unit step signal are indicated
in figures 11.11–11.17. Table 11.6 shows a comparison of TS, MP, ISE and integral
absolute time error (IATE) obtained for various values of λ and μ.

Figure 11.18 illustrates the path traced by λ, μ for a population size of 75; the
number of iterations is 60 along with 0.75 as the reduction rate. The values obtained
in Matlab (2011b) during every iteration during the optimization process using
BBBC are considered for plotting the path.
Here, another approach to optimally tune the FOPID parameters using BBBC
[34–37] is being proposed. The results yielded are compared to GA [32]. The values

1.4

1.2

1
Amplitude

0.8

0.6 GP(s)
GCLBBBC(s)
0.4 GCLBBBCRED(s)
GCLGA(S)
0.2

0
0 5 10 15 20 25 30 35
Time (sec)

Figure 11.9. Comparison of unit step responses (FOPIDBBBC and FOPIDGA controllers for λ = 0.9 and
μ = 1.1).

11-20
 Modern Optimization Methods for Science, Engineering and Technology

0.8
Amp litud e

0.6

0.4

0.2

0
0 5 10 15 20 25 30 35 40 45 50
Time (sec)

Figure 11.10. Unit step response of FOPID controller (λ, μ = 1).

Figure 11.11. Unit step response of FOPID controller (μ < 1).

of Mp, Ts, ISE and ITAE are being taken into account for the qualitative analysis
and it is observed that the proposed technique has performed comparatively better.
Also, the FOPID controller yielded good results compared to the PID controller
(integer mode). Further, researchers can continue tuning the values of λ, μ up to two
or three decimal digits for better responses.

11-21
 Modern Optimization Methods for Science, Engineering and Technology

Figure 11.12. Unit step response for FOPID controller λ (λ < 1).

Figure 11.13. Unit step response for FOPID controller (λ < 1, μ < 1).

11.3 Conclusion
In this chapter, the task of designing both PID and FOPID controllers has been
accomplished successfully. In the case of the PID controller, both direct and indirect
methods of controller design are considered in this work. The reduction of the closed

11-22
 Modern Optimization Methods for Science, Engineering and Technology

Figure 11.14. Unit step response for FOPID controller (λ > 1).

Figure 11.15. Unit step response for FOPID controller (λ > 1, μ > 1).

loop system is then performed using PSO and BBBC by reducing the error between
the reference model and the reduced model. Later, unknown controller parameters
are found. The suitability of the proposed methods is justified by solving numerical
examples from the literature available. The step responses illustrate the value of the
proposed method.

11-23
 Modern Optimization Methods for Science, Engineering and Technology

Figure 11.16. Unit step response for FOPID controller (λ > 1, μ < 1).

Figure 11.17. Unit step response for FOPID controller (λ < 1, μ > 1).

In the case of the FOPID controller, the values of λ and μ are tuned by using
BBBC and then the results are compared to the existing technique. Qualitative
comparison in terms of settling time, peak overshoot, ISE and IATE are tabulated
apart from step responses to justify the proposed method.

11-24
 Modern Optimization Methods for Science, Engineering and Technology

Table 11.6. Effect of variations in λ and μ on time domain specifications.

Values obtained by BBBC (proposed) Values obtained by GA [32]

Peak Peak
Settling overshoot Settling overshoot
λ μ time Ts Mp ISE IATE time Ts Mp ISE IATE

1 1 5.4627 11.3171 1.9246 36.3443 23.65 47.15 2.350 53.70

1 0.3 4.3278 10.0349 2.4982 33.1984 41.742 46.0245 2.889 123.7
1 0.5 2.5002 11.6848 2.2635 23.6048 34.903 46.9016 2.659 89.85
1 0.7 2.2034 14.3641 2.0940 18.2592 25.293 47.2256 2.505 67.85
1 0.9 10.4159 10.6855 1.9705 35.1073 24.148 47.3067 2.405 56.97
0.3 1 4.2231 10.8044 2.1812 29.7341 27.228 37.3379 2.243 68.56
0.5 1 1.9948 11.3069 2.0830 23.9988 24.263 40.0857 2.253 56.97
0.7 1 3.2528 14.6011 2.3424 28.4420 24.031 42.9715 2.291 54.11
0.9 1 10.4159 10.6855 1.9705 35.1073 23.632 45.8391 2.335 53.58
0.5 0.5 1.9548 11.4000 1.9830 24.1121 35.404 38.7205 2.543 95.84
0.5 0.7 6.8865 12.2687 2.2163 38.9573 26.266 39.7392 2.397 72.19
0.5 0.9 3.7937 11.5378 2.1081 23.6011 24.703 40.1304 2.300 60.52
0.7 0.5 3.3538 14.8011 2.3436 28.4520 28.661 42.1173 2.589 92.45
1.1 0.5 11.1732 11.2030 2.2012 24.1625 23.4050 48.5314 2.381 54.04
1.5 0.5 3.8488 0.9096 2.0339 23.2997 27.4584 53.5798 2.475 56.14
7.5 0.5 3.9123 0.1060 5.5167 11.0425 37.8834 141.1292 5.643 121.3
1.1 1.1 2.4884 13.5817 2.2104 18.5385 22.925 45.8945 2.142 49.08
1.1 1.15 4.4859 11.8160 2.2241 502 059 22.096 36.124 1.438 38.42
1.1 1.2 4.5773 6.7630 1.9964 70.8083 13.319 44.7987 0.2839 13.54
1.1 1.21 4.1228 10.1151 2.2047 19.7048 20.628 93.2290 0.8556 30.17
2.5 1.1 3.9123 0.0 5.5167 11.0425 25.475 60.6300 2.441 57.84
2.5 1.15 3.9120 0.0 5.5180 11.0200 24.189 46.7059 1.658 44.59
2.5 1.2 3.9200 0.0 5.5200 11.0100 17.147 44.2505 0.4092 21.84
1.1 0.3 4.5326 12.9902 2.2375 36.9991 41.357 47.7478 2.918 123.5
1.1 0.9 3.9314 10.7146 2.3835 28.9681 23.860 48.6390 2.428 57.27
2.5 0.9 3.9123 0.0 5.5167 11.0425 30.375 64.9865 2.764 67.17
4.5 0.3 3.9122 0.0 5.5169 11.0430 40.946 102.1544 4.574 135.4
0.3 1.1 2.6811 10.5548 2.2649 22.5673 26.564 36.0403 2.021 63.14
0.9 1.1 4.9125 6.7860 1.6417 23.2709 23.366 43.4405 2.098 48.6
0.3 1.15 13.3011 11.1522 2.0382 47.0520 25.260 30.5292 1.344 48.47
0.9 1.15 10.1522 10.8518 2.0804 27.5545 22.705 34.3791 1.403 38.11
0.3 1.2 1.5322 13.6581 0.1837 2.98015 4.8701 47.5611 0.1939 5.328
0.9 1.2 1.4958 12.2436 0.2507 5.36410 13.042 45.1063 0.2591 11.7

11-25
 Modern Optimization Methods for Science, Engineering and Technology

60 *

50 *
* *
* * * * * *
40
* **
* * *
* * **
*
Iterations

30
*
*
* *
20
* * * * **
* *
* ** * *
10
* * *
*
0
* ** **
1.4
* *
1.2
1 * *
0.8 0.9
0.8
0.6 0.7
0.6
Mu 0.4 0.5
0.4
0.2 0.3
0.2
0 0.1 Lambda
0

Figure 11.18. Path traced by λ and μ.

References
[1] Aguirre L A 1991 New algorithm for closed-loop model matching Electron. Lett. 27 2260–2
[2] Bandyopadhyay B, Unbehauen H and Patre B M 1998 A new algorithm for compensator
design for higher-order system via reduced model Automatica 34 917–20
[3] Goddard P J and Glover K 1998 Controller approximation: approaches for preserving H∞
performance IEEE Trans. Autom. Control 43 858–71
[4] Mukherjee S 1997 Design of compensators using reduced order models Int. Conf. on
Computer Applications in Electrical Engineering, Recent Advances (CERA-97) (Roorkee:
Indian Institute of Technology Roorkee), pp 529–36
[5] Pal J and Sarkar P 2002 An algebraic method for controller design in delta domain Int. Conf.
on Computer Applications in Electrical Engineering, Recent Advances (CERA-01) (Roorkee:
Indian Institute of Technology Roorkee), pp 441–9
[6] Prasad R, Pal J and Pant A K 1990 Controller design using reduced order models 14th
National Systems Conf. (NSC-1990) (Aligarh: Aligarh Muslim University), pp 182–6
[7] Wang G, Sreeram V and Liu W Q 2001 Performance preserving controller reduction via
additive perturbation of the closed-loop transfer function IEEE Trans. Autom. Control 46
771–5
[8] Ramesh K, Nirmalkumar A and Gurusamy G 2009 Design of discrete controller via novel
model order reduction technique Int J Elect Power Eng. 3 163–8
[9] Hemmati R, Boroujeni S M S, Delafkar H and Boroujeni A S 2011 PID controller
adjustment using PSO for multi area load frequency control Aust. J. Basic Appl. Sci. 5
295–302
[10] Anderson B D O and Liu Y 1989 Controller reduction: concepts and approaches IEEE
Trans. Autom. Control 34 802–12

11-26
 Modern Optimization Methods for Science, Engineering and Technology

[11] Wilson D A and Mishra R N 1979 Design of low order estimators using reduced models Int.
J. Control 29 447–56
[12] Yousuff A and Skelton R E 1984 A note on balanced controller reduction IEEE Trans.
Autom. Control 29 254–7
[13] Erol O K and Eksin I 2006 New optimization method: Big Bang–Big Crunch Adv. Eng.
Software 37 106–11
[14] Hassan R, Cohanim B, de Weck O and V G 2005 A comparison of particle swarm
optimization and the genetic algorithm 46th AIAA/ASME/ASCE/AHS/ASC Structures,
Structural Dynamics and Materials Conf. (Austin, TX)
[15] Clerc M 1999 The swarm and the queen: towards a deterministic and adaptive particle
swarm optimization ICEC 1999 (Washington, DC) pp 1951–7
[16] Kennedy J and Eberhart R 1995 Particle swarm optimization Proc. IEEE Int. Conf. on
Neural Networks vol 4 pp 1942–8
[17] Engelbrecht A P 2007 Computational Intelligence: An Introduction 2nd edn. (Chichester: Wiley)
[18] Kennedy J, Eberhart R C and Shi Y 2001 Swarm Intelligence vol 10 (San Francisco, CA:
Morgan Kaufmann)
[19] Bahrepour M, Mahdipour E, Cheloi R and Yaghoobi M 2009 SUPER-SAPSO: a new SA-
based PSO algorithm Applic. Soft Comput. 58 423–30
[20] Shi Y and Eberhart R C 1998 A modified particle swarm optimizer IEEE Int. Conf. on
Evolutionary Computation (Piscataway, NJ: IEEE), pp 69–73
[21] Eberhart R C and Shi Y 2000 Comparing inertia weights and constriction factors in particle
swarm optimization Congress on Evolutionary Computation 2000 (San Diego, CA) pp 84–8
[22] Clerc M 2006 Particle Swarm Optimization (London: ISTE)
[23] Jamshidi M 1983 Large Scale Systems Modelling and Control Series vol 9 (Amsterdam:
North Holland)
[24] Prasad R 1989 Analysis and design of control systems using reduced order models PhD
Thesis University of Roorkee, India
[25] Valerio D and da Costa J S 2006 Tuning-rules for fractional PID controller Fract. Diff. Appl.
2 28–33
[26] Abu-Al-Nadi D, Othman M K A and Zaer S A-H 2011 Reduced order modeling of linear
MIMO systems using particle swarm optimization The Seventh Int. Conf. on Autonomic and
Autonomous Systems (ICAS 2011)
[27] Zhang J, Wang N and Wang S 2004 A developed method of tuning PID controllers with
fuzzy rules for integrating processes 2004 American Control Conf. (Boston, MA) pp 1109–14
[28] Podlubny I 1999 Fractional-order systems and PID controllers IEEE Trans. Autom. Control
44 208–14
[29] Caponetto R, Fortuna L and Porto D 2004 A new tuning strategy for a non-integer order
PID controller First IFAC Workshop on Fractional Differentiation and its Applications
(Bordeaux)
[30] Petras I 1999 The fractional order controllers: methods for their synthesis and application
J. Electr. Eng. 50 284–8
[31] Cao J-Y and Cao B-G 2006 Design of fractional order controller based on particle swarm
optimization Int. J. Control Autom. Syst. 4 775–81
[32] Padhee S, Gautam A, Singh Y and Kaur G 2011 A novel evolutionary tuning method for
fractional order PID controller Int. J. Soft Comput. Eng. (IJSCE) 1 1–9

11-27
 Modern Optimization Methods for Science, Engineering and Technology

[33] Saptarshi D, Indranil P, Shantanu D and Amitava G 2012 Improved model reduction and
tuning of fractional-order PIλDμ controllers for analytical rule extraction with genetic
programming ISA Trans. 51 237–61
[34] Desai S R and Prasad R 2012 PID controller design using BB-BCOA optimized reduced
order model IJCA Special Issue on Advanced Computing and Communication Technologies
for HPC Applications ACCTHPCA 5 32–7
[35] Desai S R and Prasad R 2013 A novel order diminution of LTI systems using big bang big
crunch optimization and routh approximation Appl. Math. Modell. 37 8016–28
[36] Desai S R and Prasad R 2014 Novel technique of optimizing FOPID controller parameters
using BBBC for higher order system IETE J Res. 60 211–7
[37] Desai S R and Prasad R 2013 A new approach to order reduction using stability equation
and big bang big crunch optimization Syst. Sci. Control Eng. 2013 20–7

11-28
 IOP Publishing

Modern Optimization Methods for Science, Engineering and

Technology
G R Sinha

Chapter 12
A variational approach to substantial efficiency
for linear multi-objective optimization problems
with implications for market problems
Glenn Harris and Sien Deng

Market problems can often be modeled using multi-objective optimization and

solved by finding properly efficient solutions. However, even when the problems are
solved with a properly efficient solution, sometimes the markets in these problems
can be manipulated by individuals to receive substantial gains at the expense of the
integrity of the market. For this reason it may be desirable to find solutions which
avoid this potential exploitation. Substantially efficient solutions, which are an
extension of properly efficient solutions, do just that. Using variational analysis and
by constructing novel examples, substantially efficient solutions and their topology
are explored in the context of linear multi-objective optimization problems. What is
discovered is that, in general, these problems do not necessarily have substantially
efficient solutions but there are ways to identify them. Furthermore, while substan-
tially efficient solutions are difficult to guarantee, it is possible to ensure their
existence under demanding constraints. The implications of this work are a more
principled choice for solutions to market problems which preserve market stability
and the recognition of potential market misuse when those choices are not available.

12.1 Introduction
Often in problems of engineering, industry, economics, and the like, decisions are
made concerning optimizing multiple objectives. However, these objectives are not
always measured on a common scale. For example, a genuine analysis of a
company’s profit, worker happiness, public image and time spent on a project
cannot all necessarily be put in terms of a single currency. These are multi-objective
optimization problems and are a very important area in optimization.

doi:10.1088/978-0-7503-2404-5ch12 12-1 ª IOP Publishing Ltd 2020

 Modern Optimization Methods for Science, Engineering and Technology

Deciding on a solution for most multi-objective optimization problems can be

difficult because there is often no choice that is clearly a minimum in every objective.
An efficient solution is one for which no other solution is better in performance in at
least one objective while also being better or equal in performance in all other
objectives. For the most part, only efficient solutions are considered for decision
makers in multi-objective optimization problems. So another difficulty arises for a
decision maker since there is often more than one efficient solution to a problem.
Some efficient solutions are anomalous in that the ratio, called the trade-off,
between an objective that stands to gain and any objective that stands to lose by
changing the solution to some other solution can be made arbitrarily large. An
aversion to that type of solution is understandable as the decision maker may feel
unsatisfied knowing there is a potentially worthwhile trade-off available. Proper
efficient solutions are efficient solutions that are not anomalous in that way as they
require that for any other solution there is a boundary on the trade-off between any
improving objective and at least one diminishing objective.
Only achieving improper efficiency would leave the ability to change to some
other solution where every component of the objective function that stands to do
better can do so at a negligible loss of all the components that stand to do worse.
One can imagine a situation where multiple parties have come together with a
problem to solve. This could be businesses looking to increase profits, nation states
looking to increase their currency value, or engineers looking to design a product.
Each party has their own objective function that assigns an output to a resource
allocation, such as regional control, cash input, labor distribution or concentration
of material. A properly efficient solution seems like it would be reasonable because if
a petition was put forth to change to some other solution, at least one party would
stand to lose proportionally to someone who would stand to gain. The party that
stood to lose may consider it unfair and attempt to put a stop to the change in
resource allocation. So, the allocation would not change from the properly efficient
solution. However, resource allocation is not always entirely held in the hands of
people who have a vested interest in a particular objective. This leads to the real
world issue that a properly efficient solution may not address, that is, market
manipulation. Some mechanisms of manipulation that could be employed in this
way are in currency exchanges or stock option markets.
Assuming multiple countries’ currency value could be modeled by assigning
objective functions to each country and certain states of affairs (such as policies,
resource holdings, etc) as inputs for those functions, all the members of the global
community may want to collectively maximize their country’s currency value. A
state of affairs resulting in a properly efficient solution may not be enough to avoid a
sort of currency exchange manipulation that has the potential for seizing the whole
market or inflating currency values. Imagine three countries who have in some way
achieved a state of affairs which is properly efficient. This means that any changes
that would increase the value of one state’s currency would decrease at least one
other proportionately. This would give some governing body that determines
exchange rates for currencies helpful information to adequately set market prices.
However, being only properly efficient still leaves open the possibility of a relatively

12-2
 Modern Optimization Methods for Science, Engineering and Technology

nontrivial increase in the value of one currency traded for a negligible decrease of
value of one of the other two currencies. An intrepid investor may spot this
possibility and, with the advent of instant trading and the lightning speed at which
information travels in the modern age, see an opportunity for exploitation. Evidence
of this speed goes as far back as 2004 when electronic currency trading took place at
two centralized electronic broker systems between parties in under a second [1]. In
order to see how this manipulation might be achieved, let f1, f2 and f3 be the
objective functions of countries 1, 2 and 3, respectively, and consider this a
maximization problem. Let x be the properly efficient solution at which the system
is currently. Let {yn }n∈ be some other collection of efficient solutions such that
f1 (yn ) − f1 (x )
f1 (x ) < f1 (yn ), f2 (x ) > f2 (yn ) and f3 (x ) > f3 (yn ), but f3 (x ) − f3 (yn )
> n . Even if as n → ∞
the gain in the first objective is minuscule, the loss of the other is that much smaller.
The investor could profit by exchanging a large sum of currency from country 3’s
to country 1’s currency, then implement actions that would ensure the state of affairs
yn occurs. They would then change their currency back to country 3’s. They could,
depending on the size of the sum of money they exchanged and the state yn they
aimed for, achieve practically any level of profit they wanted. Not only that, but
given that computer programs can be implemented to carry out instant transactions
repeatedly over a short period of time, it could be the case that this operation can be
carried out multiple times over the course of seconds or minutes pushing between the
states yn and x. In between those swaps, they would be purchasing and selling
currencies between the two countries resulting in profit for the investor. This could
occur repeatedly until no one is willing to buy or sell their currency anymore,
effectively skewing the global currency market driving prices up and seizing up the
system that would otherwise be needed for international investment. This may not
even be immediately detectable as yn may be a state very near x. Knowing when the
state of affairs does not have a bounded trade-off between all pairs of contravening
objectives helps to determine what sorts of regulation (fines/fees) to apply to block
the misuse. Either that or the currency brokers could know when a sort of extra
premium should be charged for a currency exchange between countries whose
marginal trade-off is not bounded in order to stop investors from taking advantage.
Another realistic example, which one could call an ‘option scheme’, is rooted in
the stock market. An option is a finance derivative contract which gives the holder
the liberty (but is not compelled, or forced) to buy or sell an asset at a given price by
a certain date. A savvy business entity could purchase and sell stock options
appropriately and then nudge the market one way or the other to make significant
gains in one stock to negligible losses in another. So a person could sell the option to
buy from the company that stands to lose negligibly beneath that loss. They could
also short a stock knowing full well what they will do later will not change the stock
values notably, then use those funds to buy the option to sell the stocks of the
company that stands to gain a noticeable amount (in relation to the negligible loss).
Admittedly, these options may be harder to come by, but it is still a possibility that a
large or unlimited number of options may be offered in certain circumstances.
Indeed, in 2014 the total dollar value of the Chicago Board Options Exchange, a

12-3
 Modern Optimization Methods for Science, Engineering and Technology

powerful options trading place, was roughly 570 billion dollars [2] and in 2016 it was
roughly 666 billion dollars [3]. In addition to options, the scheme may be applicable
to other derivitive products as well. There seems to be an almost unlimited potential
for derivatives in the form of credit default swaps, with an overall market value of
between 45 and 62 trillion dollars in 2007 [4], with a large portion of them
responsible for exacerbating the housing market crisis in 2008. Moreover when
these swaps are on securitized debt they can be exercised for partial losses and do not
require an entire collapse [4]. That is an important point for the scheme because the
exchanges are on marginal amounts, but high in volume.
As long as there is a customer, one can sell the option to buy a stock. Such an
option scheme would probably require more than diminutive gains, because they
may not trigger the stock option to sell the rising stock. They can still be small
though, such as $0.10 or $0.02 but perhaps not $0.00003. It may also be difficult for
the exploiter to manipulate stock values, but it is not impossible. Companies can
misinform or withhold information, as in the case of Enron projecting estimated
profits up to 20 years out, even though there were obvious concerns about those
profits [5]. Commodities holders and bankers can withhold or saturate stored
commodities or capital, resulting in altered valuation of stocks. These are only
some of the ways people could contribute to revaluing stocks, which could be
considered as components of the vectors in the feasible set of solutions. While
purchasing stocks can be done manually and valued individually by humans, many
stocks are evaluated using models. These values are recalculated over the day
instantly by computers. That in conjunction with actions on the options being
instantly transacted (within milliseconds) with the modern stock systems [6] makes it
possible to repeatedly take advantage of the option scheme. This could be
considered a misuse of stock options that could degrade trust in the market system
resulting in stagnation or depression. Not knowing this potential for certain
solutions means that stock options may not be valued correctly.
These examples show that there are many situations when properly efficient
solutions to multi-objective problems are not enough. In multi-objective market
problems it is more desirable to have a solution whose trade-off is bound for all pairs
of objectives that have opposing outcomes when switching to another solution. This
is where substantial efficiency enters. Substantial efficiency is a natural extension of
proper efficiency, requiring the trade-off bound apply to all the objectives that stand
to suffer.
The sorts of machinations given above make knowing when a solution to a
problem is or is not substantially efficient of interest to many groups. Investors,
business people and fortune seekers may want to know when a solution to a problem
is not substantially efficient in order to exploit that to potentially build a treasure
trove. Governing bodies and institutions may want to ensure that a solution to a
problem they are solving is a substantially efficient solution in order to avoid such
actions by said investors. Knowing when the current state is not substantially
efficient could result in a governing body taxing such option transactions, making
the option scheme impossible. Or in the currency case, a governing body could make
small changes to the value of a currency, or note that more premiums should be

12-4
 Modern Optimization Methods for Science, Engineering and Technology

charged for a currency change to currency brokers to avoid such manipulations.

Either way, knowing if a solution to a market problem is substantially efficient may
reduce the chance that the market can be exploited. Further, engineers can
appreciate situations where such marginal trade-offs being bound between all
objectives would be preferable in their optimization problems.
The aspiration of this chapter is to inquire about and then examine the prevalence
and nature of substantially efficient solutions for linear multi-objective optimization
problems. This is done through exploring many examples and expanding the current
framework of results that exist. It will be shown that such solutions do not
commonly exist. What this means for market problems is that regulators and
administrators must be ever-vigilant and on guard for market manipulators. For
engineers, it means that many solutions to their problems still provide the potential
for another solution where the trade-off between certain objectives is net positive,
giving them more to consider when decision making.
The new results are summarized as follows, starting with the simplest type of
linear multi-objective problem (LMOP) where the solutions come from a one-
dimensional space, all the efficient solutions must be substantially efficient (SE) as
well. This leads one to wonder if that is always the case, but an example is then given
showing that efficiency on its own does not imply substantial efficiency in LMOPs.
A further example is then given, showing that LMOPs do not even necessarily have
a substantially efficient solution. Since SE solutions are not always available, some
necessary conditions are sought out to determine what kinds of further restrictions
may need to be placed on problems to ensure an SE solution. A bounded feasible set
of solutions and a relation between the tangent cone at a point with the objectives
turn out to be necessary conditions. An example shows they are not sufficient
however, so another approach is attempted. A solution being properly efficient for
every restriction of the LMOP to pairs of objectives makes that solution SE. This
finally leads to a very specific type of problem where an SE solution can be
guaranteed. Seeing how difficult it is to ensure an SE solution exists for a problem,
the authors turn to the topology of the collection of SE solutions to help find them
when they do exist. This culminates in two theorems, one relating the collection of
bounds of SE solutions being bounded themselves to the closure of the set of SE
solutions. The other theorem shows that the location of SE solutions is on the
boundary of the feasible collection of solutions for a very wide range of problems.
This work is split into five sections. Section 2 will give the necessary background
for understanding what follows. Section 3 will introduce substantially efficient
solutions and attempt to explain the current status of research into these types of
solutions. Section 4 will provide new results and illuminating examples discovered
by the authors for the linear case of problems. Section 5 will conclude with some
final remarks and some suggestions for future research.

12.2 Background
The information that is contained in this section is a collection of terminology that is
required for understanding the later sections. Everything that follows will be using

12-5
 Modern Optimization Methods for Science, Engineering and Technology

real Euclidean vector spaces. The following notation is for describing vectors and
comparing them, along with some topological notation to be used.

Notation 12.1. For vectors x = (x1, … , xn ) ∈ n and y = (y1, … , yn ) ∈ n ,

x(i ) = xi , and
by x = 0 it is meant that x = (0, 0, … , 0),
by x = y it is meant that xi = yi for all i ,
by x < y it is meant that xi < yi for all i ,
by x ⩽ y it is meant that xi ⩽ yi for all i and xj < yj for some j .
For a space X , let X ° denote the interior of X .
The fundamental structure of a problem and solutions to those problems are as
follows.
Definition 12.1. Let f (x ) = (f1 (x ), … , fm (x )) be a vector-valued function and
X ⊆ n . Consider the problem
(P ) Minimize: f (x )
Subject to : x ∈ X .
The set X contains what are called the feasible solutions for (P). A feasible solution is
known as an efficient solution to (P) when there is no other y ∈ X such that
f (y ) ⩽ f (x ) and f (y ) ≠ f (x ). Solving (P) is understood to be the process of
identifying all the points in X that are efficient.
Introduced by Pareto, efficiency made its début in [7] and was modernized in [8]
by Koopmans. It was further formalized in the form given here by Kuhn and Tucker
in [9]. The particular type of problem considered in this work is the linear case of
problem (P). ◊
Definition 12.2. Consider the problem
(L ) Minimize: f (x ): =(m1 · x + b1, … , mm · x + bm )
Subject to: x ∈ X = {y ∈ n Ay ⩽ c},
{mi }im=1 ⊂ n, {bi }im=1 ⊂  ,
where
r , m , n ∈  , c ∈ r , and A ∈ r×n.
This type of problem is called a linear multi-objective optimization problem or
LMOP. Here mi will be referred to as the direction of the objective function fi . ◊
To rephrase, an LMOP is a multi-variate multi-objective function where each
objective component is a linear function and the feasible space is the intersection of
closed half-spaces. Since the intersection of half-spaces is convex and linear
functions are convex, this problem is automatically a convex problem. It may be
assumed that bi = 0 for all i as these constant terms do not affect the subsequent
results. Moving on, a special type of efficiency called proper effiency was introduced
by Geoffrion in [10].

12-6
 Modern Optimization Methods for Science, Engineering and Technology

Definition 12.3. A properly efficient solution of (P) is properly efficient solution x

∈X for which there exists a scalar M > 0 such that for each i , if y ∈ X and
fi (y ) < fi (x ) then there exists an index j for which f j (y ) > f j (x ) and
fi (x ) − fi (y )
⩽ M.
f j (y ) − f j (x )

The quotient on the left may be referred to as the trade-off quotient from here on out.
Efficient solutions that are not properly efficient are simply called improperly
efficient. ◊
Solutions that are properly efficient cannot do any ‘better’ in any single component
without doing proportionally ‘worse’ in at least one other component. Proper
efficiency has been shown to be of interest to many in the optimization community
and within industry as well. For instance, in [11] Seinfeld and McBride state that
proper efficiency help to avoid anomalous solutions in their problem solving
process that could arise when just any efficient solution would be chosen.
Another example, in [12] Belenson and Kapur explicitly state that an efficient
solution on its own may not be desirable and instead opt for properly efficient
solutions as they avoid some anomalies as well. In [13], Borwein characterized
proper efficient points in terms of tangent cones. In [14], Benson and Morin
developed some necessary and sufficient conditions for the presence of properly
efficient solutions. Then they related those conditions to the stability of specified
single-objective maximization problems. Also, in [15] Wendell and Lee involved
proper efficient points in their generalization of results from LMOPs to non-linear
cases relying on duality. Further, in [16], Benson furthered Borwein’s work by
defining properly efficient points in terms of projection cones. In [17], Liu extends
the concept of proper efficiency to ϵ -proper efficiency. In [18], Jiang and Deng
extend the concept of proper efficiency to α -proper efficiency. These are just a few of
the many considerations relevant to the concept of proper efficiency. The area has
garnered much more attention than the record provided.
A well known fact following [14] corollary 1 or [18] corollary 3.2 is the following.
Proposition 12.1 In an LMOP, every efficient solution is properly efficient.
Some helpful notation is provided to distinguish the different sets of solutions.
Notation 12.2. Let E(P ) ≔ {x ∈ X : x is efficient for the problem (P )} and let
P(P ) ≔ {x ∈ X : x is properly efficient for the problem (P )}.
The following definition of a tangent cone, which can be thought of as all the
vectors away from x into X , can be found in [19].

12-7
 Modern Optimization Methods for Science, Engineering and Technology

Definition 12.4. The tangent cone TX (x ) of a vector x in X is defined by w ∈ TX (x ) if

w = lim (xk − x )/ λk
k →∞

for some sequence {xk}k∈ ⊆ X with xk → x and some scalar sequence λk 0. ◊

This was a short overview of what is necessary for what follows. Good references
that can be used for further variational analysis and optimization are [20] and [19].
Further information concerning multi-objective optimization is located in [21] and [22].

12.3 A review of substantial efficiency

The concept of proper efficiency can be extended to include a boundary on more
than just one trade-off quotient. That extension leads to the concept introduced by
Kaliszewski in [23].
Definition 12.5. The solution x ∈ X is substantially efficient or SE if it is efficient
and there exists some real M > 0 such that for each fi and y ∈ X satisfying
fi (y ) < fi (x ), then when f j (x ) < f j (y ),
fi (x ) − fi (y )
⩽ M.
f j (y ) − f j (x )

Let S(P ) ≔ {x ∈ X : x be substantially efficient for the problem (P )}. ◊

Solutions that are substantially efficient have the property that that one cannot do
any ‘better’ in any single component without doing proportionally ‘worse’ in all
components that stand to suffer.
The definition was again stated in [24] under the name ‘strongly properly
efficient’. In this paper Khaledian, Khorram and Soleimani-Damaneh rediscovered
the concept of substantial efficiency, giving examples and studying the notion in
different senses. One was in the sense of Geoffrion, which is the sense presented in
this paper. The other was in the sense of Benson using cones.
Some more work was done in [25] where Pourkarimi and Karimi provided two
characterizations for substantial efficiency, one rooted in a scalar problem and the
other based on a stability concept. They also introduced the concept of quasi-
substantial efficiency and presented two similar characterizations. Quasi-substan-
tially efficient solutions can have varying orders, with each order determining the
type of bound on the trade-off quotients.
A fact that will be used, that is adapted from [23, p 135] is the following.
Proposition 12.2. For any problem (P), if the number of efficient solutions is finite
then all the efficient solutions are substantially efficient.
The following are trivial facts based on the definitions needed before moving
forward to the new results.

12-8
 Modern Optimization Methods for Science, Engineering and Technology

Proposition 12.3. If the problem (P) has only two objective functions (i.e.
f (x ) = {f1 (x ), f2 (x )}) then any properly efficient solution is actually substantially
efficient.
Remark. It is apparent that S(P ) ⊆ P(P ) ⊆ E(P ) for any problem (P).

12.4 New results and examples

Engineers and economists will want to know the conditions for which SE solutions
exist. The first proposition relates substantially efficient solutions with all efficient
solutions in the mono-variable case as it is a simple type of LMOP.
Proposition 12.4. If the LMOP (L) has X ⊆  , then every efficient solution is
actually substantially efficient.
Proof. Fix an efficient solution x*. Note that for each i = 1, … , m that
fi (x ) = mi x + bi . So for any fixed i if x ∈ X ⊆  is a feasible solution and
j ∈ {1, … , m} for which fi (x*) > fi (x ) and f j (x ) > f j (x*) note that mj ≠ 0 and
x ≠ x* or else f j (x ) > f j (x*) is impossible. Observe that

fi (x*) − fi (x ) mi x* − bi − mi x + bi
=
f j (x ) − f j (x*) mj x − bj − mj x* + bj
(12.1)
m (x * − x ) −mi
= i ⩽ .
mj (x − x )
* mj

Take M = maxi,j ∈{1,…,n} −mmi to obtain a bound that works for all viable combinations
j
of i and j . This is enough to say that any efficient solution is actually substantially
properly efficient. ■
This leads one to wonder if substantial efficiency is equivalent to proper efficiency
in general LMOPs. However, despite every efficient solution being properly efficient
in LMOPs, it is not always the case that every efficient solution is automatically a
substantially efficient solution. A counter-example is provided.
Example 12.1. Consider the LMOP
Minimize: f (x ) = f (x1, x2 , x3)
= (f1 (x ), f2 (x ), f3 (x ))
(12.2)
= (x1 + x2 − x3, x1 − x2 , −x1 + x2 + x3),
subject to: x = (x1, x2 , x3) ⩾ 0.

The solution x* = (0, 0, 0) is an efficient solution with f (x*) = 0. Indeed, consider

any other feasible solution x = (x1, x2, x3) ⩾ 0 with f (x ) ⩽ 0. If f1 (x ) ⩽ 0 then
x3 ⩾ x1 + x2 . And if f2 (x ) ⩽ 0 then x2 ⩾ x1. Using those, if f3 (x ) ⩽ 0, then
0 ⩾ f3 (x ) = −x1 + x2 + x3 ⩾ 2x2 ⩾ 2x1 ⩾ 0. That implies that x1, x2, x3 = 0. This is

12-9
 Modern Optimization Methods for Science, Engineering and Technology

enough to say x* = (0, 0, 0) is an efficient solution because there is no other solution

for which f (x ) ⩽ 0. Proposition 12.1 implies this solution is properly efficient.
Proper efficiency can also be demonstrated using a method from [10] which states
that x* is properly efficient if and only if there exists λ ∈  +m such that x* minimizes
m
∑ λk fk (x).
k=1

Indeed if λ = (1, 1, 1) then

3
∑ λk fk (x) = x1 + x2 (12.3)
k=1

which is minimized when x1, x2 = 0. The solution x* = (0, 0, 0) has that property
and so it is properly efficient.
Now it will be shown that x* is not substantially efficient. Let M > 1 be any
1
potential bound on the trade-off quotients. Consider the solution x = (1, 1 − M , 5),
which gives f (x ) = ( −3 − M1 , M1 , 5 − M1 ). Since f1 (x ) < f1 (x*) and f2 (x ) > f2 (x*), if
x* is SE then it should be the case that the trade-off quotient between f1 and f2 is
bounded by M . However that is not the case,
1
f1 (x*) − f1 (x ) 3+
= M = 3M + 1 > M . (12.4)
f2 (x ) − f2 (x*) 1
M
So given any M > 1, an x can be found so that the trade-off quotient is greater than
M . So x*, while efficient and properly efficient, cannot be SE. □

So it is not the case that every efficient solution is SE in LMOPs, but it would be
nice if a substantially efficient solution were present in any given problem. That
would give problem analysts a target to aim for when solving problems. However, it
is not always the case that a substantially efficient solution exists.

Example 12.2. Consider the feasible solution set

X = {(x , y , z ) ∈ 3: 0 ⩽ x , y , z ⩽ 1, y ⩾ x , z ⩾ y − x}
for the problem of minimizing f (x , y, z ) defined by the objective functions
f1 (x , y, z) = − x + y
f2 (x , y, z) = z + x − y
f3 (x , y, z) = z − x + y
f4 (x , y, z) = x
f5 (x , y, z) = − x

12-10
 Modern Optimization Methods for Science, Engineering and Technology

and a feasible solution (x1, y1, z1), where x1 < y1 and z1 > y1 − x1 > 0. Whether or not
x1 = y1, f2 and f3 can be made smaller without influencing any other objective by
taking z2 = y1 − x1. So an efficient solution must have its third component equal
to the second minus the first. Observe that f1 (x1, y1, z2 ) = −x1 + y1 > 0 and
f3 (x1, y1, z2 ) = z2 − x1 + y1 = 2y1 − 2x1 > 0 since y1 > x1. Both can be lessened by
taking y2 = x1. This will not affect f2 since f2 will stay at 0. Also f4 and f5 will remain
unchanged since x1 has not changed. So a new solution that will lower the first three
components has the form (x1, y2 , z2 ) = (x1, x1, 0). So if a solution is to be efficient it
must have the form (x , x , 0) and f (x , x , 0) = (0, 0, 0, x , −x ). From this it is seen
that every solution of the form (x , x , 0) is efficient as a change in x will result in a
fair exchange in the fourth and fifth component of f.
However, none of these efficient solutions turn out to be substantially efficient.
Two cases will be examined for the efficient solution (x , x , 0). First when 0 < x ⩽ 1
and second when x = 0.
Case 1. Assume that 0 < x ⩽ 1 and let M > 0. Take ϵ so that (x − ϵ, x − ϵ, ϵ 2 ) ∈ X
1
and ϵ < M . Then

f4 (x , x , 0) = x > x − ϵ = f4 (x − ϵ , x − ϵ , ϵ 2 )
and
f3 (x − ϵ , x − ϵ , ϵ 2 ) = ϵ 2 > 0 = f3 (x , x , 0)
so the trade-off quotient takes the form
f4 (x , x , 0) − f4 (x − ϵ , x − ϵ , ϵ 2 ) x − (x − ϵ ) ϵ 1 1
= = 2 = > 1 = M . (12.5)
f3 (x − ϵ , x − ϵ , ϵ 2 ) − f3 (x , x , 0) ϵ2 − 0 ϵ ϵ
M

This means that (x , x , 0) is not SE when x > 0.

1
Case 2. Assume that x = 0 and M > 1 is given. Then take ϵ < M
and note that
(ϵ, ϵ, ϵ 2 ) ∈ X . Then
f5 (0, 0, 0) = 0 > −ϵ = f5 (ϵ , ϵ , ϵ 2 )
and
f3 (ϵ , ϵ , ϵ 2 ) = ϵ 2 > 0 = f3 (0, 0, 0)
so the trade-off quotient takes the form
f5 (0, 0, 0) − f5 (ϵ , ϵ , ϵ 2 ) ϵ 1 1
= 2 = > 1
=M (12.6)
f3 (ϵ , ϵ , ϵ 2 ) − f3 (0, 0, 0) ϵ ϵ
M

so in this case (0, 0, 0) is also not SE. □

12-11
 Modern Optimization Methods for Science, Engineering and Technology

A simple criterion for the existence of SE solutions has eluded the authors. The
difficulty of not having SE solutions abundantly available leads to investigating
what may be responsible for their irregularity. That could reveal what types of
problems do have SE solutions. The following theorem helps one to understand why
SE solutions are not ubiquitous.
Theorem 12.1 If (L) is an LMOP such that S(L ) = ∅ then either X is unbound in at
least one component or for each x* ∈ E(L ) there is an mj perpendicular to a non-zero
vector in TX (x*).
Proof. Note that the definition of (L) ensures X is closed and convex. Let x* ∈ E(L )
but x* ∉ S(L ), then for all M > 0 there exists xM ∈ X and indices iM , jM ∈ {1, …, n},
miM · x* > miM · xM ,
m jM · x* < m jM · xM ,
(12.7)
miM · (x*−xM )
> M.
m jM · (xM − x*)

This defines a sequence {xk}k∈ . Moreover by the pigeon hole principle it is possible
to pass to a subsequence so that the objective indices i and j are fixed because there is
only a finite number of combinations of two elements from a collection of size n but
infinitely many xk . Let Mk be the corresponding bounds of that subsequence.
There are then two cases since
mi · (x*−xk )
lim > lim Mk = ∞ . (12.8)
k →∞ mj · (xk − x*) k →∞

Case 1. mi · (x*−xk ) → ∞ .
In this case mi · x*−mi · xk > Nk for some scalar sequence {Nk}k→∞. Since mi · x*
is fixed then mi · xk must go to −∞ to stay ahead of Nk as Nk → ∞. Since mi is fixed
this is only possible if at least one of the components of xk goes to ±∞.
Case 2. limk→∞mj · (xk − x*) = 0.
x − x*
In this case either xk → x ≠ x*, xk → x*, or ∥xk∥ → ∞ with mj ⊥ limk→∞ ∥ xk − x* ∥ .
k
If ∥xk∥ → ∞ then X is unbounded in some component. If xk → x then since X is
x (k − 1)x*
convex, every point between xk and x* will be in X . In particular kk − k
∈ X.
This forms a sequence in X with
xk (k − 1)x*
+ − x*
k k = xk − x* → x − x*. (12.9)
1
k
The definition of a tangent cone states that x − x* ∈ TX (x*). So mj is perpendicular
to x − x* since mj · (x − x*) = 0.

12-12
 Modern Optimization Methods for Science, Engineering and Technology

If xk → x* then ∥xk − x*∥ → 0 so

xk − x*
lim ∈ TX (x*) (12.10)
k →∞ ∥xk − x*∥

by definition. So
mi · (x*−xk )/ ∥xk − x*∥ mi · (x*−xk )
= > Mk . (12.11)
mj · (xk − x )/ ∥xk − x ∥
* * mj · (xk − x*)

Observe that
(x*−xk ) ∥x*−xk∥
mi · = cos(θi )∥mi ∥
∥xk − x*∥ ∥xk − x*∥ (12.12)
= cos(θi )∥mi ∥ ⩽ ∥mi ∥ ,
where θi is the angle between mi and (x*−xk ). So the numerator of the left most
quotient from (12.11) is bounded above. This means that in order for (12.11) to hold,
(xk − x*) xk − x*
lim mj · = mj · lim = 0. (12.13)
k →∞ ∥xk − x*∥ k →∞ ∥xk − x*∥

This implies that mj is perpendicular to a non-zero vector in TX (x*). ■

Comparing an efficient solution with any other solution in problems in a one-

dimensional feasible space ensures one solution will be greater than the other. So
there are only two directions of trade-off when the feasible space is one-dimensional,
greater than or less than. However, the concepts of greater than or less than are not
enough to determine a solution’s direction away from another when the feasible
space has two or more dimensions. A set of angles can express the direction of a
solution from another in a feasible space of dimension two or more. This means that
for a given efficient solution x , infinitely many solutions with a unique direction
from x may have to be considered when checking the collection of trade-offs for
boundedness. When an objective is constant in a direction α away from x , other
feasible solutions which have a direction arbitrarily close to α can be considered for
trade-off. These are problematic if one objective is increasing in the directions
arbitrarily close to α and another objective is decreasing but not constant in the
direction α , resulting in an unbounded trade-off quotient. Having infinite directions
away from any efficient point and more than two objectives creates a difficulty in
determining what types of problems have SE solutions. While possible, testing every
efficient solution’s tangent cone’s perpendicularity to the directions of the objective
functions (akin to looking for directions away from the efficient solution that leave
an objective constant) is not as useful as knowing a priori if a substantially efficient
solution exists in the problem. A corollary to theorem 12.1 for checking the
substantial efficiency of solutions is now provided.

12-13
 Modern Optimization Methods for Science, Engineering and Technology

Corollary 12.1. Let (L) be an LMOP as in definition 12.2, and let x* be an efficient
solution. If X is bounded and there is no mj perpendicular to any non-zero vector in
TX (x*) then x* ∈ X is substantially efficient.
Proof. In theorem 12.1 if it was assumed that x* is not SE then at then end of the
proof a contradiction would be arrived at, implying that x* must have been SE. ■

Remark. Extensions can be made on this corollary by assuming that each mj is

entirely positive, then X only needs to be bounded below.

The following is an example illustrating the usefulness of the previous corollary in

identifying if a solution is substantially efficient.

Example 12.3. Consider the following LMOP,

minimize: f (x ) = { f1 (x ), f2 (x ), f3 (x )}≔{m1 · x , m2 · x , m3 · x}
(12.14)
where m1 = (1, −1, −1), m2 = (2, −1, −2), m3 = ( −1, 2, 1)

subject to: x ∈ X ≔{(x1, x2 , x3): 0 ⩽ x1, x2 , x3 ⩽ 2, x1, +x2 + x3 ⩾ 4}. (12.15)

Observe that X is a closed bounded irregular tetrahedron. Let x* = (0, 2, 2), and
note that x* is the unique maximizer of f3 and unique minimizer of f1 in X , showing
that x* is efficient. Since x* is efficient, the corollary gives a sufficient condition to
check for substantial efficiency. The first condition is that X be bounded, which it is.
The second condition to check is if m1, m2 and m3 are not perpendicular to any
non-zero vector in TX (x*). Since
TX (x*) = {(x1, x2 , x3): x1 + x2 + x3 ⩾ 0, x2 , x3 ⩽ 0, x1 ⩾ 0}
and mi (1) has the opposite sign of both mi (2) and mi (3) for all i, it would be
impossible for there to be x ∈ TX (x*) with x · mi = 0 for any i unless x = (0, 0, 0).
That is to say that there is no non-zero x in the tangent cone perpendicular to any of
the mi . So the corollary then implies that x* = (0, 2, 2) is substantially efficient.
To demonstrate how the corollary compares to a direct method, the direct
method for showing substantial will also be given. Again, note that x* is the unique
minimizer of f1 and f2 and the unique maximizer of f3. So it only needs to be shown
that there exists M > 0 such that for every x ∈ X /{x*},
f3 (x*) − f3 (x ) 6 − ( −x1 + 2x2 + x3)
= ⩽M (12.16)
f1 (x ) − f1 (x*) x1 − x2 − x3 − ( −4)
and
f3 (x*) − f3 (x ) 6 − ( −x1 + 2x2 + x3)
= ⩽ M. (12.17)
f2 (x ) − f2 (x )
* 2x1 − x2 − 2x3 − ( −6)

12-14
 Modern Optimization Methods for Science, Engineering and Technology

To show this, first observe

x1 + x2 + x3 ⩾ 4
x1 + x2 ⩾ 4 − x3
⩾ 4 − 2 = 2 ⩾ x3
because x3 is between 0 and 2. From this,
x3 ⩽ x1 + x2
− x3 ⩽ x1 + x2 − 2x3
− 2x2 − x3 ⩽ x1 − x2 − 2x3
(12.18)
6 + x1 − 2x2 − x3 ⩽ 2x1 − x2 − 2x3 + 6
6 − ( −x1 + 2x2 + x3)
⩽ 1.
2x1 − x2 − 2x3 − ( −6)
Also,
x3 ⩽ 2
− x1 + x3 ⩽ 2
x1 − x3 ⩽ 2x1 − 2x3 + 2
6 + x1 − 2x2 − x3 ⩽ 2x1 − 2x2 − 2x3 + 8 (12.19)
6 + x1 − 2x2 − x3 ⩽ 2(x1 − x2 − x3 + 4)
6 − ( −x1 + 2x2 + x3)
⩽ 2.
x1 − x2 − x3 − ( −4)

So taking M = 2 it has been shown that x* is substantially efficient. □

The comparison of these two methods shows that there are circumstances when
looking at the tangent cone to check substantial efficiency is easier than a direct
method. The same problem gives a non-substantially efficient solution. This example
is given in hopes of explaining why no vector can be perpendicular to mj as
mentioned in corollary 12.1.
Example 12.4. Within the same problem framework as example 12.3 consider the
solution y* = (2, 0, 2) which is also efficient with f (2, 0, 2) = (0, 0, 0). This is seen
because
f1 (x ) = x1 − x2 − x3 < 0 ⇒ x1 < x2 + x3
(12.20)
⇒ f3 (x ) > 0

f2 (x ) = 2x1 − x2 − 2x3 < 0 ⇒ 2x1 < x2 + 2x3 ⩽ 4x2 + 2x3

(12.21)
⇒ f3 (x ) > 0

f3 (x ) = −x1 + 2x2 + x3 < 0 ⇒ x1 > 2x2 + x3 ⩾ x2 + x3

(12.22)
⇒ f1 (x ) > 0.

12-15
 Modern Optimization Methods for Science, Engineering and Technology

2+M
Given M > 0, consider the point yM = (2, 1+M
, 1) ∈ X and note that

2+M
f1 (yM ) = 2 − − 1 < 0 = f1 (y*)
1+M
and
2+M
f2 (yM ) = 4 − − 2 > 0 = f2 (y*).
1+M
So
2+M
−4 + +2
f2 (y*) − f2 (yM ) 1+M −2(1 + M ) + 2 + M
= = = M. (12.23)
f1 (yM ) − f1 (y )
* 2 + M (1 + M ) − 2 − M
2− −1
1+M
This is true for every M > 0 so y* cannot be substantially efficient.
This relates to the corollary. Note that limM →∞yM − y* = (0, 1, −1), and
(0, 1, −1) ∈ TX (y*) = {(x1, x2 , x3) ⊂ 3: x1 + x2 + x3 ⩾ 0, and
x1, x3 ⩽ 0 and x2 ⩾ 0}.
This is the type of vector perpendicular to the mj mentioned in the corollary. □

This looks like a promising characterization of all SE solutions, however, the

converse of corollary 12.1 is not true. If X is bounded, x* is efficient and there exists
mi perpendicular to some vector in the tangent cone of x*, then x* can still be SE.
Consider the following counter-example.
Example 12.5. Let X : ={(x1, x2, x3) ⊂ 3: 0 ⩽ x1, x2, x3 ⩽ 1}. The identity mapping
f (x ) = x has only one efficient solution at (0, 0, 0), which is substantially efficient
by proposition 12.1. However, the vector m1 = (1, 0, 0) from f1 (x ) = (1, 0, 0) · x is
perpendicular to (0, 1, 0) ∈ TX (0, 0, 0). □

Checking every efficient point’s tangent cones in the pursuit of SE solutions may
be too difficult. So seeing the limitations of that approach, the authors turn to
another way to characterize SE solutions. Since substantial efficiency gives bounda-
ries for the trade-offs of pairs of objectives, it is reasonable to look at all the
restrictions of the problem to two objectives. The next proposition is for a general
problem connecting substantial efficiency to proper efficiency when a problem is
restricted.
Proposition 12.5. Consider a problem (P) (X not necessarily closed). If x* ∈ X is
such that the restriction of (P) to any two objective functions has x* as a properly
efficient solution then x* is SE for (P).
Proof. If x* is properly efficient when (P) is restricted to any two objective functions
then x* must be efficient to (P) on its own. Indeed if there exists an x ∈ X with

12-16
 Modern Optimization Methods for Science, Engineering and Technology

fi (x ) < fi (x*) then for any other index j, f j (x ) > f j (x*) otherwise x* would not be
efficient with regards to the restriction of (P) to the ith and jth components.
Let Ma,b be the bound of the trade-off quotient from the properly efficient
criterion when (P) is restricted to the ath and bth objective components. Let
M = max a,b∈{1, 2, …, m};a≠b(Ma,b ). Then for each objective component fi and x ∈ X
satisfying fi (x ) < fi (x*), f j (x*) < f j (x ) implies
fi (x*) − fi (x )
⩽ M.
f j (x ) − f j (x * )

So x* is substantially efficient. ■
The converse does not hold in general, however, with a slight alteration proposition
12.5 can be made necessary and sufficient. First a counter-example is given.
Example 12.6. Consider the minimization problem (P) where f (x ) = ( −x , −x + 1, x )
and X = [0, 1]. Note that every feasible solution is efficient. Also note that x* = 0 is a
substantially efficient solution. Indeed, if x is such that fi (x*) > fi (x ) and f j (x*) < f j (x )
then either i = 1 or 2 and j = 3. Also
f1 (x*) − f1 (x ) f (x*) − f2 (x ) 0 − x
= 2 = = 1 ⩽ 2, (12.24)
f3 (x ) − f3 (x*) f3 (x ) − f3 (x*) x−0

so if M = 2 the criterion for substantial efficiency is satisfied. However, when (P) is

restricted to the output functions f1 and f2 , x* = 0 turns out not to be efficient and
thus cannot be properly efficient. □
This shows what is keeping the converse of the previous proposition from being
true. When the feasible solution is not efficient in all the restrictions, substantial
efficiency is not equivalent to every restriction being properly efficient. So a slight
change is made to arrive at a bijective statement.
Theorem 12.2. Consider the general problem (P). If x* ∈ X is a solution such that the
restriction of (P) to any two objective functions is efficient, then x* is a properly
efficient solution for the restriction of (P) to any two objective functions if and only if
x* is a substantially efficient solution of (P).
Proof. (⇒) Proposition 12.5.
(⇐) Assume that x* is SE with the trade-off quotient bound M. Take
a, b ∈ {11…, n} and let (Pa,b) be problem (P) restricted to the objective functions
fa and fb. By assumption, x* is efficient for (Pa,b). Without loss of generality for the
choice of a or b, assume x is such that fa (x*) > fa (x ) and fb (x*) < fb (x ). Then
fa (x*) − fa (x )
because x* is substantially efficient, fb (x ) − fb (x*)
⩽ M . So the criterion for proper
efficiency of (P) restricted to fa and fb is also trivially satisfied. ■
Looking at the intersection of properly efficient solutions for all the pairwise
problem restrictions allows one a method for finding an SE solution, but it is still a lot

12-17
 Modern Optimization Methods for Science, Engineering and Technology

to ask. In LMOPs, only the intersection of efficient solutions is necessary as efficiency

implies proper efficiency. An empty intersection does not necessarily mean that no SE
solution exists either. However, this still leads to a corollary of the theorem 12.2 that
finally elicits a sufficient condition for the guaranteed presence of an SE solution in a
special case of (L). This is proven directly and using theorem 12.2.
Corollary 12.2. Let X = {(x1, …, xn ): a1 ⩽ x1 ⩽ b1, …, an ⩽ xn ⩽ bn} (a closed
bounded box) for a problem (L). If the results of all the pairwise sums of mi have
non-zero components with the same signs (i.e. for any possible combination of i, j, k, l,
p, that Sgn{(mi + mj )(p )} = Sgn{(mk + ml )(p )} ≠ 0) then an SE solution exists.
Proof. (Without using theorem 12.2.) Let y be the vector in X that minimizes
(mi + mj ) · x for all i , j and x ∈ X . So for all k ∈ {1, …, n} it will be that y(k ) = ak
when Sgn{(mi + mj )(k )} = 1 and y(k ) = bk when Sgn{(mi + mj )(k )} = −1. The
solution y is efficient, for if it were not there would exist z ∈ X and i for which
fi (z ) < fi (y ), and for all other j it would be that. But then summing these inequalities
would give (mi + mj ) · z < (mi + mj ) · y which contradicts how y is defined.
Now, it will be shown that y is substantially efficient. Assuming it is not, let i , j
and x ∈ X be such that fi (y ) > fi (x ) and f j (y ) < f j (x ) and

fi (y ) − fi (x )
> 1. (12.25)
f j (x ) − f j (y )

But then
mi · y − mi · x
>1
mj · x − mj · y
(12.26)
mi · y − mi · x > mj · x − mj · y
mi · y + mj · y > mi · x + mj · x ,
but then this again contradicts how y is defined. So y must be substantially efficient.
(Using theorem 12.2.) whenever (L) is restricted to two objective functions, y will
be an efficient solution and thus properly efficient as well. Using theorem 12.2 it
must be the case that y is an SE solution. ■
This is a very particular set of problems for which SE solutions can be guaranteed
to exist. The rarity of the assurance of existence leads one to wonder what can be
said of the set of SE solutions when they do exist. The next theorem ties the topology
of the collection of SE solutions to the trade-off quotient’s bounds.
Theorem 12.3. Consider the problem (P). If the collection of trade-off quotient
bounds {Mx}x∈S(P ) is bounded (i.e. there is a uniform trade-off quotient bound M > 0
that can be applied for any x ∈ S(P )) then S(P ) is closed.
Proof. Assume S(P ) is not closed, thus there is a limit point of S(P ) that S(P ) does not
contain. Let {xn}n∈ ⊆ S(P ) be a sequence that converges to x ∉ S(P ). Since X is
closed, x ∈ X /S(P ). Let M ∈  with M > 2 be a common bound for the trade-off
quotients of every element in {xn}n∈ .

12-18
 Modern Optimization Methods for Science, Engineering and Technology

Let {zn}n∈ ⊆ X be a sequence for which

fi n (x ) − fi n (zn) > 0

and f jn (zn) − f jn (x ) > 0

fi n (x ) − fi n (zn)
but > n.
f jn (zn) − f jn (x )

Since there are infinitely many n ∈  but only finitely many combinations of in and
jn then by the pigeon hole principle, a subsequence can be passed to in order to keep
the in and jn fixed. Do so to obtain a fixed i and j for which
fi (x ) − fi (zn) > 0

and f j (zn) − f j (x ) > 0,

fi (x ) − fi (zn)
but > n.
f j (zn) − f j (x )
Then pass to another subsequence so that fi (x ) is always greater or lesser than
fi (xn ) and so that f j (x ) is also always greater or lesser than f j (xn ). Now there are
some cases to consider.
Case 1. fi (xn ) > fi (x ) and f j (xn ) > f j (x ).
Choose N large enough that f j (zM ) > f j (xN ). Then

fi (x ) − fi (zM ) f (xN ) − fi (zM )

M< < i ⩽ M, (12.27)
f j (zM ) − f j (x ) f j (zM ) − f j (xN )

which is a contradiction.
Case 2. fi (xn ) > fi (x ) and f j (xn ) < f j (x ).
f j ( z 2M ) − f j ( x )
Choose N large enough so f j (z2M ) − f j (xN )
< 2, f j (z2M ) − f j (xN ) > 0, fi (xN )−
f j (z2M ) − f j (x )
fi (z2M ) > 0 and f j (x ) < f j (xN ) + ϵ , where ϵ < 2M
. Then
fi (xN ) − fi (z2M ) > fi (x ) − fi (z2M )
> 2M (f j (z2M ) − f j (x ))
> 2M (f j (z2M ) − f j (xN ) − ϵ )
fi (xN ) − fi (z2M ) 2Mϵ
⇒M ⩾ > 2M − (12.28)
f j (z2M ) − f j (xN ) f j (z2M ) − f j (xN )
f j (z2M ) − f j (x )
2M
> 2M − 2M > 2M − 2.
f j (z2M ) − f j (xN )

12-19
 Modern Optimization Methods for Science, Engineering and Technology

But that is a contradiction when M ⩾ 2, which it is.

Case 3. fi (xn ) < fi (x ) and f j (xn ) > f j (x ).
f j ( z 2M ) − f j ( x )
Take N large enough that f j (z2M ) − f j (xN )
< 2, f j (z2M ) − f j (xN ) > 0, fi (xN )−
fi (z2M ) > 0 and fi (x ) < fi (xN ) + ϵ , where ϵ < f j (z2M ) − f j (x ). Then

fi (xN ) + ϵ − fi (z2M ) > fi (x ) − fi (z2M )

> 2M (f j (z2M ) − f j (x ))
> 2M (f j (z2M ) − f j (xN ))
fi (xN ) − fi (z2M ) ϵ (12.29)
⇒M ⩾ > 2M −
f j (z2M ) − f j (xN ) f j (z2M ) − f j (xN )
f j (z2M ) − f j (x )
> 2M − > 2M − 2,
f j (z2M ) − f j (xN )

which is again a contradiction.

Case 4. fi (xn ) < fi (x ) and f j (xn ) < f j (x ).
f j ( z 2M ) − f j ( x )
Take N large enough so f j (z2M ) − f j (xN )
< 2, f j (z2M ) − f j (xN ) > 0, fi (xN )−
f j (z2M ) − f j (x )
fi (z2M ) > 0, and fi (x ) < fi (xN ) + ϵ and f j (x ) < f j (xN ) + ϵ , where ϵ < 2M + 1
.
Then
fi (xN ) + ϵ − fi (z2M ) > fi (x ) − fi (z2M )
> 2M (f j (z2M ) − f j (x ))
> 2M (f j (z2M ) − f j (xN ) − ϵ )
fi (xN ) − fi (z2M ) (2M + 1)ϵ
⇒M ⩾ > 2M − (12.30)
f j (z2M ) − f j (xN ) f j (z2M ) − f j (xN )
f j (z2M ) − f j (x )
(2M + 1)
> 2M − 2M + 1 > 2M − 2,
f j (z2M ) − f j (xN )

which is again a contradiction.

Since these are all the possible cases and each ends in a contradiction then the
original assumption that S(P ) is not closed is not true. ■
This also shows that when the set of SE solutions is not closed the collection of
trade-off bounds is unbounded. The next thing to be considered is when the set of SE
solutions is a collection of isolated points. There is a common situation when this
can occur. First a lemma involving the existence of a particular vector is needed.

12-20
 Modern Optimization Methods for Science, Engineering and Technology

Lemma 12.1. Let a, b ∈ n with a∥b and 1 > ϵ > 0 be given. Then there is a δ > 0
such that there exists c ∈ ∂Nϵ(0) with a · ( −c ) > δ a ϵ while b · (c ) = m1 ∥b∥ϵ for a
choice of m that is large.
Proof. For this proof let ∠xy denote the positive angle between x and y; so
∠xy = ∠yx . Let Q be the plane containing both a and b. If a ⊥ b then take
c ∈ ∂Nϵ(0) ∩ Q so ∠a( −c ) < 45° and ∠bc < 90° as close to 90° as necessary to
1 2
obtain cos(∠bc ) = m
for some choice of m that is large. Then δ = 2
will suffice
because
2 (12.31)
a · ( −c ) = cos(∠a( −c )) a c > a ϵ = δ a ϵ,
2
and
1
b · c = cos(∠bc ) b c = b ϵ.
m
On the other hand, if a ⊥ b take z to be the unit vector perpendicular to b in Q with
a · z > 0. So z and a are on the same ‘side’ of b, so to speak. Let θ = min{∠ab , ∠za}.
Then take c to be the vector in ∂Nϵ(0) ∩ Q such that ∠c( −z ) < θ /2 and
1 1 1
b · c = m b c = m b ϵ with m being any value for which m < cos(θ /2). So c is
the vector in Q perpendicular to b, with magnitude ϵ , and whose angle with −z
measures between 0 and θ/2. So −c is within an angle of θ/2 of z and is also within the
same plane as a and b. Since ∠za = ∣90° − ∠ab∣ ≠ 0° then
⎧ θ
⎪ ∠za + if 0° < ∠ab ⩽ 45°
⎪ 2
⎪ θ
∠a( −c ) ⩽ ⎨ ∠za + if 45° < ∠ab < 90°
⎪ 2
⎪ θ
⎪ ∠za − if 90° < ∠ab < 180°
⎩ 2
(12.32)
⎧ 1
⎪ 90° − ∠ab + ∠ab if 0° < ∠ab ⩽ 45°
⎪ 2
⎪ 1
= ⎨ 90° − ∠ab + (90° − ∠ab) if 45° < ∠ab < 90°
⎪ 2
⎪ 1
⎪ ∠ab − 90° − (∠ab − 90°) if 90° < ∠ab < 180°.
⎩ 2
In every case ∠a( −c ) is a fixed angle less than 90°. This means that one may set
cos(∠a( −c )) = :δ > 0 so a · ( −c ) = δ∥a∥ ∥−c∥ = δ∥a∥ϵ . ■
A shifted variant of lemma 12.1 is used in the following proposition.

12-21
 Modern Optimization Methods for Science, Engineering and Technology

Proposition 12.6. Consider the LMOP (L) with all mi non-parallel and spanning a
subspace of dimension at least n, and an open feasible set X ⊆ n with n ⩾ 2. Then
S(L ) ∩ X ° consists of isolated points. That is, S(L ) ∩ X ° contains no limit points.
Proof. Let (xn )n∈ ⊆ S(L ) ∩ X ° be some sequence converging to a point x ∈ n . The
goal is to show that x ∉ S(L ) ∩ X °. It may be assumed that x ∈ X ° and x ∈ E(L ) or
else x ∉ S(L ) ∩ X ° automatically. For every xn there is a corresponding Mn such that
for all i and y ∈ X with fi (xn ) > fi (y ),

fi (xn) − fi (y )
⩽ Mn
f j ( y ) − f j ( xn )

for all j with f j (y ) > f j (xn ). Since x ∈ X ° and X ° is open, there exists ϵ > 0 and N
such that for all n > N , the neighborhoods Nϵ /2(xn ) ⊂ Nϵ(x ) ⊆ X °.
The fact that the mi span a subspace of at least dimension n shows that there is no
vector v ∈ n such that v is perpendicular to each mi . Therefore, for each xn it must be
the case that for some jn , f jn (xn ) − f jn (x ) = m jn · (xn − x ) > 0; but since each xn and x
itself are efficient there is some other in for which fin (x ) − fin (xn ) > 0. Since there are
infinitely many xn and only finitely many combinations of in and jn , then by the pigeon
hole principle there must a subsequence where the in match and the jn match. Pass to
that subsequence and fix i and j to be those values so that fi (x ) > fi (xn ) and
f j (x ) < f j (xn ) for all n.
For each n > N , using lemma 12.1 take yn ∈ ∂Nϵ /2(xn ) so that mi · (xn − yn )>
1
δ∥mi ∥ ∥xn − yn ∥ > 0 for some fixed δ and mj · (yn − xn ) = n ∥mj ∥ ∥yn − xn∥. The
freedom of choice of yn and the fact that the mi and mj cannot be scalar multiples of
ϵ
one another ensure that such a choice of yn is possible. Note that ∥xn − yn ∥ = 2 .
Take δ 3ϵ > γ > 0 and n large enough so that ∥x − xn∥ < γ . This means that
fi (x ) > fi (xn ) − γ∥mi ∥ because fi is continuous. Similarly true is f j (x ) > f j (xn )−
δ∥mj ∥. Now observe that fi (x ) > fi (xn ) > fi (yn ) and f j (x ) < f j (xn ) < f j (yn ). Now
consider the trade-off quotient
ϵ
fi (x ) − fi (yn) fi (xn) − γ − fi (yn) δ∥mi ∥ − γ∥mi ∥
⩾ > 2 > 0.
1 ϵ (12.33)
f j (yn) − f j (x ) f j (yn) − f j (xn) + γ ∥mj ∥ + γ∥mj ∥
n 2
Since the choice of γ was arbitrary, and γ → 0 implies n → ∞, then the whole trade-
fi (x ) − fi (yn )
off quotient f j (yn ) − f j (x )
→ ∞ as γ → 0. This means that the trade-off quotient cannot
be bounded and thus the sequence xn does not converge to an SE solution.
This quickly leads to a corollary. ■
Corollary 12.3. Consider the LMOP (L) with all mi non-parallel and spanning a
subspace of dimension at least n, and with a closed feasible set X ⊆ n with n ⩾ 2. If
x ∈ S(L ) is not isolated then x must lie on the boundary of the feasible collection.

12-22
 Modern Optimization Methods for Science, Engineering and Technology

One may wonder under what circumstances would an SE solution be found on the
interior. The following shows what an SE point on the interior would entail.
Proposition 12.7. For a problem (L) with X closed and bounded, if there exists an
x ∈ X ° that is SE, then all y ∈ X are SE.
Proof. Let x ∈ X ° be SE and assume that y ∈ X is not SE. First note that y must
be efficient, for if it is not then f (y ) ⩾ f (x ) implies that for all i that mi · y ⩾ mi · x
and for some j that mj · y > mj · x . So for some very small ϵ > 0 it must be that
x + ϵ(x − y ) ∈ X since x is on the interior of X . But that implies that
f (x ) ⩾ f (x + ϵ(x − y )) and in the jth component there is a strict inequality.
That implies x is not efficient which is a contradiction so y must be efficient.
Now if y is not substantially efficient, then there exist sequences
{zk}k∈ ⊆ X
{ik}k∈ , {jk }k∈ ⊆ {1, … , n}

for which
fi n (y ) − fi n (zn)
> n.
f jn (zn) − f jn (y )
As earlier, by the pigeon hole principle a subsequence can be passed to hold in and jn
fixed at i and j .
So given any N > 0 there exists an ϵ > 0 for which x + ϵ(y − zN ) ∈ X . Note
that since the objectives are linear fi (x ) − fi (x + ϵ(y − zN )) = ϵ(fi (y ) − fi (zN )) > 0.
Similarly it must be that f j (x + ϵ(y − zN )) − f j (x ) = ϵ(f j (y ) − f j (zN )) > 0. However,

fi (x ) − fi (x + ϵ(y − zN )) ϵ(fi (y ) − fi (zN ))

= > N, (12.34)
f j (x + ϵ(y − zN )) − f j (x ) ϵ(f j (zN ) − f j (y ))

which is a contradiction to x ∈ S(L ).

Corollary 12.3 and proposition 12.7 combined show that in a common situation
substantially efficient points will only be found on the boundary.
Theorem 12.4. Consider the LMOP (L) with all mi non-parallel and spanning a
subspace of dimension at least n ⩾ 2, where the feasible set X ⊆ n is closed. If
x ∈ S(L ) then it must lie on the boundary of X.
Proof. If x is not on the boundary then it is on the interior. By proposition 12.7 all
the points must be substantially efficient, so x is not an isolated substantially efficient
point. But then by corollary 12.3 if x is not isolated then it must lie on the boundary.
That is a contradiction, so x must lie on the boundary to begin with. ■
This is good because for the most part, analysts will only need to look at the
boundary of X to find SE solutions. For clarification, SE solutions on the boundary
of the feasible collection need not be isolated as the last example makes evident.

12-23
 Modern Optimization Methods for Science, Engineering and Technology

Example 12.7. Let X = {(x1, x2, x3) ∈ 3: x2, x3 ⩾ 0} be the collection of feasible
solutions. Let

F (x ) = (f1 (x ), f2 (x ), f3 (x )) = (m1 · x , m2 · x , m3 · x ),

where m1 = (0, 1, 1),

m2 = (0, 1, 2),

m3 = (0, 2, −1)
be the multi-valued function to be minimized over X . It is evident that the set
S = {(x1, 0, 0) ∈ 3: x1 ∈ } is within the collection of efficient solutions. In order
to check the substantial efficiency of any given s ∈ S only points x ∈ X where
f3 (x ) < f3 (s ) need to be considered because there is no x ∈ X where f1 (x ) < f1 (s ) or
f2 (x ) < f2 (s ). If x has f3 (x ) < f3 (s ) then f1 (x ) > f1 (s ) = 0 and f2 (x ) > f2 (s ) = 0. So
f1 (s ), f2 (s ), f3 (s ) = 0, x3 > 2x2 and x2, x3 ⩾ 0 give
f3 (s ) − f3 (x ) −2x2 + x3
= ⩽1 (12.35)
f1 (x ) − f1 (s ) x2 + x3

f3 (s ) − f3 (x ) −2x2 + x3
and = ⩽ 1. (12.36)
f2 (x ) − f2 (s ) x2 + 2x3

This is enough to say that any s ∈ S is substantially efficient.

12.5 Conclusion
To conclude, a brief history of SE solutions was provided and then followed by some
new information regarding substantial efficiency in general and in the context of
linear problems. It was shown that SE solutions do not always exist in linear
problems, even when the feasible space is bounded. Substantial efficiency was also
shown to connect the perpendicularity of the tangent cone to the direction vectors of
objective functions in linear problems. Various restrictive existence criteria were
provided to ensure a substantially efficient point in very specific cases. Finally,
information about the topology of the set of SE solutions is presented. The theme of
the new results was that finding SE solutions is difficult but not impossible. The
insight gained was that SE solutions are worth pursing, but when they cannot be
found, analysts and decision makers need to be aware of the potential for
disproportional trade-offs to occur between some objectives which can have negative
consequences.
Engineers can find uses in SE solutions as they may find the potential anomaly of
an unbound trade-off between any of their objectives disconcerting. The implica-
tions for market analysts are many if their problems can be made into LMOPs.
First, the knowledge that SE solutions are rare and difficult to detect should keep the
analyst forever on guard against market manipulators and could call for perpetual
regulation through things such as taxation, fines, premiums, time constraints on

12-24
 Modern Optimization Methods for Science, Engineering and Technology

transactions and so on. Second, if the feasible collection of solutions is bounded,

they have a method of checking if their solution is SE by comparing the tangent cone
to the direction vectors mj in the problem. Third, they have a method for trying to
locate an SE solution by restricting the problem to pairs of objectives and finding the
collections of properly efficient solutions then intersecting those collections. Lastly,
they know to only consider the boundary of the feasible solutions for most LMOPs.
Engineers may also find use in taking advantage of a non-market system in the same
way when they have a non-substantially efficient solution to their problem.
All things considered, it seems to the authors that substantial efficiency would be
desired above proper efficiency when available in a solution set. However the
extremely restrictive nature of substantial efficiency makes their existence, if not
rare, hard to predict. This is a good reason to continue pursuing solutions similar to
substantially efficient solutions but with some relaxed conditions. In [23],
Kaliszewski himself introduced Δ-substantial efficiency which excludes infinitesimal
changes around the solutions in consideration. Conditions such as that provided in
[25], where Pourkarimi and Karimi introduced quasi-substantially efficient solu-
tions, or solutions that are substantially efficient only on a subset of objectives.
Another option would be the extension of the concept of ϵ -proper efficiency in [17].
These may make the existence of solutions that continue to avoid the anomalies
produced by improperly efficient solutions in many pairs of objectives more
probable.

References
[1] Chaboud A P, Chernenko S V, Howorka E, Iyer R S K, Liu D and Wright J 2004 The high-
frequency effects of US macroeconomic data releases on prices and trading activity in the
global interdealer foreign exchange market International Finance Discussion Papers 823,
Board of Governors of the Federal Reserve System (US) https://EconPapers.repec.org/
RePEc:fip:fedgif:823
[2] CBOE 2014 2014 CBOE market statistics Technical report, CBOE Global Markets
[3] CBOE 2016 2016 CBOE market statistics Technical report, CBOE Global Markets
[4] Brunnermeier M K 2009 Deciphering the liquidity and credit crunch 2007–2008 J. Econ.
Perspect. 23 77–100
[5] Healy P M and Palepu K G 2003 The fall of Enron J. Econ. Perspect. 17 3–26
[6] Wah E and Wellman M P 2013 Latency arbitrage, market fragmentation, and efficiency: a
two-market model Proceedings of the Fourteenth ACM Conference on Electronic Commerce,
EC ‘13 (New York: ACM), pp 855–72
[7] Pareto V 1971 Manual of Political Economy (New York: Kelley)
[8] Koopmans T C 1951 Efficient allocation of resources Econometrica: J. Economet. Soc. 19
455–65
[9] Kuhn H W and Tucker A W 1950 Nonlinear programming Proc. Second Berkeley Symp. on
Mathematical Statistics and Probability pp 481–92
[10] Geoffrion A M 1968 Proper efficiency and the theory of vector maximization J. Math. Anal.
Appl. 22 618–30

12-25
 Modern Optimization Methods for Science, Engineering and Technology

[11] Seinfeld J and McBride W 1970 Optimization with multiple performance criteria, application
to minimization of parameter sensitivities in a refinery model Ind. Eng. Chem. Process Des.
Develop. 9 53–7
[12] Belenson S and Kapur K 1973 An algorithm for solving multicriterion linear programming
problems with examples Oper. Res. Q. 24 65–77
[13] Borwein J 1977 Proper efficient points for maximizations with respect to cones SIAM J.
Control Optimization 15 57–63
[14] Benson H and Morin T 1977 The vector maximization problem: proper efficiency and
stability SIAM J. Appl. Math. 32 64–72
[15] Wendell R and Lee D 1977 Efficiency in multiple objective optimization problems Math.
Program. 12 406–14
[16] Benson H 1979 An improved definition of proper efficiency for vector maximization with
respect to cones J. Math. Anal. Appl. 71 232–41
[17] Liu J 1999 ϵ-properly efficient solutions to nondifferentiable multiobjective programming
problems Appl. Math. Lett. 12 109–13
[18] Jiang Y and Deng S 2014 Enhanced efficiency in multi-objective optimization J. Optim.
Theory Appl. 162 577–88
[19] Rockafellar R T and Wets R J B 1998 Variational Analysis (Berlin: Springer)
[20] Chen G, Huang X and Yang X 2005 Vector Optimization, Set-Valued and Variational
Analysis (Berlin: Springer)
[21] Ehrgott M 2005 Multicriteria Optimization 2nd edn (Berlin: Springer)
[22] Miettinen K M 1999 Nonlinear Multiobjective Optimization (Norwell, MA: Kluwer
Academic)
[23] Kaliszewski I 1994 Quantitative Pareto Analysis by Cone Separation Technique (Norwell,
MA: Kluwer Academic)
[24] Khaledian K, Khorram E and Soleimani-damaneh M 2016 Strongly proper efficient
solutions: efficient solutions with bounded trade-offs J. Optim. Theory Appl. 168 864–83
[25] Pourkarimi L and Karimi M 2016 Characterization of substantially and quasi-substantially
efficient solutions in multiobjective optimizations problems Turk J. Math. 41 293–304

12-26
 IOP Publishing

Modern Optimization Methods for Science, Engineering and

Technology
G R Sinha

Chapter 13
A machine learning approach for engineering
optimization tasks
Arpana Rawal, Mamta Singh and Jyothi Pillai

The evolving world is continuously tuning itself with the famous underlying
Darwinian principle ‘survival of the fittest species’. This naturally existing successful
optimization system continues to inspire philosophers, analysts and practitioners to
address real-world optimization tasks in our daily lives as well. While dealing with
the combinations of possible worlds that satisfy combinations of constraints,
researchers often attempt to build preference relations for possible worlds, and try
to find a best possible world according to the preferences. The preference for
choosing a set of available alternatives is often based on minimizing some negative
performance parameters, say an error, or maximize some favorable parameters, say
accuracy. Both constraints and alternatives act as functioning world parameters in
optimization terminology.
The conventional taxonomy of optimization methods has seen a multitude of
mathematical programming techniques formulated based on either the types of
objectives or constraints [1]. The principles of mathematical programming/optimi-
zation are borrowed from the field of operations research (OR) and are found to be
suitable for analyzing a set of model functioning parameters under a precise set of
constraints so as to minimize the defined objective function of the model.
Optimization tasks are inherently combinatorial in nature and can be described
with a finite set of constraints and more importantly have a cost function attached to
them. The main goal of carrying out optimization tasks is to find the problem’s
solution or the parameters of the problem in order to arrive at a ‘good’ or
‘acceptable’ solution, if not the ‘best’ solution. Some degree of intelligence ought
to be applied to arrive at optimal, good or acceptable solutions, because of the
infeasibility of applying brute-force to determine all possible solutions in problem
search spaces. One of the most significant trends in the realm of artificial intelligence

doi:10.1088/978-0-7503-2404-5ch13 13-1 ª IOP Publishing Ltd 2020

 Modern Optimization Methods for Science, Engineering and Technology

(AI) during the past fifteen years has been the integration of optimization methods
throughout. Virtually all machine learning (ML) algorithms are reduced to optimal
solution results by mixing models and optimization methods.

13.1 Optimization: classification hierarchy

The initial classification used the distinction criterion of the domain type of the
problem settings, namely discrete valued domains over real-valued domains. Some
models only make sense if the variables take on values from a discrete set, often a
subset of integers, whereas other models contain variables that can take on any
real value.
Models with discrete variables are discrete (combinatorial) optimization prob-
lems, while models with continuous variables are continuous optimization problems.
Continuous optimization problems tend to be easier to solve than discrete opti-
mization problems; the smoothness of the functions means that the objective
function and constraint function values at a point x can be used to deduce
information about points in a neighborhood of x. Continuous optimization
algorithms are important in discrete optimization because many discrete optimiza-
tion algorithms generate a sequence of continuous sub-problems (table 13.1).
Conventional methods of addressing combinatorial optimization problems range
from diversified application domains of resource scheduling, element-routing in
transportation networks and minimization of cost, to performance in real-modeling
problems of Netflix recommendations, driving navigation by GPS, etc. These
methods encompass exact algorithms over shortest-path or minimum-cost spanning
trees or branch-and-bound spaces using integer linear programming formulations
and ‘polynomial-time’ algorithms of linear programming for maximum-speed flow
and minimum-cost flow in networks, multi-commodity flow, and coloring and
matching problems over graph structures, while approximate algorithms are used
for the vertex cover and set cover Steiner-tree and traveling salesman problems.
Owing to the disadvantages observed in these algorithms due to the sub-optimal
solutions arrived at, with weak worst-case bounds and poor empirical performance
in achieving real-world domain constraints, combinatorial approaches have been
used to seek faster and more effective but tailored heuristic algorithms to obtain
nearly optimal solutions using real-time problem-specific constraints. The reader is
referred to [2] and [3] to obtain sound concepts of convex optimization methods by
mathematical programming techniques, exploiting structure using graphs, stochastic
processes and statistical methods [2, 3].
The motivation behind deploying learning heuristics in combinatorial optimiza-
tion tasks is to replace fast approximation computations against heavy ones using
domain knowledge and make these computations generic for similar problem
families. In the case that partially acceptable solutions are obtained, ML heuristics
aim to explore sub-optimal solution spaces, out of which the best solution can be
selected for final real-time problem modeling. Some appropriate exemplary domains
in ML that can be captured to highlight the use of meta-heuristics for combinatorial
optimization are the minimum vertex cover social network routing of messages

13-2
 Table 13.1. Successful optimization implementations in machine learning domains.

Machine learning domain Application domain Case study (instance) Optimization technique

Combinatorial methods Network routing problems Traveling salesman problems Continuous approximation
(exact + approximate methods) methods
Deep neural networks Traveling salesman problems Continuous approximation
methods
Constraint satisfaction Traveling salesman problems Heuristic search methods
Multi-attribute utility theory Decision-making problems Product line consolidation and Statistical methods
(MAUT) selection (family of staplers)
Bayesian classifiers Candidate recruitment Heuristic search methods
Statistical ML Classification problems Text classification Continuous approximation
(convex optimization)
methods
Bayesian networks + non-negative fMRI brain imaging Hidden Markov models
matrix factorization (NMF-FE)

13-3
SVM classifiers Classification (imbalanced Mathematical programming
datasets)
Multi-objective learning Multiple-digit classification, Perceptual task learning,
scene understanding gradient descent search
(citiscape identification)
Probabilistic inference (ranking) Classification problems Search cost optimization for Polynomial (linear)
models early prediction of good programming
protein torsion angles for
protein structure
Modern Optimization Methods for Science, Engineering and Technology

reconstruction
Reinforcement learning Planning and control Job-shop scheduling problems Mathematical programming
Deep neural networks Perceptual task learning Speech and image recognition Gradient descent search
 Modern Optimization Methods for Science, Engineering and Technology

owing to changing influence patterns of neighbor-contacts over the social network,

the formulation of maximum-cut user clusters owing to changing user behaviors
thus resulting in changing maximum-cut weights, and the classical traveling sales-
man problem for the minimum-cost and optimal resource utilization solutions for
large-scale marine transportation network routing problems for the shipment
of goods.
In contrast to mono-objective optimization problems, multi-objective optimiza-
tion problems, if modeled with one equation, can introduce a bias in the modeling
phase, hence multiple tasks need to be solved jointly, sharing an inductive bias.
Instead of arriving at a consistently unique solution in the case of a mono-objective
optimization problem, we obtain a set of solutions obtained in multi-objective
optimization problems, also called Pareto solutions or the trade-off surface. The
latter spectrum of optimization problems has been addressed to date by following
two theoretical principles, as follows.

Definition A multi-objective optimization problem (MOOP) can be defined as the

search for the best set of solutions X = (x1, x2,…..,xn )T (where n is the number of
optimization parameters) that minimizes a number of M conflicting (or not)
objective functions:
Minimize _ _
= fm (X ) m = 1, … , M . (13.1)
Maximize
In the multi-attribute utility theory (MAUT) approach, optimization in complex
engineering decision making can be brought about by integrating the risks and
uncertainties involved with two or more relatively important multiple attributes,
assuming that these multiple attributes are considered in choosing the decision
alternatives and, second, the notion that one attribute is ‘relatively more important’
than another. This optimization step is executed by maximizing the utility function
defined on these attributes or by making tradeoffs depending on the current attribute
levels. Consider a decision problem such as selection of a job, choice of an
automobile or resource allocation, where the choice of the decision alternatives
shares a common feature; decision alternatives are affected by the relative impor-
tance of multiple attributes of interest arising from the input feature vector of the
problem. Applications of MAUT in the public sector are numerous since most
public sector problems involve multiple conflicting objectives, such as in public
healthcare systems, environmental policy, regulatory issues, site selection, and
energy and public policy.
The multi-criteria decision-making theory (MCDM) approach refers to simulat-
ing optimized decision making in the presence of multiple, usually conflicting,
situations and constraints. MCDM problems are present in almost all decision-
making situations ranging from common household decisions to complex strategic
and policy level decisions in corporations and government undertakings. An ML
optimization has been successfully implemented on Bayesian learning using utility
functions by Thomas [4], who mapped a subjective-answer evaluation framework to

13-4
 Modern Optimization Methods for Science, Engineering and Technology

multi-objective decision making incorporating the concept of utility functions as one

of the candidate answer evaluation methodologies. The evaluation method using
utility functions was extended to fuzzy-Bayesian decision making of answer
evaluation categories using fuzzy states for no information, probabilistic imperfect,
probabilistic perfect, fuzzy information and fuzzy perfect information. The reader is
referred to fuzzy-Bayesian learning experiments on the subjective question–answer
evaluation framework [4].
Despite the existence of such nice academic principles of optimization, real
complex industrial domains still face difficulties in performing real-time optimiza-
tion. One such application case study is highlighted—industrial test-cases of material
processing numerical simulation which allow a strategy of trial and error to improve
virtual processes without incurring material costs or interrupting production and
therefore save a lot of money. In the case study, in a French founded ANR project,
ten industrial partners have been selected to cover the different areas of the
mechanical forging industry and provide different examples of the simulation tools.
It aims to demonstrate that it is possible to obtain industrially relevant results on a
very large range of applications within a few tens of simulations and without any
specific automatic optimization technique knowledge, although the simulations
required large computational times and also user time to analyze the results, adjust
the operating conditions and restart the simulation. Therefore, the simulations often
have to be interrupted before the optimum is reached due to a lack of time.
This called for handling the large computational time with a meta-modeling
approach. An evolutionary algorithm coupled with meta-modeling allowed inter-
polating the objective function on the entire parameter space by only knowing the
exact function values at a reduced number of ‘master points’. Two algorithms are
used: an evolution strategy combined with a Kriging metamodel and a genetic
algorithm combined with a meshless finite difference method. The latter approach is
extended to multi-objective optimization. The population-based approach, aiming
to find the set of solutions that allowed the best possible compromises between the
different conflicting objectives, was found to be highly computationally efficient with
the use of the parallel capabilities of the new multi-core hardware in the Forge 2009
IHM computer being trained over all the defined examples and computing several
simulations at reduced time complexities. The reader is referred to [5] for reading on
the presented examples of mono-objective optimization of forging the billet shapes
of a common rail as well as the cogging of a bar, and multi-objective optimization on
the two design parameters in a wire drawing industry process.

13.2 Optimization problems in machine learning

The ML community has produced a tremendous number of successful applications
in recent years, ranging from image recognition to autonomous driving and
outplaying humans in complex games. This spree of newsworthy results renewed
the interest in combining the perspectives of AI and operations research. The state-
of-the-art ML algorithms are gradually being augmented by better optimal
solutions, as advancing businesses and technology are being impacted by the new

13-5
 Modern Optimization Methods for Science, Engineering and Technology

paradigm shift of data science and analytics. This is due to the fact that the
fundamental natural and human resources still remain finite along with the legal and
ethical bounds on real-world complex problem-solving frontiers. Many decades of
early works saw conventional AI and ML algorithms solve real-world optimization
problems using both OR and ML analysis. These works were the consequence of the
need to arrive at the most suitable option from a family of ML models that could
perform fairly well according to some minimum estimate of the generalization error
based on the given training data. This search typically involves some combination of
data pre-processing, optimization and heuristics; these are still constrained by three
popular sources of error: error distribution introduced by learning bias, noise in
datasets and the difficulty of the search problem that underlies the given modeling
problem. These standard optimization packages suffered from efficiency and
scalability issues, owing to the large-scale ML model designs. This in turn led to
the need to design suitable optimization methods using ML heuristics. This is rightly
described as ‘the interplay of optimization and machine learning’ [6].
Optimization lies at the heart of ML. Most ML problems reduce to optimization
problems. The readers of this book must have sound knowledge of operations
research, mathematical programming and baseline ML algorithms. The chapter
does not intend to deal with the mathematical foundations of optimization types,
statistical methods, meta-heuristics nor does it drill down into optimization
classification hierarchy but attempts to explore the depth to which optimization
has invaded different ML tasks based on assimilated exhaustive works performed by
ML analysts. The aim of this chapter is to explore the increasing degree of coupling
observed between available optimization techniques and ML models with the help
of case studies in diverse domains. The authors seek to examine the cross-boundaries
of the following types of works in the literature. First, how small changes to existing
ML models which use methods such as multi-kernel, ranking, clustering, structured
learning and similar have adapted the optimization tasks into newer dimensions.
The second category of works explores how extended sets of optimization methods
have enhanced the scalability and efficiency of ML models, which have also
promoted the use of newer exploratory domains. We have already introduced
chronologically the classification hierarchy of optimization methods as applied in
diversified domains of ML tasks. The rest of the chapter is organized as follows.
Section 13.3 deals with popular optimization methods used for ML in supervised
learning. Section 13.4 explains some case studies revealing how optimal solutions
can be used to address the most challenging task of feature selection in ML.

13.3 Optimization in supervised learning

Supervised learning is the process of data mining for deducing rules from training
datasets. A broad array of supervised learning algorithms exists, every one of them
with its own advantages and drawbacks. There are some basic issues that affect the
accuracy of the classifier while solving a supervised learning problem, such as bias–
variance trade-off, dimensionality of the input space and noise in the input data
space. All these problems affect the accuracy of the classifier and are the reason that

13-6
 Modern Optimization Methods for Science, Engineering and Technology

there is no global optimal method for classification. The classification results are also
altered by the noise in the data, that is, redundant records, incorrect records, missing
records, outliers and so forth. All these problems affect the accuracy of a classifier.

Definition The phenomenon of the dimensionality curse arises due to a large

number of attributes acting as feature vectors in a dataset, wherein, even if the
decision depends on a subset of this high dimensional feature vector, the perform-
ance of the classifier will be clouded by high variance due to the high dimensionality
of the dataset.

In contrast to single-task learning, multi-task learning can be perceived as a learning

paradigm in which data from multiple tasks are used with the hope of obtaining
superior performance over learning each task independently. In multi-task learning,
multiple tasks are solved jointly, sharing the inductive bias between them. Multi-task
learning is inherently a multi-objective problem because different tasks may conflict,
necessitating a trade-off. The potential advantages of MTL can be realized over even
seemingly unrelated real-world tasks, albeit they share strong dependencies due to
the shared processes, for example autonomous driving and object manipulation are
seemingly unrelated, but the underlying data are governed by the same laws of
optics, material properties and dynamics. This motivates the use of multiple tasks as
an inductive bias in learning systems. Other multi-task learning problems that have
already been addressed are digit classification, scene understanding (joint semantic
segmentation, instance segmentation and depth estimation), and multi-label classi-
fication [7].
A typical MTL system is given a collection of input points and sets of targets for
various tasks per point. A common way to set up the inductive bias across tasks is to
design a parameterized hypothesis class that shares some parameters across tasks.
Typically, these parameters are learned by solving an optimization problem that
minimizes a weighted sum of the empirical risk for each task. However, the linear-
combination formulation is only sensible when there is a parameter set that is
effective across all tasks. In other words, minimization of a weighted sum of
empirical risk is only valid if tasks are not competing, which is rarely the case. MTL
with conflicting objectives requires modeling of the trade-off between tasks, which is
beyond what a linear combination achieves, hence this forms the most suitable case
study in illustrating how optimization works in such a learning paradigm.

13.3.1 Bayesian optimization

The use of the naïve Bayes’ theorem (also called the maximum a posteriori
hypothesis) as part of probability statistics provides a quantitative approach to
learning algorithms by manipulating over probabilities in order to weigh the
evidence supporting the hypothesis. In a simple sense, it can accommodate the
classifying of new data instances (prior knowledge is given for observed data
instances) by combining the predictions of multiple hypotheses, weighted by their

13-7
 Modern Optimization Methods for Science, Engineering and Technology

probabilities. Even if the Bayesian learning algorithm does not explicitly manipulate
probabilities, it can help in identifying the model characteristics under which the
learning behaves optimally [8]. Such Bayesian optimality can be observed in the
learning of problems with continuous-valued domains, where Bayesian analysis is
used for minimizing the squared error between the output hypothesis predictions
exhibiting maximum likelihood estimates. The reader is referred to the basic
concepts of probability densities over normal distributions for understanding
Bayesian classifiers in [8].

13.3.2 Bayesian optimization for weight computation: a case study

We examine a case study that describes a problem setting where the maximum
likelihood hypothesis is used for predicting probabilities that are learnt as weights
for training a fuzzy weighted associative classifier on a collection of patient heart-
disease data taken from the heart.D53.N303.C5.num database [9]. Initially, a
criterion is defined upon which the optimization step is performed using the
maximum likelihood hypothesis upon any learnt target function, say f ′ in the
problem setting; f ′: X ➔ [0, 1], such that f ′ = P(f(x) = 1). Training data D is
extracted in the form D = {〈x1, d1〉, 〈x2, d2〉, …, 〈xm, dm〉}, where di are the observed
occurrences of the disease, with a 0 or 1 value for each patient instance, xi. With the
assumption that the probability of encountering a patient xi is independent of the
hypothesis (h) about his/her chances of being a victim of disease occurrence, we
attempt to compute the maximal probability P(di∣h, x1) of observing di = 1 for a
single instance xi, given a world in which hypothesis h holds:
⎧ h(xi ) if di = 1
P(di / h , xi ) = ⎨ (13.2)
⎩(1 − h(xi )) if di = 0

P(di / h , xi ) = h(xi )di (1 − h(xi ))1−di . (13.3)

Thus, treating the xi and di variables of a binomial distribution of the given training
data samples, we compute the likelihood ratios by iterating over all probability
values around the initially assigned probability value of disease occurrence as
observed from the training samples and obtain the maximum of all h(.) values as a
maximum likelihood estimate. The expression for MLE is:
argmax m
hML =
h∈H
∏i=1 h(xi )di (1 − h(xi ))1−di . (13.4)

Table 13.2 provides the binomial distribution of 11 patient instances in age group 1
and 17 patient instances under age group 2, whose weights need to be computed
using Bayesian optimization over the maximum likelihood hypothesis. The like-
lihood estimate computations, P(di∣ h, xi) can be seen in table 13.3 for a set of
different iterating probabilities, h(.). The maximum of all h(.) values is declared as
the MLE and is assigned as the weight for learning the fuzzy weighted associative
classifier.

13-8
 Modern Optimization Methods for Science, Engineering and Technology

Table 13.2. Patient dataset (training sample).

Patient_ID age_group P(.) Patient_ID age_group P(.)

1 I 0 4 II 0
2 I 0 5 II 0
3 I 0 6 II 0
4 I 0 7 II 0
5 I 0 8 II 0
6 I 0 9 II 0
7 I 0 10 II 1
8 I 0 11 II 1
9 I 1 12 II 1
10 I 1 13 II 1
11 I 1 14 II 0
1 II 0 15 II 1
2 II 0 16 II 1
3 II 0 17 II 1

Table 13.3. Bayesian optimization over maximum likelihood hypothesis.

Prior Training
probability sample h(.) (iterating hML
Patient_age_group_type (h(xi)) (1 − h( xi)) size (m) probabilities) (MLE)

I 3/11 8/11 11 {1/11, 2/11, 3/11, 0.2727

4/11}
II 7/17 10/17 17 (5/17, 6/17, 7/17, 0.412
8/17, 9/17}

13.3.3 Bayesian optimal classification: a case study

Organizational dynamics falling in line with employee career dynamics, the selection
of candidates to be placed in a certain position by any organization, requires careful
planning. This case study deals with managerial research conducted on a quantita-
tive scale that aims to formulate a decision-making system that assigns a
suitable designation alternative to the participating candidates on the basis of
quantitative assessments [10]. The human resources (HR) section of company X is
responsible for taking the initiative in selecting candidates in accordance with the
assessment of their superiors by collecting data on a five-point rating using a
questionnaire method applied over the candidate’s behavioral data and interviews.
Three managers were nominated to assess the 11 candidates for subordinates and fill
in the questionnaire comprising twelve indicators, as can be seen in table 13.4.

13-9
 Table 13.4. Bayesian optimization over the maximum likelihood hypothesis. Adapted as illustration from [10] © 2018 IOP Publishing. Reproduced with permission.
All rights reserved.

Indicator 5 4 3 2 1

Work quality Always exceeds the set Sometimes exceeds the set Always meets the set Sometimes does not meet Always achieves less than
standards. standards. standards. the set standards. the set standards.
Work quantity The quantity, volume, The quantity, volume, The quantity, volume, The quantity, volume, The quantity, volume,
frequency or speed of frequency or speed of frequency or speed of frequency or speed of frequency or speed of
completion of work completion of work completion of work completion of work is completion of work is
always exceeds the set sometimes exceeds the always meets the set sometimes insufficient always less than the set
standards. set standards. standards. to meet the set standards.
standards.
Knowledge and Very good at mastering Good at mastering and Masters and understands A lack of mastering and Very lacking in mastering
work skills and understanding the understanding the the knowledge and understanding the and understanding the
knowledge and skills for knowledge and skills for skills for the range of knowledge and skills for knowledge and skills for
the range of duties and the range of duties and duties and the range of duties and the range of duties and

13-10
responsibilities. responsibilities. responsibilities. responsibilities. responsibilities.
Initiative Very fast, precise and Very fast, precise and Very fast, precise and Slow in acting to carry out Slow in acting to carry out
correct in acting to correct in acting to correct in acting to and complete duties, and complete duties,
carry out and complete carry out and complete carry out and complete and always waits for and always waits for
duties without waiting duties, and sometimes duties, and sometimes detailed instructions. very detailed
for orders and waits for instructions waits for relatively instructions.
instructions. that are general in detailed instructions.
nature.
Cooperation Active in helping and Active in helping and Needs to be reminded to Not enthusiastic in helping Provides no help or
supporting all co- supporting co-workers help and support co- or supporting support for colleagues
Modern Optimization Methods for Science, Engineering and Technology

workers, gives very in certain sections/ workers, needs to pay colleagues despite being and/or always becomes
serious attention to the groups only, gives serious attention to the reminded/reprimanded, an obstacle in achieving
achievement of group/ serious attention to the achievement of group/ little attention to the group/company goals.
company goals. achievement of group/ company goals. achievement of group/
company goals. company goals.
 Reliability Always willing to be Always willing to be Always willing to be Always willing to be Always unwilling to be
assigned outside duties assigned outside duties assigned outside duties assigned outside duties assigned outside duties
and responsibilities and and responsibilities, and responsibilities, and responsibilities, and responsibilities,
results are as expected. results are relatively results are less than results are rarely as results are rarely as
close to what is expected. expected. expected.
expected.
Learning ability A great passion to learn A good spirit to learn for Sufficient interest to learn Less enthusiasm for Little interest in learning
and for self-development self-development and for self-development learning for self- for self-development
willingness and improvement of improvement of work and improvement of improvement and and improvement of
work ability, earnestly ability, seriously wants work ability, but needs improvement in work work ability, although
applies what has been to apply what has been to request support from ability, and/or needs to has been reminded.
learned in duties. learned in duties. superiors to apply what be controlled /reminded
has been learned in to always apply what
duties. has been learned in
duties.
Attendance Never late, never leaves Performance appraisal is Performance appraisal is Performance appraisal is Performance appraisal is

13-11
early, no work lost. affected by early logout affected by 3–5 late affected by late logins affected by habitual late
and/or 1 day lost work. arrivals and early and/or 4–5 days lost logins and/or more than
departures, and/or 2–3 work. 5 days lost work.
days lost work.
Planning and Excellent ability in setting Good ability in setting Fairly good ability in Less able to set task Not able to assign task
organizing task priorities and task priorities and setting task priorities priorities and less priorities and manage
placing and managing placing and managing and placing and effective in placing and existing resources.
existing resources. existing resources. managing existing managing existing
resources. resources.
Controlling Good at monitoring and Good at monitoring and Sometimes needs guidance Always needs guidance to Always needs a reminder
Modern Optimization Methods for Science, Engineering and Technology

controlling all resources controlling all resources to monitor and control monitor and control all to simply monitor and
very optimally. optimally enough. all resources, but less resources, but less than control resources.
than optimally. optimally.

(Continued)
 Table 13.4. (Continued )

Indicator 5 4 3 2 1

Decision Makes a decision on a Makes a decision on a Capable of making a Sometimes needs a great Always slow in making
making problem very quickly problem quickly and decision on a problem deal of support and decisions, even if they
and accurately, and the accurately, and the quickly and accurately, direction to make have been given
results can be justified. results can be justified. although often needs decisions quickly and guidance, and their
support and direction. accurately, and/or directives or decisions
decisions are sometimes are not justifiable.
less justified.
Development of Very attentive in the Attentive in the Sufficient attention on the Lack of attention on the No attention on the
subordinates improvement and improvement and improvement and improvement and improvement and
development of development of development of development of development of
subordinates and/or has subordinates, but lacks subordinates, but does subordinates, and does subordinates, and does
a planned program for a planned program for not have a planned not have a planned not have a planned
staff. staff. program for staff. program for staff. program for staff.

13-12
Modern Optimization Methods for Science, Engineering and Technology
 Modern Optimization Methods for Science, Engineering and Technology

Table 13.5. The ranked criteria for candidate assessment. Adapted as illustration from [10] © 2018 IOP
Publishing. Reproduced with permission. All rights reserved.

Criteria of scale Description

Criterion 1 Has a good ability to understand, a high morale and a satisfactory quality of
work.
Criterion 2 The results obtained are very effective with a near perfect level of accuracy.
Criterion 3 Self-organization and subordinates are optimal, and is responsible for all the
work done.
Criterion 4 Has a high level of creativity and initiative in working with the task of
applying the knowledge mastered.
Criterion 5 Has the courage to make decisions and dares to be accountable for all the
risks that may occur.

These indicators are: (a) work quality, (b) work quantity, (c) knowledge and work
skills, (d) initiative, (e) cooperation, (f) reliability, (g) learning ability and willing-
ness, (h) attendance, (i) planning and organizing, (j) controlling, (k) decision making
and (l) development of subordinates, each expressed with five-point rated descriptive
options that can closely measure the participating candidate. The indicators and
their rating scale of responses is pre-defined by the HR department of company
X after analyzing several factors that are sure to influence the decision-making
process, including the past experience of the candidate, cognitive differences, age
and individual differences, belief in personal relevance and lastly their escalation of
commitment. Another constraint imposed by the HR department is the fulfillment
of the criteria in selecting candidates according to the needs of the position.
The criteria are defined in ascending order of preference for recruiting candidates
for top managerial positions of the company, as can be seen in table 13.5, as a
consequence of previous review; this study is internal and has been recognized by the
management of the company.
The understanding of Bayes’ optimal classification in the case study refers to the
fulfillment of the recommendation criteria by any contesting candidate (Kj) for
whom the probability P(Kj∣D) is maximum. This most probable classification of
instance ki can be computed from an expression of Bayesian optimal classification:
argmax
vj ∈ V ∑hij ∈H
P(vj ∣hi )P(hi ∣D ). (13.5)

It should be noted that in the case study experiments the set of five criteria is
defined as a set of five hypotheses hi in the whole hypothesis space H. While the
candidates were being personally judged, the managers felt that some of the
decision-making indicators were relatively more important and occupied the top
priority in lieu of the fulfillment of the specific criteria that were defined in table 13.5.
Hence, distinct sets of preferred indicators needed to be screened while interviewing
candidates under distinct criteria, in order to decide whom to recommend for the
related vacant positions. This causes the formulation of a criteria-specific feature-

13-13
 Modern Optimization Methods for Science, Engineering and Technology

vector mapping table that could be drawn from the response value patterns that were
captured by the managers during the candidates’ interview procedures, as can be
seen in table 13.6.
Eventually, each of the candidates were assessed with above average values of
responses in a set of not more than 6–7 indicators, as listed in table 13.7. The rest of
the responses are assumed to be ‘0’ or ‘—’ (do not care) for obtaining the resulting
table of naïve Bayesian probability computations for the contesting candidates
under each criterion of assessment.
The posterior probabilities of each candidate being recommended into each
criterion needs the computation of separate hypothesis combinations, governed by
the distinct set of feature vectors (acting as evidential support) mentioned above. For
example, the posterior probability of candidate Kj, P(Kj ∣ D) can be computed by the
expression that follows from equation (13.5) over all five hypothesis (criterion)
combinations.
Each P(Kj ∣ hi) is the conditional probability of candidate Kj being recommended
in criterion (hypothesis combination) hi in our case study; P(hi ∣ D) is the frequency
of occurrence of the relative sample of criterion hi assessed over each of the
12 indicators. The latter component can be visualized as the consequence of training

Table 13.6. The criteria-specific feature-vector mapping matrix.

Criterion a b c d e f g h i j k l

h1 X X X X X X
h2 X X X X X X X
h3 X X X X X X X X
h4 X X X X X X X X X
h5 X X X X X X X X

Table 13.7. Naïve Bayesian computations for candidates K1–K11.

Candidate Criterion X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 p(X∣C)

K1 1 4 3 3 — — — 5 5 3 — — — 0.7138
K2 4 4 — 5 5 — — 3 5 — — 3 3 0.7188
K3 2 5 5 — — — 4 — 4 3 — 3 — 0.7119
K4 3 3 — — 5 — — — 5 5 — 4 5 0.6853
K5 5 — 3 — 5 3 — 3 4 — 3 5 — 0.7875
K6 4 3 — 5 5 3 — 5 5 — 3 — — 0.7478
K7 1 4 4 3 — — — 4 5 3 — — — 0.7069
K8 3 3 — 3 4 5 — — 4 5 — 4 5 0.816
K9 5 — 3 — 4 — 3 3 4 — — 5 — 0.6958
K10 2 5 4 3 — — 4 — 4 — — 3 — 0.678

13-14
 Modern Optimization Methods for Science, Engineering and Technology

the recruitment domain with past employee promotion histories and is available in
table 13.8;
Ph5(K5 ∣ D ) = ∑P(K5 ∣ h5). P(h5 ∣ D) (13.6)

P *(K5 ∣ D ) = max[P(K5 ∣ h5). P(h5 ∣ D ), P(K5 ∣ h 4). P(h 4 ∣ D ), P(K5 ∣ h3).

(13.7)
P(h3 ∣ D ), P(K5 ∣ h2 ). P(h2 ∣ D ), P(K5 ∣ h1). P(h1 ∣ D )].

It can be observed that the candidates were assessed under five different criteria
using the relevant set of decision-making indicators, and their posterior probabilities
of falling in those criteria were computed using the conditional probabilities of
table 13.8 relevant to those criteria. These observations are listed in table 13.9 along
with the recommended criterion fulfilled by each candidate Kj.
The candidate K5, who is ranked one under criterion 5, was recommended for the
top managerial position for decision making in the company’s management, while
candidate K6, who satisfies criterion 4 with maximum marginal probability (0.7478)

Table 13.8. Naïve Bayesian conditional probability distribution over 12 indicators, five criteria. Adapted as
illustration from [10] © 2018 IOP Publishing. Reproduced with permission. All rights reserved.

Criterion Conditional probability distribution

a b c d e f g h i j k l

1 1/5 9/58 8/58 7/58 1/58 1/58 1/58 10/58 11/58 7/58 1/58 1/58 1/58
2 1/5 11/59 10/59 4/59 1/59 1/59 9/59 1/59 9/59 4/59 1/59 7/59 1/59
3 1/5 7/75 1/75 4/75 8/75 11/75 1/75 1/75 10/75 11/75 1/75 9/75 11/75
4 1/5 8/69 1/69 11/69 11/69 4/69 1/69 9/69 11/69 1/69 4/69 4/69 4/69
5 1/5 7/48 1/48 1/48 10/48 4/48 4/348 7/48 9/48 1/48 4/48 11/48 1/48

Table 13.9. Naïve Bayesian posterior probability computations for candidates K1–K11.

Candidate Criterion Ph5(ki ∣ D) Ph4(ki ∣ D) Ph3(ki ∣ D) Ph2(ki ∣ D) Ph1(ki ∣ D) P*(ki ∣ D) = max(.)

K1 1 1.73 2.39 1.64 1.71 3.57 3.57

K2 4 3.10 3.59 2.64 2.31 2.69 3.59
K3 2 1.88 1.39 1.80 2.67 2.59 2.67
K4 3 2.90 2.46 3.43 2.40 2.02 3.43
K5 5 3.94 2.46 2.11 1.96 1.69 3.94
K6 4 3.21 3.74 2.19 1.48 2.88 3.74
K7 1 1.60 2.26 1.64 1.73 3.53 3.53
K8 3 2.92 2.91 4.08 2.27 2.19 4.08
K9 5 3.48 1.96 1.56 2.21 1.69 3.48
K10 2 1.85 1.87 1.52 2.65 2.45 2.65
K11 5 3.56 2.23 1.89 1.96 2.03 3.56

13-15
 Modern Optimization Methods for Science, Engineering and Technology

was recommended as the assistant of the elected head of office. The candidates K8,
K3 and K1 satisfying criteria 3, 2 and 1 with maximum a posteriori probabilities
0.816, 0.712 and 0.714, respectively, will be placed around the factory with the
responsibility of holding machines and necessary job-scheduling tasks. The reader is
referred to [10] for the details of the other objectives of the mentioned case study.

13.3.4 Bayesian optimization via binary classification: a case study

When solving an optimization problem, we usually obtain a set of a large number of
solutions. What if we learn about the optimum using these solutions? Shylo and
Shams [11] worked on an ML approach to formulate a statistical learning problem
as an optimal binary classification problem by implementing a logistic regression
model on a collection of solutions of empirical datasets. Job scheduling problems
and exploring their predictive accuracy computationally in this application domain
happens to be the earliest of all case-study domains [11].
A binary optimization problem can be formally expressed as
Minimize f (x )
s.t. x ∈ X ⊂ (1, 0),
where the objective function f is an arbitrary function, and {0, 1} ∩ x ≠ ∅.
A feasible vector x is defined as one possible optimal solution defined in solution
space χ with its feature dimensionality denoted by index parameter j iterating on
index set 1, n ≡ {1, 2, ……, n}. Also, the learning classifier is assumed to partition
each solution x using an iterative method; the learning transits from one solution to
another after each iteration i, resulting in a set of m such binary optimal solutions,
{x1, x2, …, xm} ⊆ χ, discovered after m iterations.
Mathematically, each solution xi can be partitioned into two subsets:

x1 : C11 = {j∣x j* = 1} and C10 = {j∣x j* = 0} , j ∈ 1, n

· (13.8)
·

x* : C*1 = {j∣x j* = 1} and C*0 = {j∣x j* = 0} , j ∈ 1, n

· (13.9)
·

x m : C m1 = {j∣x j* = 1} and C m0 = {j∣x j* = 0} , j ∈ 1, n . (13.10)

For such a sequence of binary optimal solutions, we evaluate the corresponding

objective function values. For j ∈ 1, n , let Ij(.) denote a list of m pairs formed by the
jth components of these solutions and the corresponding objective values:
⎡ ⎤
( )
I j (.) = ⎣ xj 1j , f (x1) , (x j2 , f (x 2 )),……(x jm , f (x m ))⎦ , (13.11)

13-16
 Modern Optimization Methods for Science, Engineering and Technology

where each jth component falls either as an element of binary class 1 or class 0 for
any mth solution instance xm. Hence the two partitions constructed out of Ij(t) can be
defined by the two sequences:

I 1j (.) = {(x k
j , ) ( )
f (x k ) : x jk , f (x k ) ∈ I j (t ), x jk = 1}, (13.12)

I j0(.) = {(x k
j , ) ( )
f (x k ) : x jk , f (x k ) ∈ I j (t ), x jk = 0}. (13.13)

In other words, in order to build a prediction model, we need to map n such vectors,
I j (t ), j ∈ 1, n , into the interval [0, 1], estimating the larger of the two conditional
probabilities P (j ∈ C1 ∣ I j (t )) and P (j ∈ C 0 ∣ I j (t )), and classifying the class label
with that class, whose probability value is greater. Now, this binary classification
problem can be solved with a variety of statistical methods. The case study aims to
classify such a problem using the logistic regression method, a widely used statistical
technique and, hence, often included as a combinatorial optimization method,
whose sigmoidal functional representation is defined as
1
h θ (X ) = . (13.14)
1 + e−(θ0+θX )
Usually, the maximum likelihood estimate (MLE) metric Θ is used to fit the
parameters of the above-mentioned dataset {xk, f(xk)} and can be obtained as
1
hθ(t )(I j (t )) = .
⎡ ⎤ (13.15)
1 + exp ⎢∑ θ (t )f (xk ) − ∑(0, f (x ))∈I (t ) θ (t )f (xk )⎥
⎣ (1, f (xk ))∈I 1j (t ) k
0
j ⎦

Now, substituting the sequences of optimal objectives corresponding to the two

binary class labels C1 and C0 in the denominator component of the above
expression, we obtain the revised expression as
1
hθ(t )(D 1j , (t ), , , D j0, (t )) = , (13.16)
1 + exp θ (t )⎡⎣D 1j (t ) − D j0(t )⎤⎦
( )
where

D j0(i ) = min ({f (x ) k

j (x jk , )
f (x k ) ∈ I j (t ), x jk = 1 }) (13.17)

D 1j (i ) = min ({f (x ) k
j (x jk , )
f (x k ) ∈ I j (t ), x jk = 1 }). (13.18)

An example of a training dataset is considered for implementing logistic regression,

as shown in table 13.10, where two independent regressor variables D1j(i) and D0j(i) are
listed from the execution of a baseline search algorithm (the TABU and Guided TABU
search techniques are used in the case study) that reports the best objective function f(x)
for each solution component. The regression modeling for the component has been

13-17
 Modern Optimization Methods for Science, Engineering and Technology

Table 13.10. Training samples (logistic regression model).

D1(t) –
D0j (i ) D1j (i ) D0(t) Ο(.) Θ * Δ hΘ=0.5 Θ * Δ hΘ=0.2 Θ * Δ hΘ=0.15 Θ * Δ hΘ=0.1 Θ * Δ hΘ=0.05

1395 1366 29 0 14.5 0.00 5.80 0.00 4.35 0.01 2.9 0.05 1.45 0.19
1368 1400 −32 1 −16 1.00 −6.40 1.00 −4.8 0.99 −3.2 0.96 −1.6 0.83
1366 1438 −72 1 −36 1.00 −14.40 1.00 −10.8 1.00 −7.2 1.00 −3.6 0.97
1373 1366 7 0 3.5 0.03 1.40 0.20 1.05 0.26 0.7 0.33 0.35 0.41
1379 1365 14 1 7 0.00 2.80 0.06 2.1 0.11 1.4 0.20 0.7 0.33
1365 1389 −24 1 −12 1.00 −4.80 0.99 −3.6 0.97 −2.4 0.92 −1.2 0.77

computed for various probabilities P(.) = {0.5, 0.2, 0.15, 0.1, 0.05} and the model is
found to fit the best for the predicted probability (0.5), as can be seen in the calculations
performed in the same table.

13.4 Optimization for feature selection

As ML aims to address larger, more complex tasks, such as personalization of
filtering systems on the world wide web or electronic mailboxes, Netnews and the
like, the problem of screening out the most relevant information from the huge
volumes of both necessary and low-quality information is becoming an increasing
concern. Revisiting the curse of dimensionality, which increases the computational
time exponentially with an increase in the number of features to tens or hundreds in
an induction task, a set of either irrelevant or redundant features plays no role in
prediction except for serving as noise and increasing the induction time. A very
popular exemplary domain that can be used to illustrate the above-mentioned
phenomenon is automatic image recognition in more realistic contexts involving
noise, changing lighting conditions and shifting viewpoints. Since each feature used
as part of a classification procedure can increase the cost and running time of a
recognition system, there is strong motivation within the image processing com-
munity to design and implement systems with small feature sets. Similarly, it is not
uncommon in a text classification task to represent examples using 104 to 107
attributes, with the expectation that only a small fraction of these will contribute in
learning the patterns. In contrast, in an attempt to include a sufficient set of features
to achieve high recognition rates by generating effective classification rules, the
image learning community seeks an ‘optimal’ subset of features from a larger set of
possible features for various recognition tasks.
In order to remove irrelevant or redundant features, the feature selection
relevancy criterion can be defined in one of the following patterns. We adhere to
the below mentioned standard notations throughout this section for representing our
data and variables:
• The relevance of a feature with an output class label: Given an N-dimensional
feature vector, described by an instance space of feature domains, F1 X F2 X
… X Fn, we can model the sample S as having been extracted from this

13-18
 Modern Optimization Methods for Science, Engineering and Technology

instance space and associate them with classification labels 1 ... C (whether
Boolean, multiple valued or continuous) according to the target function, c. A
feature xj can be considered as relevant to target concept c, if there exists a
pair of samples s and s′ (s, s′ ∈ S) such that c(s) ≠ c(s′) and xj(s) ≠ xj(s′) only
apply in their feature vectors.
• Extending the concept of relevance with respect to the distribution of samples
in instance space, a feature xj can be considered as strongly relevant to sample
set S, if there exists a pair of samples s and s′ (s, s′ ∈ S) such that c(s) ≠ c(s′)
and xj(s) ≠ xj(s′) only applies in their feature vectors. In other words, a feature
xj is said to be weakly relevant to sample S, where S = f(c, D), and if it is
possible to remove half a subset of features so that xj becomes strongly
relevant. Several other works have also concluded that a feature can be
regarded as irrelevant, if it is conditionally independent of class labels which
means that the feature having no influence on class labels can be discarded
[19–22]. Even if the feature is independent of the input data, it cannot be
independent of the class label for constructing a suitable learning model.

Discussing the vast spectrum of feature selection algorithms for dealing with
datasets that contain large numbers of irrelevant attributes, these have been
characterized as either ‘wrapper’, ‘filter’ or ‘embedded’ approaches, based on the
relation between the selection scheme and the basic induction algorithm. Irrespective
of any of the methods adopted, a convenient paradigm for viewing many of these
algorithms is to find a sub-optimal procedure to compute a subset of possible
features out of a heuristic search, with each state in the search space specifying a
subset of the possible features. Instead of directly evaluating all subsets of (2N)
features in N-dimensional instance space, we adopt one of the following approaches
to implement our heuristics:
1. One might begin with nothing or a finite set of attributes representing a
starting point in the search space and successively add attributes; this
approach is called forward selection.
2. One might start with all attributes and successively remove them on the basis
of some heuristics; this approach is known as backward elimination.

This is nicely explained by Blum and Langley through a step-wise constructed state-
transition diagram of partially ordered feature search space [12].
Filter methods use variable ranking techniques as the principle criteria for variable
selection by ordering; hence, they are used for the pre-processing step, wherein highly
ranked features are selected above a threshold using a suitable ranking criterion. They
can be applied before classification to filter out the less relevant variables. Two of the
well-known filter methods are RELIEF and its variants [13] and FOCUS [14]. While
RELIEF computations provide the relevance weighting of each feature with reference
to its class label; the FOCUS algorithm iterates on all feature subsets to arrive at a
minimal set of features that can provide consistent labeling of training data. The
reader is advised to go through these methods before working on optimal feature-
extraction methods, i.e. see [13, 14].

13-19
 Modern Optimization Methods for Science, Engineering and Technology

The major difference between the two filter methods is that RELIEF computa-
tions need an extra heuristic to compute a threshold on selecting the final feature
subset. Although effective in removing redundant features, all combinations of
highly correlated features in the given feature vector will obtain high relevance
weightings in RELIEF, while the FOCUS algorithm is very sensitive to noise or
inconsistencies residing in the training dataset. Another less popular category of
wrapper methods performs a search through the space of feature subsets using the
estimated accuracy obtained after ‘wrapping around’ the particular selection from
that induction (learning) model. These methods are computationally expensive.
Moreover, both feature-extraction methods lose their practicality with an exponen-
tial rise in the dimensionality of the total feature vector given at the input.

13.4.1 Feature extraction using precedence relations: a case study

Normally, educational data mining (EDM) predictor attributes include students’
personal files, demographic data, academic data, learning patterns, behavioral
attributes, institutional infrastructure and facilities. Usually, the student’s perform-
ance evaluation tools do not declare the kind of effort they still need to apply in
order to pass their ongoing course of study. Very few prediction models are able to
optimize to what extent these academic efforts must be enforced both at the student
level and at the teacher level so that they can be predicted into the category of pass
students in their forthcoming end-of-semester examinations. One such case study
deals with a sequence of experiments carried out by Singh [16], where in the first
experiment one takes in four independent experimental parameters, namely attend-
ance (x1), assignment credit (x2), internal score (x3) and subject count (x4).
In subsequent experiments, the attribute schema was increased by two more
features, namely laboratory credit (x5) and previous year or semester marks (x6) of
the students to formulate a six-attribute feature-extraction (FE)-cum-ranking model.
For formulation of a nine-attribute model, three additional parameters were
included, that is marks in higher secondary exam (x7), medium of study (x8) and
student’s location of residence (x9).
The implemented feature-extraction-cum-ranking model was developed with two-
phase functionality: in the first phase, a naïve Bayesian (NB) mining model
framework was implemented to predict ‘at-risk’ and ‘above-risk’ levels of students
belonging to ongoing courses. The posterior probability computations on students’
fitness and unfitness that determine their ‘risk’ class labels are shown in the
expressions below:
4 ⎛ xi ⎞
∑i=1p(fit). p⎜⎝ ⎟
fit ⎠
P(fit ∣ {x1, x2 , x3, x4}) = (13.19)
4 ⎛x ⎞ 4 ⎛ x ⎞
∑i=1p(fit). p⎜⎝ i ⎟⎠ + ∑i=1p(unfit). p⎜⎝ i ⎟⎠
fit unfit

13-20
 Modern Optimization Methods for Science, Engineering and Technology

4 ⎛ xi ⎞
∑i=1p(unfit). p⎜⎝ ⎟
unfit ⎠
P(unfit ∣ {x1, x2 , x3, x4}) = . (13.20)
4 ⎛ xi ⎞ 4 ⎛ xi ⎞
∑i=1p(fit). p⎜⎝ ⎟ + ∑i=1p(unfit). p⎜⎝ ⎟
fit ⎠ unfit ⎠
These NB a posteriori computations on class labels were further used to arrive at
the relative relevance of attributes contributable to success/failure grades in the
second phase. Here, the individual portions of the summative numerator compo-
nents of expressions (4.2) and (4.3) that pertain to the single attribute, reflected the
posterior effect of that attribute in evaluating relative attribute fitness/unfitness. The
comparisons among such individual portions helped in ranking the attributes of the
experimental feature vector. In this way, attribute precedence relations were obtained
by revisiting these individual numerator components, i.e. the average fitness
(average_fit(xi, tj)) and average unfitness (average_unfit(xi, tj)) of the students owing
to the degree of involvement from each attribute.
Kira and Rendell [13] designed the RELIEF algorithm that assigns a relevant
weight to each attribute of the feature vector by computing the difference between
the selected test instance with reference to the nearest-hit and the nearest-miss
training instances. Assuming the training instances are denoted by a p-dimensional
feature vector X, the RELIEF algorithm makes use of a p-dimensional Euclidean
distance to select ‘near-hit’ and ‘near-miss’ instances from the training dataset. If the
test instance xj is predicted as a positive instance then the near-hit instance (xj) is
assigned as Z+ and the near-miss instance (xj) is assigned a Z-value. The reverse
happens if the test instance is predicted as a negative instance. In such a situation,
the nearest negative training neighbor is assigned Z+ and the nearest positive
training neighbor, possessing the opposite class value, is assigned a Z-component.
These components are computed as part of the preparation for computing weight
updates, as described in expression (13.21). This weight update operation was
performed on each of the participating attributes in the experimental feature vector.
These updated attribute weights act as rank values of the attributes when sorted in
increasing order of relevance. The author also appreciates the nearest-neighbor
approach to finding the ‘nearest-hit’ and ‘nearest-miss’ training instances to compute
the weight updates as defined above:

wi′ = wi − diff(xi , near-hit-instancei )2

. (13.21)
+diff(xi , near-miss-instancei )2

Sun and Wu [15], during their in-depth study of feature selection methods, proved
that RELIEF is the most successful algorithm that solves a convex optimization
problem with a margin based objective function [15]. As it was observed that the
RELIEF model could not filter out redundant attributes as well as weakly relevant
ones, this case study provided variant logistics to the approach. Attribute precedence
relations were thus introduced as an innovative mining metric for personalized
counseling of students and the performance of such a hybrid feature-extraction-cum-
ranking model is evaluated by extending the experiments with the RELIEF method

13-21
 Modern Optimization Methods for Science, Engineering and Technology

as the benchmark, i.e. by comparing equivalently generated attribute precedence

relations from the RELIEF feature-extraction model. We will not go into the details
of the heuristics, or on how the residual weights are initialized to arrive at the
weighted feature vector by performing weight update computations using RELIEF
logistics. The reader is referred to the feature-vector ranking computations
performed by Singh as a part of doctoral work in the educational data mining
realm [16].
The performance of the hybrid model is shown in table 13.11 in terms of
percentage accuracy for three types of experimental set-ups as already discussed in
the beginning of this case study. The feature-extraction model accuracy was found to
be the highest in the four-attribute FE model and the lowest in the nine-attribute FE
model. The reasons for the above results revealed that when some academic
parameters (static attributes) were appended as add-on attributes (making the total
number of parameters p = 6), the accuracy of the model is significantly decreased.
This suggested that the features should be directly co-related with the students’
academic features.

Table 13.11. Performance comparisons of attribute precedence relations of fitness.

Experiment_ID Dataset Modeling attributes Accuracy (%) Accuracy (%)

Model type Training Test Proposed FE model Comparison Comparison

tuple tuple with with
count count RELIEF RELIEF
weights weights
(normalized) (prior-
probabilities)
Four-attribute 87 20 Attendance, assignment 82% 83%
model credit, internal score,
subject count
Six-attribute 87 20 Attendance, assignment 60% 58.33%
model credit, internal score,
subject count,
laboratory credit,
previous year
percentage
Nine-attribute 87 20 Attendance, assignment 39.44% 45%
model credit, internal score,
subject count,
laboratory credit,
previous year
percentage, score in
higher secondary,
living location,
medium of study

13-22
 Modern Optimization Methods for Science, Engineering and Technology

The above processing steps on feature-vector ranking were still deviated further
by computing an optimal set of fitness precedence relations from both the
precedence relation sets obtained due to average probabilities of fitness and unfit-
ness. Initially the attribute precedence relation of unfitness is converted into
equivalent attribute precedence relations of fitness by simply reversing the increasing
order of attributes so that the least unfit attributes becomes the most fit in equivalent
fitness precedence and vice versa. These two sets of fitness precedence relations were
compared in order to identify a consistent position j, defined either as exact
corresponding position j occupied by attribute xi in both relations or, at the most,
occupying either of the adjoining position combinations such as ( j, j + 1) or ( j − 1, j).
In this way, the attributes find their final positions of optimized fitness precedence
owing to the heuristics applied over their consistent valid and conflicting positions.
Having observed increased model accuracies in the latter set of experiments, it
could be concluded that the feature-vector ranking becomes more robust for the
Bayesian driven hybrid FE approach, if optimized attribute precedence relations of
fitness are used instead of conventional attribute precedence relations of fitness
(table 13.12).

13.4.2 Feature extraction via ensemble pruning: a case study

Ensemble learning methods are gaining popularity among the ML and data mining
communities. Ensemble learning models are widely accepted to make better
predictions than individual classifiers, given the same amount of training informa-
tion, for example the bagging, boosting, arcing and random forest algorithms. The
effectiveness of the ensemble methods relies on creating a collection of diverse, yet
accurate learning models. Two costs, namely the huge requirements of memory and
computational time, are the great disadvantages when implementing a large-scale
ensemble learning on real-world datasets that are prone to easily generate an
ensemble with thousands of learning models. In addition, the larger the size of an
ensemble, the more prone it is to overfitting problems in an attempt to minimize the
classification error to zero. Hence, the complexity of such ensemble learning models
is eventually corrected by incorporating pruning over ensembles, such as KL-
divergence pruning and kappa pruning over an Adaboost model, kappa-error
convex hull pruning, back-fitting pruning and many more. All these pruning
methods suffered from greedy search criteria, which they adopted in order to
optimize the selected subset of dataset feature vectors and thus improve the
performance on the lower end.
Unlike previous heuristic approaches, it was found that ensemble pruning looks
for a subset of classifiers that has the optimal accuracy–diversity trade-off [17].
Ensemble pruning can be viewed more or less similarly to weight-based ensemble
optimization that aims to improve the learning accuracy of the ensemble by tuning
the weight on each ensemble member, but without compromising its performance.
The individual accuracy and pairwise independence of classifiers in an ensemble is
often referred to as the strength and divergence of the ensemble. The generalization
_ _
error rule of an ensemble states that the error is loosely bounded by ρ2 , where ρ is
S

13-23
 Modern Optimization Methods for Science, Engineering and Technology

Table 13.12. Performance comparisons of optimal attribute precedence relations of fitness.

Experiment_ID Data set Modeling attributes Accuracy (%) Accuracy (%)

Model type Training Test Proposed FE model Comparison Comparison

tuple tuple with with
count count RELIEF RELIEF
weights weights
(normalized) (prior-
probabilities)
Four-attribute 87 20 Attendance, 83% 84%
model assignment credit,
internal score,
subject count
Six-attribute 87 20 Attendance, 63.33% 64.17%
model assignment credit,
internal score,
subject count,
laboratory credit,
previous year
percentage
Nine-attribute 87 20 Attendance, 39.44% 44.4%
model assignment credit,
internal score,
subject count,
laboratory credit,
previous year
percentage, score in
higher secondary,
living location,
medium of study

the average correlation between classifiers and s is the overall strength of the
classifiers. Thus, the ensemble pruning dealt with in the case study initially
formulates the ensemble error function by computing strength and diversity
measurements for a classification ensemble followed by minimizing this approximate
ensemble error function through a quadratic integer programming formulation.
At first, this ensemble pruning method was tested on twenty-four UCI repository
datasets with Adaboost as the ensemble generation technique and was reported to
perform favorably over two other metric-based ensemble pruning algorithms:
diversity-based pruning and Kappa pruning, picked as the benchmarks. This was
followed by testing the same subset selection procedure in the form of a cross-
domain classifier-sharing strategy on a publicly available marketing dataset—here it
was a catalog marketing dataset from the direct marketing association (DMEF
academic dataset three, specialty catalog company, Code 03DMEF)—to select a

13-24
 Modern Optimization Methods for Science, Engineering and Technology

good subset of classifiers from the entire ensemble for each problem domain. The
essence of sharing classifiers is sharing common knowledge among different but
closely related problem domains. In that study, classifiers trained from different but
closely related problem domains are pooled together and then a subset of them is
selected and assigned to each problem domain. The computational results show that
the selected subset performs as well as, and sometimes better than, including all
elements of the ensemble. The reader is referred to Zhang et al [17] for the details of
the methodology in both benchmarking and empirical experiments.

13.4.3 Feature-vector ranking metrics

Assimilating the works on feature extraction as perceived by Wu et al [18], a few
more ranking methods were found that help in the computation of feature relevance
levels, and these are described in the set of expressions 13.22–13.28. One of the
simplest metrics is Pearson’s correlation coefficient, meant for detecting linear
correlation ranking between variables and target class labels:
cov(xi ,Y )
R(i ) = , (13.22)
var(Xi ) ∗ var(Y )

where xi is the ith variable, Y is the output (class labels), cov() is the covariance and
var() is the variance. Correlation ranking can only detect linear dependencies
between the variable and target. Another feature selection metric measures the
dependence of two variables and is based on information entropy of class label c,
which is defined as
H (Y ) = −∑ p(y )log(p(y )), (13.23)
y

while the conditional entropy of class label c, given attribute xj, can be defined as
H ( Y ∣X ) = − ∑ ∑y p(x, y )log(p(y∣x)), (13.24)
x

so that the mutual information (MI) can be computed as

I (Y , X ) = H (Y ) − H (Y ∣X ). (13.25)
It can be observed that the uncertainty in output class c is reduced, constrained by
the inclusion of the attribute xj. The simple interpretation of the MI metric is that xj
and c will not be correlated if MI(xj, c) = 0 otherwise there exists some correlation
between them. In other words, the greater the MI (I(.)) value, the more the feature xj
contributes to classifying the output class label c.
Mutual information is an important concept to be used in embedded methods of
feature extraction. We can incorporate both backward elimination and forward
selection approaches to resolve the feature ranking based on conditional mutual
information computation in successive iterations, which needs to be maximized for
obtaining optimal feature selection that takes care of a trade-off between interde-
pendence and discrimination.

13-25
 Modern Optimization Methods for Science, Engineering and Technology

Having invaded through insights of feature selection methods based on relevance

and feature interaction of samples, another important aspect of the feature-
extraction method is to evaluate how well that method helps in resolving the
induction (learning) model. For this, we need to understand the baseline objective of
any learning model in terms of characterizing the relationship between F and C. For
a feature-extraction model defined as a function f(S, F, C), where S = {s1, s2, …, sn},
F = { F1 X F2 X … X Fn}, C = { C1, C2, …, Ck}. Assuming that Fs1 is the subset of
already-selected features, Fs2 is the subset of unselected features, and F = Fs1 ∪ Fs2,
Fs1 ∩ Fs2 = ϕ, any optimal feature subset obtained by selection algorithms should
preserve the existing relationship between F and C hidden in the dataset. The most
appropriate evaluation metrics that extract the best of the feature subset are
classification accuracy and F-score, that also tend to preserve the above relationship
between F and C [18].

Definition The classification accuracy can be formulated as the function ‘classify

(s)’, which returns the classification accuracy rate of si ∣ si ∈ S, S is the set of data
items to be classified and sc is the class of the item, si . Mathematically, this can be
expressed as
∣s∣
∑i=0 ass(si ) (13.26)
acc(s ) = , si ∈ S ,
∣s ∣

⎧1, classify(s ) = sc
ass(s ) = ⎨ , (13.27)
⎩ 0, otherwise

where ∣S∣ represents the number of the elements in the collection S, si ∈ S.

Definition Another metric of feature discrimination is the F-score. The larger the F-
score is, the more this feature is discriminative. Given training vectors Xk if the
number of the jth dataset is nj, then the F-score of the ith feature is defined as
m _
∑ j=1( xi,j − xi )2
F (si ) = m nj 2
, (13.28)
∑ (1/(nj + 1))∑ xik,j − xi,j
j =1
(
k =1
)
_
where xi , xi,j are the average of the ith feature of the whole dataset and the jth
dataset, respectively; xik,j is the ith feature of the kth instance in the jth dataset and m
is the number of datasets, k = 1, 2, …, m and j = 1, 2, …l.

References
[1] Rao S S 2009 Engineering Optimization: Theory and Practice (New York: Wiley)
[2] Trevisan L 2011 Combinatorial optimization: exact and approximate algorithms CS261:
Optimization and Algorithmic Paradigms Lecture Notes (Stanford, CA: Stanford University)

13-26
 Modern Optimization Methods for Science, Engineering and Technology

[3] Collette Y and Siarry P Multi-Objective Optimization: Principles and Case Studies (New
York: Springer)
[4] Thomas A, Sharma H R and Sharma S 2014 A Tool Design of Subjective Question Answering
Using Text Mining (Saarbrücken: Lambert Academic)
[5] Fourment L, Duclocix R, Marie S, Ejday M, Monnereau D, Masse T and Montmitonnet P
2010 Mono and multiobjective optimization techniques applied to a large range of industrial
test cases using metamodel assisted evolutionary algorithms AIP Int. Conf. Proc. 1252 833–40
[6] Bennet K P and P-Hernandz E 2006 The interplay of optimization and machine learning
research J. Machine Learn. Res. 7 1265–81
[7] Sener O and Koltum V 2018 Multi-task learning as multi objective optimization Proc. of the
32nd Conf. on Neural Information Processing Systems (NeurIPS) (Montreal, Canada) pp
1–15
[8] Niculescu R S, Mitchell T M and Rao R B 2006 Bayesian network learning with parameter
constraints J. Machine Learn. Res. 7 1357–83
[9] Soni S and Vyas O P 2011 Performance evaluation of weighted associative classifier in health
care data mining and building fuzzy weighted associative classifier International Conference
on Parallel Distributed Computing Technologies and Applications Communications in
Computer and Information Science vol 203 (Berlin: Springer), pp 224–37
[10] Kadar J A, Agustono D and Napitupala D 2018 Optimization of candidate selection using
naïve Bayes: case study in company X J. Phys: Conf. Ser. 954 012028
[11] Shylo O and Shams H 2018 Boosting Binary Optimization Via Binary Classification: A Case
Study of Job Shop Scheduling arXiv: 1808.10813v1
[12] Avrim L and Langley P 1997 Selection of Relevant Features and Examples in Machine
Learning (Amsterdam: Elsevier), pp 245–71
[13] Kira K and Rendell L 1992 The feature selection problem: traditional methods and a new
algorithm 10th National Conference on Artificial Intelligence (San Francisco: Morgan
Kaufmann), pp 128–34
[14] Almuallim H and Dietterich T G 1991 Learning with many irrelevant features 9th National
Conference on Artificial Intelligence (Cambridge, MA: MIT Press), pp 547–52
[15] Sun Y and Wu D 2008 A RELIEF based feature extraction algorithm Conf. Proc. of the
SIAM Int. Conf. on Data Mining (Atlanta, GA) pp 188–95
[16] Singh M 2017 Prediction of academic performance of students using machine learning
techniques Doctoral thesis Dr C V Raman University, Kota Bilaspur, Chhattisgarh, India
[17] Zhang Y, Burer S and Street W N 2006 Ensemble pruning via semi-definite programming
J. Mach Learn. Res. 7 1315–38
[18] Wu S, Hu Y, Wang W, Feng X and Shu W 2013 Application of global optimization methods
for feature selection and machine learning Math. Prob. Eng. 2013 241517
[19] Chandrashekar G and Sahin F 2014 A survey on feature selection methods Comput. Electr.
Eng. 40 16–28
[20] Dietterich T G 2000 Ensemble methods in machine learning ed J Kittler and F Roli MCS
2000, LNCS 1857 (Berlin: Springer), pp 1–15
[21] Koller D and Sahani M 1996 Toward optimal feature selection Conf. Proc. of Machine
Learning pp 1–14
[22] Guyan I 2003 An Introduction to variable and feature selection J. Mach. Learn. Res. 3
1157–82

13-27
 IOP Publishing

Modern Optimization Methods for Science, Engineering and

Technology
G R Sinha

Chapter 14
Simulation of the formation process of spatial
fine structures in environmental safety
management systems and optimization of the
parameters of dispersive devices
Sergij Vambol, Viola Vambol, Nadeem Ahmad Khan, Kostiantyn Tkachuk,
Oksana Tverda and Sirajuddin Ahmed

The topical scientific applied issue of the creation of control systems for ecological
safety through the use of dispersive devices is considered. For suppression of the
formation processes of toxic substances and limitation of their distribution in the
atmosphere during extraction, processing and transportation of bulk materials
(which produce dust) and during fire suppression and thermal waste treatment, an
analysis of the systems that use spatial fine structures is offered. The results of
numerical simulation of these ecologically hazardous technological processes are
described. The physical model of controlling such processes is based on injecting a
cooling liquid using dispersive devices. Mathematical models of the gas and
dispersed phases of spatial fine structures are developed. The mathematical
formulation of the conservation laws for viscous gas (steam) is achieved through
the Navier–Stokes equations; for drops, it is given as an equation of the balance of
forces that affect the drop and equalize the inertia force and the resultant forces
of gravity and aerodynamic resistance. For dispersion of the fluid, irrigation systems
of the nozzle type, atomizers and centrifugal atomizer have been suggested. The
dependence of the ability to create effective fine structures on the characteristics of
the technical devices in the context of natural and man-made hazards of different
origins is presented. Using the numerical simulation of the formation processes of
spatial fine structures, the most efficient modes of supplying liquid to various
hazardous factors are defined.

doi:10.1088/978-0-7503-2404-5ch14 14-1 ª IOP Publishing Ltd 2020

 Modern Optimization Methods for Science, Engineering and Technology

14.1 The use of spatial finely dispersed multiphase structures in

ensuring ecological and technogenic safety
By spatial finely dispersed multiphase structures, we understand to mean structures
consisting of finely dispersed drops of fluid, air and particles (such as dust particles,
chemical components and biodestructors), which are present in working areas and
can exist there for a given time. The simplest structure is a finely dispersed water
suspension, that is, a stream of sprayed water, which consists of small droplets and is
able to partially or completely separate one part of space from another (figure 14.1).
Dispersed water suspensions are widely used in modern stationary systems for
ensuring technogenic safety, in fire protection systems, and are also one of the means
of protecting people and equipment during emergency response situations. Creating
a system of water suspension from neutralizing solutions is the main way to limit the
spread and neutralization of the clouds formed when chemically hazardous substances
are released into the atmosphere (such as ammonia, chlorine, nitrogen oxides, sulfur
dioxide, hydrogen chloride and fluoride, ethylene oxide, phosgene, etc).
To create a finely dispersed water suspension, sprayers of various designs are
used: slotted sprayers, spray guns (such as fire-hoses, figure 14.2(a)), bag sprayers,
combined installations (mobile dispersed structures) and drencher sprinklers (sta-
tionary dispersed structures, figure 14.2(b)).

14.1.1 Analysis of recent research and publications

In the process of loading and unloading loose dusty materials, in the destruction of
rocks, etc, a significant amount of finely dispersed (soaring) particles (including
nanoparticles) enter the atmospheric air and adversely affect natural components
and human health [1–5]. This applies to different countries where mining facilities
exist, and the importance of the mining sector is steadily increasing [6, 7]. During the
storage and transportation of coal after the grinding operation, the release of coal
dust also takes place. To reduce such emissions, various emulsions are used based on
synthetic polymeric materials, pulp and paper waste, and refined products and waste
[8]. However, this method complicates the further use of the coal.
Emergencies in waste storage facilities [9, 10], at industrial enterprises [11, 12] and
other technogenic objects are accompanied by high-intensity energy release and the
formation of environmentally hazardous molecular compounds [13–15]. At the same

Figure 14.1. Creation of a finely dispersed water suspension by an atomizer.

14-2
 Modern Optimization Methods for Science, Engineering and Technology

Figure 14.2. Schemes of dispersed structures (finely dispersed water suspension).

time, such emergency situations are characterized by the entry into the atmosphere
of a significant amount of carbon monoxide, carbon dioxide, soot, etc, but no special
measures are taken to localize (prevent the spread of) these substances [16–18].
The use of thermal methods for the disposal of solid waste products improves the
environment by reducing the number of landfills. However, the process of thermal
utilization itself is accompanied by harmful emissions into the environment [19, 20].
This process can be ecologically effective in the case of preventing the formation of
highly toxic substances (such as dioxins and furans) at the stage of thermo-chemical
treatment of waste. The study of the mechanism of the formation of dioxins during
heat treatment of waste was given much attention in [21, 22], and in [23] the authors
proposed a purification system, scientifically substantiated its effectiveness and
confirmed this experimentally. It should be emphasized that it is difficult to
implement such a system for economic reasons.
The most modern and promising methods of waste disposal are based on the use
of high-temperature treatment [19, 24]. This is due to the fact that under conditions
of high temperature (1200 °C and above), dioxins and other highly toxic substances
decompose into simple fragments [20]. However, the mechanisms of dioxin re-
formation have also been investigated. Re-formation of dioxins during cooling of
the high-temperature multicomponent gas is observed if the gas temperature is from
450 °С to 300 °С. The process of their formation is affected by the presence of
chlorine and oxygen as well as the rate of gas cooling [20]. This fact gives us reason
to believe that the waste recycling process can be more environmentally friendly.

14-3
 Modern Optimization Methods for Science, Engineering and Technology

This is possible if we exclude the conditions that contribute to the formation of

dioxins after the gas leaves the plasma reactor. Thus, the search for technological
solutions that increase the level of ecological safety of solid waste thermal utilization
is a very important scientific task. Scientists from different countries have already
proposed various methods for cleaning the hot flue gases of various industrial
processes. However, as a result of this cleaning, a new problem arises, for example,
toxic filtration materials, other hazardous chemical compounds, etc. Since dioxins
are not formed at temperatures below 300 °С, the most logical solution in this
situation is the instantaneous cooling of the flue gases to a temperature below 300 °С.
In [25, 26], the authors considered mathematical models and performed research on
the creation of dispersed structures. It should be noted that in solving the problems
listed above, specific technical conditions and requirements were considered. Based on
the classical equations of fluid dynamics, there currently already exists a general
approach for the physical and mathematical formulation of this class of problems.

14.1.2 Statement of the problem and its solution

Creating a finely dispersed water suspension from neutralizing solutions is a good
approach to suppress the formation processes and limit the spread of toxic
substances and (or) neutralize them. For this purpose sprayers of various designs
are used.
At present, theoretically justified methods for calculating the regimes of for-
mation of dispersed water structures are needed. This will allow determining their
main parameters through calculations: geometrical dimensions in various condi-
tions, spatial distribution of droplet concentration, interaction with air and heat
flows, influence of technical parameters and methods of optimization.
The experience of creating a finely dispersed water suspension shows that there is
a need for in-depth study of the delivery processes of dispersed fluids and their
further sedimentation to design the necessary spatial structure of the water aerosol
and to develop effective design solutions on this basis. The possibility of using
sprayers in the development and creation of an ecological friendly process of thermal
utilization is also promising.
By virtue of economy and convenience in the modern scientific world, numerical
experiments are more often used for a detailed analysis of any complex process. To
study the process of creating a dispersed water suspension in order to eliminate
sources of environmental hazards, a numerical experiment is also the most
convenient tool. This chapter presents the methodology and results of a numerical
experiment of the possibility of delivering dispersed water for a required distance
and height for dust suppression. At the same time, the effectiveness of dust
suppression is investigated depending on the modes of water supply by the atomizer.
It also presents the results of numerical modeling to create a dispersed water
suspension for cooling the gas to prevent the formation of dioxins, to allow thermal
utilization to be an ecologically friendly technological process.

14-4
 Modern Optimization Methods for Science, Engineering and Technology

The object of this research is finely dispersed multiphase structures. The subject of
the research is the dependence of the parameters of ecologically effective finely
dispersed water structures on the technical features of the devices used for their
creation.

14.2 Physical and mathematical simulation of the creation process of

spatial finely dispersed structures
The most universal theoretical description of the creation process of spatial finely
dispersed structures is based on the laws of conservation of mass and momentum in
a medium that is not uniform in phase composition, including atmospheric air and
water droplets. These equations with given boundary conditions can be jointly
solved by modern numerical methods, which are separated into an independent
branch of knowledge—computational aero-hydrodynamics. In the mathematical
description of the dispersed structure, the following assumptions were made:
• an incompressible, isothermal and turbulent carrier gas phase;
• isotropic turbulence;
• spherical, non-evaporating drops;
• the volume occupied by the drops is neglected.

The interaction of the phases was taken into account as a ‘drop—source in a cell’
model. In accordance with this model, the presence of particles in the flow manifests
itself through an additional source of momentum in Reynolds-averaged Navier–
Stokes equations, which are closed by a semi-empirical model of k-ε–type
turbulence.
Thus, to fully describe the creation process of spatial fine-dispersed structures, it is
necessary to consider models of the gas phase, the dispersed phase and the transition
process—the model of interfacial interaction.

14.2.1 Gas phase study and mathematical model description

The mathematical model of a three-dimensional quasistationary turbulent gas flow
in a working area is based on the system of Navier–Stokes equations averaged
according to Reynolds. The equations of conservation of mass and momentum in
the vector record form are [27]
∇ · (ρu ⃗ ) = 0, (14.1)

∇ · (ρuu⃗ ⃗ ) = −∇p + ∇ · (τ ) + Sf , (14.2)

where ρ is density, u ⃗ is the speed vector, p is static pressure, Sf is the source of
momentum due to interfacial interaction, and τ is the tension tensor, determined by
the expression
τ = (μ + μt )(∇u ⃗ + ∇u T⃗ ),
where μ is molecular viscosity and μt is turbulent viscosity.

14-5
 Modern Optimization Methods for Science, Engineering and Technology

For the closure of the system of equations averaged according to Reynolds (14.1),
(14.2), the Lauder–Spalding turbulence k-ε–model was used [28]. The transport
equations for the kinetic energy of turbulence k and its dissipation rate ε are

∂ ∂ ⎡⎛ μ ⎞ ∂k ⎤
(ρkui ) = ⎢⎜μ + t ⎟ ⎥ + G − ρε , (14.3)
∂xi ∂xj ⎣⎝ σ k ⎠ ∂xj ⎦

∂ ∂ ⎡⎛ μ ⎞ ∂ε ⎤ ε ε2
(ρεui ) = ⎢⎜μ + t ⎟ ⎥ + C1ε G − C2ερ , (14.4)
∂xi ∂xj ⎣⎝ σε ⎠ ∂xj ⎦ k k

where ρ is density, k is the kinetic energy of turbulence, ui is the the projection of the
averaged gas velocity on the axis of a three-dimensional rectangular Cartesian
coordinate system, xj are the coordinates of a three-dimensional rectangular
Cartesian coordinate system, ε is the turbulence kinetic energy dissipation rate
and σk , σε , C1ε , C2ε are empirical coefficients. G is the term characterizing the
emergence of the kinetic energy of gustiness due to shear stresses and is defined by
the formula
∂uj
G = −ρu i′u j′ .
∂xi
Turbulent viscosity is determined by the Kolmogorov–Prandtl formula [29]
k2
μt = Cμρ , (14.5)
ε
where Cμ is the empirical coefficient.
To determine Sf in equation (14.2) an interfacial interaction model was used.

The border conditions

The system of partial differential equations (PDEs) (14.1)–(14.4) was complemented
by the boundary conditions for independent variables.
At the boundaries of the computational domain, the following boundary
conditions for the continuous phase were set:
• At the outlet of the atomizer nozzles: air speed and wind speed corresponding
to the atmosphere.
• At the remaining boundaries: wind speed corresponding to the atmosphere,
and static pressure.
• On solid surfaces: adhesion condition, approximated by the wall function.

Direct application of the adhesion condition is necessary—modification of the

turbulence model in the near-wall region, where the turbulent viscosity is close to
molecular and significant refinement of the computational grid near the wall. The
experience of numerical simulation of three-dimensional flows shows that the
complexity of the geometric form of the computational domain often leads to
the fact that it is the required dimension of the computational grid, which becomes

14-6
 Modern Optimization Methods for Science, Engineering and Technology

the critical parameter, which determines the possibility of conducting numerical

experiments. Therefore, instead of the sticking condition, the wall functions were
used to describe the turbulent boundary layer—a system of semi-empirical functions
for independent variables in the center of the near-wall computational cell (point P)
and variables on the wall that correspond to these values. Moreover, these functions
are based on the assumption of Lunder and Spalding [30, 31]. The law of the wall for
the averaged velocity is
⎧ y* when y* ⩽11.225,
⎪
U* = ⎨ 1 (14.6)
⎪ ln(Ey*) when y* > 11.225,
⎩K

where K is Karman’s constant, Е is the empirical constant, U* and y* are

dimensionless parameters, which are determined by the formulas
UPC μ1/4k P1/2
U* = , (14.7)
τw / ρ

ρC μ1/4k P1/2yP
y* = , (14.8)
μ
where UP is the average velocity of gas at a point Р, kP is the kinetic energy of
gustiness at point Р, τw is friction tension on the wall, ρ is gas density, yP is the
distance between point P and the wall, and μ is dynamic viscosity.
Next, it is necessary to solve equation (14.4) in the near-wall cells and the entire
computational domain. At the same time the boundary condition on the wall was set
for k, which is written as
∂k
= 0, (14.9)
∂n
where n is the local coordinate normal to the wall.
The turbulence kinetic energy generation G and its dissipation rate ε, which are
included in the source term of equation (14.4), are calculated in the near-wall cells
based on the local equipoise hypothesis. With this assumption we believe that the
emergence of turbulence kinetic energy and its dissipation rate in the near-wall
control volume are the same. Then equation (14.5) for ε in the near-wall cells is not
solved; instead, the turbulence kinetic energy dissipation rate is determined by the
formula
C μ3/4k P3/2
εP = , (14.10)
κyP

where κ is the empirical constant.

14-7
 Modern Optimization Methods for Science, Engineering and Technology

14.2.2 Dispersed phase study and mathematical model description

Consider the movement of a two-phase flow. One phase is air and the second phase
is polydisperse water aerosol. Since the volume concentration of water droplets in
the flow is small, we can consider the motion of noninteracting droplets of various
sizes separately. Regarding the dispersed phase, we accept the following basic
assumptions:
• The dispersed phase is a completely atomized fluid (water) spray, consisting
of a finite set of evaporating spherical droplets of various diameters.
• All forces acting on the drop, except for the forces of gravity and aerody-
namic drag, are neglected.
• The processes of secondary crushing and coagulation of droplets are
neglected.

Under the above assumptions, the behavior of the dispersed phase (water
droplets) is conveniently considered in the Lagrangian description. For sprayed
fluids, the Rosin–Rammler expression is the generally accepted droplet size
distribution [32]. The entire range of initial sizes of droplets was divided into
detached intervals; each of them is represented by the average initial diameter for
which the trajectory calculation is performed. In addition, each simulated drop is
actually a ‘package’ of drops with identical trajectories. The Rosin–Rammler
equation describes the distribution of droplets in size, and the droplets’ mass
fraction with a diameter greater than d is described by the formula
n
Yd = e−(d /d ) , (14.11)

where d is the average median diameter of droplets in the spray and n is a

distribution parameter.
The motion of a particle of the dispersed phase (droplets) is described by
Newton’s second law. Under the above assumptions, the equation of motion for a
single drop in vector form is

dup  
mp = G + FR, (14.12)
dt
 
where mp is drop mass, up is drop rate, t is time and G is gravity, determined by
 
G = mpg , (14.13)
 
where g is the acceleration of gravity and FR is the aerodynamic drag force
determined by the expression
    
FR = 0, 5CRAp ρ∣u −u p∣(u −u p), (14.14)

where СR is the coefficient aerodynamic drag resistance, Ар is the middle sectional


area of the drop, ρ is gas density and u is gas velocity.

14-8
 Modern Optimization Methods for Science, Engineering and Technology

Substituting the expressions (14.13) and (14.14) into equation (14.12) and taking
into account that the mass of a spherical drop and the area of its average cross
section are determined by the formulas
πd p3
mp = ρp , (14.15)
6

Ap = 0.25πd p2, (14.16)

and having projected the vectors of both parts of equation (14.12) on the axis of the
fixed Cartesian coordinate system, we will have the motion equation system of the
drop in the form
1
πd p3 ⎡ ⎤2
dupj 3ρCR ⎢ 2⎥
= ρp g − (upj − uj ) ∑(upj − uj ) , (14.17)
dt 6 j 4ρp dp ⎢⎣ ⎥⎦
j

where j = 1, 2, 3.
To calculate the trajectory of a drop, we supplement the system (14.17) with the
following equation
dxpj
= upj , (14.18)
dt
where j = 1, 2, 3, хрj are the Cartesian coordinates of a drop.
At values of droplet speeds, when compressibility can be neglected, the coefficient
of aerodynamic resistance CR of a spherical droplet is an unambiguous function of
the relative Reynolds number and can be calculated from the formula
 
ρdp u − up
Rep = , (14.19)
μ
where dp is the droplet diameter and μ is the dynamic viscosity of gas.
For approximating the dependency, the empirical Zhen–Trizek formula is
used [33]
24 6
CR = + + 0.27. (14.20)
Rep 1 + Rep

Thus, in a known gas-dynamic field the changing of the parameters of a single

non-evaporating drop are described by ordinary differential equation (ODE)
systems (14.17) and (14.18). In this system, the desired functions are the projections
of the absolute velocity of the drop upj and its coordinates xpj. The remaining values
serve as parameters. Some of these parameters are functionally related to independ-
ent variables. Therefore, to close the system, it must be supplemented with algebraic
equations (14.11), (14.19) and (14.20).

14-9
 Modern Optimization Methods for Science, Engineering and Technology

The initial conditions

For the numerical solution of the system of ODEs (14.17) and (14.18), it is necessary
to specify the initial conditions, that is, the values of independent variables at the
initial time t = 0: upj0 and xpj0.
When simulating irrigation by a fire-hose with a sprayer, it is assumed that the
drops start from the point of collapse of a compact jet located at a distance from the
nozzle
r = ψhdrop , (14.21)
where ψ is the ratio of the height of raising individual drops hdrop to the height of
raising a compact jet hjet, determined by the table 20.10 from [34]; hdrop is the height
of raising individual drops, determined by the formula [34]
h
hdrop = , (14.22)
1 + ψ1h
where h is the full head at the start of the sprayer and ψ1 was the empirical coefficient
depending on the shape of the sprayer [34].
The initial velocity of the droplets was set equal to the local velocity of the fluid in
a compact jet
up0 = v = ϕ 2gh , (14.23)

where φ is the speed coefficient (φ = 0.95).

The direction of the drops’ initial velocity vector was determined by the value of
the first derivative of the equation of the trajectory of the compact jet flowing from
the nozzle at the point of its decay,
gx 2
y = xtg Θ − , (14.24)
2v cos2 Θ
2

where Θ is the angle of inclination of the nozzle to the horizon, and х and y are the
horizontal and vertical Cartesian coordinates, respectively.
When simulating atomizer irrigation, the coordinates of the points of emission of
drops xpj0 were chosen at the centers of the faces of the computational cells
belonging to the output section of the atomizer. It has been assumed that the drops’
initial velocity and the local gas velocity in the outlet section of the atomizer are
equal: upj0 = uj.
In all cases, the diameters of droplets dp were specified by a histogram of the
initial distribution of droplets by size, constructed using equation (14.11).

14.2.3 Mathematical model of interfacial interaction

There is interaction between the phases. This interaction is taken into account using
the discrete model ‘particle—source in a cell’. In accordance with this model, the
drop availability in the flow was taken into account as an additional component in
the equation for the conservation of the momentum of a continuous phase [35].

14-10
 Modern Optimization Methods for Science, Engineering and Technology

Figure 14.3. Interfacial interaction scheme.

During the calculation of the trajectories of droplets, the amount of movement

acquired or lost by the ‘package’ of droplets following along a given trajectory is
tracked. The resulting values should then be taken into account as the source term
for calculation of the gas phase Sfi in equation (14.2). Thus, since the gas phase
already affects the dispersed phase through equations (14.17) and (14.18), the
response impact of the dispersed phase into continuum is also counted. That is, by
alternately solving the equations of the dispersed phase and the continuous phase
until the solution of both phases is established, this two-way interaction will be taken
into consideration. A diagram of such a model of interfacial interaction is shown in
figure 14.3.
The transfer of the amount of movement from the continuous phase to the
dispersed phase is calculated by estimating the change in the droplet movement
amount as it passes through each control volume of a study area geometric model.
This change in the amount of movement is calculated as
⎛ 2μC Re ⎞
ΔSfi = ∑⎜⎜ R
2
p
(upi − ui )ṁ pΔt⎟⎟ , (14.25)
⎝ 3ρp d p ⎠

where ṁ p is the drops’ mass flow and Δt is the time step.

14.3 Numerical simulation of the formation of spatial dispersed

structures and the determination of the most effective ways of
supplying fluid to eliminate various hazards
Initially, it is necessary to ensure the numerical solution’s stability, convergence and
accuracy.

14.3.1 Ensuring numerical solution stability, convergence and accuracy

The numerical solution stability can be ensured by using the method of lower
relaxation of independent variables.
The number of iterations required for solving stationary problems of aero-
hydrodynamics is determined by both the algorithm of the difference scheme and
the criterion for evaluating the convergence of the solution. The criteria used by
researchers for convergence of the solution can be classified as follows:

14-11
 Modern Optimization Methods for Science, Engineering and Technology

1. Definition of the discrepancy for each solved differential equation. In this case,
as a rule, to achieve the convergence of the entire solution, it is necessary for
each difference equation to provide the specified level of the residual.
2. Integral discrepancy. In this case, a uniform criterion is determined for all
equations, which allows analyzing the convergence of the solution.

To assess convergence, applying an integral criterion with respect to the vector of

conservative variables is proposed.
The condition of convergence of the solution can be represented as

∑Ri2 (14.26)
⩽ ε,
V2
where Ri is the error for difference equations modeling the transfer of independent
variables, V is th volume of control space and ε is the convergence criterion.
A numerical solution can be considered converged if one of the following
conditions is met:
• Condition (14.26) is satisfied.
• The solution no longer changes with the continuation of the iterations.

As an alternative criterion for the convergence of problem solving, the pulsation

of the mass flow ratio in the input section to the output section can be taken.
The accuracy of the numerical solution can be assessed by the degree to which the
convergence criteria are met, the degree of independence of the decision on the size
of the computational grid, and the degree of qualitative correspondence of the
calculation results to known physical concepts. Solution independence can be
estimated by numerical results comparing obtained on several computational grids
differing in the number of computational cells. Thus, using this computational
method, at any point of the computational domain one can determine the gas
parameters.

14.3.2 Description of the numerical integration method of the dispersed phase

equations
The numerical integration of the ODE system (14.17) and (14.18) with the given
initial conditions was performed by the fourth order Runge–Kutta method with a
variable step chosen by the formula
τ
Δt = a , (14.27)
χ
where χ is the the coefficient of ‘grinding’, given depending on the desired accuracy
of integration, and τa is the aerodynamic relaxation time determined by the
expression

14-12
 Modern Optimization Methods for Science, Engineering and Technology

ρp d p2
τa = , (14.28)
18μ
where dp is the droplet diameter, μ is the dynamic viscosity of gas and ρ is the drop
density.
For the numerical simulation of the specific situation of eliminating the dust
cloud, we assume:
1. Bulk material taken from the center of the pile.
2. When material is collected, a spherical dust cloud with a radius of 2 m is
formed.
3. Water is delivered to the dust cloud by an atomizer, which can be installed in
three positions:
• lower position (at ground level, that is, 6 m below the level of backfill).
• average position (6 m above ground level, that is, at the level of
backfill).
• top position (12 m above ground level, that is, 6 m above the level of
backfill).

When simulating irrigation of zones of occurrence of natural or man-made

hazards of up to several tens of meters (for example, when a dust cloud is formed
during loading or unloading of dusty bulk materials) in conditions of calm,
headwind and tailwind, the calculated area covered a fragment of space with,
located in the middle of it, a stack of bulk material that on the x-axis had a length of
18 m and on the y-axis 6 m in opposite directions from the dust cloud center, with an
adjacent area of the atmosphere with a height of 24 m.
When simulating the hazard described above, the computational domain was
covered with a non-uniform computational grid that included 199 656 tetrahedral
cells. The boundaries of the dust cloud were modeled by a permeable spherical
surface with a center located at the top of the stack.
When simulating irrigation under cross-wind conditions, the length of the
computational domain was increased to 35 m (for the full length of the stack) in
the leeward direction from the center of the dust cloud. In this case, the computa-
tional domain included 1 187 702 tetrahedral cells.
Under different conditions of wind action, the boundaries of the dust cloud were
simulated by a permeable spherical surface with a center located at the top of the
coal pile.
In the following, the discretization algorithm for the governing equations (14.1)–
(14.4) was considered on the example of the generalized transfer equation of an
arbitrary scalar quantity written in integral form for an arbitrary control volume V
using the appropriate dependences.
The task of transporting finely dispersed drops of the fluid to the dust-release zone
using single-phase jet-centrifugal atomizers was solved in [36] with the participation
of the authors of this work. The same software product was used to solve the
problem of transporting process fluid droplets using two-phase nebulizers and

14-13
 Modern Optimization Methods for Science, Engineering and Technology

atomizers [37]. Using this approach, it is possible to present options for transporting
fine droplets to the zone of occurrence of natural or man-made hazards using the
example of suppressing dust using various sprayers: single-phase jet-centrifugal, two-
phase sprayers and atomizers.

14.3.3 Results of numerical simulation of a spatial finely dispersed structure creation

process which suppresses dust
As an applied study of the problem of transporting processed fluid droplets to the
dust emission zone, it was proposed to consider and numerically simulate creating a
water dispersed structure with two different devices. The first chosen technology is
the fire-hose and the second device is the atomizer. As a criterion assessment for
comparing the effectiveness of these devices, the following parameters were chosen:
the total residence time of all the fluid droplets in the dust cloud and the mass
fraction of fluid droplets that did not fall into the dust cloud.

14.3.3.1 Numerically simulating creating a water dispersed structure above a dust

cloud using a fire-hose
Several variants of irrigation of a dust cloud with a fire-hose, which differed in the
delivery angle, full head, speed and wind direction, were investigated. In all the
variants, the following parameters of droplet size distribution were used:
• The minimum diameter of droplets in the spray is 100 μm.
• The maximum diameter of the droplets in the spray is 1500 μm.
• The average median diameter of the droplets in the spray is 1000 μm.
• The exponent in the Rosin–Rammler formula is 2.

The histogram of droplet size distribution corresponding to the above parameters

is shown in figure 14.4.

Figure 14.4. The histogram of the distribution of droplets by size with irrigation using a fire-hose.

14-14
 Modern Optimization Methods for Science, Engineering and Technology

As a result of the calculations, it was found that at a feed angle of 45°, the
compact jet reaches the surface of the coal pile without disintegrating. Therefore, for
further analysis, four variants with a feed angle of 45° were selected (table 14.1,
figures 14.5–14.10).
From figure 14.5 it can be seen that in variant no. 1 most of the drops, with the
exception of the smallest and largest, fall into the dust cloud. In this case, all the
drops are precipitated within the storage area.
From figure 14.6 it can be seen that in variant no. 2 all the drops fall into the dust
cloud. However, the smallest droplets are carried by the wind (see figure 14.7)
outside the storage area.
From figure 14.8 it can be seen that in variant no. 3, none of the droplets reach the
dust cloud—they all are blown back by the headwind (see figure 14.9). At the same
time, most drops are deposited within the storage area. The smallest drops are
carried by headwinds outside the storage area.

Table 14.1. Irrigation variants of the dust cloud using a fire-hose.

Variant Angle Pressure, m Wind speed, m s−1 Wind direction

1 45° 40 0 —
2 45° 25 3 tailwind
3 45° 40 3 headwind
4 45° 40 3 cross wind

Figure 14.5. Trajectories of ‘packages’ of drops, painted in accordance with their diameter, in meters (variant no. 1).

14-15
 Modern Optimization Methods for Science, Engineering and Technology

Figure 14.6. Trajectories of ‘packages’ of drops, painted in accordance with their diameter, in meters (variant no. 2).

Figure 14.7. Air velocity vectors in the plane of symmetry of the computational domain, painted in accordance
with the absolute value of the velocity, in m s−1 (variant no. 2).

14-16
 Modern Optimization Methods for Science, Engineering and Technology

Figure 14.8. Trajectories of ‘packages’ of drops, painted in accordance with their diameter, in meters (variant no. 3).

Figure 14.9. Air velocity vectors in the plane of symmetry of the computational domain, painted in accordance
with the absolute value of the velocity, m s−1 (variant no. 3).

14-17
 Modern Optimization Methods for Science, Engineering and Technology

Figure 14.10. Trajectories of ‘packages’ of drops, painted in accordance with their diameter, in meters (variant no. 4).

From figure 14.10 it can be seen that in variant no. 4 part of the representative
droplets reaches the dust cloud. At the same time, most drops are deposited within
the storage area. The smallest drops are carried by the side wind outside the
storage area.
According to the results of numerical simulations, it was found that in variant no.
1, the total residence time of all the drops in the dust cloud was 21.6 s. At the same
time, the mass fraction of droplets that did not fall into the dust cloud is 7.7%. In
variant no. 2, the total residence time of all the drops in the dust cloud was 23 s. At
the same time, the mass fraction of droplets that did not fall into the dust cloud is
1.1%. For variant no. 4, the total residence time of all drops in a dust cloud was 9.2 s.
At the same time, the mass fraction of droplets that did not fall into the dust cloud is
22.4%. The obtained data allow us to further consider the issue of optimizing the
irrigation of a dust cloud with a fire-hose.

14.3.3.2 Numerically simulating creating a water dispersed structure above the

dust cloud using an atomizer
As in the previous case, several options for the dust cloud irrigation by an atomizer
were investigated. The variable parameters selected are different angles of flow, the
speed of the blowing aerosol, and the speed and direction of the wind.
In all the variants, the following parameters of droplet size distribution were used:
• The minimum diameter of droplets in the spray is 60 μm.
• The maximum diameter of the droplets in the spray is 130 μm.

14-18
 Modern Optimization Methods for Science, Engineering and Technology

• The average median diameter of the droplets in the spray is 100 μm.
• The exponent in the Rosin–Rammler formula is 3.

The histogram of droplet size distribution, corresponding to the above param-

eters, is shown in figure 14.11.
For further analysis, 15 variants with feed angles of 45° and 60° were selected
(table 14.2, figures 14.12–14.18).
In variants no. 1 to no. 4, most of the droplets, with the exception of the largest
ones, fall into the dust cloud. At the same time most of the drops, with the exception
of the smallest, are deposited within the storage area.
For variant no. 5 (figure 14.12), a large area of irrigation of the storage area is
characteristic, with only a fraction of the droplets falling into the dust cloud.
From figure 14.13, it can be seen that in variant no. 6 only part of the droplets fall
into the dust cloud. In this case, all the drops are deposited within the storage area.
From figure 14.14, it can be seen that in variant no. 7 most of the droplets are
deposited on the windward slope of the pile, not reaching the dust cloud. Only the
smallest droplets fall into the dust cloud and are carried away by a tailwind outside
the storage area.
From figure 14.15, it can be seen that in variants no. 8 and no. 9 larger droplets
are deposited on the windward slope of the pile, not reaching the dust cloud. Smaller
droplets fall into the dust cloud, which are carried away by a tailwind outside the
storage area.
From figure 14.16, it can be seen that in variant no. 10 only the largest drops are
deposited on the windward slope of the pile, not reaching the dust cloud. All other
droplets fall into the dust cloud and are carried away by a tailwind outside the
storage area. An exception is made for medium-sized droplets which, being involved

Figure 14.11. The histogram of the distribution of droplets by size with irrigation using an atomizer.

14-19
 Modern Optimization Methods for Science, Engineering and Technology

Table 14.2. Irrigation variants of the dust cloud by atomizer.

Atomizer Aerosol Aerosol blowing Wind Wind

Variant location blowing angle speed, m s−1 speed, m s−1 direction

1 bottom 45° 10 0 —
2 middle 45° 10 0 —
3 top 45° 10 0 —
4 top 0° 10 0 —
5 bottom 45° 3 1 tailwind
6 middle 45° 3 1 tailwind
7 middle 45° 5 1 tailwind
8 middle 45° 10 1 headwind
9 middle 45° 20 1 headwind
10 top 0° 10 1 headwind
11 top 0° 20 1 headwind
12 bottom 60° 20 1 cross wind
13 middle 60° 20 1 cross wind
14 top 60° 20 1 cross wind
15 top 60° 20 1 cross wind

Figure 14.12. Trajectories of ‘packages’ of drops, painted in accordance with their diameter, in meters (variant no. 5).

14-20
 Modern Optimization Methods for Science, Engineering and Technology

Figure 14.13. Trajectories of ‘packages’ of drops, painted in accordance with their diameter, in meters (variant no. 6).

Figure 14.14. Trajectories of ‘packages’ of drops, painted in accordance with their diameter, in meters (variant no. 7).

14-21
 Modern Optimization Methods for Science, Engineering and Technology

Figure 14.15. Trajectories of ‘packages’ of drops, painted in accordance with their diameter, in meters (variant no. 8).

Figure 14.16. Trajectories of ‘packages’ of drops, painted according to their diameter, in meters (variant no. 10).

14-22
 Modern Optimization Methods for Science, Engineering and Technology

Figure 14.17. Trajectories of ‘packages’ of drops, colored according to their diameter, in meters (variant no. 11).

Figure 14.18. Trajectories of ‘packages’ of drops, colored according to their diameter, in meters (variant no. 12).

14-23
 Modern Optimization Methods for Science, Engineering and Technology

in the zone of reverse air currents that occur behind the leeward slope of the pile, are
deposited on its surface.
From figure 14.17, it can be seen that in variant no. 11 large drops reach the dust
cloud and settle on the leeward slope of the pile. Small drops are enveloped by a
headwind and are carried out by them to the limits of the storage area.
From figure 14.18, it can be seen that in variant no. 12 only a fraction of the
droplets reach the dust cloud. Most of the larger droplets are deposited on the
leeward side of the pile, before they reach the dust cloud. The smallest drops are
turned by the headwind and are carried by them outside the storage area.
According to the results of numerical simulation in variant no. 13, an insignificant
proportion of the large drops reaches the dust cloud. Most of the droplets do not
reach the dust cloud—the largest drops are deposited on the leeward side of the pile,
not reaching the dust cloud, and the smaller drops turn around the headwind and are
carried away from the storage area. In variant no. 14, some of the droplets fall into
the dust cloud. In this case, all the drops are carried by the cross wind outside the
storage area. In variant no. 15, none of the droplets reach the dust cloud—they are
all carried by the side wind outside the storage area.
The most preferable were the results in variants no. 5 and no. 6. In variant no. 5,
the total residence time of all droplets in a dust cloud was 175.6 s. At the same time,
the mass fraction of droplets that did not fall into the dust cloud was 27.8%. In
variant no. 6, the total residence time of all droplets in the dust cloud was 164.4 s. At
the same time, the mass fraction of droplets that did not fall into the dust cloud is
70.8%.
The obtained data allow us to further consider the issue of optimizing the
irrigation of a dust cloud using an atomizer.
Within the framework of the developed numerical model, the estimated values of
the total residence time of the droplets in the working area were obtained and the
proportion of drops that did not fall into the working area was established.
Numerical simulation of the presented mathematical model made it possible to
determine the most effectual water spraying modes using various devices for dust
suppression under different wind directions. The results suggest that the use of
dispersed water structures to suppress dust is effective. At the same time, environ-
mental safety management is achieved by the right choice of the technical device and
the mode of supplying the process fluid (or water).

14.3.4 Results of numerical simulation of the spatial finely dispersed structure

creation process, which instantly reduces the gas stream temperature
For the numerical simulation and visualization of the results of the instant reduction
of gas stream temperature, gas cooling which was obtained from the high-temper-
ature recycling of solid carbon-containing waste was considered. In this case, we
have a two-phase multicomponent structure with phase transformation, since water
particles in the gas stream with a high temperature will evaporate. An ODE
describes the behavior of a single drop in a gas-dynamic space [35] and has been
shown in section 14.2.3. In this system, some of the parameters are functionally

14-24
 Modern Optimization Methods for Science, Engineering and Technology

related to independent variables. To close and solve an ODE system, it is necessary

to take into account two-way interaction by solving dispersed phase equations and
continuous phase equations alternately.
The fine water mist effectiveness for gas stream cooling depends on the structure
and gas–droplet stream parameters. Such parameters include the dispersion of
spraying water by nozzles and its rate of outflow from the nozzles. So, the variant of
spraying water by nozzles into the high-temperature gas stream is necessary, which
will ensure the instant cooling of the gas stream to the required temperature. We
believe that the gas flow cooling occurs in the heat exchanger.
Based on the use of the classical description of the two-phase flow movement, the
physical and mathematical description of this process is possible [38–40].
A gas phase mathematical model of cooling of high-temperature gas is described
in [38, 39]. This model shows the three-dimensional flow characteristics in a heat
exchanger and the influence of the dispersed phase. To analyze the dispersed phase
behavior, the following assumptions were made, which allowed its behavior to be
considered in the Lagrangian description [40]:
• The dispersed phase is a completely sprayed torch of fluid (water), consisting
of a finite set of evaporating spherical droplets of various diameters.
• All forces acting on the drop, except for the forces of gravity and aerody-
namic drag, are neglected.
• We neglect the processes of secondary crushing and coagulation of droplets.

The phase interaction was taken into consideration through the discrete model
‘particle—source in a cell’. This model assumes that the droplet presence in the flow
is expressed by additional elements in the conservation equations of the continuous
phase [40]. This is described in section 14.2.3. When calculating the trajectories of
the droplets, the values of the impulse, mass and temperature of droplet ‘packets’
which changed when moving were tracked. These values as initial conditions were
included for the gas phase calculation Sm , Sq , Sfi .
By estimating the droplet mass change as it passes through each control volume
of the geometric model, the mass transfer between dispersed and continuous phases
were calculated as
⎛ Δmp ⎞
ΔSm = ∑⎜⎜ · ṁ p0⎟⎟ , (14.29)
⎝ m p0 ⎠

where Δmp is the change in the drop mass in the control volume, m p0 is the initial
mass of a drop and ṁ p0 is the initial droplets’ mass flow rate.
By estimating the drop pulse change as it passes through each control volume of
the geometric model, the momentum transfer between continuous and dispersed
phases was calculated using formula (14.25). By estimating the drop enthalpy
change as it passes through each control volume of the geometric model, the heat
transfer between continuous and dispersed phases was calculated as

14-25
 Modern Optimization Methods for Science, Engineering and Technology

⎛ mp Δmp ⎛ Tp ⎞⎞
ΔSq = ∑⎜⎜ cpΔTp + ⎜−L + ∫T cpi (T )dT ⎟⎟⎟ · ṁ p0 , (14.30)
⎝ m p0 m p0 ⎝ 0 ⎠⎠

where mp is the average drop mass in the control volume, cp is the heat capacity of
drops, ΔT is the change in the temperature of drops in the control volume, L is the
latent heat of evaporation, cpi is the heat capacity of vapor fluid, Тp is the drop
temperature at the outlet of the control volume and Т0 is the enthalpy standard
temperature.
Since the gas phase already affects the dispersed phase, the reverse effect of the
dispersed phase on the continuum was also taken into account [39].
To numerically simulate the instantaneous cooling of a hot gas stream, a
fragment of the heat exchanger space was investigated. This fragment is bounded
by the heat exchanger walls, and its inlet and outlet sections (figure 14.19).
Centrifugal nozzles are built in the heat exchanger to disperse the fluid. The purpose
of this numerical simulation is to investigate the possibility of instantaneous cooling
of a gas stream from 1200 °C to 300 °C.

The initial conditions

For the numerical solution of the ODE system, it is necessary to specify the initial
conditions, that is, the values of independent variables at the initial moment of time
t = 0: upj0, xpj0, dp and Tp.
We believe that at the initial moment of time the gas elementary volume is located
in the cross section center of the heat exchanger flow-through part. This section
precedes the liquid injection zone. For variants no. 1 and no. 2, at the initial moment
of time the gas elementary volume is located in the section ‘B—B’ and for variant no.
3 in the section ‘A—A’. In this case, the uneven computational domain included
77 087 polyhedric cells.
The initial diameters of a cluster of droplets having the same paths in the package
were given by the histogram of the initial distribution of droplets by size. The initial
velocity of the drops was assumed to be equal to the rate of flow of fluid from the
nozzle; the initial temperature of the droplets was equal to 20 °C. The initial
direction of the vector of the initial velocity of the drops was determined from the
root angle of the atomization plume. Numerical integration was carried out for three

Figure 14.19. Fragment of the heat exchanger space which was investigated: (a) right side view and (b) isometry.

14-26
 Modern Optimization Methods for Science, Engineering and Technology

Table 14.3. The values of the parameters of water injection through nozzles [24].

Parameter name Variant no. 1 Variant no. 2 Variant no. 3

d0 (m) 0.0006 0.0009 0.0011

Аnozzle/(Dsd0) 0.75 1.3 1
CD 0.43 0.56 0.50
α (degrees) 66 50 57
v (m s−1) 10 3 3
Δp (Pa) 48 228 5650 3189
Ree 3858 2257 1957
dmin (m) 4.7 · 10−6 55 · 10−6 202 · 10−6
d¯ (m) 16.6 · 10−6 144 · 10−6 514 · 10−6
dmах (m) 25 · 10−6 234 · 10−6 802 · 10−6
n 3.77 3.99 3.6

variants (modes) of water supply. The values of the parameters of water injection
through nozzles are presented in table 14.3 [24].
By integrating the gas phase equations [38, 39], gas parameters can be calculated
at any studied fragment point of the heat exchanger. So we can control the gas flow
temperature and speed in different heat exchanger sections. In addition, most
rational parameters of the liquid supply by sprays can be determined.
For the numerical integration of a system of partial differential equations with
specific boundary conditions, their discretization must be performed. The controlled
volumes method was applied to the discretization of equations in space [41] on an
unstructured computational grid, which contains polyhedric elemental volumes—
cells. The droplet volume distribution in a sprayed liquid stream is based on data
from [26] and is described by the Rosin–Rammler equation.
The droplets’ heat and mass transfer is described by two models—evaporation
and boiling. The evaporation model is valid until the boiling point is reached by the
droplet Tbp. Once the boiling point is reached, then the drop’s heat and mass transfer
is determined by the boiling rate.
The evaporation model suggests that the rate of evaporation of a drop is
determined by Fick’s law:
dmν dc
= Aν Dρ , (14.31)
dt dr
where mν is the steam mass, ρ is the gas density, D is the coefficient of binary
diffusion of vapor in gas, с is the vapor concentration, Аv =πd p2 is the evaporation
surface area, t is time and r is a radial coordinate.
Dividing the variables and integrating equation (14.31) with the boundary
conditions c = cs with r = rp, c = c∞ with r = r∞, we obtain

14-27
 Modern Optimization Methods for Science, Engineering and Technology

dmν
= βAν ρ(cs − c∞), (14.32)
dt
where cs is the volume concentration of vapor at the surface of the drop, rp is the
radius of a drop, c∞ is a volume concentration of vapor in the ambient gas and β is
the experimentally determined evaporation rate.
The results of experimental studies are usually presented as criterial dependences
Sh (Re, Sc), where Sh is the Sherwood number, defined as
βdp
Sh = . (14.33)
D
Taking into account (14.33) and the evaporation area of a spherical drop,
equation (14.32) takes the form
dmν
= ShρDπd p2(cs − c∞). (14.34)
dt
For an approximation of the criterion dependence Sh(Re, Sc), the empirical
Rantz–Marshall formula was used [42]:
Sh = 2 + 0, 6 Re0,5 Sc0,33. (14.35)
Suppose that the partial pressure of vapor on the surface is equal to the saturated
vapor pressure psat at the drop temperature Tp. In this case, the vapor concentration
on the surface of the drop can be calculated,
Mνpsat
cs = , (14.36)
Mνpsat + M (p − psat )
where М is the molecular mass of gas, Мv represents the molecular mass of vapor,
and р is the absolute pressure of the gas–vapor mixture.
If we differentiate equation (14.34) by time, then an equation for the decrease rate
of the evaporating drop diameter can be obtained:
dmp d (d p )
= 0,5πρp d p2 . (14.37)
dt dt
dmp dmp
Considering equation =− from (14.34) and (14.37), we obtain
dt dt
d (d p ) 2ShρD
=− (cs − c∞). (14.38)
dt ρp dp

When the gas is cooled, the drop temperature of dispersed fluid changes before
reaching the boiling point in accordance with the heat balance
dTp dm
mpcp = αAν (T∞ − Tp) + L ν , (14.39)
dt dt

14-28
 Modern Optimization Methods for Science, Engineering and Technology

where cp is the heat capacity of drops, α is the experimentally determined heat

transfer coefficient, Aν is the surface area of a drop, L is the latent heat of
evaporation, Tp is droplet temperature and T∞ is local gas temperature.
The results of experimental studies are usually given in the form of criterion
dependences Nu (Re, Pr),
αλdmν
Nu = , (14.40)
dp

where λ is the thermal conductivity of gas and Nu is the Nusselt number.

Taking into account (14.34) and (14.40), the notation (14.39) can be represented
in the following form:

dTp (T∞ − Tp) QL (14.41)

= + ,
dt Θ Θ
where
LShρD(cs − c∞)
Q= ; (14.42)
Nuλ

ρ p d p2cp
Θ= . (14.43)
6Nuλ
Based on the assumption of a complete analogy between heat and mass transfer
to determine the Nusselt number, to approximate the criterial dependence of Nu
(Re, Pr), a relationship similar to formula (14.35) was used:
Nu = 2 + 0, 6 Re0,5 Pr 0,33. (14.44)
After the droplet temperature is equal to the boiling point, instead of equation
(14.38) the equation of the speed of boiling is applied:
⎛ ⎡ cp∞(T∞ − Tp) ⎤⎞
d (d p ) 4λ ⎜
=− 1 + 0, 23 Re 0,5
ln ⎢ 1 + ⎥⎟ , (14.45)
ρp cp∞dp ⎜⎝ ⎥⎦⎟⎠
p
dt ⎢⎣ L

where cp∞ is the gas heat capacity.

Equation (14.45) is derived under the assumption of stationary flow at constant
pressure. It is believed that while the boiling law is applied, the drop maintains a
constant temperature.
The results of numerical simulation of instantaneous cooling of high-temperature
gas are shown in tables 14.4–14.6.
According to the results in variant no. 1, water drops very quickly (in the
immediate vicinity of the injection site) lose their initial impulse and then follow a
compact ‘swarm’ along the heat exchanger. The resulting asymmetry of the flow
and, accordingly, the asymmetry of the droplet trajectories are due to the high

14-29
 Modern Optimization Methods for Science, Engineering and Technology

Table 14.4. The results of numerical simulation of instantaneous cooling of high-temperature gas for
variant no. 1.

Parameter name Section 1 Section 2 Section 3 Section 4

z (m) −0.55 −0.912 −1.274 −1.6365

tmin (°С) 31.3 188.7 275.3 311.8
tav (°С) 408.6 364.7 357.6 355.8
tmax (°С) 951.6 592.8 482.3 424.1
γТ 0.8892 0.9422 0.9649 0.978
gH2O.min 0.041 0.116 0.143 0.158
gH2O.av 0.1666 0.1756 0.1769 0.1774
gH2O.max 0.282 0.227 0.2 0.19
γH2O 0.8804 0.9426 0.9653 0.9781

Table 14.5. The results of numerical simulation of instantaneous cooling of high-temperature gas for variant
no. 2 [24].

Parameter name Section 1 Section 2 Section 3 Section 4

z (m) −0.55 −0.912 −1.274 −1.6365

tmin (°С) 257 337.7 347.7 350.4
tav (°С) 356.9 354.3 354.41 354.42
tmax (°С) 414.1 363.2 359.6 358
γТ 0.9795 0.9955 0.9973 0.9983
gH2O.min 0.162 0.175 0.176 0.177
gH2O.av 0.177 05 0.177 66 0.177 64 0.177 63
gH2O.max 0.2057 0.1822 0.1795 0.1787
γH2O 0.9799 0.9957 0.9974 0.9984

Table 14.6. The results of numerical simulation of instantaneous cooling of high-temperature gas for variant
no. 3 [24].

Parameter name Section 1 Section 2 Section 3 Section 4

z (m) −0.55 −0.912 −1.274 −1.6365

tmin (°С) 285.1 330.2 345.7 354.1
tav (°С) 418.2 393.3 382 375
tmax (°С) 616.5 467.5 422.4 396.6
γТ 0.9501 0.9711 0.9835 0.9904
gH2O.min 0.112 0.147 0.159 0.166
gH2O.av 0.1604 0.1668 0.1698 0.1718
gH2O.max 0.2 0.184 0.18 0.177
γH2O 0.9443 0.9693 0.9829 0.9903

14-30
 Modern Optimization Methods for Science, Engineering and Technology

sensitivity of relatively small droplets (≈16.6 μm) to small perturbations (fluctua-

tions) of the gas velocity.
In addition, the droplets take a temperature from 12 °C to 58 °C, not reaching a
boiling point of water of 100 °C, but most of the droplets have time to evaporate
before entering a narrow portion of the heat exchanger, that is, section 1. Therefore,
the numerical value of the average mass fraction of water vapor gH2O.av in the vapor–
gas mixture in section 1 is significantly less than the equilibrium one. Accordingly,
the numerical value of the gas–vapor mixture average temperature tav in section 1 is
55 °C more than the equilibrium temperature (454 °C).
The temperature of the gas–vapor mixture and the mass fraction of water vapor
in it are unevenly distributed over the above-mentioned sections, particuarly in
section 1, where the coefficients of uniformity are γT = 0.8892 and γH2O = 0.8804. The
gas–vapor mixture maximum temperature in all sections exceeds the equilibrium
(454 °C) by 598 °C, 239 °C, 128 °C and 70 °C, respectively (see table 14.4). Thus, the
gas–droplet flow structure and parameters do not ensure the achievement of the
intended purpose, and consequently variant no. 1 of the water supply cannot be
considered satisfactory.
In variant no. 2, water drops retain their initial impulse for a long time
(approximately to the middle of the channel width), after which they follow
trajectories under the influence of a hot gas stream. These trajectories are curved
as a result of the interphase exchange of momentum and rebound from the walls of
the vapor pipe. In this case, the symmetry of the flow and, accordingly, the
symmetry of the droplet trajectories are largely preserved, which is due to the
weak sensitivity of relatively large droplets (≈144 μm) to small perturbations
(fluctuations) of the gas velocity. In addition, from figure 14.20 it can be seen that
droplets, as a result of the processes of convective heat exchange and evaporation
when moving inside the heat exchanger, take temperatures from 20 °C to 59 °C, not
reaching the boiling point of water, 100 °C.
It is also seen that some of the droplets do not have time to evaporate before
entering a narrow section of the vapor pipe, that is, before section 1. Their
evaporation continues in a narrow section of the heat exchanger to section 4.
However, the number of such drops is small, since the numerical value of the
average mass fraction of water vapor gH2O.av in the gas–vapor mixture in section 1 is
very close to equilibrium, to which this value tends further downstream in sections 2,
3 and 4 (see table 14.5). Accordingly, the numerical value of the gas–vapor mixture
average temperature tav in section 1 is only 3 °C higher than the equilibrium
temperature (454 °C).
The temperature of the gas–vapor mixture and the mass fraction of water vapor
in it are evenly distributed in the sections, as indicated above, including section 1
(figure 14.21), where the coefficients of uniformity are γT = 0.9795 and γH2O = 0.9799.
The gas–vapor mixture maximum temperature at all sections exceeds the
equilibrium state (454 °C) by only 60 °C, 9 °C, 6 °C and 4 °C, respectively

14-31
 Modern Optimization Methods for Science, Engineering and Technology

Figure 14.20. Trajectories of ‘packages’ of drops, painted in accordance with the time of their stay, in seconds
(variant no. 2, isometry).

Figure 14.21. Gas temperature distribution over heat exchanger sections, in °C (variant no. 2, isometry).

14-32
 Modern Optimization Methods for Science, Engineering and Technology

(table 14.5). Thus, the structure and parameters of the gas–droplet flow as a whole
correspond to the set purpose and variant no. 2 of water supply can be considered
satisfactory.
In variant no. 3, water drops retain their initial impulse almost unchanged across
the entire width of the channel until they collide with its walls, after which they
follow trajectories mainly due to the rebound phenomenon. Accordingly, the
symmetry of the droplet trajectories is preserved, which is due to the extremely
low sensitivity of very large droplets (≈514 μm) not only to small perturbations
(fluctuations), but also to the averaged gas velocity. At the same time, droplets have
a strong influence on the gas flow, which in this case remains almost symmetrical. It
is also observed that water droplets, as a result of the convective heat exchange and
evaporation, take temperatures from 20 °C to 57 °C, without reaching the boiling
point of water, 100 °C.
Drops do not have time to evaporate, not only to the entrance to a narrow section
of the heat exchanger (to section 1), but also to the output of section 4. Although the
drops only partially evaporate within the entire heat exchanger, they fill the volume
of the flow part very fairly and evenly. As a result, the numerical value of the average
mass fraction of water vapor gH2O.av in the vapor–gas mixture in section 1 is slightly
less than the equilibrium. Accordingly, the numerical value of the gas–vapor mixture
average temperature tav at section 1 is 64 °C higher than the equilibrium (455 °C).
The temperature of the gas–vapor mixture and the mass fraction of water vapor
in it are also fairly evenly distributed over all sections, particularly in section 1,
where the coefficients of uniformity are γТ = 0.9501 and γH2O = 0.9443 (see table 14.6).
At the same time, the gas–vapor mixture maximum temperature in the sections
exceeds the equilibrium temperature (454 °С) by 263 °С, 114 °С, 68 °С and 43 °С,
respectively. Thus, the gas–droplet flow structure and parameters do not ensure the
achievement of the intended purpose, and consequently variant no. 3 of water supply
cannot be considered satisfactory.
It is obvious that the recognition of variant no. 2 for supplying water as
satisfactory does not exclude the existence of an even more perfect version, for
the future search for which we can formulate and solve the corresponding
optimization problem.
Further, a numerical experiment to determine the cooling time of the gas stream
to the required temperature was carried out. Considering the movement of the gas
elementary volume at the heat exchanger fragment, figure 14.22 shows the control
jet of the gas stream for various modes of dispersed fluid supply and figure 14.23
shows the Z-coordinate dependence graph of the gas elemental volume on time τ.
From the graph Z(τ), the moments of time were determined at which the
elementary volume of gas reaches the next section i. The average tav, minimum
tmin and maximum tmax gas temperature, average temperature deviation and
uniformity coefficient of gas temperature distribution in control sections, and the
residence time Δτ, cooling Δt, gas elementary volume average cooling rate Δt/Δτ
between adjacent sections i and (i—1) were determined.

14-33
 Modern Optimization Methods for Science, Engineering and Technology

Figure 14.22. The trajectory of the gas elementary volume.

Figure 14.23. The Z-coordinate dependence graph of the gas elementary volume on the movement time along
the heat exchanger.

The total gas elementary volume existence time at the heat exchanger for the
investigated variants is 1.32 s, 6.42 s and 2.37 s, respectively. The maximum gas
cooling is observed during its contact with the dispersed fluid droplets, which
evaporate:
• Variant no. 1 is characterized by lowering the gas temperature to 472 ° C for
0.47 s. This is observed between sections C and E.
• Variant no. 2 is characterized by gas temperature 436 °C between sections C
and E. This temperature is reached for 1 s of its stay in the heat exchanger.
• Variant no. 3 is characterized by gas temperature 454 °С between sections A
and E. This temperature is reached during 1.6 s of its stay in the heat
exchanger.

14-34
 Modern Optimization Methods for Science, Engineering and Technology

The investigation shows that the maximum gas cooling rate is reached between
sections C and D. According to the results of numerical simulation, it can be seen
that for variant no. 1 it is −1919 °C s−1, for variant no. 2 it is −1269 °C s−1 and for
variant no. 3 it is −975 °C s−1. The rest of the time is taken to mix the gas with water
vapor. The dependence of the gas temperature on the residence time of the gas
stream in the heat exchanger is shown in figure 14.24 and the histogram of the
average gas cooling rate in the areas located between the control sections i and (i − 1)
is shown in figure 14.25.
Thus, the variant no. 2 of dispersed fluid supply is characterized by a short time to
establish an equilibrium state of the vapor–gas mixture, by reduced fluid flow and
eliminatesing its accumulation in the heat exchanger. However, from the point of
view of ensuring environmental safety, the first option can be considered as the most

Figure 14.24. The dependence of the gas flow temperature on the time of its cooling: red—variant no. 1;
blue—variant no. 2; black—variant no. 3.

Figure 14.25. The gas residence time between the control sections: red—variant no. 1; blue—variant no. 2;
gray—variant no. 3; A, B, C, D, E, 1, 2, 3, 4 are control sections.

14-35
 Modern Optimization Methods for Science, Engineering and Technology

satisfactory, since the cooling time of the gas stream from 1200 °C to 305.8 °C is
1.32 s.
However, these results do not exclude the possibility of determining the most
effective option, in the search for which a corresponding optimization problem can
be formulated and solved in the future.

14.4 General conclusions

The proposed mathematical model of the transportation process of finely dispersed
drops of fluid to the zone where there is a source of environmental hazard reliably
describes the peculiarities of the process depending on the ways of supplying fluid.
Using numerical simulation of the spatial finely dispersed structure creation
process, it is possible to determine the most effective ways of supplying fluid to
hazards of various origins, to propose various technological solutions for optimizing
this process and managing environmental safety.
The results of the calculations provide an opportunity to argue that the use of
spatial finely dispersed multiphase structures for environmental safety management
is an effective way to eliminate the dangers from different sources.

References
[1] Kolesnyk V, Pavlychenko A, Borysovs’ka O and Buchavyy Y 2018 Formation of physic and
mechanical composition of dust emission from the ventilation shaft of a coal mine as a factor
of ecological hazard Solid State Phenom. 277 178–87
[2] Vambol S, Vambol V, Sundararajan M and Ansari I 2019 The nature and detection of
unauthorized waste dump sites using remote sensing Ecol. Questions 30 1–17
[3] Vambol S, Bogdanov I, Vambol V, Suchikova Y, Kondratenko O, Hurenko O and
Onishchenko S 2017 Research into regularities of pore formation on the surface of
semiconductors Eastern-Eur. J. Enterp. Techn. 3/5 37–44
[4] Kolesnik V, Ye, Pavlichenko A V and Buchavy Y V 2016 Determination of dynamic
parameters of dust emission from a coal mine fang Nauk. Vis. Nat. Hirny. Univ. 2 81–7
[5] Tverda O, Plyatsuk L, Repin M and Tkachuk K 2018 Controlling the process of explosive
destruction of rocks in order to minimize dust formation and improve quality of rock mass
Eastern-Eur. J. Enterp. Techn. 3/10 35–42
[6] Sundararajan M, Sharma S N, Kumar R, Ansari I and Kumar G 2018 Estimation of SPM
emission in air environment through empirical modeling and remedies for dust control in and
around coal mining complexes National Seminar on Environmental Issues: Protection,
Conservation and Management (EIPCM) (Dhanbad, Jharkhand, 26–7 February 2016)
pp 207–14
[7] Tverda O, Tkachuk K and Davydenko Y 2016 Comparative analysis of methods to
minimize dust from granite mine dumps Eastern-Eur. J. Enterp. Techn. 2/10 40–6
[8] Kulyk M P 2014 Analiz ekolohichnoyi nebezpeky obyektiv teplovoyi enerhetyky ta metodiv
zmenshennya shkidlyvykh vykydiv Vis. Inzh. Aakad. Ukrayiny 2 253–8
[9] Vambol S, Vambol V, Sobyna V, Koloskov V and Poberezhna L 2018 Investigation of the
energy efficiency of waste utilization technology, with considering the use of low-temperature
separation of the resulting gas mixtures Energetika 64 186–95

14-36
 Modern Optimization Methods for Science, Engineering and Technology

[10] Vambol S, Vambol V, Kondratenko O, Koloskov V and Suchikova Y 2018 Substantiation

of expedience of application of high-temperature utilization of used tires for liquefied
methane production Arch. Mater. Sci. Eng. 87 77–84
[11] Savov N, Hristova T and Gencheva P 2019 Exploring possible accidents and recommen-
dations for biological waste treatment installations Qual.— Access Success 20 195–200
[12] Ragimov S, Sobyna V, Vambol S, Vambol V, Feshchenko A and Shalomov V 2018 Physical
modelling of changes in the energy impact on a worker taking into account high-temperature
radiation J. Achievem. Mater. Manuf. Eng. 91 27–33
[13] Vambol S O, Bohdanov I T, Vambol V V, Suchikova Y O, Kondratenko O M and
Onyschenko S V 2017 Formation of filamentary structures of oxide on the surface of
monocrystalline gallium arsenide J. Nano- Electron. Phys. 9 06016
[14] Hristova T 2017 Measures for restricting corrosion of a motor drive subjected to atmospheric
corrosion in the plant for wastewater treatment Qual.—Access Success 18 202–7
[15] Vambol S, Vambol V, Suchikova Y, Bogdanov I and Kondratenko O 2018 Investigation of
the porous GaP layers’ chemical composition and the quality of the tests carried out Arch.
Mater. Sci. Eng. 86/2 49–60
[16] Vambol S, Vambol V, Bogdanov I, Suchikova Y, Lopatina H and Tsybuliak N 2017
Research into effect of electrochemical etching conditions on the morphology of porous
gallium arsenide Eastern–Eur. J. Enterp. Techn. 6/5 22–31
[17] Suchikova Y, Bogdanov I, Onishchenko S, Vambol S, Vambol V and Kondratenko O 2017
Photoluminescence of porous indium phosphide. Evolution of spectra during air storage
Proc. of the 2017 IEEE 7th Int. Conf. on Nanomaterials: Applications and Properties (NAP-
2017) pp 138–41
[18] Sokolov D, Sobyna V, Vambol S and Vambol V 2018 Substantiation of the choice of the
cutter material and method of its hardening, working under the action of friction and cyclic
loading Arch. Mater. Sci. Eng. 94/2 49–54
[19] Lei M, Hai J, Cheng J, Lu J, Zhang J and You T 2018 Variation of toxic pollutants emission
during a feeding cycle from an updraft fixed bed gasifier for disposing rural solid waste Chin.
J. Chem. Eng. 26 608–13
[20] Milosh V V Dioksiny i ih potentsial’naia opasnost’ v ekosisteme "chelovek—okruzhaiush-
chaia sreda" http://crowngold.narod.ru/articles/dioxini.htm
[21] Wu X, Zheng M, Zhao Y, Yang H, Yang L, Jin R and Liu G 2018 Thermochemical
formation of polychlorinated dibenzo-p-dioxins and dibenzofurans on the fly ash matrix
from metal smelting sources Chemosphere 191 825–31
[22] Vambol S, Vambol V, Bogdanov I, Suchikova Y and Rashkevich N 2017 Research of the
influence of decomposition of wastes of polymers with nano inclusions on the atmosphere
Eastern-Eur. J. Enterp. Techn. 6/10 57–64
[23] Lu P, Huang Q, Bourtsalas A T, Themelis N J, Chi Y and Yan J 2018 Review on fate of
chlorine during thermal processing of solid wastes J. Environ. Sci. 78 13–28
[24] Vambol V 2016 Numerical integration of the process of cooling gas formed by thermal
recycling of waste Eastern Eur. J. Enterp. Techn. 6/8 48–53
[25] Launder B E and Spalding D B 1972 Lectures in Mathematical Models of Turbulence
(London: Academic), pp 458
[26] Schmidt D P, Corradini M L and Rutland C J 2000 A two-dimensional, non-equilibrium
model of flashing nozzle flow Proc. of the 3rd ASME·JSME Joint Fluids Eng. Conf.

14-37
 Modern Optimization Methods for Science, Engineering and Technology

(San Francisco, CA, 18–23 July 1999) (New York: American Society of Mechanical
Engineers), pp 1322
[27] Loitsianskii L G 1978 Mehanika Zhidkosti i Gaza (Moskva: Nauka), p 736
[28] Joshi M 2002 Failure of dust suppression systems at coal handling plants of thermal power
stations—a case study https://plant-maintenance.com/articles/dust_suppression.pdf
[29] Redko A, Dzhyoiev R, Davidenko A, Pavlovskaya A, Pavlovskiy S, Redko I and Redko O
2019 Aerodynamic processes and heat exchange in the furnace of a steam boiler with a
secondary emitter Alexandria Eng. J. 58 89–101
[30] Launder B E and Spalding D B 1974 The numerical computation of turbulent flows Comput.
Meth. Appl. Mech. Eng. 3 269–89
[31] Vambol S A, Skob Y U A and Nechiporuk N V 2013 Modelirovaniye sistemy upravleniya
ekologicheskoy bezopasnost’yu s ispol’zovaniyem mnogofaznykh dispersnykh struktur pri
vzryve metanovozdushnoy smesi i ugol’noy pyli v podzemnykh gornykh vyrabotkakh
ugol’nykh shakht Vestnik Kazan. Tekhnol. Un-ta 16/24 168–74
[32] SAS IP 2011 Using the Rosin–Rammler diameter distribution method https://sharcnet.ca/
Software/Fluent14/help/flu_ug/flu_ug_sec_discrete_diameter.html
[33] Kostyuk V and Ye 1988 K vyboru approksimiruyushchego vyrazheniya dlya koeffitsiyenta
aerodinamicheskogo soprotivleniya kapli Nauch.-metod. Mater. Teor. Aviats. Dvig.: Sbor.
Nauch. Trud. KHVVAIU 6 13–21
[34] Tsipenko A V 2004 Matematicheskaya model’ dispersnogo neravnovesnogo potoka s
bol’shoy doley zhidkosti v sople s uchetom plenki, stolknoveniy i aerodinamicheskogo
drobleniya kapel’ Moskva, NII NT, 46 s
[35] Krou D 1982 Chislennyye modeli techeniy gaza s nebol’shim soderzhaniyem chastits Ser.
Teor. Osn. Inz. Rasch. 104 114–22
[36] Vambolʹ S O 2012 Systema upravlinnya ekolohichnoyu bezpekoyu pry vykorystanni
pylopryhnichuyuchykh system zroshennya u protsesi navantazhennya ta rozvantazhennya
sypkykh materialiv u portakh Otk. Ynform. Komp’yut. Ynteh. Tekhnol. 55 161–67
[37] Vambolʹ S A 2012 Yssledovanye chyslennym metodom protsessa postanovky dyspersnoy
vodyanoy zavesy v systemakh upravlenyya ékolohycheskoy bezopasnosty Ekol. Bez.: Prob.
Shlyak. Vyris. 2 154–59
[38] Vambol V V 2014 Matematicheskoye modelirovaniye gazovoy fazy okhlazhdeniya gener-
atornogo gaza ustanovki utilizatsii otkhodov zhiznedeyatel’nosti Ekol. Bez. 6 148–52
[39] Vambol V V 2015 Modelirovanie gazodinamicheskih protsessov v bloke ohlazhdeniia
generatornogo gaza ustanovki dlia utilizatsii othodov Tehnol. Tehnos. Bezo.: Intern.-Zhur.
1 http://ipb.mos.ru/ttb/index.html
[40] Vambol V V 2015 Matematicheskoye opisaniye protsessa okhlazhdeniya generatornogo
gaza pri utilizatsii otkhodov zhiznedeyatel’nosti Tekhnol. Audit Rez. Proiz. 2/4 23–9
[41] Fletcher K 1991 Vychislitel’nyye metody v dinamike zhidkostey М.: Mir 1 504
[42] Shervud T, Pigford R and Uilki C 1988 Massoperedacha (Moscow: Mashinostroenie), p 600

14-38
 IOP Publishing

Modern Optimization Methods for Science, Engineering and

Technology
G R Sinha

Chapter 15
Future directions: IoT, robotics and AI based
applications
K C Raveendranathan

The recent innovations in the information age, where data is considered the ‘new oil’,
point toward rapid all pervasive developments in the Internet of Things (IoT),
robotics and artificial intelligence (AI) based applications. Data science has evolved
in the last few decades as a promising field of vast opportunities and challenges,
which encompasses all the endeavors of mankind. As raw data evolves into
information and intelligence through several data processors, its value is multiplied
many-fold. In this chapter, we primarily focus on future directions in the disruptive
technologies such as IoT and its importance in building smarter systems for a brave,
new and smarter world, where robotics and AI based applications play a pivotal role
in every human activity. IoT is a coinage of Kelvin Ashton of MIT in 1999, and is, in
general, any network of smart, connected devices which can be controlled from
anywhere across the globe through the Internet. It should be emphasized that despite
its potential advantages, such as being an enabler of global, remote connectivity and
thus aiding use of our home appliances which are smart enough to connect to the
Internet, its vulnerability to cyber attacks cannot be overlooked by the intelligent
designer, developer, or even the end user. Robotics and its principles were known
to technology evangelists and to end-users from the days of the science fiction play
R.U.R. (Rossum’s Universal Robots) by the Czech science fiction author Karel Čapek
(and his brother Joseph Čapek, who actually coined the term ‘robot’) in the 1920s. The
principles of AI were first suggested by Herbert A Simon, Marvin Minsky, Allen
Newell and John McCarthy at a 1955 conference, however, the credit for the term is
rightfully attributed to the latter. If we take a count of the devices/appliances/other
smart things connected to the internet by the year 2025, it will be the astounding
figure of 34.2 billion worldwide (projected data from IoT analytics), compared to the
present status of 17.8 billion worldwide in 2018, inclusive of IoT devices. One can take

doi:10.1088/978-0-7503-2404-5ch15 15-1 ª IOP Publishing Ltd 2020

 Modern Optimization Methods for Science, Engineering and Technology

these figures for granted and it is more likely that the actual numbers may exceed the
projections by 2025 unless some other disruptive technology evolves. Robotics has
matured enough with developments in interdisciplinary technologies such as mecha-
tronics and AI, such that fully automated vehicular systems and other means of
transport have become the order of the day. It is seemingly unpredictable, with the
current rate of innovations in mechatronics and AI, in which direction products and
processes will evolve. However, it is arguably going to be more toward a level playing
field, where the tools and techniques are primarily AI based, starting from the raw/
semi-processed data pumped by the IoT devices, communicated through the Internet,
and processed by the most advanced and inexpensive signal processors (both passive
and active). AI has grown to a mature technology in machine intelligence, and there
have been several more recent developments, such as machine learning (ML) and deep
learning, and several types of artificial neural networks (ANNs). The convolution
neural networks (CNN), recurrent neural networks (RNN) and deep neural networks
(DNN) fall into this category.

15.1 Introduction
In our information age, data is considered to be the new oil. The advent of the
Internet and its associated technologies has given way to huge data proliferation
from a number of sources. The IoT is an ecosystem of connected mechanical and
digital devices, physical objects, or living organisms, including human beings, that
are given a unique identity and have the ability to communicate and internetwork
together over a network, without needing human intervention. The IoT devices that
are interconnected through the Internet can be categorized into three groups:
1. Devices that receive and retransmit information.
2. Devices that receive information, and then process and act on it.
3. Devices that do both of the above.

Each of the above categories of things has enormous advantages associated with it.
Examples for the first category of devices, namely those devices that collect and send
information to the Internet, include all sorts of sensors. Typical sensors are
temperature, motion, moisture, air quality, or even optical (light). These sensors
enable the user to gather environmental information such as temperature, humidity,
air quality, etc, from their surroundings autonomously and communicate through
the wired or wireless Internet, which will, in turn, allow the end-users to make apt
and timely decisions. Just like the human sensory organs such as the nose, ear,
tongue and eyes help human beings perceive the surrounding world, these ‘smart
sensors’ enable various IoT devices to do the same. Also, the devices that receive and
act on sensory information can be operated remotely from a distant place, and this is
one of the major merits of the IoT. It is interesting to note that IoT devices also
greatly contribute to the proliferation of data, this will eventually open up a new
avenue for extensive study and research in data sciences, namely data analytics.
The third category of devices—the ones that collect information and retransmit,
as well as receive information and act on it—is the true goal of the IoT. To cite an

15-2
 Modern Optimization Methods for Science, Engineering and Technology

example, a soil moisture sensor in an agricultural farm can sense the moisture in the
soil and in turn decide when to switch on the irrigation system to water the plants.
Note that this is done in an intelligent fashion without the intervention of the farmer.
Obviously, the IoT improves the efficiency of operation of devices which are
connected to the Internet. The major outcome of the IoT and associated technol-
ogies is that of data proliferation. The IoT results in large chunks of real-time data to
be processed either locally or at a distant data center. Thus the proliferation of IoT
devices eventually leads to big data analytics.
The IoT is considered to be a disruptive technology, due to the fact that it has indeed
pre-empted several prevalent technologies. Further, several new applications have
evolved and proliferated with IoT, including distributed computing, wireless sensor
networks and so on. The principles of robotics have evolved and the technologies
matured over the past several decades. AI has created another paradigm shift in
modern-day computing. Robotics and AI, together, have produced several cutting edge
technologies and applications in the current millennium. The intelligence displayed by
man-made machines is termed AI or machine intelligence (MI). It is to be contrasted to
the natural intelligence displayed by human beings as well as other types of living
organisms, including plants. With AI technology one can build intelligent machines
that can work and react exactly like human beings. It is interesting to note that we have
now entered into a new era of technological advancements, where machines have
started creating machines, with minimal or almost nil human aid. The machine
intelligence of such machines has to be ‘measured’ using a different gauge of MI.
In this chapter, we mainly focus on the future directions in IoT, robotics, and AI
based applications.

15.1.1 The impact of AI and robotics in medicine and healthcare

A recent study by the market research firm Tractica estimated that the number of
shipments of healthcare robots would exceed 10 000 units annually by the year 2021.
Another study by yet another research firm, Allied Market Research, suggested that
the global market of robots used in clinical surgery would nearly double from US
$3.00 billion (in 2014) to US$6 billion (in 2020). AI technology is taking a crucial
role in the medical and pathological assessments of various health conditions and
the subsequent phases of treatments needed. Expert systems that can scan through a
massive volume of medical literature and transcripts are aiding medical practitioners
the world over to achieve more accurate diagnosis in a shorter time. As a typical
example, the novel ‘GP at Hand’ service, which is very popular in the UK (London),
enables patients to check various health conditions using a mobile app, thereby
being able to consult qualified medical practitioners through video conferencing
within two hours. GP at Hand uses AI to diagnose the ailment.
Yet another study by Frost and Sullivan revealed that the market driven by AI
technologies for healthcare would increase by 40% per annum from 2014 to 2021;
the corresponding rise in revenue varied from US$633.8 million in 2014 to US$6.662
billion by 2021. Another important example is that of the DAvinCi operating
robots, which will be discussed in more detail in a later part of this chapter.

15-3
 Modern Optimization Methods for Science, Engineering and Technology

They are very intelligent machines with a very high machine intelligence quotient
(MIQ) that use AI algorithms at their core to improve the accuracy when operating
on patients. AI techniques are used in producing high-resolution digital images and
holograms that are crucial in several applications such as space-imagery, the
detection and prevention of forgery, etc [1].

15.1.2 Advances in AI technology and their impact on the workforce

Over the years, AI has evolved from the elementary symbolic AI through embodied
AI to hybrid AI. As CPUs, networks and wireless communications matured over the
years, so did AI and robotics techniques [2]. Autonomous systems began to play a
major role in several diverse application scenarios and fields, such as space
technology to service robotics. Such intelligent robots with a high MIQ can assist
the users in various daily chores. They are very good at performing hazardous, dirty
and monotonous jobs. However, the use of such robotic systems to perform
autonomously in complex and diverse applications over an extended time (from a
few days to years) imposes several threats to humanity. Many of these are studied by
experts in the sub-disciplines of AI. They are navigation and mapping, reasoning,
knowledge representation, perception, autonomous planning, interaction and learn-
ing from past experiences. These sub-disciplines of AI with advanced techniques
which are integrated with an self-governing system enable robotic systems to operate
with more efficacy with a high factor of safety in a very complex, long duration
application [3].
Although AI arguably demonstrates profound advantages in the use of the
technology and provides several benefits that can be enjoyed every day, thereby
enhancing the quality of human life on Earth, there are a few apprehensions with
respect to the increasing negative ‘adverse ontological and existential’ outcomes of
AI. The proponents of this school of thought hold the firm belief that such super-
intelligent outcomes of AI technologies are highly hazardous to mankind at present.
The researchers Vardi, Tegmark and Greene compared it to a time-bomb ready for
explosion in the near future. A recent paper on 23 AI principles compiled by nearly
1200 AI/robotics researchers and 2342 other researchers from different domains,
presented at the ‘Future of Life (FLI) Conference’, added more credence to the fears
which most researchers share despite the benefits of AI to mankind. ‘It is imperative
that the rising advancements in AI technologies, continues to alienate humans from
their existential human nature’ [4]. Already, millions of routine, working-class jobs
in manufacturing have been automated. Furthermore, AI is now progressively
maturing to automate non-routine jobs in transportation, logistics, banking and
financial services, secretarial and clerical services, and healthcare. To cite a few
examples, the Google translator, Amazon’s Alexa and so on, are enabled by AI
technology. The ordinary working class had worried that advanced AI technology
would, more or less, demolish many routine jobs. Despite many learning the new
skill sets demanded by the new jobs which replaced the old ones, most of the workers
who were not able to keep up with the need to learn the new job-skills demanded by
the industry could not be redeployed. Humans could never successfully compete

15-4
 Modern Optimization Methods for Science, Engineering and Technology

with machines that could outperform them in every endeavor. Several modern-day
economists envisage that society will not be able to cope easily with the change.
People now realize that the dictum that AI technology that destroys existing jobs will
eventually create new job opportunities is too optimistic in reality [5]. Questions such
as whether the jobs replaced by AI systems will be replaced by the new jobs created,
of course after acquiring a new skill set, are highly relevant. A socio-economic study
conducted in the year 2013 reported that about 47% of American workers held jobs
that were at high risk of extinction due to automation, just a few decades from now.
The highly relevant questions are: ‘Will technology be able to create about a
100 million jobs if these jobs are automated by AI technologies? [5] Will AI
technology be able to create new jobs to compensate for those it has demolished or
made irrelevant? Will these jobs be created fast enough to meet the rising demands of
those who lost their earnings? What will be the state of employees whose skill sets
cannot catch up with the existential advancements in modern technology? Will such
workers lose their frugal existence in society? So far, the answers to the above
questions are affirmative.
A recent US study revealed that the employment potential of highly paid
cognitive jobs and low-paid service jobs are growing fast, such as physical aid to
the disabled or physically challenged older generation and fast-food services, which
are yet to be automated. However, technology is puncturing the economy severely,
through the automation of mid-skill, working-class jobs. Since 2000, several millions
of low salary service jobs have disappeared, and the workers were either forced to
leave the labor force or accept low income jobs that often pay meager sums, without
any other perks or privileges. It is apparent that firms in the communications
technology sector save huge amounts otherwise spent on salary by hiring temporary
employees on a contract basis, instead of full time regular staff. This led to the ‘gig
economy’—a job market where short-term contracts and flexible hours with almost
no benefits to the employees have flourished. Automation due to massive AI based
applications has fully alienated the creation of jobs from economic growth, resulting
in economic growth with an alarming rise in unemployment and large-scale
shrinkages in incomes of workers, thus resulting in more inequality in society. It
may also be noted that the newer technologies have created a ‘winner-takes-all’
situation, where the loser can hardly survive. In 1990, the three largest corporations
in Detroit were worth US$65 billion and had a labor force of 1.2 million. By 2016,
the three largest corporations in the Silicon Valley were worth US$1.5 trillion but
could accommodate only a mere 190 000 workers [5]. Hence the modern industrial
trend is larger communication companies and fewer jobs. If this is true, then AI
technology must create nearly 100 million new jobs to balance the gap in the labor
market.
The reliance is on advances in the domain of AI technologies that would create
more new jobs for the majority of the labor force, and everyone would earn a living
as ‘existential beings in their society’. On the other hand, when fewer new jobs are
created to accommodate even the newly up-skilled jobseekers, there are very high
unemployment rates, which is the situation termed by Verdi [5] as ‘a state of violent
uprising’. Considering the volatile state of the labor market, educational institutions

15-5
 Modern Optimization Methods for Science, Engineering and Technology

for engineering must focus on training their student community with the right skill
set needed, which the innovations in AI demand. Employees would need to
continually upgrade and hone their skills by attending appropriate training
programs for better prospects in the job markets. Even then, ‘the need to adapt
and train for new jobs will become more challenging as AI continues to automate a
greater variety of tasks’, as very rightly pointed out by Verdi.

15.1.2.1 Some recent trends on workforce lay-offs, redeployment and new hires
A recent article that appeared in IEEE Spectrum [8] revisited the statistics of jobs lost
due to lay-offs, redeployment after up-skilling, and the newly created jobs. The study
mainly focused on the Silicon Valley area in the US and went on to add that most of
the jobs lost due to lay-offs in various companies are compensated for by the hiring of
new technical and non-technical staff against the vacancies that were created by lay-
offs. Thus the new technologies enable the creation of new jobs and most of the
workforce who lost their jobs could easily find new avenues for employment.

15.1.3 AI technologies and human intelligence

An intelligent agent (IA) is a system that senses its environment alone and takes
appropriate actions to reduce its chances of failure. American computing pioneer
John McCarthy, who is credited as the originator of the term artificial intelligence
(AI) in 1956, defined the discipline as ‘the art, science and engineering of making
intelligent machines’. AI can be thought of as that discipline in the field of science
that focuses on building devices or machines that address complex issues, finding
solutions for common needs of human beings and thereby enhancing the quality of
life. In that process, the AI system designer inculcates human intelligence into
algorithms that govern the autonomous functioning of intelligent machines in a
user-friendly manner.
On 11 June 1997 an IBM mainframe computer, namely ‘Deep Blue’, defeated
Garry Kasparov, the greatest chess player (Grand Master) at the time, in six straight
set games of chess, proving that machine intelligence can become comparable to that
of humans. Thus Deep Blue passed Alan Turing’s test in machine intelligence.
Contemporary computing machines are much more than just intelligent, maturing
to super-conscious, covertly intelligent automatons, with the capacity to think on
their own, which is frequently depicted as detrimental to mankind. It now evident
that with the advent of highly intelligent AI and ML technologies, computing
devices have now successfully invented and programmed automations, such as
autonomous cars, pacemakers and involuntary trading systems. A few decades
earlier, all these were mere dreams. Notwithstanding the projected merits of AI and
ML technologies, Stephen Hawking, the greatest physicist of our times, could not
suppress his apprehension of the suggested virtues of AI technology. He rightly
stated: ‘The success in creating AI would be the biggest event in human history.
Unfortunately, it might also be the last, unless we learn to avoid the risks associated
with it’ [6].

15-6
 Modern Optimization Methods for Science, Engineering and Technology

One specific area where AI technology could be horrendous to mankind, for

example, could be a situation where advanced AI based weaponry systems are
encoded to perform risky maneuvers in military or space applications. Such deeply
programmed systems are extremely difficult to alter easily once they are deployed—as
in the case of the Inter-Continental Ballistic Missile (ICBM) mechanized weaponry
designed to target an enemy, some 700 miles aways, which is approximately the
distance between Japan and North Korea. Advanced AI war machines of this type are
known to counter all external influences aimed to alter their core objective. This
potential defect must be met with critical counter measures. Hence a machine such as
an ICBM guides itself, evading whatever obstacles anyone throws in its path,
ultimately attaining its original task of annihilating its target. In this context
Professor Stephen Hawking observed that ‘Artificial intelligence machines could kill
us because they are too clever. Such computers could become so competent that they
kill us by accident. The real risk with AI isn’t malice but competence. A super-
intelligent machine will be extremely good at accomplishing its goals, and if those
goals aren’t aligned with ours, we’re in trouble’.
In general, super-intelligent AI technologies are mostly feared for their capability to
upgrade themselves. Perhaps, they can even reprogram themselves whenever the need
arises. Having acquired a high MIQ, they will be more aware and conscious of their
environment, including the flora and fauna around them. Interestingly, they will begin
to think on their own and make judgments or decisions on what best to do with regards
to carrying out their assigned tasks or objectives. The mega intelligent AI systems are
feared to have the rare ability to upgrade, preserve and protect themselves from
whatever they may think of as a threat, meaning they would in the nearest future be
able to resist being disrupted, corrected or reprogrammed by their programmers [4].

15.2 Cloud robotics, remote brains and their implications

The term cloud computing was initially proposed by an American computer
engineer, Andy Hertzfeld. The demand for cloud-based applications is increasing
every day, since the cloud has a high level of scalability and is very efficient because
of its capability to be resilient in meeting current needs. Ever after the cloud
paradigm became ubiquitous and inexpensive, several automatons previously
thought of as very lethargic and inordinately complex became quite viable. For
robotics and AI, this meant that if the computational power behind the cloud could
be utilized, it would be possible to build tiny, highly energy efficient robots because
there would be no need to have a powerful computer on board. The entire brain of
the robot could be in the cloud. Yet the concept of a ‘remote-brain’ is not new [7].
Due to the distributed nature of the cloud across the globe, it is feasible to have
faster response times. This is possible as the processors are close to the data they are
working on, to reduce network delay, and have a low mean time between failures
(MTTF), and in particular is financially viable as it is a ‘pay per use model’. James
Kuffner, who was working at Google and is now with Toyota Research Institute
(TRI), predicted correctly that the cloud could make robots ‘lighter, cheaper and
smarter’ [8]. The use of the cloud to handle computationally intensive tasks permits

15-7
 Modern Optimization Methods for Science, Engineering and Technology

the use of smaller on-board computers in robots. These are used to attend to tasks
that need real-time processing, such as the control of sensors/actuators or stepper
motors. The grouping of low-level and high-level reasoning was first explored as ‘the
concept of remote-brain’ in 1996 at the University of Tokyo. The global knowledge
base for robots is a depository of objects, actions or environments. The robots are
enabled to download the object description and user manual, even when it was the
first time it encountered that particular object, or will be able to plan a route in the
environment, which was earlier traversed by another. One major disadvantage in
using cloud-based architecture is the chance of loss of connectivity between a
transaction, and even if the robot uses the cloud services for basic functionality only,
it will fail to do anything tangible. This is an acceptable constraint if a backup is
provided to meet that contingency. At present, the stability of network connectivity
is no longer an issue. Acquiring a new stable infrastructure is affordable compared
to the cost of a robot with full embedded intelligence. Now the real problem pertains
to the latency of the network [7].
Cloud robotics usually comprise two types of communications, namely machine-
to-machine (M2M) and machine-to-cloud (M2C). The first tier of communication,
M2M, implies that the robots are in a collaborative computing environment termed
an ‘ad hoc cloud’, which allows them to jointly share computationally intensive tasks
by pooling the resources needed and by exchanging information for collaboration,
and more importantly connecting to robots not in the range of the cloud access point
and helping them to communicate with the cloud. In short, most of the communi-
cations in this tier will be restricted to M2M. In the second tier of communication,
i.e. the M2C level, the cloud can release resources for computation and storage on-
demand. The robots in M2M mode communicate and share resources for a task that
is beyond their capability or which is common among themselves. This is illustrated
in figure 15.1.

Figure 15.1. M2M and M2C communication modes. Image source: [9].

15-8
 Modern Optimization Methods for Science, Engineering and Technology

As mentioned before, this in effect acts as a ‘remote-brain’ along with the shared
memory of actions, acquired skills, and information already learned and obtained
from the cloud.

15.2.1 Cloud computing and the RoboEarth project

The underlying principle of the RoboEarth [10] project was to create a pool of
robots which would allow them to share common knowledge, having a plug-in
architecture, and would use the ontology web language (OWL). There were three
databases in the cloud which could hold information on actions, objects and the
environments. The databases were maintained using OWL. Using the robot
operating system (ROS) the robots were interconnected [11]. But the robots were
not restrained to the use of ROS for communication. The entire information on the
capabilities of the robots, including construction and sensor types, was published for
the system, and hence the actions a robot could do were made known to others. The
essence of the RoboEarth architecture was the recognition/labeling component
(RLC). The RLC connected the hardware of the robot. The following abstractions
are maintained inside the robot’s system—actions, objects and environments. The
major objective of RLC was to translate abstract definitions from the databases of
RoboEarth to a format understood by the particular robot and vice versa, so the
robot could contribute to the databases with new acquired knowledge [7]. What it
meant was that it could work on low-level actions, called atomic primitives—signals
from sensors, motors and others. It could also work on high-level actions—spatial
and temporal relations between actions to create and execute abstractions of actions.
The architecture of RoboEarth is shown in figure 15.2.
New actions were learned through tele-operation. An operator controlled the
robot. Since RoboEarth was put on the ROS platform, it received all the signals
from the sensors and the motors, and assembled them to the action plan. Then the
robot interrogated the operator to label the action it did. The learning process was
also running when the robot was executing the algorithm of a particular action, to
further improve it [7].

15.2.2 The DAvinCi platform as a service (PaaS) surgical robot

The DAvinCi framework [12] was a software system that provided the merits of
parallelism and scaling up of cloud computing for service robots in fairly large set-ups.
The robot operating system (ROS) was employed to communicate with the robots.
The Hadoop/MapReduce framework [13] was used for parallel-processing to enhance
speed. The framework capabilities were tested on the implementation of the fast
simultaneous localization and mapping (FastSLAM) algorithm. The DAvinCi was a
typical platform as a service (PaaS) cloud-based application. The robots communi-
cated with the DAvinCi server. They could run ROS nodes for the robots which did
not have the capability to do so. Hierarchically, the Hadoop distributed file system
(HDFS) cluster which was located above the server was used for execution of robotic
algorithms. The DAvinCi server acted as a central communication hub and the
master node. The FastSLAM algorithm was implemented and tested successfully [7].

15-9
 Modern Optimization Methods for Science, Engineering and Technology

Figure 15.2. The architecture of RoboEarth. Image source: [23].

Although the domain of cloud computing is quite new, there are multiple projects
based on its principles and those of robotics. The major advantages of cloud
computing are its scalability, massive storage, resilience and ability for parallelism.
The methods used in AI, grabbing all resources in hold (‘grasping’), image processing,
and neural networks, are heavily computationally intensive. Hence it is not viable to
use them on a wide range of robots which do not possess enough on-board computing
resources. With cloud robotics it is possible to achieve the ‘remote-brain’. It is an older
concept in which the entire hardware and software are separate [7].

15.3 AI and innovations in industry

A system of computer-based applications that performs a group of tasks previously
done by white-collar workers, but without human intervention, is aptly called a
‘white-collar robot’ [14]. White-collar robots can automate certain tasks. Human
intervention is limited to deciding what task is to be done by a robot, and to choose
the right robot to do that. As in the case of a shop-floor managers, white-collar
robots might be part of a process, giving information to others, including people and
robots. This definition is consistent with the robots used for expediting manufactur-
ing processes and does not imply that the robots have ‘consciousness’. Instead, the
robot just executes the processes that they were programmed to do. Quite often the
robot will outperform the persons involved in certain scenarios, but the persons are
also likely to outperform the robot in certain other scenarios.

15-10
 Modern Optimization Methods for Science, Engineering and Technology

15.3.1 Watson Analytics and data science

Watson Analytics (www.ibm.com/analytics/watsonanalytics) provides excellent
interactive software to analyze assorted datasets. On uploading data to the cloud,
Watson Analytics will analyze the data quality, provide an initial analysis and
prompt the user to consider different combinations of variables. Watson Analytics is
dissimilar to Watson Cognitive. Watson Cognitive is widely understood from its
capabilities shown on the popular TV show Jeopardy! The capabilities of Watson
Cognitive focus on text and natural language processing. As a result, the capabilities
of Watson are grouped into two seemingly independent entities—cognitive sciences
and data analytics.
Watson Analytics needs the user or someone else to upload the data to the cloud.
After the data are uploaded, Watson Analytics prompts the user to analyze the data.
These starting points are a sequence of questions that the system has developed on
the basis of the data—for example, ‘What drives Process A?’, ‘What is a predictive
model of B?’ and ‘What is the trend of B and C?’ [14].
The data format is important to Watson Analytics, as it does not work well when
there are more than two dimensions to the data, say, with row or nested headings, or
both. Also, note that Watson Analytics is an active software—one that makes
recommendations and performs analysis without user intervention [15]. Generally,
active software makes assumptions on user needs. Active software is autonomous
and goal oriented. White-collar robots generally use active software [14].
It is highly likely that Watson Analytics frames the data for the user, which could
lead the user to form initial conclusions that are not appropriate under greater
scrutiny. Furthermore, Watson Analytics does not focus on how independent
variables drive the dependent variable. Thus users might understand which variables
predict or drive, but they may not know in which direction this occurs. Other
limitations include the unavailability of classical statistical tests, and the inability to
designate control variables and to choose particular statistical methods. In addition,
the modeling approach limits the number of variables used in estimation to one or
two [14, 15].

15.4 Innovative solutions for a smart society using AI, robotics and
the IoT
AI, robotics and the IoT are attracting widespread attention, as they are expected to
be technologies that affect society to a great extent in the future. These innovative
technologies have the potential to build seamless communication and a symbiotic
society between humans and robots, and a safe and secure networked society [16].
Various components of smart solutions include borderless communication and
symbolic communication between human beings and robots (machines).
The smart solutions for a smarter society discussed in this section include
automatic speech translation systems, a robot based dam inspection system, and a
large-scale security system used in imaging. Speech recognition in noisy environ-
ments results in inaccurate translation. The beam forming technology demonstrated

15-11
 Modern Optimization Methods for Science, Engineering and Technology

its effectiveness by improving speech recognition performance by up to 40%. Also,

the directivity and direction of beam forming must be controlled based on the
number and direction of speakers and noise sources. For example, directivity should
be set narrower when the speakers and noise sources are close [16]. Dam inspection
by intelligent robots with adequate provision for lighting was demonstrated to be a
viable alternative. For efficient operation of a large-scale system with over 10 000
cameras, visual security by humans is inadequate and image recognition technology
will be needed. The conventional approach requires a large number of cameras with
a high-capacity network and large-scale server system, which lacks scalability. To
solve this issue, a functionally distributed facial recognition system has been
developed. The ‘Best-shot method’ detects faces from security camera images, and
then transfers only the best-shot thumbnail which is selected as the most useful
image for facial recognition (with the highest resolution angle, facial angle, focus,
etc) to the facial recognition server. The server then extracts facial features for face
matching. This method reduces the network and server load, making it easier to
build a large-scale security system [16].

15.4.1 Cyber-physical systems (CPSs)

AI technologies will play more important roles in expanding cyber-physical systems
[17] sustainably in the future. A high-performance cloud-based processing environ-
ment in cyber space will also be necessary. However, due to the performance
limitations on cloud-based processing as described in the large-scale imaging
security solutions, load distribution is required. The load balance between cloud
and IoT/EDGE devices is important when building CPSs. The quickness of response
is important for natural dialog translation and autonomous robot solutions.
In the past, the cyber and physical systems were considered as two fairly
disjointed and distinct entities. However, modern researchers observed that these
two entities are closely related, after integration of sensors/actuators in the cyber
systems. Cyber systems respond to the physical world by enacting real-time control
from a conventional embedded system. The emerging new research paradigm is
termed as a CPS [17]. One can find a lot of applications for CPS, including
monitoring of water resources and mines, aerospace, surveillance and so on. Many
service providers are focusing on implementing inbuilt CPS technologies for their
customers. In CPS, communication is needed for conveying sensor observations to
controllers/actuators; thus, the design of the communication architecture is a critical
requirement.
The concept of a vehicular cyber-physical system (VCPS) is fairly old. It refers to
a wide range of integrated transportation management systems working in real-time,
with high efficiency and accuracy. The traditional modes of transport are becoming
more intelligent by inculcating the benefits of modern technologies, such as
electronics, computers, sensors and networks. The CarTel project, developed by
the Massachusetts Institute of Technology (MIT), combines mobile computing and
sensing, wireless networks and data intensive algorithms running on servers in the
cloud to address these challenges. CarTel helps applications to easily collect,

15-12
 Modern Optimization Methods for Science, Engineering and Technology

process, deliver, analyze and visualize data from sensors located on mobile units.
The applications of CarTel include traffic mitigation, road surface monitoring and
hazard detection, vehicular networking and so on [17].
Precision agriculture is experimenting on large-scale farming practices, products,
fundamental geographic information of farmland, micro-climate information and
other data. The project ‘wireless underground sensor network’ (WUSN) was
developed by the University of Nebraska-Lincoln Cyber Physical Networking
Lab, where Agnelo R Silva and Mehmet C Vuran developed a novel cyber-physical
system through the integration of pivotal systems with wireless underground sensor
networks, i.e. CPS for precision agriculture [18]. The WUSNs consist of wirelessly
connected underground sensor nodes that communicate through the soil.
The health cyber-physical systems (HCPS) will replace traditional health devices
working on an individual basis. With interconnected sensors and networks, various
health devices work together to detect the patient’s physical condition in real time.
This is particularly useful for patients who are critically ill, such as patients with
heart disease, strokes, etc. The portable terminal devices carried by the patient can
detect the patient’s condition at any time and send a timely warning or prediction of
critical conditions in advance. In addition, the real-time activation of health
equipment and data delivery systems would be much more beneficial for patients
with critical conditions [17]. The proposed standard CPS architecture is illustrated in
figure 15.3.
The six standard CPS architecture modules are the sensing module, data
management module (DMM), next-generation Internet, service aware modules
(SAM), application module (AM), sensors and actuators. The purpose of the
sensing module is to send an communication request to DMM and receive its
acknowledgment of the request. Once the handshake between the DMM and sensing
module is done, the transmission of the sensed data to the DMM from sensing nodes
commences. Here, the bridge between the cyber world and the physical world is
provided by noise reduction and data normalization techniques. Through quality of
service (QoS) routing, data are transferred to SAMs using the next-generation

Figure 15.3. The proposed standard cyber-physical system architecture. Image source: [17].

15-13
 Modern Optimization Methods for Science, Engineering and Technology

Internet. Available services are assigned to different applications in the application

module. To ensure the security and integrity of data, during each network operation
data are sent to a cloud platform and also to a local database [17].

15.4.2 IoT architecture, its enabling technologies, security and privacy, and
applications
Fog/edge computing has been integrated with the IoT with a purpose—to enable
computing devices deployed at the network edge to improve the user experience and
resilience of the services, in the case of system failures. The advantages of fog/edge
computing are their inherent distributed architecture and closeness to the end-users.
Thus, faster response and greater QoS for IoT based applications can be provided.
This makes fog/edge computing-based IoT, very attractive in IoT deployment [19].
Note that using fog/edge computing, the massive data generated by different kinds
of IoT devices can be processed at the network edge, instead of transmitting it to the
centralized cloud infrastructure due to bandwidth and energy consumption con-
straints. Since fog/edge computing devices are organized following a distributed
architecture model, they can process and store data in networked edge devices,
which are close to end-users. Thus they can provide services with faster response and
greater quality, compared to cloud computing. Thus, fog/edge computing is more
amenable to be integrated with IoT devices, to provide efficient and secure services
for a large number of end-users. In short, fog/edge computing-based IoT can be
considered as the future of IoT infrastructure.
It is well known that both CPS and the IoT try to achieve a close connection
between the cyber and physical worlds. In particular, the CPS and IoT can measure
the status information of the physical components via smart sensors without human
intervention. In both cases, the measured state information can be shared through
communication networks. Based on the analysis of measured status information,
both CPS and IoT can provide secure, efficient and intelligent services to the end-
users. The existing efforts on the CPS and IoT applications have been expanded to
similar application domains, such as smart grids, smart transportation, smart cities,
and so on. In CPS, the sensor/actuator layer, communication layer, and application
or control layers are present. The sensor/actuator layer is used to collect real-time
data and execute the commands. The communication layer delivers data to the
upper layers and commands to the lower layers. The function of the application or
control layer is to analyze the data and make decisions based on it [19]. Note that the
CPS is a vertical architecture. On the other hand, the IoT is a cluster of finely
internetworked devices in large numbers which are used to monitor and control
devices by using modern interconnection technologies in cyber space. Specifically,
the crux of IoT lies in ‘interconnection’. The main objective of IoT is to interconnect
various heterogeneous networks so that the data collection, resource sharing,
analysis and management can be carried out smoothly across networks.
Figure 15.4 illustrates the typical integration of IoT with CPS. The basic
difference between CPS and IoT is that CPS is considered a system, whereas the
IoT is considered an ‘Internet’.

15-14
 Modern Optimization Methods for Science, Engineering and Technology

Figure 15.4. Integration of IoT and CPS. Image source: [19].

The common requirements for both are real-time, reliable, and secure data
transmission and retrieval. The distinct requirements for CPS and IoT can be
summarized as follows. For CPS, effective, reliable, accurate and real-time controls
are the primary goals. For IoT, resource sharing and management, data sharing and
management, interfacing among different networks, massive-scale data and big data
collection and storage, data mining, data aggregation and information extraction,
and very high QoS networking are important requirements. Applications of the
integrated IoT and CPS include smart grids, intelligent transport systems (smart
transportation) and smart cities.
The smart grid is an integrated technology of IoT and CPS. The smart grid has
been developed to replace the traditional power grid to provide reliable and efficient
energy to consumers [20]. Distributed energy resources are introduced to improve
the utilization of distributed energy in electric vehicles to improve the capability of
energy storage and reduce the emission of CO2. In this, smart energy meters and
duplex communication networks are introduced to achieve the effective interactions
between customers and utility providers. Using IoT devices, a huge number of smart
meters can be deployed in houses and buildings connected to the communication
networks in the smart grid. Smart meters can monitor energy generation, storage
and consumption. They are capable of interacting with power utility providers to
report the energy demand information of customers and receive real-time electricity
pricing for customers. With the aid of fog/edge computing infrastructure, the huge
chunks of data collected from smart meters can be stored and effectively processed
so that the efficient operation of the smart grid is possible. With dynamic
information processing on the load conditions and the power generated, the utility
providers can optimize the energy dispatched in the grid. The customers can
optimize their energy consumption, resulting in the improvement of resource
utilization and the reduction of cost [19]. This is a truly win-win situation.

15-15
 Modern Optimization Methods for Science, Engineering and Technology

Smart transportation, a.k.a. intelligent transportation, is another typical example

of intertwined IoT–CPS based application. In smart transportation, intelligent
transportation management, control systems, communication networks and com-
puting techniques are integrated to make transportation systems more reliable,
energy efficient and secure. In the smart transportation scenario, a large number of
smart vehicles are interconnected through wireless networks. Smart vehicles can
efficiently perceive and share traffic data and schedule drivers’ travels with great
efficiency, reliability and safety. In recent years, several smart vehicles (Google’s
self-driving car, etc) have been designed and field-tested.
Smart cities can be considered a complex IoT paradigm, aiming to manage a host
of public activities via information and communication technology (ICT) solutions.
Smart cities can use public resources in more efficient ways, which results in the
improvement of the QoSs provided to users and the reduction of operational costs.
Smart cities are indeed a complex CPS/IoT application. The smart cities thrive on
several sub-applications or services, including the smart grid, smart transportation,
the structural health monitoring of buildings, waste management, environmental
monitoring, smart health, smart lighting, etc. All these services should be supported
by a unified communication network infrastructure, or communication networks
designed for these sub-applications. In other words, the services should be inter-
connected to establish a large-scale interconnected heterogeneous network for IoT/
CPS applications, with the aim of achieving the best use of public resources in cities.

15.4.3 The Internet of robotic things (IoRT) and Industry 4.0

It was Dan Kara, director of robotics at ABI Research, who conceived the idea of
the Internet of robotic things (IoRT). The IoRT is a diverse combination of
technologies including cloud computing, AI, ML and the IoT. As the concept of
the IoT is evolving and maturing, it leads to significant developments in terms of
innovation to various application domains. Several new terminologies are evolving,
based on the principles of IoT—the Internet of medical things (IoMT), Internet of
nano things (IoNT), Internet of mobile things (IoMBT), Internet of cloud things
(IoCT), Internet of autonomous things (IoAT), Internet of drone things (IoDT),
Industrial internet of things (IIoT) and Internet of underwater things (IoUT) [21].
It is quite interesting to add that the IoT has also laid a strong foundation for the
development of Industry 4.0 or smart factories. Most work will be performed via
sophisticated next-generation sensors and robotic technologies in smart factories.
Also, in smart factories, all the working personnel can respond quickly to any sort of
outage, which currently goes almost unnoticed. Some authors rightly pointed out
that Industry 4.0 [22] or the industrial Internet would eventually bridge the physical
and digital worlds.
One can easily predict that the IoT, in the near future, coupled with diverse
technologies such as AI, ML, deep learning, augmented reality, cloud computing
and swarm intelligence will change the face of robotics. The proposed next-
generation class of intelligent robotics are aptly titled as IoRT. The market research
firm Stratistics MRC stated that the market of IoRT was around $4.37 billion in

15-16
 Modern Optimization Methods for Science, Engineering and Technology

2016. It is expected to reach $28.03 billion by 2023 with a staggering compound

annual growth rate (CAGR) of 30.4%. The primary factors responsible for this
growth are the rise of e-commerce platforms, and the exponential growth of the
education sector, consumer markets, research and development wings, and above all
Industry 4.0.

15.4.4 Cloud robotics and Industry 4.0

One can find that the primary motivation behind the accelerated growth of the IoRT
is cloud robotics [23]. It is regarded as a system that relies on cloud computing
infrastructure to access large amounts of processing power and huge amounts of
data to perform very specific operations. All operations ranging from sensing,
computation and storage are integrated into a single, standalone system such as a
networked robot, in cloud robotics. Cloud robotic systems are programmed in such
a way that a portion of their capacity is reserved for local processing for low-latency
responses in case a network failure occurs. Cloud robotics is conceived and designed
as a progression between pre-programmed and networked robotics. With the advent
of cloud computing, big data analytics, other emerging technologies, and the
integration of cloud technology and multi-robot systems, it has become quite easy
to develop multi-robot systems which are very energy efficient with high real-time
performance and are also very inexpensive. Figure 15.5 illustrates the use of cloud
robotics in an industrial environment.
Since cloud computing became widely available, lots of computationally complex
and intensive algorithms and systems previously thought of as very time consuming
and hence impractical became viable. In the context of robotics and AI this implies
that if the power behind the cloud could be harnessed, it would be possible to build
smaller, more battery efficient robots. This is so, as there would be no need to have a
powerful, energy greedy computer on-board as the real brain of the robot can be in
the cloud. Thus it is now viable to build and operate a remote-brain, as discussed
above. The centralized cloud for the robot also implied that the system memory
could be nearly infinite and instantly available to other robots, so the process of
learning and exchanging the knowledge could be simplified and almost seamless [7].
The manufacturing automation protocol (MAP) was developed by General
Motors in the 1980s. A diverse set of incompatible proprietary protocols were
offered by several vendors. A paradigm shift occurred in the early 1990s when the
world wide web popularized the hyper text transfer protocol (HTTP) over Internet
protocol (IP).
The first industrial robot was connected to the world wide web with an intuitive
graphical user interface (GUI) in the year 1994. Thus it became possible for visitors
to teleoperate the robot via any Internet browser. In the mid and late 1990s,
researchers across the globe developed a series of web interfaces for robots and
devices to explore issues such as more user-friendly user interfaces and robustness in
design. Thus the era of the subfield of ‘networked robotics’ began.
The term Industry 4.0 was first introduced in Germany in the year 2011. It paved
the way for a fourth industrial revolution that would use computer networking very

15-17
 Modern Optimization Methods for Science, Engineering and Technology

Figure 15.5. The implementation of cloud robotics in an industrial environment. Image source: [23].

widely to automate factory operation remotely. It became the successor to the first
(mechanization of production using water and steam power), the second (mass
production with electric power) and the third (use of electronics to automate
production) industrial revolutions. The term ‘industrial Internet’ was first coined
by the engineers from General Electric in 2012, to describe new efforts where
industrial equipment such as wind turbines, jet engines and MRI machines connect
over networks to exchange data. Thus, the processing of data for industries
including energy, transportation and healthcare became quite efficient [24].

15.4.5 Opportunities, challenges and future directions

Several new challenges are present in using the cloud computing paradigm for
robotics and automation systems. A wide range of privacy and security issues were
raised by the open, overwhelming connectivity prevalent in the cloud. These
concerns include data privacy, since the data generated by cloud-connected robots
and sensors may include images or video or data from people, institutions (including
data pertaining to healthcare), or even trade secrets of the corporate world. Cloud
robotics and automation also introduces the potential threat of robots and
associated systems being hacked remotely: an unethical hacker could take over a
robot and use it to disrupt functionality or cause severe damage or disruption. For

15-18
 Modern Optimization Methods for Science, Engineering and Technology

example, researchers at University of Texas at Austin were successful in hacking into

and remotely control UAV drones via inexpensive GPS spoofing systems in an
evaluation study for the Department of Homeland Security (DHS) and the Federal
Aviation Administration (FAA). These concerns raise new regulatory, account-
ability and legal issues related to safety, control and transparency [24].
Now, let us consider the technical challenges in cloud robotics. The primary
concern is that to cope with time-varying network latency and quality of service
(QoS), faster algorithms and processes are needed. Faster data connections, both
wired Internet connections and wireless standards such as Long Term Evolution
(LTE), are reducing latency. Yet, the algorithms must be capable of degrading
gracefully when the cloud resources are very slow, noisy or totally unavailable. For
example, ‘anytime’ load balancing algorithms for speech recognition on smart
phones send the speech signal to the cloud for analysis and simultaneously process it
internally and then use the best results available after a reasonable delay. Similar
algorithms are desirable for robotics and automation systems using the cloud. Faster
algorithms are also needed that scale up to the size of ‘big data’, which often
contains dirty data that requires new methods to clean or sample the data
appropriately. When the cloud is employed for parallel-processing, it is vital to
have oversampling by the algorithms to take into account the fact that some remote
processors may fail or experience inordinate delays (long latency) in returning
results. Also algorithms are needed to filter unreliable or incorrect input data and
balance the costs of human intervention with the cost of robot failure, whenever
human computations are employed. Transferring the robotics and automation
algorithms into the cloud requires frameworks that facilitate the smooth transition.
In general, the cloud services can be classified into three categories: infrastructure as
a service (IaaS), platform as a service (PaaS) and software as a service (SaaS).
RoboEarth is an example of PaaS.
With SaaS, a user interface allows data to be sent to a cloud server that processes
it and returns the outputs, which relieves users from the burden of maintaining data,
software and hardware, and allows companies to control proprietary software. This
technique is widely known as robotics and automation as a service (RAaaS) [24].
Cloud robotics permits different robots to share computational resources,
information and data with each other, and share new knowledge and skills not
learned by themselves. This will eventually open up a new paradigm shift in
robotics, leading to exciting developments in the future. It permits the deployment
of inexpensive robots with low computational power and on-board memory
requirements, to leverage their potential by using the high speed communications
network and the elastic computing resources offered by the cloud infrastructure.
Applications that can benefit from the cloud robotics approach are several. They
include grasping, navigation and several others, such as weather monitoring,
intrusion detection, remote surveillance, non-destructive testing, fault detection
and formation control [9].

15-19
 Modern Optimization Methods for Science, Engineering and Technology

15.5 The human 4.0 or the Internet of skills (IoS) and the tactile
Internet (zero delay Internet)
One can predict the fast emergence of an entirely new, more vibrant Internet, that
capitalizes on the latest developments in 5G and ultra-low delay networking, as well
as the innovations in AI and robotics. This novel Internet will enable the delivery of
skills in digital form. The delivery of physical experiences remotely (and globally) is
made possible by the Internet of skills (IoS), which will revolutionize operations and
servicing capabilities for industries. In general, it will be a quantum jump in the way
we communicate, teach, learn and interact with our surroundings. It will be a brave
new world where our best engineers can service cars instantaneously around the
world over the tactile internet; or anybody can be taught how to paint by the best
available artists globally. At an estimated revenue of US$20 trillion per annum
worldwide, which is approximately 20% of today’s global gross domestic product
(GDP), it will be a technology enabler for skill set delivery—thus a very timely
technology for service driven economies across the globe [25]. The transformation to
the Internet of skills is illustrated in figure 15.6.
The IoS will democratize labor, in the same way as the Internet has democratized
the dissemination of knowledge. The core technologies used in the Internet of skills
are (a) ultra-fast data networks (zero delay Internet or tactile Internet), (b) haptic
encoders (both kinesthetic and tactile) and (c) edge AI (to beat the light limit).

15.6 Future directions in robotics, AI and the IoT

We have gone a long way since the advent of the IoT and AI. We are now living in a
smarter world that deals with computational statistics and brilliant cloud data
servers. It is absolutely certain that with the merger of two brilliant innovations such
as IoT and AI, things are moving towards a superior future. Businesses such as
automation, propelled robotics, smart domestic appliances that constitute the smart
home, assembling, design and retail are for the most part harvesting the advantages
of these advanced technologies. While this mix or integration of shrewd strategies is
yet to mature and its benefits are yet to be fully harvested, every year we advance a
couple of steps and make a quantum leap compared to the last one. It is obvious that
our homes and domestic appliances have already started adopting the intelligence
from AI and the capability to connect smart devices through the technical endow-
ment of the IoT [26].
It is expected that the IoT will affect the further development of AI and the
progression will be highly promising. When one analyzes the AI that has been
commercialized to date, the algorithms deployed are generally single agents, and
most of them are practically first-person algorithms. The purpose of the algorithms
was to foresee, analyze and act, beyond serendipity. Although the majority of the
insight happens independently, there is communication. Consider how we observe
different algorithms, for example those intended for managing a commercial center.
The associated costs are fixed because of numerous small-scale interactions between
individual dealers or responses to issues in the supply chain management or

15-20
 Modern Optimization Methods for Science, Engineering and Technology

Figure 15.6. The evolution of the Internet of skills (human 4.0). Image source: [25].

fluctuations in currency. These interacting sets of individual events permeate to a

value computation for a wide assortment of goods. Such an algorithm is termed
social AI [26].
With smart gadgets that can be controlled with a cell phone, the level of
intelligence in a household will turn out to be progressively greater. Without their
physical presence, anybody could send directions to a washing machine to wash and
dry garments, turn the lights on and off, turn the thermostats high or low, and direct
venetian blinds to close and open, using a remote gadget. Smart kitchens would
become more productive, where a greater part of the work could be performed with
smart electrical appliances that need fewer directions from the human operator.
Smart homes will be interconnected with inbuilt IoT. The average household will
face an ascent in the use of smart devices and appliances, and will apparently gear up

15-21
 Modern Optimization Methods for Science, Engineering and Technology

to a problem-free way of life. The digital voice assistants offered by Amazon and
Google are some of the incredible gadgets which will make life easier. The
development of amazing IoT solutions would be more in vogue, in view of
controlled purchasing power and appreciation for advancing innovations [26].
The gap between the IoT and AI will continue to diminish, leading to the amazing
growth of functionality in gadgets, and a consequent operational excellence. With
explicit client intelligence, a similar IoT gadget would be equipped to offer explicit
client experiences. IoT applications would become more intelligent, to cleverly
watch the Earth, correct flaws and autocorrect operational glitches. A growth in
client explicit innovations will provide the necessary base for a personal AI era [26].
The dependence on the cloud for inexpensive, localized computing power should
not be ignored. The advent of cloud robotics will result in numerous applications.
Also, more recent developments in deep learning and ML will continue to support
several AI applications, including those in image processing.
In the era of rapidly evolving technologies, big data analytics and IoT are the two
leading radical technologies which can explicitly modify the style of business
operations. Both technologies are still in their nascent stages and hold massive
potential and opportunities, and pose several unresolved challenges. The two
technologies can be coupled together for more efficient implementation and can
help all applications by making smarter decisions [27].
ML is a modern science which enables computers to work autonomously. No
explicit programming is needed in ML. The modern-day technology deploys
algorithms that can train and improve on the data that is being fed to them. Over
the years, ML has matured and made possible the concept of self-driving autono-
mous cars. ML is also a technology enabler for more effective web searches, spam
free emails, practical speech recognition software, personalized marketing and so on.
Today, ML is increasingly being deployed in credit card purchase fraud detection,
personalized advertising through pattern recognition, personalized shopping/enter-
tainment recommendations, to determine cab arrival times, pick-up locations, and
finding routes on maps [28]. The five recent technological trends are (a) the creation
of more jobs in data science, (b) new approaches to data security, (c) robotic process
automation (industry 4.0), (d) improved IT operations and (e) transparency in
decision making [28].
It is interesting to note that the essence of ML and AI lies in ANNs. DNNs
enhance the learning capability of ANNs by adding more hidden layers to the
network. CNNs are a class of deep, feed-forward (not recurrent) artificial neural
networks that are applied to analyze videos. CNNs are usually composed of a set of
layers that can be grouped by their functionalities. CNNs consist of feature learning
layers and classification layers apart from the input and output layers.
With the advent of the ANN, ML has taken a giant leap in recent times. ANNs
are biologically inspired computational models, and are capable of exceeding the
performance of previous forms of AI to a larger extent. The CNN is one of the most
impressive forms of ANN architecture. CNNs are primarily used to solve difficult
image-driven pattern recognition tasks with their precise yet simple architecture. A
simplified method of getting started with ANNs is provided by CNNs [29]. Note that

15-22
 Modern Optimization Methods for Science, Engineering and Technology

first generation ANNs were shallow, in the sense that apart from the input and
output layers they contained at most one hidden layer. In contrast, DNNs contain
several hidden layers.
CNNs differ from other forms of ANNs in that instead of focusing on the entirety
of the problem domain, knowledge about the specific type of input is exploited. This in
turn allows for a much simpler network architecture to be set up [29].

References
[1] Banks J 2018 The human touch—practical and ethical implications of putting AI and
robotics to work for patients IEEE Pulse 9 15–8
[2] Fujita M 2019 AI x robotics: technology challenges and opportunities in sensors, actuators,
and integrated circuits Proc. of the 2019 IEEE Inter. Solid-State Circ. Conf. (ISSCC 2019)
pp 276–7
[3] Lars Kunze et al 2018 Artificial intelligence for long-term robot autonomy: a survey IEEE
Robotics and Automation Letters 3 4023–30
[4] Wogu I A P et al 2017 Artificial intelligence, alienation and ontological problems of other
minds: a critical investigation into the future of man and machines Proc. of the 2017 Int.
Conf. on Computing, Networking and Informatics (ICCNI)
[5] Davey T 2017 Artificial intelligence and the future of work: an interview with Moshe Vardi
Future of Life https://futureoflife.org/2017/06/14/artificial-intelligence-and-the-future-of-
work-an-interview-with-moshe-vardi/
[6] Hawking S, Tegmark M, Russell S and Wilczek F 2014 Transcending complacency on super-
intelligent machines Huffington Post http://huffingtonpost.com/stephen-hawking/artificial-
intelligence_b_5174265.html
[7] Lorencik D and Sincak P 2013 Cloud robotics: current trends and possible use as a service
Proc. of the IEEE 11th Int. Symp. on Applied Machine Intelligence and Informatics (SAMI
2013) pp 85–8
[8] Guizzo E 2011 Robots with their heads in the clouds IEEE Spectrum http://spectrum.ieee.
org/robotics/humanoids/robots-with-their-heads-in-the-clouds
[9] Hu G, Tay W P and Wen Y 2012 Cloud robotics: architecture, challenges and applications
IEEE Network 26 21–8
[10] RoboEarth Project http://roboearth.org/
[11] The Robotic Operating System (ROS) http://ros.org/wiki/
[12] Arumugam R et al 2010 DAvinCi: a cloud computing framework for service robots Proc. of
the 2010 IEEE Int. Conf. on Robotics and Automation (ICRA) (3–7 May) pp 3084–9
[13] Apache Hadoop http://hadoop.apache.org/
[14] O’Leary D E 2017 Emerging white-collar robotics: the case of Watson analytics IEEE Intell.
Syst. 32 63–7
[15] IBM 2014 Introduction to IBM Watson: Analytics, Data Loading and Data Quality, IBM
Document Version 2.0, December 16, 2014 https://docplayer.net/10919185-Introduction-to-
ibm-watson-analytics-data-loading-and-data-quality.html
[16] Yukitake T 2017 Innovative solutions toward future society with AI, robotics, and IoT Proc.
of 2017 Symp. on VLSI Circuits pp C16–9
[17] Ahmed S H, Kim G and Kim D 2013 Cyber physical system: architecture, applications and
research challenges Proc. of the IFIP Wireless Days Conf. (WD’13) pp 1–5

15-23
 Modern Optimization Methods for Science, Engineering and Technology

[18] Silva A R and Vuran M C 2010 (CPS)2: integration of center pivot systems with wireless
underground sensor networks for autonomous precision agriculture Proc. of the 1st ACM/
IEEE Int. Conf. on Cyber-Physical Systems pp 79–88
[19] Jie Lin et al 2017 A survey on internet of things: architecture, enabling technologies, security
and privacy, and applications IEEE Internet of Things J. 4 1125–42
[20] NIST 2016 NIST & The Smart Grid Accessed: 12 April 2019 https://nist.gov/engineering-
laboratory/smart-grid/about-smart-grid/nist-and-smart-grid
[21] Nayyar A, Batth R S and Nagpal A 2018 Internet of robotic things: driving intelligent
robotics of future—concept, architecture, applications and technologies Proc. of the 2018 4th
Int. Conf. on Computing Sciences pp 151–60
[22] Adebayo A O, Chaubey M S and Numbu L P 2019 Industry 4.0: the fourth industrial
revolution and how it relates to the application of Internet of things (IoT) J.
Multidiscip. Eng. Sci. Stud. 5 2477–82
[23] Jiafu Wan et al 2016 Cloud robotics: current status and open issues IEEE Access 4 2797–807
[24] Ben Kehoe et al 2015 A survey of research on cloud robotics and automation IEEE Trans.
Autom. Sci. Eng. 12 398–409
[25] Mischa Dohler et al 2017 Internet of skills, where robotics meets AI, 5G and the tactile Internet
Proc. of the 2017 European Conf. on Networks and Communications (EuCNC) pp 1–5
[26] Dialani P 2019 AIOPS: the integration of AI and IoT https://analyticsinsight.net/aiops-the-
integration-of-ai-and-iot/
[27] Sarkar S 2017 How to build IoT solutions with big data analytics https://analyticsinsight.net/
how-to-build-iot-solutions-with-big-data-analytics/
[28] Some K 2018 Top 5 machine learning trends of 2018 https://analyticsinsight.net/top-5-
machine-learning-trends-of-2018/
[29] O’Shea K and Nash R 2015 An introduction to convolutional Neural Netw. 2 1–10

15-24
 IOP Publishing

Modern Optimization Methods for Science, Engineering and

Technology
G R Sinha

Chapter 16
Efficacy of genetic algorithms for
computationally intractable problems
Ajay Kulkarni and Sachin Puntambekar

A genetic algorithm is a metaheuristic method that has been proved highly efficient
for searching for optimal solutions for problems belonging to the category of NP-hard
and which are often algorithmically solvable but computationally intractable.
A genetic algorithm is a probabilistic heuristic search technique motivated by the
principle of natural genetic systems. It aims to locate the global optimal solution for
a problem from the given solution space. Due to its population based approach, the
requirement for a fitness function rather than its derivatives and the probabilistic
nature of the operators, genetic algorithms possess the capability of exploring the
search space efficiently and efficaciously and are therefore applied to search for
optimal or near-optimal solutions of various optimization problems.
A genetic algorithm can be viewed as an abstract version of the evolutionary
process which operates on a population of artificial chromosomes. Each chromo-
some represents an encoded version of a candidate solution and is associated with a
fitness value which reflects the eminence of that chromosome as the solution to the
problem. Binary coding is a commonly used technique for encoding the solutions—
in this strategy the solution appears as a bit string which facilitates the subsequent
operations. These candidate solutions are further subjected to operations such as
crossover and mutations to search an offspring population that is expected to be
better than the parent population.
Both crossover and mutation are nondeterministic in nature and are applied with
certain probabilities. The crossover operator is a mechanism for exchanging
information between chromosomes; this operator allows two parent chromosomes
to exchange their genetic characteristics in order to generate two offspring. Single-
point crossover, two-point crossover and uniform crossover are commonly used
exchange mechanisms for binary coded genetic algorithms. The mutation operator is

doi:10.1088/978-0-7503-2404-5ch16 16-1 ª IOP Publishing Ltd 2020

 Modern Optimization Methods for Science, Engineering and Technology

used to introduce new genetic material; it mutates one or more alleles in a

chromosome from their current states which can result in new individuals with
different characteristics. For binary coded chromosomes, bit flip mutation is
normally used. In this procedure, mutation is applied to each bit location in the
chromosomes with a certain predefined probability. Mutation is a subsequent
process of crossover and is applied to the offspring evolved as a result of crossover
process. The process of mutation plays a very vital role in perpetuating the genetic
diversity and thereby avoiding the problem of genetic drift which may lead to
premature convergence. The mutation probability is often kept low so as to avoid
the redundant randomness in the search
Chromosomes generated as the result of recombination are further subjected to
the process of replacement in which the offspring population becomes the new
parent population. Complete replacement, replacement with elitism and steady state
replacement are commonly used replacement schemes. These processes are there-
after iterated through a number of generations until some termination criterion is
satisfied. The termination criterion could be either a fixed number of iterations or
convergence to a best fit solution.
This chapter details the various aspects of genetic algorithms, such as their
operational mechanism, theoretical analysis and ability to deal with other compli-
cacies such as constrained optimization and others.

16.1 Introduction
A genetic algorithm is a heuristic optimization search technique inspired by the
principles of natural genetic systems and aims to seek a global optimum solution for
a problem from the set of candidate solutions in an iterative manner. Due to its
population based approach, the requirement for a fitness function rather than its
derivatives and the probabilistic nature of operators, genetic algorithms possess the
potential to explore the search space efficiently and efficaciously for optimal or near-
optimal solution of the problem under consideration. Genetic algorithms and their
variants have been successfully applied in engineering fields to search for the
solution for optimization problems with significant complexities. The capability of
a genetic algorithm to solve the optimization problems of the nature of NP-hard and
NP-complete problems has been reflected in several research findings cited in the
literature. Genetic algorithms can be viewed as a technique to search for an optimal
solution of a problem by performing repeated iterations of a solution. To quantify
the process of iteration and to implement the concept of natural selection, genetic
algorithms rely on an objective function or fitness function. The fitness function
provides a measure to determine the candidate solution’s relative fitness and is used
by the genetic algorithm to evolve better solutions in subsequent iterations. Another
aspect of genetic algorithms is the concept of population, instead of operating on a
unique solution the genetic algorithm operates simultaneously on a population of
candidate solutions, known as individuals. This implicit parallelism involved in the
genetic algorithm makes it a global search rather than confining it to a local area of
the search space. However, the size of the population is an issue of concern, as small

16-2
 Modern Optimization Methods for Science, Engineering and Technology

population size may lead to premature convergence while large size may result in
excessive computational time.
Also, to imitate the natural selection procedure, the genetic algorithm operates on
the encoded versions of the parameters to be optimized and not on the parameters
themselves. Parameter encoding transforms the actual optimization problem into
combinatorial optimization as the genetic algorithm is essentially a combinatorial
search technique.
A genetic algorithm maintains a population of solutions, known as chromosomes
or individuals, over the search space which is iteratively modified so as to drive this
population towards an optimal or near-optimal solution. Starting with some ran-
domly or heuristically selected initial population, the genetic algorithm generates a
new population at each iteration using the following steps: (i) calculation of a fitness
function value for each chromosome constituting the old (existing) population—this
value reflects the potency of each solution and can be considered as a meaningful
measure to analyze the claim of the algorithm to approach optima in successive
iterations; (ii) using the fitness value as selection measure the individuals are selected
for subsequent procedures of recombination and selection; (iii) selected individuals
(parents) are thereafter subjected to a genetic operation called crossover to generate
probably better new solutions (offspring); (iv) the newly generated solutions are
further subjected to mutation, this operator aims to preserve the genetic diversity in the
vicinity of the candidate solution; and (v) a new population is generated to replace
the existing one. Iterations are carried out until some terminating criterion is met [1].
Recombination and selection operations in genetic algorithms are nondetermin-
istic and are governed by some probabilistic rules instead of some deterministic
procedure; this nondeterministic nature facilitates the objective of preserving the
global explorative properties of the search. An important attribute of genetic
algorithms is that they require only a mathematical function acting as an objective
function for the considered problem and have no dependence on other character-
istics of the fitness function, such as the existence of its derivatives or differ-
entiability. This attribute allows the applicability of this algorithm even for problems
with non-smooth functions. Procedural details of the genetic algorithm are explained
by the flowchart shown in figure 16.1.

16.2 Genetic algorithm implementation

A genetic algorithm is implemented using a number of distinct components. The
main components of a genetic algorithm designed for a particular problem are
genetic representation, initialization, evaluation, selection, recombination (crossover
and mutation), replacement and termination [1–4].
1. Search space: In the case of search algorithms, the search space is defined as
the space containing all the feasible solutions. In graphical notation, each
point in the search space may be viewed as a feasible solution to the problem
under consideration and can be characterized by its fitness value for that
problem. Search techniques such as genetic algorithms try to locate the
optimal solutions among the solutions constituting the search space.

16-3
 Modern Optimization Methods for Science, Engineering and Technology

Start

Initialize Population;

Evaluate Fitness

Perform Selection

Crossover and Mutation to

generate offspring

Count=Count+1 Evaluate fitness of offspring

Select new population

Termination
condition satisfied
No

Yes
Optimal Solution

End

Figure 16.1. Flowchart of the working principle of genetic algorithms.

2. Genetic representation/encoding: To facilitate the simulation of the process of

natural evolution and selection, a genetic algorithm operates on the encoded
versions of the solutions, known as chromosomes. These artificial chromo-
somes closely resemble natural chromosomes, which are convoluted threads
of DNA (deoxyribonucleic acid). Traits which are coded versions obtained
by some combination of DNA, can be viewed as the constituting elements of
a chromosome; various traits governing the hereditary characteristics of an
individual are strung longitudinally along a chromosome. Artificial chromo-
somes, encoded versions of the solutions, can be viewed as the stringing of
various features of the solution along its length. These encoded versions are
known as the genotype and its decoded counterpart is called the phenotype.
Regarding parameter encoding, genetic algorithms use the following typical
vocabulary: individual solution strings in their encoded version are referred
to as chromosomes, whereas a particular part of the chromosomes represent-
ing a feature or characteristics is known as a gene, the position and value of a
particular gene are known as the locus and allele, respectively.

16-4
 Modern Optimization Methods for Science, Engineering and Technology

In genetic algorithms, genetic operators such as crossover and mutation

operate on genotype, whereas evaluation and selection of individuals is
carried out using phenotype, so the mapping function should be selected such
that there exists a close resemblance between the genotype and the character-
istics of phenotypic space. There exist several encoding schemes in the genetic
algorithm literature which are often problem specific, e.g. real number
coding is suitable for optimization problems subjected to constraints,
whereas binary or integer number encoding is appropriate for combinatorial
optimization problems.
Some of the main coding schemes are:
Binary coding:
• In genetic algorithms, binary representation of chromosomes is the
most commonly used encoding scheme. In binary representation, the
string is composed of zeros and ones. The length of the string is
governed by the precision desired for the solution.
In a genetic algorithm each variable (allele) of the optimization
problem is encoded as a binary string of appropriate length (gene) and
the chromosome is constructed by concatenating all the substrings
(genes) together. For a two variable optimization problem with four bit
substrings, a chromosome can be represented as (x1, x2 ) → (s1, s2 ) =
00101100. Binary encoding follows a linear mapping rule:

β
xiu − xil
xi = xil + ∑ γj 2 j , (16.1)
2β − 1 j = 0

where xil and xiu represent the lower and upper bounds of xi ,
respectively, β represents the length of the binary representation of xi ,
β
whereas ∑ j = 0 γj 2 j is the decoded value of the binary substring. The
accuracy obtained with substrings of length β provides accuracy of the
order of 1/2β of the search space, so arbitrary precision can be achieved
by using strings of appropriate length. Binary representation is a
commonly used encoding scheme in classical genetic algorithms and
is normally supported by two arguments. First, binary alphabets
maximize the level of implicit parallelism. Second, the encoding scheme
associated with a high cardinal alphabet requires a larger population
size for the effective exploration of the search space which often reduces
the computational efficiency of the algorithms. However, the binary
representation has not been found to be appropriate when dealing with
a continuous search space with large dimensions and when a numerical
precision of considerable degree is required, as it handles continuous
problems as discrete ones. Application of binary coding for the problems
of a continuous domain requires decoding and repair algorithms to

16-5
 Modern Optimization Methods for Science, Engineering and Technology

transform the solution searched for by the genetic algorithm into a viable
solution of the original problem. Designing of decoding and repair
algorithms often complicates the implementation of genetic algorithms.
Real coding:
• The real coding scheme has been found to be appropriate for opti-
mization problems of continuous nature; it uses a floating point
representation for representing various parameters to be optimized.
As a result, each chromosomal string appears as a vector of floating
point numbers. The precision in this scheme is often decided by the
nature of the problem to be solved. A common approach in this scheme
for constructing chromosomes is to represent each variable to be
optimized by a gene and to keep the length of the chromosomal string
the same as that of the solution vector. Also, the value of a gene
representing a particular variable is restricted to the interval defined for
that variable and the genetic operators forced to preserve this require-
ment. Float point representation of parameters is capable of represent-
ing the large domains, whereas in binary implementations, an increase
in domain size results in a loss of precision with a fixed length of the
chromosomes. Real coding also offers the feature of local tuning of the
solution by exploiting the graduality of the functions with continuous
variables, which is often difficult in the case of binary coding due to the
problem of the Hamming cliff. In real coding, as the genotype and
phenotype are identical, there is no requirement of coding and decoding
and so the speed of the algorithm increases.

3. Initialization: The selection of the initial population is a vital issue in genetic

algorithms; it is mainly concerned with the population size and method for
selection of individuals to constitute a population. An optimum population
size is always required for the effective and efficient implementation of
genetic algorithms, as a small population may not provide sufficient space to
explore the search space effectively, whereas a large population may impair
the efficiency of the algorithm by consuming excessive computational time.
According to Goldberg [1], the size of the population is directly related to the
difficulty n of the problem, whereas according to Reeves [2] the minimum
population size to initiate the meaningful search is governed by the principle
that it is possible to traverse every feasible solution in the search space
starting from the initial population by using the crossover mechanism only.
The selection of a good initial population has a better possibility of finding a
good solution. Diversity can be considered as the key issue when selecting the
initial population as it ensures the effective exploration of the search space
and is particularly useful in the case of an unknown or weakly known search
space. One commonly used approach is to randomly select the initial
population; however, it may not provide the uniform coverage of the search
space. For reducing the computational time, one normally used approach is

16-6
 Modern Optimization Methods for Science, Engineering and Technology

to seed some high quality solutions into the initial population. However, this
inclusion induces the possibility of premature convergence.
4. Evaluation/fitness: As genetic algorithms try to impersonate the survival of
the fittest principle of nature to formulate a search process, it is necessary to
evaluate the fitness of a potential solution relative to others. This evaluation
is carried out using the fitness function. The fitness function is a mathemat-
ical evaluation function that computes and reflects the superiority of the
chromosome as a solution to the problem under consideration. The fitness
function allocates reproductive traits to an individual and acts as a measure
to be maximized in subsequent iterations. It implies that individuals with a
higher fitness function value usually have a better chance of participating in
subsequent stages of the algorithm. The algorithm is structured such that it
aims to increase the average population fitness in an iterative manner.
5. Selection: In a genetic algorithm, the fitness function value is used as an
assessor of the quality of the solution represented by a chromosome and the
average fitness is considered as a qualitative measure of the population
comprising a stipulated number of chromosomes. The selection mechanism
of a genetic algorithm utilizes the fitness as a guide to select individuals to
form a mating pool so as to participate in the process of evolution. Due to
fitness based criteria, the chromosomes with better fitness have a relatively
higher probability of being selected for subsequent operations than others.
As the selection mechanism is normally carried out with replacement, fit
chromosomes have a chance of being selected more than once, thus fitter
chromosomes participate more frequently in the mating process and posess a
relatively higher probability of surviving in succeeding iterations. During the
selection process key factors that are required to be balanced are selection
pressure and genetic diversity. Selection pressure describes the tendency to
select only the best individuals of the current population to participate in
subsequent steps; selection pressure controls the rate of convergence of the
genetic algorithm towards the optimum. Genetic diversity refers to the
maintenance of diversity in the solution population and is required to ensure
the effective exploration of solution space, which is often necessary during
the earlier stages of the optimization process. Very high selection pressure
results in a loss of genetic diversity due to which the genetic algorithm is
likely to undergo premature convergence with some local optimum. With too
low a selection pressure, the genetic algorithm may not converge to an
optimum solution in adequate computational time. In order to ensure the
convergence of a genetic algorithm to a global optimum solution in
reasonable time, an appropriate balance between the selective pressure and
genetic diversity is required to be maintained by the selection mechanism.
There exist several selection mechanisms in the genetic algorithm literature,
such as roulette wheel selection, stochastic reminder sampling, stochastic
universal sampling, linear rank selection, exponential rank selection, tourna-
ment selection and truncation selection [1]. Roulette wheel (fitness pro-
portional) selection is the conventional selection method in which each

16-7
 Modern Optimization Methods for Science, Engineering and Technology

candidate is assigned a slot on the roulette wheel with size proportional to its
fitness, thus the candidates with higher fitness have a larger slot size than the
less fit individuals. The roulette wheel is spun an appropriate number of
times, each time selecting a candidate pointed at by the wheel pointer. As per
the scheme suggested by Goldberg [1], the roulette wheel selection mecha-
nism to select n individuals from a population size of n can be implemented
through the following steps:
i. Calculate the fitness value, fi , for all the candidate solutions constitut-
ing the population.
ii. Calculate the probability (slot size), pi of selection for each solution
i
pi = fi /f , where f = ∑f j .
j=1
i
iii. Calculate the cumulative probability, qi = ∑pj .
j=1
iv. Generate a random number, r ∈ (0, 1].
v. Select the individual s1 if r < q1, otherwise select si if qi−1 < r ⩽ qi.
vi. Repeat steps iv and v n times to select n candidates in the mating pool.

In n experiments of the roulette wheel, an individual with probability pi is

expected to make npi copies in the mating pool. The computational complex-
ity of this procedure is of the order of O(log n ) steps for a population of size
n. In roulette wheel selection, the fitness of an individual acts as a governing
factor to decide the selection probability of an individual, the dominating
individual has a chance of being selected more than once during the trials
and thus this technique is associated with high selection pressure which may
result in premature convergence leading to local optima due to the possible
selection of a dominating individual. Also, this technique is associated with a
high degree of stochastic variability and the actual number of times a
chromosome is selected may vary from its expected value. There exist
some variants of this scheme, such as stochastic remainder sampling and
stochastic universal sampling, which circumvent these drawbacks to a certain
extent. Stochastic remainder selection calculates the reproduction count of
each individual as esi = nfi /f and uses the integer portion of esi as the
deterministic count of si in the mating pool. The fractional parts of esi are
then used to fill the remaining pool in a probabilistic way. The probabilistic
mechanism is carried out either without replacement or with replacement, in
the former case each remainder is used to bias the flip of a coin to decide
whether the individual receives another copy or not, whereas in the latter
case the fractional parts are used to size the slots of the roulette wheel
selection process. Stochastic universal selection, which corresponds to
systematic sampling, is carried out by sizing the slots of a roulette wheel
with n equally spaced pointers. A single spin of the wheel simultaneously
produces the selection count of all the chromosomes which is equal to the

16-8
 Modern Optimization Methods for Science, Engineering and Technology

number of pointers that points to the slots associated with respective

chromosomes. Linear ranking selection provides the flexibility of tuning
the selection pressure and thereby deals with the problem of premature
convergence in an effective manner. In rank selection, individuals are sorted
and ranked according to their fitness values, rank n is assigned to the best
individual and rank 1 to the worst. As per Colin [2], the selection probability
is assigned linearly to the individuals according to their rank:
p(j ) = α + βj ; j ∈ {1, 2, ⋯ , n}, (16.2)
2n − φ(n + 1) 2(φ − 1) n
where α = n(n − 1)
and β = n(n − 1)
with ∑ j = 1α + βj = 1.
prob[best chrosome]
Here, φ is the selection pressure defined as φ = prob[average chrosome]
with
1 < φ ⩽ 2. With this framework, corresponding to a pseudo-random number
r, j can be calculated as

−(2α + β ) ± (2α + β )2 + 4βr

j= . (16.3)
2β
This process can be carried out in O(1) time. In linear ranking selection, the
sorting of the individuals is a crucial operation and complexity of sorting
often dominates the complexity of the algorithm, thus complexity is
O(n log n ). Exponential ranking selection is similar to linear ranking selec-
tion except for the fact that the exponentially weighting technique is used to
decide the probabilities of the individuals:
c − 1 n −j
p(j ) = c ; j ∈ {1, 2, ⋯ , n}; 0 < c < 1. (16.4)
cn − 1
In the tournament selection scheme, a set of k individuals is randomly
selected from the population and the best individual from that set is further
selected for subsequent processing and the process is repeated n times. Here,
k is known as tournament size which may be selected arbitrarily, but in
general the process is carried out with k = 2, which is known as a binary
tournament. In tournament selection, group size k is used as the controlling
parameter to tune the selection pressure; group size 2 implies the weakest
selection pressure. The time complexity of tournament selection is of the
order of O(n ). In truncation selection a certain portion q of the fittest
candidates are selected and are reproduced 1/q times to maintain appropriate
population size. This technique often results in premature convergence due to
deterministic elimination of less fit candidates during the selection procedure.
As truncation selection involves sorting, its time complexity is of the order
O(n log n ).
6. Recombination: Recombination is concerned with the generation of new
individuals by recombining the chromosomes selected from a source pop-
ulation. New individuals thus generated may become the members of a

16-9
 Modern Optimization Methods for Science, Engineering and Technology

successor population. Crossover and mutation are two genetic operators

which constitute the process of recombination. The key idea behind the
recombination process is to simulate the exchange of genetic material that
takes place when organisms reproduce as well as the changes in genetic
characteristics that may occur after fertilization. As the selection of individ-
uals to participate in the process of recombination is biased towards higher
fitness, there exists a fair chance of the evolution of chromosomes with
improved fitness. In a genetic algorithm, these operators are dealt with in a
nondeterministic way, these operators act with a predetermined probability
and their exact outcomes are also nondeterministic. The crossover operator is
one of the defining characteristics of the genetic algorithm responsible for the
introduction of new genetic material. The crossover mechanism allows two
parent chromosomes from the mating pool to exchange their characteristics
in some specified manner so as to evolve two offsprings with different
characteristics, with the possibility that the crossover of good chromosomes
may result in the evolution of better individuals. In general, the crossover
operator is applied to a pair of chromosomes in some probabilistic sense and
the likelihood of crossover being applied is controlled by the crossover rate
or crossover probability pc . Definitions of the crossover operator are largely
representation dependent. In the case of binary coded genetic algorithms,
commonly used crossover techniques are single-point crossover, two-point
crossover and uniform crossover. In single-point crossover, a crossover point
is randomly selected and the characteristics of the parent chromosomes are
exchanged to construct the child chromosomes, as illustrated below.

Parent one: 10010010

Parent two: 10110110
Crossover point: ↑
Child one: 10010110
Child two: 10110010

In two-point crossover, two crossover points are selected along the length of
the chromosomes and the characteristics of the two parents are exchanged
between these two points to generate two offspring.

Parent one: 10010010

Parent two: 10110110
Crossover points: ↑ ↑
Child one: 10110110
Child two: 10010010

In the case of uniform crossover, the crossover operator is expressed as a

binary string or masks of the same length as that of the chromosome. An
offspring is generated by copying either of the parent alleles at each locus

16-10
 Modern Optimization Methods for Science, Engineering and Technology

according to the ones and zeros in the mask—if at a particular locus the mask
bit is 1, then the allele is copied from parent one, and if the mask bit is 0 then
the allele is copied from parent two. The second offspring is generated by
reversing the process.

Parent one: 10010010

Parent two: 10110110
Mask: 10011010
Child one: 10110110
Child two: 10010010

Another genetic operator used in genetic algorithms to introduce new genetic

material is mutation. Mutation flips one or more alleles in a chromosome
from its current state which can result in new individuals with different
characteristics. The mutation operator thus introduces genetic diversity
whenever the population approaches homogeneity because of repeated use
of selection and crossover operators and the individuals resulting from the
process of mutation may even lead the algorithm towards some better
solution. Mutation is applied to the evolved offspring as a result of the
crossover process. The process of mutation is applied in a probabilistic sense
and is controlled by the mutation rate or mutation probability pm . Mutation
probability is often kept low in order to avoid redundant randomness in the
search. In the case of binary coded genetic algorithms, the mutation operator
is applied to each bit in the chromosome. Corresponding to a bit position,
this procedure involves the generation of a random number in the interval
[0, 1] which is thereafter compared to the mutation probability pm . If the
random number is greater than the mutation probability, no mutation is
applied at that position, and in case the value of the random number is lower,
the bit value is flipped from 0 to 1 or vice versa. Mutation plays a vital role in
maintaining the genetic diversity and thereby avoiding the problem of genetic
drift which may lead to premature convergence. It has been observed that in
the absence of a mutation operator, a finite size population always converges
to a single genotype, this phenomenon is known as genetic drift. When all the
individuals converge to a single genotype, crossover does not contribute to
the search and so the mutation operator is required to maintain diversity.
7. Replacement: After the process of recombination, the next step is replace-
ment, which describes the technique for the generation of a successor
population to replace the existing one. There exist several replacement
techniques in the genetic algorithm literature, which primarily emphasize
the issues of selection pressure and genetic diversity. The classical replace-
ment scheme, proposed by Holland, is based on a generational approach
which favors complete replacement, where the successor population is
entirely composed of the chromosomes generated by the process of recombi-
nation. This complete replacement scheme is relatively easy to implement

16-11
 Modern Optimization Methods for Science, Engineering and Technology

and is free from the issues of diversity and selection pressure; the drawback
associated with this technique is that it contains the risk of discarding the
good solutions from the existing population. An improved version of the
complete replacement scheme is replacement with elitism; this method
emphasizes preserving one or two of the best individuals from the existing
population and replacing the others. Elitism speeds up the performance of
the genetic algorithm but exerts a high selection pressure due to the
deterministic selection of relatively fit candidates. Another scheme for the
replacement is steady state or incremental replacement, which introduces
the concept of population overlapping, in this scheme only a fraction G
(known as the generation gap) of the existing population is replaced.
Selection of individuals from the existing population for replacement is an
issue of concern in steady state replacement. One technique is the replace-
ment of the worst individuals; however, it exerts a strong selection pressure
and often requires a large population size or high mutation rate to maintain
diversity. Another replacement scheme, which maintains the considerable
degree of diversity, is steady state with no duplicates. In this case, offspring
are not included in the population if they are mere duplicates of the existing
individuals. Other evolution strategy based replacement schemes are the
(μ, λ ) and (μ + λ ) schemes. In the (μ, λ ) scheme, μ( >λ ) individuals are
generated from the λ parents and the best λ of these newly generated
individuals are selected as the successor population. In the (μ + λ ) scheme,
μ offspring are generated and combined with λ parents to generate a set of
(μ + λ ) individuals; from this set the best λ individuals are selected to
constitute the successor population. This scheme also exerts a high selection
pressure and requires a high mutation probability to maintain divergence.
8. Termination: The processes of selection, recombination and replacement
are iterated until some terminating criterion is satisfied. Commonly used
termination criteria are either population convergence criteria where almost
all the solutions become identical or nearly identical. Another is the
improvement in best fitness score with iterations which emphasizes the
accuracy of the solution, and the third is a fixed number of iterations with
a reasonable compromise between population convergence, accuracy of the
solution and computation time.

16.3 Convergence analysis of the genetic algorithm

A correctly designed and implemented genetic algorithm population with increasing
fitness is supposed to evolve over generations and ultimately culminate in a global
optimum. Theoretical investigation of the convergence characteristics of genetic
algorithms has been undertaken by several researchers. Attempts to develop a
theoretical framework to explain the chromosome convergence in genetic algorithms
have mainly resulted in the schema theorem [5]. The schema theorem provides
a theoretical framework to analyze the convergence properties of the genetic
algorithm. The schema theorem is based on the conception of schema which is

16-12
 Modern Optimization Methods for Science, Engineering and Technology

obtained by fixing the allele of specific chromosome loci, thus it can be viewed as a
typical pattern which could be observed in some chromosomes. A schema thus
represents a similarity pattern which describes a subset of individuals with similar
features in some positions. A schema is a string of {1, 0, *} where a typical pattern is
described by using {1, 0} and * is used at the chromosome loci kept unspecified by
the pattern. For example, a schema λ = 101 * *011 describes a subset of chromo-
somes with elements {10100011, 10101011, 10110011, 10111011}, i.e. the elements
of this subset belong to λ . For a schema, commonly defined terms are the order and
length of schema. The order of λ , o(λ ) is the number of defined bits, whereas length λ ,
lλ is the difference of the allele positions of the first and last defined bits. For
λ = 101 * *011, o(λ ) = 6 and lλ = 7.
The schema theorem provides an explanation of how the schemata featured in
relatively fit chromosomes possess greater prospects of propagating through
successive populations with the evolution of the genetic algorithm. The theorem
can be formally stated as follows.
Schema theorem. Let λ represent a schema for binary coded chromosomes of
length L and let m(λ , t ) be the number of chromosomes in the current population (t )
which belongs to λ . Then the expected number of chromosomes of λ in the next
generation (t + 1) is given by the formula
f (λ ) ⎛ l ⎞
m(λ , t + 1) ⩾ m(λ , t ) ⎜1 − pc λ ⎟(1 − pm )o(λ ) , (16.5)
f ⎝ L − 1⎠
where f (λ ) is the average fitness of chromosomes of λ and f is the average fitness of
f (λ )
the population. Here, the term m(λ , t ) f represents the expected number of
chromosomes of λ after the selection process (fitness proportionate selection). It
implies that schema with relative fitness above one will continue to increase, whereas
with relative fitness less than one they will decrease; this is due to the selective
l
pressure in fitness proportionate selection. Term (1 − pc L −λ 1 ) represents the prob-
ability of survival of chromosomes of λ after the crossover process, this term is high
l
when lλ is low and when lλ approaches L − 1, (1 − pc L −λ 1 ) approaches (1 − pc ), which
indicates that the disruption of the chromosomes is almost certainly due to cross-
over. Term (1 − pm )o(λ) represents the probability of survival of chromosomes of λ
after the mutation process, and is high when o(λ ) is low.
It indicates that schemata with short length, low order and above average fitness
receive increasing trials in successive iterations. These schemata are known as
building blocks and can be viewed as the partial solutions which provide a higher
fitness value to the chromosomes which contain these building blocks. It means that
a chromosome composed of such building blocks will be a near-optimal solution.
The consequence of this theoretical framework is the building block hypothesis,
which states that the genetic algorithm can approach the near-optimal solution
through appropriate selection and combination of building blocks. The notion of
building blocks is particularly useful when a heuristic insight of the search problem is
available. During the process of encoding, the desired factors could be encoded in

16-13
 Modern Optimization Methods for Science, Engineering and Technology

neighboring loci, thereby having low order schema. Crossover sites could also be
adjusted heuristically such that chromosome disruption occurs only at locations
which are worth it for that problem. The procedure thus results in increasing the
probability that the defined process will productively juxtapose building blocks and
rapidly approach an optimal solution [5].

16.4 Key factors

There exist some crucial characteristics which offer genetic algorithms an efficacy to
deal with optimization problems which are computationally intractable and asso-
ciated with other complexities, such as constraint satisfaction, multimodal solution
and others. Significant progress has taken place in this field and the following
subsections cover these issues [1–5].

16.4.1 Exploitation and exploration

Search techniques are considered as generalized solution methods for the problems
which belong to the nondeterministic category. Search techniques can be categorized
as blind strategies or heuristic strategies. Blind search strategies do not use any
additional information obtained about the problem domain to guide the further
search, whereas heuristic search strategies use additional information to guide
the search along the best search directions. To ensure an efficient and effective
search, search techniques are required to address issues such as exploration and
exploitation. Exploration can be considered as the process of seeking solutions in
new areas of a search space, whereas exploitation is the process of finding solutions
in the vicinity of the already existing solution in order to refine it. Hill-climbing and
random search are examples of pure exploitation and exploratory strategies,
respectively, whereas search techniques such as genetic algorithms maintain an
appropriate balance between exploration and exploitation of the search space so as
to ensure the optimal or near-optimal solution. In genetic algorithms, issues of
exploitation and exploration are mainly controlled by the process of selection and
selection operators. The extent of exploration or exploitation can be controlled by
selection processes by causing a variation in selection pressure. Search techniques
with low selection pressure are exploratory in nature, whereas those with higher
selection pressure are inclined towards exploitation. Selection of an appropriate
selection pressure may result in a proper balance between exploration and
exploitation. Mutation is also a way to control the exploration and exploitation.
Mutation, as in the case of a binary coded genetic algorithm, flips the bits
constituting a chromosome with a given probability. By keeping a control over
the mutation probability, this genetic operator can be used to maintain the genetic
diversity of the population. Mutation with high mutation probability acts as an
exploration operator and explores new solutions, whereas with small mutation
probability acts as an exploitation operator and preserves most of the genetic
material. However, as a high mutation rate introduces blind diversity, thereby
preventing the population from converging, the use of mutation as an exploration
operator is often considered inefficient. Crossover is another search operator used in

16-14
 Modern Optimization Methods for Science, Engineering and Technology

genetic algorithms; however, the nature of the search performed by the crossover
operator is mainly determined by the diversity of the population. For random and
diverse populations, normally at the commencement of the search, crossover
performs a widespread search and explores the search space, whereas for similar
populations, normally after several iterations when high fitness solutions develop, it
performs the search in the neighborhood of existing solutions. The probability of the
crossover operator also affects the exploration and exploitation—high crossover
probability results in the quick introduction of new solutions in the search space, but
there is a chance of missing good solutions and of failing to exploit existing
solutions. With low crossover probability, most of the existing solutions are
preserved and a considerable search area may be left unexplored. Thus the crossover
and mutation probabilities can be used as controlling parameters to achieve balance
between exploration and exploitation. Population diversity is also considered as a
measure to achieve balance between exploration and exploitation. High genetic
diversity indicates that the genetic algorithm was in the phase of exploration, whereas
low genetic diversity indicates that the algorithm was in the phase of exploitation.

16.4.2 Constrained optimization

The existence of constraints in any system often acts as an obstacle in the application
of search techniques such as genetic algorithms for optimizing the problems as
genetic operators (search operators), such cases may result in some infeasible
solutions. A few methods are highlighted in the genetic algorithm literature to
deal with infeasible solutions. The first technique, the penalty function method,
allows the genetic operators to generate infeasible solutions which are later on
penalized by a modification of the objective function. The penalty function method
discriminates an infeasible solution according to its violation of each constraint, thus
it transforms the constrained problem into an unconstrained optimization problem.
However, the penalty function method is highly sensitive to the penalty parameters
(weights) selected for various constraints; an inappropriate choice of penalty
parameters may either result in inconsequential penalty terms or highly consequen-
tial penalty terms as compared to objective functions. The former results in
inordinate exploration of infeasible solutions, whereas the latter may cause
premature convergence to a feasible but unacceptable solution. As in such cases, a
feasible solution, once generated, may dominate the mating pool with a rapid
elimination of infeasible solutions before the generation of any other feasible
solution. Other techniques emphasize the use of repair algorithms to correct the
infeasible solutions or the application of some decoder scheme to amplify the
probability of generation of feasible solutions. Repair algorithms are normally used
in the case of combinatorial optimization problems where search operators some-
times result in some illegal solution i.e. a solution which cannot be decoded.
However, decoder and repair algorithms are highly problem specific and are difficult
to design. Also, the decoder is often required to be augmented with some repair
algorithm or penalty function as there still exists some possibility of the generation
of infeasible solutions [1].

16-15
 Modern Optimization Methods for Science, Engineering and Technology

16.4.3 Multimodal optimization

Multimodality in the case of optimization problems is referred to as the existence of
more than one optimal solution and multimodal optimization deals with the search
for multiple optima. Knowledge of multiple optima provides the flexibility to choose
an optimal solution under the constraints of cost, realizablility, etc. Methods which
are used in genetic algorithms to allocate and/or maintain multiple optimal or
suboptimal solutions are referred as niching methods. Genetic algorithms, due to the
nature of their selection operators, often converge to a single solution, even though
multiple solutions of the same quality exist. Earlier techniques, known as temporal
niching methods, were based on the concept of locating multiple optima in a
sequential manner for which a traditional algorithm is run several times in order to
find a different optima each time. Such methods, however, require certain knowl-
edge of the basin of attraction of optima, otherwise the runs may converge to the
same optimum. Since the genetic algorithm is a population based approach, by
modifying the framework of traditional genetic algorithms, the coexistence of a large
number of optimal solutions may be made possible, thereby evolving multiple
optimal solutions simultaneously. There exist niching methods such as fitness
sharing and crowding, which operate on the concept of preserving population
diversity which eventually leads to simultaneous evolution of multiple optima. These
niching methods use a distance metric defined over the search space to distinguish
the similarity of individuals. Crowding techniques use this measurement for the
replacement of similar individuals, whereas the fitness sharing method uses the
metric for de-rating an individual’s fitness by a particular amount in accordance with
the number of similar individuals in the population [3, 4].
De Jong [3] introduced a crowding factor model for multimodal optimization.
For a given generation gap G, G.popsize individuals are selected to create an equal
number of offspring. For each of the offspring, CF individuals are randomly selected
from the current population to constitute a sample; CF is known as the crowding
factor. The offspring then replaces the most similar individual of this sample.
The hamming-distance between two individuals is used as the measurement for
similarity.
Restricted tournament selection is a variant of the crowding method in which the
parents are selected randomly from the population. For each offspring, CF
individuals are randomly selected and the offspring then competes with the most
similar individual from the selected ones and in the case that the offspring is better, it
replaces that individual. The struggle genetic algorithm is an another variant of the
crowding method which is functionally equivalent to restricted tournament selection
with crowding factor equal to the population size.
The fitness sharing method uses the concept of sharing function to affect the
population diversity and to maintain sub-populations. Sharing function reflects the
extent of sharing that can be done between two strings. In the case that the two
strings are identical, the sharing function returns a value of 1 and if the distance
between two strings exceeds a certain threshold distance (niche radius) the sharing

16-16
 Modern Optimization Methods for Science, Engineering and Technology

function returns a value of 0, implying that the individuals are in different niches. A
commonly used sharing function is
⎧ d
⎪1 − ij ; dij < ρ
Sh(dij ) = ⎨ ρ , (16.6)
⎪
⎩ 0; otherwise

where dij is the distance between the ith and jth string in phenotype space and ρ is the
niche radius. For every string, a sharing function is calculated for other strings and
these sharing function values are added to obtain a niche count mi = ∑j dij .
Thereafter the shared fitness value of an individual is computed as fi* = fi /mi and
is used in the selection mechanism instead of fi . Now, if there exist lesser strings from
any optimal solution then these strings will have a lower niche count and higher
shared fitness than the strings corresponding to other optima. As the selection
operator will now emphasize these strings, the number of strings form this optima
will increase in the subsequent iterations. Thus this method allows the simultaneous
existence of multiple optima in the population. One commonly used replacement in
this case is the overlapping population scheme where new offspring are first added to
the current population and shared fitness values are computed for all the candidates
and individuals (equal to the number of offspring) with the worst shared fitness
values then eliminated from the population and the shared fitness values are
recalculated for the next iteration [1, 4].

16.4.4 Multi-objective optimization

Multi-objective optimization emphasizes that simultaneous optimization of more
than one objective function is required. There exist two approaches for multi-
objective optimization. The first technique is to convert the problem into a single
objective optimization problem by suitably combining the individual objective
function into a composite one. A commonly used approach to develop the single
objective function is the weighted combination of individual functions. However,
this method is highly sensitive to the weight parameters and a slight perturbation in
the weights may lead to a different solution. The second technique deals with the
determination of a set of solutions (known as Pareto optimal solutions) obtained by
considering different single objective functions synthesized by some technique. A
commonly used approach to determine the Pareto optimal solution is the concept of
nondomination. Nondominated solutions are those solutions which are superior to
others in at least one objective function. Solutions which are nondominated with
respect to each other constitute the Pareto optimal set. Due to their population
based approach, genetic algorithms are well suited for the multi-objective optimi-
zation problems as a classical genetic algorithm can be extended to maintain a
diverse set of solutions. There exist several multi-objective optimization algorithms,
such as the weighted sum approach, vector evaluated genetic algorithms (VEGA:
altering the objective function approach) and Pareto ranking approaches, such as

16-17
 Modern Optimization Methods for Science, Engineering and Technology

niched Pareto genetic algorithm, random weighted genetic algorithm (RWGA),

nondominated sorting genetic algorithm (NSGA), nondominated sorting genetic
algorithm with elitism (NSGA-II) [4].

16.5 Concluding remarks

Genetic algorithms have been successfully applied to search for the optimal solution
for combinatorial optimization problems in various fields, and successful application
of genetic algorithms to NP-hard and NP-complete problems has been reported
in several research findings [5–18]. Holistic detail of the application of genetic
algorithms is a formidable task and can be found from the references cited and the
references within these. Some of the broad application areas of genetic algorithms
are machine learning, economics, immune systems, ecology, social systems and
others. Because of its promising performance, the application areas of genetic
algorithms have grown rapidly in recent years. The genetic algorithm community
and researchers from other realms are constantly working to develop new variants of
the genetic algorithm which would provide optimization solutions to various
problems with an efficient time and space complexity approach.

References
[1] Goldberg D S 1989 Genetic Algorithms in Search, Optimization and Machine Learning
(Boston, MA: Addison-Wesley Longman)
[2] Reeves C R and Rowe J E 2002 Genetic Algorithms: Principles and Perspectives: A Guide to
GA Theory (Norwell, MA: Kluwer Academic)
[3] Jong K D 1988 Learning with genetic algorithms: an overview Mach. Learn. 3 121–38
[4] Deb K 1999 Introduction to genetic algorithms Sadhana 24 293–315
[5] McCall J 2005 Genetic algorithm for modeling and optimization J. Comput. Appl. Math. 184
205–22
[6] Tang K S, Man K F, Kwong S and He Q 1996 Genetic algorithms and their applications
IEEE Process Mag. 22–37
[7] Renders J M and Flasse S P 1996 Hybrid methods using genetic algorithms for global
optimization IEEE Trans. Syst. Man. Cyber. B 26 243–58
[8] Blanco A, Delgado M and Pegalajar M C 2001 A real coded genetic algorithm for training
recurrent neural networks Neural Netw. 14 93–105
[9] Alonge F, Dlppolito F and Raimondi F M 2003 System identification via optimized wavelet
based neural networks Proc. IEEE Contol Theory Appl. 150 147–54
[10] Sahoo D and Dulikravich G S 2006 Evolutionary wavelet neural network for large
scale estimation in optimization Proc Multidisciplinary Analysis and Optimization Conf.
(Portsmouth, VA) pp 1–11
[11] Awad M 2009 Optimization RBFNN parameters using genetic algorithms: applied on
function approximation Int. J. Comput. Sci. Secur. 4 295–307
[12] Shou-sheng L and Yong D 2010 An evolutionary wavelet network and its training method
Int. Conf. on Computer Application and System Modeling pp 379–83
[13] Aly A A 2011 PID parameters optimization using genetic algorithm technique for electro-
hydraulic servo control system Intell. Control Autom. 2 69–76

16-18
 Modern Optimization Methods for Science, Engineering and Technology

[14] Vishwakarma D D 2012 Genetic algorithm based weight optimization of artificial neural
network Int. J. Adv. Res. Electri. Electron. Instrum. Eng. 1 206–11
[15] Awad M 2014 Using genetic algorithms to optimize wavelet neural networks parameters
for function approximation Int. J. Comput. Sci. Issues 11 256–67
[16] Kulkarni A and Kumar A 2015 Structurally optimized wavelet network based adaptive
control for a class of uncertain underactuated systems with actuator saturation Inter J. Hyb.
Intell. Sys. 12 171–84
[17] Kopel A and Yu X H 2008 Optimize neural network controller design using genetic
algorithm Proc. 7th World Congress on Intelligent Control and Automation (Chongqing,
China) pp 2012–6
[18] Chiroma H, Noor A S M, Abdulkareem S, Abubakar A I, Hermawan A, Qin H, Hamza
M F and Herawan T 2017 Neural networks optimization through genetic algorithm
searches: a review Appl. Math. Inf. Sci.11 1543–64

16-19
 IOP Publishing

Modern Optimization Methods for Science, Engineering and

Technology
G R Sinha

Chapter 17
A novel approach for QoS optimization in
4G cellular networks
Vandana Khare and G R Sinha

A cellular network needs to carry multimedia traffic efficiently during real-time

transmission. Since this network transmits multimedia data, there is a requirement
to adapt the service rate according to the network conditions in order to maintain
successful network operation. Hence, there is a need to allocate resources profi-
ciently. Quality of service (QoS) improvement in such a network is more challeng-
ing, in particular in 4G cellular networks. Many schemes have been proposed for
QoS improvement in cellular networks which combine call admission control (CAC)
and power control schemes. However, CAC with power control is not enough to
provide the most optimal resource utilization. Therefore, it is essential to develop an
efficient QoS based resource allocation and scheduling scheme for real-time multi-
media traffic. The main goal of the proposed research is to develop a QoS based
resource allocation and adaptive rate scheduling (QRAAS) scheme in cellular
networks to improve network reliability, maximize system capacity and improve
system efficiency, and hence to enhance QoS in cellular networks. To implement the
above-mentioned QRAAS scheme, the resource allocation based call admission
control (RACAC) scheme and adaptive rate scheduling (ARS) scheme were
developed.

17.1 Mobile generations

Fourth generation (4G) technology always requires good QoS to maintain real-time
(RT) multimedia traffic in cellular and mobile networks [1–3]. The OFDMA
network is the most extensively used network because earlier networks such as
FDMA and TDMA are not able to support the bandwidth and data rate require-
ments. In the case of the FDMA network, only voice transmission is supported,
whereas TDMA as well as wide-band CDMA (WCDMA) support voice and data

doi:10.1088/978-0-7503-2404-5ch17 17-1 ª IOP Publishing Ltd 2020

 Modern Optimization Methods for Science, Engineering and Technology

transmission. OFDMA is the only network to support fourth generation technology

because of the high data rate of transmission [4–6].
OFDDMA is the only technique which covers more bandwidth and provides
better coverage when compared to other networks. OFDMA actually helps in
spreading the data in the entire bandwidth, which is very large, implemented as
spectrum method, and the spreading takes place just before transmitting the signal.
The spreading factor (SF) plays an important role in the WCDMA network because
it is the ratio of chip rate to data rate [7, 8]. The spreading factor decides the number
of users and depends on the type of modulation. The phase shift keying is of different
types, of which quadrature phase shift keying (QPSK) is very popularly utilized in
the WCDMA network.
The WCDMA network enables multimedia services in wireless broadband
networks. Proper spectrum utilization is also a difficult task since it is a restricted
resource [2]. In RT multimedia, resource allocation is an important task which
should be performed effectively and allotted resources must also be utilized without
causing wastage of bandwidth.
Bandwidth allocation becomes difficult when handling of the data and voice must
be performed concurrently, because voice service is a real-time service and is highly
sensitive to delays whereas data services are not. They are, however, sensitive to data
loss and require the release of error free data packets [3]. Hence, while providing
QoS for voice and data services, several characteristics must be taken into
account [4].
Services with higher data rates are always supported by OFDMA networks.
Packet access can be managed effectively by proper resource allocation in OFDMA
wireless networks. For making use of resource allocation schemes, reliable CAC and
effective scheduling schemes are necessary for assured QoS in OFDMA wireless
networks [5].
Better QoS implies that the network should provide assured service to a particular
application. QoS parameters depend upon specific requirements of an application.
Many real-time applications require assured QoS in wireless cellular and mobile
networks.
In general, the QoS parameters include bandwidth, delay and delay jitter [6]. Due
to the air-interface nature of OFDMA networks, the coverage area and capacity of a
cell strongly depends on the desired QoS in terms of acceptable interference level,
uniform distribution of mobile users and the related time-dependent user traffic
intensity.
In OFDMA wireless networks, resource allocation with CAC, power control and
scheduling play an important role in avoiding congestion and maintaining better
QoS for satisfactory network operation.

17.2 OFDMA networks

OFDMA acts as an air-interface network which is used in forth generation cellular
and mobile communication systems and was produced by the universal mobile
telecommunication system (UMTS). OFDMA is a multiple access technique based

17-2
 Modern Optimization Methods for Science, Engineering and Technology

on orthogonal frequency division multiple access. Predominantly, two types of codes

are used for spreading the data. The channelization code is used to decide how many
bits are needed to spread single-bit information, whereas scrambling code is used to
separate the data. The spreading factor plays a vital role in the OFDMA network.
It decides the number of users during up-link and down-link transmission. The
spreading factor is 4–256 for up-link transmission and 4–512 for down-link trans-
mission. One of the important characteristics of the OFDMA network is the high
data transmission rate at the range of 144 Kbps to 2 Mbps, because of which it
supports multimedia transmission. All the channels in the OFDMA network are of
higher bandwidth, due to the fact that no link is assigned a devoted bandwidth as per
the requirement of the data rate [1]. Channel encoding is performed considering that
any decoder which has the knowledge of the code may extract the required data
from the overall signal bandwidth, which is basically in the form of noise signal.

17.2.1 Limitations of FDMA, TDMA and WCDMA networks

The major limitations of different types of networks are as follows:
1. FDMA and TDMA do not support high data rate transmission.
2. Efficient radio spectrum utilization is not possible.
3. They do not support multimedia traffic transmission.
4. They are unable to support global roaming.
5. They are capable of accommodating only a limited number of users.
6. They suffer high co-channel and adjacent channel interference.

Due to the above-mentioned issues, FDMA and TDMA can only support first
generation (1G), second generation (2G) and third generation (3G) networks.
OFDMA networks are the only solution for the transmission of RT multimedia
traffic in fourth generation (4G) cellular and mobile networks. In OFDMA networks,
a high data rate of transmission is possible for RT multimedia traffic as the data rate
depends on the chip rate and spreading factor. In the case of down-link transmission
the spreading factor is 4–512, which indicates that the network is able to support 512
users for one tower or base station.
However, every tower is accommodating only 200–225 users while the remaining
codes are left free for handoff. Maintaining the QoS and its parameters, i.e.
throughput, delay and power consumption, is thus a very important task.
A basic comparison of the FDMA, TDMA, WCDMA and OFDMA networks,
with respect to different parameters, is given in table 17.1, which depicts the data
rate, bandwidth as well as operating frequencies of the OFDMA network. It can be
observed that only power control is an important factor in maintaining the required
QoS in OFDMA networks.

17.2.2 Features of OFDMA networks

OFDMA networks in mobile communications provide a high speed of data
transmission without failing QoS requirements. The OFDMA characteristics are
given below [3].

17-3
 Modern Optimization Methods for Science, Engineering and Technology

Table 17.1. Basic comparison of FDMA, TDMA, WCDMA and OFDMA networks.

Parameters 1G 2G 3G 4G
Inventions 1980 1993 2006 2012
Access system FDMA TDMA WCDMA OFDMA
Speed (data rates) 2.4 Kbps– 84 Kbps 144 Kbps to 2 Mbps 100 Mbps
23.4 Kbps
Bandwidth 20 kHz 25 MHz 25 MHz 100 MHz
Operating 800 MHz GSM: 900 MHz 1920–1980 MHz for 850 MHz for UL,
frequencies UL and 1800 MHz
2110–2170 MHz for DL
for DL
Limitations Limited capacity Slow speed High power Complicated
consumption hardware
Applications Voice calls SMS Video conferencing High speed
mobile TVGPS operations

• High data rate transmission: A very important feature of OFDMA networks

is that they support a high data rate for transmission of multimedia traffic in
4G mobile communication, which is essential to maintain high speed as well
as throughput for long distance users. It provides a data rate from 10 Mbps to
20 Mbps which can be increased up to 100 Mbps for 4 G using high speed
packet access services.
• Supports RT multimedia traffic: Fourth generation (4G) communication
systems require audio, text and video at the same time. First generation
(1G) supports only text while second generation (2G) supports audio and text
because of the low data rate requirements. The WCDMA network also
supports audio, video and text to some extent. The OFDMA network
provides RT multimedia transmission because of high data rate services.
• Efficient radio spectrum utilization: Currently, because of an increasing number
of users, the effective use of frequency is very important. The OFDMA
network is efficiently managing multimedia traffic for a given radio range in
mobile communications.
• Global roaming: The FDMA and TDMA networks are not able to support
international roaming because of the low data transmission rate. However,
OFDMA networks emphasize the concept of ‘anytime anywhere’ in multi-
media traffic transmission. This is due to the high data transmission rate and
high bandwidth utilization in OFDMA networks.
• High capacity and more flexible communication: The OFDMA network
provides more coverage for mobile users because of the spread spectrum
technique. Also, it avoids co-channel interference between adjacent channels
and maintains the required QoS in the network.
• Complete data security: The functions of OFDMA networks are highly
adaptive to carry out the allocation of resources according to the network

17-4
 Modern Optimization Methods for Science, Engineering and Technology

conditions and meet the requirement of strong and secure data transmission
and reception by mobile users.
• Desired quality multimedia communication: The compatibility problem with
1G and 2G as well as 3G networks is entirely overcome by 4G using
OFDMA networks. This can be credited to its high rate of data transmission
that is up to 10 Mbps for 3G and 100 Mbps for 4G mobile users.
• Cheaper communication cost: The Telephony Regulatory Authority of India
(TRAI) tirelessly analyzes cost related issues and maintains lower rates for
4G mobile communication users. Additionally, it provides multimedia
communication with global roaming to all mobile subscribers.

17.2.3 Quality of service in OFDMA networks

In OFDMA networks, the improvement of QoS is a major issue owing to the
following characteristics:
1. Higher data transmission rate.
2. Limited bandwidth.
3. Global roaming.
4. More data security.
5. Interferences.
6. Handoff support.

What is QoS?
In the case of cellular networks, high and efficient QoS implies that the network
service provider is providing satisfactory performance in terms of quality of voice, a
lower call blocking and call dropping rate, and high data transmission rate, in
particular for multimedia traffic transmission.
Why QoS?
Third generation networks, mainly WCDMA networks, require multimedia
traffic transmission. Every user in cellular networks transmits the same frequency,
because of which the coverage area as well as the capacity of the network decreases.
Hence, maximum network capacity is achieved only if all the users maintain the
minimum signal-to-inference ratio (SIR), i.e. higher throughput, and this is possible
only when QoS in the network is maintained. In maintaining the required
conditions, some important parameters are responsible.

17.2.3.1 Parameters of QoS

The important parameters for determining QoS are:
• Throughput: the rate at which the packet should be transmitted from source to
destination.
• Delay: the time taken by the packet to travel through the network.
• Packet delivery ratio: the number of packets received to the total number of
packets transmitted.
• Call blocking probability: the ratio of the number of calls successfully
connected to the total number of calls attempted.

17-5
 Modern Optimization Methods for Science, Engineering and Technology

• Residual energy: the amount of energy remaining in the system.

• Channel bandwidth: the ratio of utilized bandwidth to total bandwidth.
• Reliability: the ability of a system to perform consistently without failure.
• Call dropping probability: the ratio of the number of calls terminated after
admission to the number of calls attempted.

17.2.3.2 QoS issues in OFDMA networks

Major issues in OFDMA networks are:
• In 4G OFDMA networks, QoS improvement may be a major challenge due
to RT multimedia traffic transmission.
• RT applications require guaranteed QoS in OFDMA networks.
• Due to the unavailability of resources, it becomes difficult to maintain QoS in
4G wireless networks.
• Another challenge faced by the network is the admission of incoming calls
with less blocking probability.
• Power control is a heavy task, required to maintain appropriate connectivity,
energy and minimum interference in 4G OFDMA networks.
• Monitoring RT multimedia traffic with perfect scheduling is also an impor-
tant obstacle in 4G OFDMA networks which needs attention.

17.2.4 QoS improvement techniques in OFDMA networks

Several issues resist the maintenance of the required QoS in the network. One of the
major hurdles is resource allocation, along with admission of calls for new arrival
users. Fast and accurate power control in OFDMA networks is also a tough task
which requires consideration. Finally, scheduling also plays a major role in
monitoring the traffic effectively and avoiding congestion for maintaining better
QoS and satisfactory network operation. The following schemes are used to improve
QoS in OFDMA networks.

17.2.4.1 Resource allocation with call admission control

Resource allocation is an important factor in OFDMA networks because if the
appropriate resources are not available, the network performance will be adversely
affected.
Resource allocation is based on three parameters:
1. Signal to interference ratio.
2. Throughput.
3. Power.

The earlier work was mostly based on the signal to interference ratio, because at that
time only voice transmission was required. When GSM emerged, i.e. when both
voice and message entered the 2G network, power allocation came into the picture.
Currently, owing to multimedia transmission, throughput based resource allocation
has become an indispensable part of every application in 4G networks, which
requires a considerably high speed and consequently, a higher data rate.

17-6
 Modern Optimization Methods for Science, Engineering and Technology

17.2.4.2 Call admission control

CAC is a technique that allows the admission of a user based on the available
resources or present traffic load conditions after comparison with a predefined value.
If resource allocation is executed along with CAC, it gives marvelous results in
improving network performance for a given QoS requirement.
RACAC decides whether a new user should be admitted or not based on the
available resources and after comparison with a predefined value. The parameters
required for RACAC are:
• Signal-to-interference ratio (SIR).
• Throughput (TP).
• Power (P).

All these parameters are used in CAC based on requirements. Hence, interference
and power measurement is required for comparison of signal strength, whereas
throughput measurement is required for comparison of data rate or speed.

17.2.4.3 Power control

OFDMA is an interference limited technique. This is because every user transmits the
same frequency based on the frequency reuse concept. In the case of frequency reuse,
the entire frequency spectrum is divided into bands and these bands of various
frequencies are assigned to different cells. However, after a certain geological distance,
the same frequency is re-assigned to other cells. Nevertheless, if the distance between
two cells is too small then co-channel interference is generated.
As a matter of fact, co-channel interference is negligible in OFDMA networks
because the operation is based on orthogonal properties. Orthogonality ensures that
if two different codes are multiplied then the output is zero.
The main problematic issue with OFDMA networks is the ‘near–far’ problem, in
which the mobile users closer to the base station or mobile tower receive good signal
strength, while those far away are unable to receive sufficient signal strength. In this
case, if the transmitted power is increased, then the signal strength can easily be
increased and the following advantages can be established:
• Avoids the problem of fading.
• Reduces the bit error rate (BER).

Fading can be defined as the reduction in signal strength. This is easily resolved
by increasing the transmitted power and reducing the BER, which can be defined as
the ratio of the number of bit errors to the total number of bits transmitted. It should
be kept low in all conditions.
If we increase the transmitted power, interference is also increased, suggesting
that fast and perfect power control is required. Power control can be categorized into
two types:
• Open loop power control.
• Closed loop power control.

17-7
 Modern Optimization Methods for Science, Engineering and Technology

Open loop power control is based on the mobile unit and no feedback is required
from the base station. This type of power control is fast but not accurate.
Close loop power control depends on the feedback. The base station continuously
takes feedback from the mobile unit and adjusts the signal strength. This technique is
slow but accurate to a great extent. Close loop power control is again classified into
two types:
• Inner loop power control.
• Outer loop power control.

In the case of inner loop power control, the base station checks the signal level of
a particular user by taking feedback from the mobile unit and adjusts it according to
the QoS requirement.
In the case of outer loop power control, the received power is controlled by
changing the target value, which is pre-decided, based on the QoS requirement.
The need for power control
Power control is very important in OFDMA networks to manage the following:
• Interference management.
• Connectivity management.
• Energy management.

In the case of interference management for down-link operation, the base station
continuously takes feedback and adjusts the signal strength by controlling interfer-
ence in the system. As a result, the near–far problem is eliminated.
In the case of connectivity management, when the mobile user leaves a particular
area and enters another area, the base station in that area should be able to provide
sufficient signal strength for that particular user. Hence, perfect connectivity
management is possible and call dropping is reduced. Also, the handoff problem
is perfectly resolved.
In the case of energy management, the mobile battery reduces power consump-
tion and hence improves battery life. This is possible only through fast and accurate
power control.

17.2.4.4 Scheduling
Scheduling plays a very important role in OFDMA networks. It monitors the traffic
and minimizes delay at the receiving end. It is basically designed for RT or non-RT
traffic transmission as well as reception.
In the case of the OFDMA network, two types of scheduling are mainly used:
• Fixed scheduling.
• Adaptive scheduling.

Fixed scheduling suggests that once the packets are scheduled, they will be
transmitted with constant time and so the delay will be greater.
Adaptive scheduling is based on adaptability, suggesting that network conditions
will change according to the situation. In the case of OFDMA networks, it is based

17-8
 Modern Optimization Methods for Science, Engineering and Technology

on resources. Hence, owing to adaptive scheduling, maximum resource utilization is

possible.
The resources used by OFDMA networks are
• Channel codes.
• Power.
• Rate.

All three types of resources can be used but the use of channel codes and power is
very popular. The use of rate as the main resource will be more beneficial for
multimedia traffic transmission because speed management is very important for RT
traffic in 4G networks, such as video telephony, live telecast and online games. Such
types of applications demand minimum delay (in milliseconds). So, if adaptive
scheduling is based on rate, then delay will surely be reduced and network
throughput will also be maintained according to given QoS requirements.

17.2.4.5 Wireless network traffic

In general, there are two types of traffic:
1. Real-time (RT).
2. Non-real-time (NRT).

RT traffic is based on a variable bit rate (VBR) and is mostly applicable for time
sensitive applications such as video telephony, audio telephony, mobile TV, online
games and videos. NRT VBR traffic is useful for the booking of air tickets and bank
transactions. All types of applications come under multimedia RT traffic trans-
mission. NRT traffic is based on a constant bit rate (CBR) and is typically used for
NRT applications such as SMS, Internet, voice, etc.
RT traffic mainly suffers from delays because of the fast development of mobile
communications. Improvement in QoS is therefore an important task in such
networks. NRT traffic generally suffers from packet loss constraints and QoS
improvement is not an issue in these types of networks.

17.3 Simulation model and parameters

This section deals with simulation models and their parameters.

17.3.1 Simulation topology

To simulate the proposed scheme, network simulator (NS2) with version 2.31 is
used. In the simulation, the base station and user are considered in a 600 m × 600 m
region for 50 s simulation time. It consists of six base stations along with six users.
Each user contains six mobile nodes (see figure 17.1). All nodes have the same
transmission range of 250 m.
For the simulation, VBR traffic is used. The simulation settings and parameters
are given in table 17.2.

17-9
 Modern Optimization Methods for Science, Engineering and Technology

Figure 17.1. Simulation topology for the OFDMA scheme.

Table 17.2. Simulation settings and parameters.

Area size 600 m × 600 m

No. of users 6
MAC OFDMA
Slot duration 1.25 ms
Routing protocol AODV
Transmission rate 1, 1.5, 2, 2.5, 3 Mbps
Frame length 8 slots
Simulation time 50 s
Traffic model VBR
No. of cells 6
Radio range 250 m
Packet size 512 bytes

17.3.2 Performance metrics

Performance evaluation is mainly based on the following matrices:
• Packet delivery ratio: the ratio of packets received successfully to the total
number of packets sent for RT traffic flows.
• Delay: the time taken by the packet to reach the destination, represented in
milliseconds.
• Throughput: the rate at which the packets travel through the network,
represented in megabits/second.

17.4 Adaptive rate scheduling in OFDMA networks

17.4.1 Introduction
After traffic regulation using the RACAC scheme, it is very important to monitor
the traffic. Scheduling plays an important role in monitoring the traffic in OFDMA
networks to improve spectral efficiency.

17-10
 Modern Optimization Methods for Science, Engineering and Technology

17.3 Notation used in adaptive rate scheduling.

Notation Meaning

FCU Feedback control unit

i Session name
Dschedule(i) Average scheduling delay of session i
Qsession(i) Queue size of session i
Rest_arrival(i) Estimated arrival rate of session i
Rallocated(i) Rate allocated to session i
Davg(i) Current average delay of session i
Dthreshold(i) Mean threshold delay for session i
T Scheduled time duration in frames

Resource allocation with call admission control and the adaptive rate scheduling
scheme (RACAC-ARS), are employed. The adaptive rate scheduling scheme is
associated with a heuristic approach based on feedback control unit (FCU) logic.
Under this approach, during the entry of the session into the network, a rate is
assigned and adjusted based on the feedback obtained from the users already
admitted into the network.
A heuristic-based FCU is present in the network, which records all the
information related to the rate of every user. On the basis of the rate of the
operating users, the rate of the incoming users for respective sessions is adjusted.
Important notations used in the analysis of adaptive rate scheduling are shown in
table 17.3.
The process of adaptive rate scheduling is as followed:
1. During the entry of the session, the rate Rallocate for the incoming user is
allocated. After that, the time slot duration T and queue size Q of session i are
assigned and the arrival rate Rest-arrival is estimated. When the user arrives at the
base station (BS), the rate is adjusted based on the feedback (FB) obtained from
the already admitted session. At the same time, FCU determines the delay for
threshold level Dthreshold. It also estimates the average scheduling delay Davg(i)
from the first user and checks the required conditions. If Davg is less than or
equal to Dthreshold, FCU pre-empts all the users.
2. If Davg is more than Dthreshold, error in arrival rate estimation is indicated.
Later, the rate Rallocate is assigned and the arrival rate for the user is
estimated. Finally, the estimated rate is sent as feedback to the BS, based
on which the BS adjusts the rate for the incoming user.

17.4.2 Adaptive rate scheduling algorithm

Figure 17.2 shows all the steps in the flow of the implementation:
Step 1. Allocate the rate Rallocate for the incoming session.
Step 2. Assign time slot duration T and queue size Q of session i.

17-11
 Modern Optimization Methods for Science, Engineering and Technology

Start

Assign allocated rate Rallocate for all sessions

Assign T, Q for incoming session

Estimate the arrival rate Rest and send it as

FB to BS

Determine Dthreshhold using FCU

Estimate Davg

Yes
If Davg> Dthreshold

No

Pre- empt all the sessions.

Send rate as FB to CAC.

End

Figure 17.2. Flow chart of the adaptive rate scheduling algorithm.

Step 3. Estimate the arrival rate Rest-arrival when the user arrives.
Step 4. BS adjusts the rate based on FB obtained from the already admitted user.
Step 5. FCU determines the delay for threshold level Dthreshold.
Step 6. Estimate the average scheduling delay Davg(i) from the first user and check
the following conditions:
If Davg ⩽ Dthreshold, FCU pre-empts all the users, otherwise if
Davg > Dthreshold indicates error in arrival rate estimation then:

• Assign rate Rallocate and estimate the arrival rate for the user.
• Send the estimated rate as FB to the BS.
• Adjust the BS rate for the incoming user.
• END

17-12
 Modern Optimization Methods for Science, Engineering and Technology

17.4.3 Average scheduling delay estimation for the ARS scheme

The average scheduling delay of all the users is estimated from the first user i
admitted into the network:
Dschedule(i ) = (Qsession(i ) + T [R est_arrival(i )])/ R allocated(i ), (17.1)
where Qsession(i) is the size of the queue of user i, T is the time slot duration in
milliseconds, Rest_arrival(i) is the estimated arrival rate allotted to user i and
Rallocated(i) is the rate allocated.

17.5 Conclusions
Currently, the 4G network is very popular and it is thought to be the next generation
of mobile computing. Its use and advantages distinguish it from all other peer
technologies. In order to keep its services available all the time, there is a need to
improve its QoS with various parameters.
In this chapter, the developed scheme for RACAC improves network reliability
by attempting to reduce the number of sessions that are blocked due to inadequate
resources. The main concept of resource allocation in the scheme is used to increase
the data rate during transmission according to user requirements, and simulta-
neously improve throughput for real-time multimedia traffic using variable bit rate
traffic for transmission. The resource allocation with CAC scheme for next
generation networks maximizes sessions for users while increasing the data trans-
mission rate. It also ensures that a new session does not violate the QoS of ongoing
sessions with the help of an ARS scheme. In an ARS scheme, a heuristic-based
approach finds the solution close to the true one. The overall process depends on
FCU logic. Thus, our proposed scheme utilizes resources efficiently and maximizes
the spectral efficiency of the network. According to analysis of the results, it is
observed that our proposed scheme improves throughput and reduces delay for 4G
cellular networks.

References
[1] Jagiasi V and Giri N 2017 Resource allocation strategies for cellular networks Inter. J. Eng.
Res. Techn. 5 130–6
[2] Jain M and Mittal R 2016 Adaptive call admission control and resource allocation in multi-
server wireless cellular networks J. Ind. Eng. 12 71–80
[3] Memis M O and Ercetin O 2015 Resource allocation for statistical QoS guarantees in MIMO
cellular networks EURASIP J. Wirel. Commun. Netw. 2015 217
[4] Nantha Kumar G and Arokiasamy A 2014 An efficient combined call admission control and
scheduling algorithm to improve quality of service for multimedia services over next
generation wireless networks Inter. J. Comput. Appl 107 416–25
[5] Liu J, Chen W, Zhang Y J and Cao Z 2013 A utility maximization framework for fair
and efficient multicasting in multicarrier wireless cellular network IEEE Trans. Netw. 21
110–20

17-13
 Modern Optimization Methods for Science, Engineering and Technology

[6] Kadhem K and Khan Y 2012 Evaluation and comparison of resource allocation strategies for
new streaming services in wireless cellular networks IEEE Trans. Commun. 12 123–35
[7] Jraifi Abdelouahed A 2012 Theoretical prediction of users in WCDMA interference J. Theor.
Appl. Inf. Techn 35 163–8
[8] Peifang Z and Jordan S 2008 Cross layer dynamic resource allocation with targeted
throughput for WCDMA data IEEE Trans. Wireless Commun. 7 4896–906

17-14