(Quantum Machine Intelligence) Siddhartha Bhattacharyya, Mario Köppen, Elizabeth Behrman, Ivan Cruz-Aceves - Hybrid Quantum Metaheuristics - Theory and Applications-CRC Press (2022)

Hybrid Quantum
Metaheuristics
Series Page
Quantum Machine Intelligence Series
SERIES EDITORS
Siddhartha Bhattacharyya, Rajnagar Mahavidyalaya, Birbhum, India
Elizabeth C. Behrman, Wichita State University, USA
Hybrid Quantum Metaheuristics

Siddhartha Bhattacharyya, Mario Köppen, Elizabeth Behrman, Ivan
Cruz-Aceves
Quantum Machine Intelligence

Siddhartha Bhattacharyya, Mario Köppen, Elizabeth Behrman, Ivan
Cruz-Aceves
For more information about this series, please visit: https://www.rout-

ledge.com/Quantum-Machine-Intelligence/book-series/QMI
Hybrid Quantum
Metaheuristics
Theory and Applications
Edited by
Siddhartha Bhattacharyya
Mario Köppen
Elizabeth Behrman
Ivan Cruz-Aceves
MATLAB® is a trademark of The MathWorks, Inc. and is used with permission. The MathWorks
does not warrant the accuracy of the text or exercises in this book. This book’s use or discussion of
MATLAB® software or related products does not constitute endorsement or sponsorship by The
MathWorks of a particular pedagogical approach or particular use of the MATLAB® software.
First edition published [2022]

by CRC Press
6000 Broken Sound Parkway NW, Suite 300, Boca Raton, FL 33487-2742
and by CRC Press

4 Park Square, Milton Park, Abingdon, Oxon, OX14 4RN
CRC Press is an imprint of Taylor & Francis Group, LLC
© 2022 selection and editorial matter, Siddhartha Bhattacharyya, Mario Köppen, Elizabeth Behrman,
Ivan Cruz-Aceves, individual chapters, the contributors
Reasonable efforts have been made to publish reliable data and information, but the author and pub-
lisher cannot assume responsibility for the validity of all materials or the consequences of their use.
The authors and publishers have attempted to trace the copyright holders of all material reproduced
in this publication and apologize to copyright holders if permission to publish in this form has not
been obtained. If any copyright material has not been acknowledged please write and let us know so
we may rectify in any future reprint.
Except as permitted under U.S. Copyright Law, no part of this book may be reprinted, reproduced,
transmitted, or utilized in any form by any electronic, mechanical, or other means, now known or
hereafter invented, including photocopying, microfilming, and recording, or in any information stor-
age or retrieval system, without written permission from the publishers.
For permission to photocopy or use material electronically from this work, access www.copyright.
com or contact the Copyright Clearance Center, Inc. (CCC), 222 Rosewood Drive, Danvers, MA
01923, 978-750-8400. For works that are not available on CCC please contact mpkbookspermis-
[email protected]
Trademark notice: Product or corporate names may be trademarks or registered trademarks and are
used only for identification and explanation without intent to infringe.
ISBN: 9780367751562 (hbk)

ISBN: 9781032254616 (pbk)
ISBN: 9781003283294 (ebk)
DOI: 10.1201/9781003283294
Typeset in Nimbus font CMR10

by KnowledgeWorks Global Ltd.
Dedication
Siddhartha would like to dedicate this volume to late Prof. John

Stewart Bell, FRS, the originator of Bell’s theorem, an important
theorem in quantum physics regarding hidden variable theories.
Elizabeth would like to dedicate this volume to JPK, EJB, and JFB;
and to the memory of CFB.
Mario would like to dedicate this book to the Schrödinger’s cat.
Ivan would like to dedicate this volume to his lovely family, his wife
Mary, his children Ivan and Yusef, and his mother Bety.
Contents
Editors .....................................................................................................................xiii
Preface.....................................................................................................................xix
Contributors ..........................................................................................................xxiii
Chapter 1 An Introductory Illustration to Quantum-Inspired Metaheuristics.. 1

1.1 Introduction............................................................................. 1
1.2 Quantum-Inspired Metaheuristics ..........................................2
1.2.1 Local Search Metaheuristics ...................................... 3
1.2.2 Constructive Metaheuristics....................................... 4
1.2.3 Population-based Metaheuristics ............................... 6
1.2.4 Hybrid Metaheuristics................................................ 7
1.3 Entanglement-Induced Optimization...................................... 7
1.4 W-state Encoding of Optimization Algorithms ...................... 8
1.5 Quantum System-based Optimization .................................... 9
1.5.1 Bi-level Quantum System-based Optimization.......... 9
1.5.2 Multi-level Quantum System-based Optimization .. 10
1.6 Applications of Quantum-Inspired Metaheuristics...............11
1.7 Conclusion ............................................................................14
Chapter 2 A Quantum-Inspired Approach to Collective Combine Basic

Classifiers ...................................................................................... 21
2.1 Introduction........................................................................... 21
2.2 Bagging Method ................................................................... 21
2.3 Classifiers Based on Similarity of Objects ........................... 22
2.4 Statistical Classification Algorithms..................................... 24
2.5 Classifiers Based on Class Separability in Attribute Space.. 25
2.6 Logical Classification Algorithms ....................................... 26
2.7 Neural Networks ................................................................... 27
2.8 Methods of Combining Basic Classifiers .............................27
2.8.1 Voting ....................................................................... 27
2.8.2 Stacking.................................................................... 28
2.8.3 Ensemble Selection.................................................. 28
2.8.3.1 The <<naive>> Bayesian Classifier ..... 30
2.8.4 Quantum-Inspired Metaheuristics Method .............. 30
2.9 Conclusion ............................................................................32
vii
viii Contents
Chapter 3 Function Optimization Using IBM Q............................................37

3.1 Introduction........................................................................... 37
3.2 Function Optimization .......................................................... 38
3.2.1 Difficulties in Optimization Methods ...................... 38
3.2.2 Definition of Multi-objective Optimization
Problem (MOOP)..................................................... 38
3.2.3 Differences between SOOPs and MOOPs ............... 40
3.3 Modern Optimization Problem-Solving Techniques ............ 40
3.3.1 Genetic Algorithm ................................................... 40
3.3.2 Simulated Annealing................................................ 41
3.3.3 Particle Swarm Optimization ................................... 41
3.3.4 Bat Algorithm .......................................................... 42
3.3.5 Cuckoo Search Algorithm........................................ 42
3.3.6 Fuzzy System ........................................................... 42
3.3.7 Neural Network-Based Optimization ...................... 43
3.4 Quantum Computing and Optimization Algorithms ............ 43
3.4.1 Quantum Computing................................................ 43
3.4.2 Optimization Using Quantum Computing ............... 44
3.5 Features of IBM Q Experience .............................................45
3.6 Circuit Composer IBM Q ..................................................... 45
3.7 QISKit in IBM Q .................................................................. 47
3.7.1 Creating 5-qubit Circuit with the Help of
QISKit in IBM Q ..................................................... 47
3.7.2 Testing the Circuit Using IBM Quantum
Computer.................................................................. 48
3.8 Optimization Using IBM Q .................................................. 51
3.9 Conclusion ............................................................................53
Chapter 4 Multipartite Adaptive Quantum-Inspired Evolutionary Algorithm

to Reduce Power Losses ................................................................ 57
4.1 Introduction........................................................................... 57
4.2 Literature Review.................................................................. 60
4.3 Problem Formulation ............................................................ 64
4.4 Power Flow ........................................................................... 66
4.5 Algorithm.............................................................................. 70
4.6 Results and Discussion ......................................................... 75
4.7 Conclusions........................................................................... 88
4.8 Parameters of IEEE Benchmark Test Bus System................ 92
Chapter 5 Quantum-Inspired Manta Ray Foraging Optimization Algorithm

for Automatic Clustering of Color Images.................................... 95
5.1 Introduction........................................................................... 95
5.2 Literature Review.................................................................. 96
Contents ix
5.3 Fundamentals of Quantum Computing.................................97

5.3.1 Rotation Gate ........................................................... 98
5.3.2 Pauli-X Gate............................................................. 98
5.4 Validity Measurement of Clustering.....................................98
5.5 Overview of Manta Ray Foraging Optimization
Algorithm.............................................................................. 99
5.6 Proposed Methodology....................................................... 100
5.7 Experimental Results and Analysis .................................... 102
5.7.1 Developmental Entertainment................................ 103
5.7.2 Dataset Used .......................................................... 103
5.7.3 Clustered Images.................................................... 103
5.7.4 Sensitivity Analysis of QIMRFO........................... 103
5.7.5 Analysis of Experimental Results.......................... 107
5.8 Conclusion and Future Scope ............................................. 108
Chapter 6 Automatic Feature Selection for Coronary Stenosis Detection in

X-Ray Angiograms...................................................................... 117
6.1 Introduction......................................................................... 117
6.2 Background......................................................................... 120
6.2.1 Feature Extraction.................................................. 120
6.2.1.1 Pixel Intensity-based Features ............... 120
6.2.1.2 Texture Features .................................... 121
6.2.1.3 Morphologic Features............................ 122
6.2.2 Feature Selection.................................................... 124
6.2.3 Support Vector Machines....................................... 125
6.2.4 Quantum Genetic Algorithm.................................. 126
6.3 Proposed Method ................................................................ 130
6.4 Experiment Details .............................................................133
6.5 Results................................................................................. 134
6.6 Conclusion ..........................................................................136
Chapter 7 Quantum Preprocessing for DCNN in Atherosclerosis

Detection...................................................................................... 141
7.1 Introduction......................................................................... 141
7.2 Related Work ...................................................................... 142
7.3 Mathematical Foundations.................................................. 145
7.3.1 Quantum Computing.............................................. 145
7.3.1.1 Qubit States ........................................... 146
7.3.1.2 Qubit Operations ................................... 147
7.3.1.3 Qubit Measurements.............................. 149
7.3.2 Convolutional Neural Networks ............................ 149
7.3.2.1 Convolutional Layer .............................. 150
7.3.2.2 Pooling Layer ........................................ 150
7.3.2.3 Fully Connected Layer .......................... 151
x Contents
7.3.2.4 Activation Functions.............................. 152

7.4 Proposed Method ................................................................ 153
7.4.1 Quantum Convolutional Layer............................... 154
7.4.2 Network Architecture............................................. 155
7.4.3 Evaluation Metrics ................................................. 157
7.5 Results and Discussions...................................................... 158
7.5.1 Dataset of Coronary Stenosis................................. 158
7.5.2 Quantum Preprocessing ......................................... 159
7.5.3 Training Results ..................................................... 160
7.5.4 Detection Results ................................................... 161
7.6 Concluding Remarks .......................................................... 164
Chapter 8 Multilevel Quantum Elephant Herd Algorithm for Automatic

Clustering of Hyperspectral Images ............................................ 171
8.1 Introduction......................................................................... 171
8.2 Literature Survey ................................................................ 173
8.3 Background Concepts.........................................................174
8.3.1 Elephant Herding Optimization ............................ 174
8.3.1.1 Clan Updation........................................174
8.3.1.2 Separation Operator...............................175
8.3.1.3 Steps of EHO......................................... 175
8.3.2 Basic Concepts of Quantum Computing................ 175
8.3.3 Fuzzy C Means Clustering Algorithm ................... 176
8.3.4 Xie-Beni Index....................................................... 177
8.4 Proposed Methodology....................................................... 177
8.4.1 HSI Preprocessing.................................................. 177
8.4.2 Qubit and Qutrit Based Elephant Herd
Optimization .......................................................... 178
8.5.1 Salinas Dataset ....................................................... 181
8.5.2 Fitness Function ..................................................... 181
8.5.3 Analysis.................................................................. 181
8.6 Conlusion............................................................................ 185
Chapter 9 Toward Quantum-Inspired SSA for Solving Multiobjective

Optimization Problems................................................................ 189
9.1 Introduction......................................................................... 189
9.2 Salp Swarm Algorithm .......................................................192
9.2.1 Initialization ........................................................... 192
9.2.2 Leaders’ Specification............................................ 193
9.2.3 Updating Position .................................................. 193
9.2.4 Re-evaluation and Decision-making ...................... 194
9.3 Proposed Multiobjective Quantum-Inspired Salp Swarm
Algorithm............................................................................ 194
Contents xi
9.3.1 Delta Potential-Well Model for SSA ..................... 195

9.3.2 Salp Position Measurement.................................... 195
9.3.3 The New Algorithm Behavior................................ 196
9.4 Experimental Procedure...................................................... 197
9.4.1 Computing Environment........................................ 197
9.4.2 Performance Assessment Metrics .......................... 197
9.4.3 Multiobjective Benchmark Problems..................... 199
9.4.4 Evaluating Method and Algorithms Parameters .... 200
9.5 Experiments and Discussion...............................................200
9.6 Conclusion ..........................................................................204
Chapter 10 Quantum-Inspired Multi-Objective NSGA-II Algorithm for

Automatic Clustering of Gray Scale Images ............................... 207
10.1 Introduction......................................................................... 207
10.2 Quantum Computing Fundamental..................................... 208
10.3 Computing the Objectives .................................................. 209
10.3.1 CS-Measure (CSM) index...................................... 209
10.3.2 Davies–Bouldin (DB) Index .................................. 210
10.4 Multi-Objective Optimization............................................. 210
10.4.1 NSGA-II................................................................. 211
10.4.2 Population Initialization and Chromosome
Representation........................................................ 211
10.4.3 Creating Cluster Centroids..................................... 211
10.4.4 Genetic Operation .................................................. 212
10.4.5 Fast Non-dominated Sorting .................................. 212
10.4.6 Crowding Distance................................................. 212
10.4.7 Basic Steps of Classical NSGA-II Algorithm for
Automatic Clustering of Gray Scale Images ......... 212
10.5 Proposed Technique............................................................213
10.5.1 Quantum State Population Initialization................ 213
10.5.2 Creating Cluster Centroids in Quantum-Inspired
Framework ............................................................. 214
10.5.3 Genetic Operators in Quantum-Inspired
Framework ............................................................. 214
10.5.3.1 Quantum-Behaved Selection ................. 214
10.5.3.2 Quantum-Behaved Crossover ................ 215
10.5.3.3 Quantum-Behaved Mutation ................. 216
10.5.4 Fast Non-dominated Sorting in Quantum-Inspired
Framework ............................................................. 217
10.5.5 Crowding Distance Computation in Quantum-
Inspired Framework ............................................... 217
10.5.6 QIMONSGA-II Algorithm for Automatic
Clustering of Gray Scale Images ........................... 217
xii Contents
10.6.1 Used Dataset .......................................................... 220

10.6.2 Parameter Settings ................................................. 220
10.6.3 Performance Evaluation ........................................ 220
10.6.4 Experimental Results ............................................. 222
10.7 Discussions and Conclusion ............................................... 225
Chapter 11 Conclusion................................................................................... 231
Appendix A Automatic Feature Selection for Coronary Stenosis Detection in

X-Ray Angiograms...................................................................... 235
A.1 Matlab Code to Extract Vessel Segments ........................... 235
A.2 Matlab Code to Find Pixel Positions ..................................240
A.3 Matlab Code to Extract a Window from a Matrix .............. 244
A.4 Matlab Code to Find Row Vector ....................................... 246
Index...................................................................................................................... 249
Editors
Dr. Siddhartha Bhattacharyya received his Bachelors in
Physics, and in Optics and Optoelectronics, and Masters
in Optics and Optoelectronics from University of Cal-
cutta, India, in 1995, 1998, and 2000, respectively. He
completed PhD in Computer Science and Engineering
from Jadavpur University, India, in 2008. He is the re-
cipient of the University Gold Medal in Masters from
the University of Calcutta. He is the recipient of several
coveted awards including the Distinguished HoD Award
and Distinguished Professor Award conferred by Com-
puter Society of India, Mumbai Chapter, India in 2017,
the Honorary Doctorate Award (D. Litt.) from The Uni-
versity of South America, and the South East Asian Re-
gional Computing Confederation (SEARCC) International Digital Award ICT Edu-
cator of the Year in 2017. He has been appointed as the ACM Distinguished Speaker
for the tenure 2018–2020. He has been inducted into the People of ACM Hall of
Fame by ACM, USA in 2020. He has been appointed as the IEEE Computer Soci-
ety Distinguished Visitor for the tenure 2021–2023. He has been elected as the full
foreign member of the Russian Academy of Natural Sciences. He has been elected
a full fellow of the Royal Society for Arts, Manufacturers and Commerce (RSA),
London, UK.
He is currently serving as the Principal of Rajnagar Mahavidyalaya, Rajnagar,
Birbhum. He served as a Professor in the Department of Computer Science and En-
gineering of Christ University, Bangalore. He served as the Principal of RCC Insti-
tute of Information Technology, Kolkata, India during 2017–2019 and as a Senior
Research Scientist in the Faculty of Electrical Engineering and Computer Science
of VSB Technical University of Ostrava, Czech Republic (2018–2019). Prior to this,
he was the Professor of Information Technology in RCC Institute of Information
Technology, Kolkata, India. He served as the Head of the Department from March,
2014 to December, 2016. Prior to this, he was an Associate Professor of Information
Technology in RCC Institute of Information Technology, Kolkata, India, from 2011
to 2014. Before that, he served as an Assistant Professor in Department of Computer
Science and Information Technology of University Institute of Technology, The Uni-
versity of Burdwan, India from 2005 to 2011. He was a Lecturer in Information Tech-
nology of Kalyani Government Engineering College, India during 2001–2005. He is
a co-author of 6 books and the co-editor of 80 books. He has more than 300 research
publications in international journals and conference proceedings to his credit. He
holds 19 patents. He has been the member of the organizing and technical program
committees of several national and international conferences. He is the founding
Chair of ICCICN 2014, ICRCICN (2015, 2016, 2017, 2018), ISSIP (2017, 2018)
xiii
xiv Contents
(Kolkata, India). He was the General Chair of several international conferences like
WCNSSP 2016 (Chiang Mai, Thailand), ICACCP (2017, 2019) (Sikkim, India), and
(ICICC 2018) (New Delhi, India), and ICICC 2019 (Ostrava, Czech Republic).
He is the Associate Editor of several reputed journals including Applied Soft
Computing, IEEE Access, Evolutionary Intelligence, and IET Quantum Communi-
cations. He is the editor of International Journal of Pattern Recognition Research
and the founding Editor in Chief of International Journal of Hybrid Intelligence, In-
derscience. He has guest-edited several issues with several international journals. He
is serving as the Series Editor of IGI Global Book Series Advances in Information
Quality and Management (AIQM), De Gruyter Book Series Frontiers in Computa-
tional Intelligence (FCI), CRC Press Book Series(s) Computational Intelligence and
Applications & Quantum Machine Intelligence, Wiley Book Series Intelligent Sig-
nal and Data Processing, Elsevier Book Series Hybrid Computational Intelligence
for Pattern Analysis and Understanding, and Springer Tracts on Human Centered
Computing.
His research interests include hybrid intelligence, pattern recognition, multimedia
data processing, social networks, and quantum computing.
He is a life fellow of Optical Society of India (OSI), India, life fellow of Inter-
national Society of Research and Development (ISRD), UK, a fellow of Institution
of Engineering and Technology (IET), UK, a fellow of Institute of Electronics and
Telecommunication Engineers (IETE), India, and a fellow of Institution of Engineers
(IEI), India. He is also a senior member of Institute of Electrical and Electronics En-
gineers (IEEE), USA, International Institute of Engineering and Technology (IETI),
Hong Kong and Association for Computing Machinery (ACM), USA.
He is a life member of Cryptology Research Society of India (CRSI), Com-
puter Society of India (CSI), Indian Society for Technical Education (ISTE), In-
dian Unit for Pattern Recognition and Artificial Intelligence (IUPRAI), Center for
Education Growth and Research (CEGR), Integrated Chambers of Commerce and
Industry (ICCI), and Association of Leaders and Industries (ALI). He is a member
of Institution of Engineering and Technology (IET), UK, International Rough Set
Society, International Association for Engineers (IAENG), Hong Kong, Computer
Science Teachers Association (CSTA), USA, International Association of Academi-
cians, Scholars, Scientists and Engineers (IAASSE), USA, Institute of Doctors En-
gineers and Scientists (IDES), India, The International Society of Service Innovation
Professionals (ISSIP), and The Society of Digital Information and Wireless Com-
munications (SDIWC). He is also a certified Chartered Engineer of Institution of
Engineers (IEI), India. He is on the Board of Directors of International Institute of
Engineering and Technology (IETI), Hong Kong.
Contents xv
Dr. Elizabeth Behrman earned her Bachelor’s degree

in Mathematics, in 1979, from Brown University. She
completed her masters in Chemistry in 1981 and her
PhD in Physics in 1985, both at the University of Illi-
nois at Urbana-Champaign. As an undergraduate she was
awarded an NSF Undergraduate Research Participation
Fellowship and membership in Sigma Xi. At Illinois she
was a recipient of a graduate fellowship.
Since 2004, she has been full professor of both
Physics and Mathematics at Wichita State University in
Wichita, Kansas, where she has earned the Presidents
Distinguished Service Award in 2015, and the Academy
for Effective Teaching Award in 2012. She was also ap-
pointed Kavli Institute for Theoretical Physics Scholar, from 2006 to 2009. Prior to
this, she was a professor of Physics from 2002 to 2004, and Chair of the Physics
Department from 2003 to 2006. She was an associate director of the honors program
at Wichita State from 1999–2003. Prior to this, she was an associate professor of
Physics at Wichita State from 1994–2002. During this time she was a recipient of a
Lady Davis Fellowship, at Hebrew University, Jerusalem, Israel, 1996–1997. Before
that she was an assistant professor at Wichita State from 1990 to 1994. Prior to that,
she was an assistant professor of Ceramic Engineering at the New York State College
of Ceramics at Alfred, in Alfred, NY. She has published over 60 papers in journals
over a wide range of fields, including engineering, physics, and chemistry, reflecting
her wide range of interests including chemical kinetics and reaction pathways, ce-
ramic superconductors, glass structure, and nuclear waste vitrification. She was the
first to predict the possibility of stable spheroids and tubes made of inorganic mate-
rials in 1994. She is also one of the first and major researchers in quantum machine
learning. She and her research group have published seminal papers on temporal and
spatial quantum backpropagation, quantum Hopfield networks, quantum Bayesian
networks, and quantum genetic algorithm.
xvi Contents
Mario Köppen studied Physics at the Humboldt-

University of Berlin and received his master’s degree in
Solid State Physics in 1991. Afterwards, he worked as
a scientific assistant at the Central Institute for Cyber-
netics and Information Processing in Berlin and changed
his main research interests to image processing and neu-
ral networks. From 1992 to 2006, he was working with
the Fraunhofer Institute for Production Systems and De-
sign Technology. He continued his works on the indus-
trial applications of image processing, pattern recogni-
tion, and soft computing, especially evolutionary com-
putation. During this period, he achieved the doctoral
degree at the Technical University Berlin with his the-
sis works: “Development of an intelligent image processing system by using soft
computing” with honors.
He has published more than 150 peer-reviewed papers in conference proceedings,
journals, and books. He was active in the organization of various conferences as a
chair or member of the program committee, including the WSC online conference
series on Soft Computing in Industrial Applications, and the HIS conference series
on Hybrid Intelligent Systems. He is founding member of the World Federation of
Soft Computing and since 2016 Editor-in-Chief of its Elsevier Applied Soft Comput-
ing journal. In 2006, he became JSPS fellow at the Kyushu Institute of Technology in
Japan, and Professor at the Network Design and Research Center (NDRC) in 2008,
and Professor at the Graduate School of Creative Informatics of the Kyushu Institute
of Technology in 2013, where he is conducting research in the fields of soft comput-
ing, especially for multi-objective and relational optimization, digital convergence,
and human-centered computing.
Contents xvii
Ivan Cruz-Aceves received a PhD in Electrical Engi-

neering from University of Guanajuato in 2014. He
works at the Mexican National Council on Science and
Technology (CONACYT) under the project Catedras-
CONACYT assigned to the Center for Research in Math-
ematics (CIMAT) since 2014. He is a member of the
Mexican National System of Researchers level 1 and his
main research interests are: biomedical signal and im-
age analysis, computational intelligence, and evolution-
ary computation.
Preface
A metaheuristic is a heuristic (partial search) algorithm that is more or less an effi-
cient optimization algorithm to real-world problems. Hybrid metaheuristics refer to a
proper and judicious combination of several other metaheuristics and machine learn-
ing algorithms. The hybrid metaheuristics have been found to be more robust and
failsafe owing to the complementary character of the individual metaheuristics in the
resultant combination. This is primarily due to the fact that the vision of hybridization
is to combine different metaheuristics such that each of the combination supplements
the other in order to achieve the desired performance. Typical examples use fuzzy-
evolutionary, neuro-evolutionary, neuro-fuzzy evolutionary, rough-evolutionary ap-
proaches to name a few. Recently, chaos theory has also found wide applications in
evolving efficient hybrid metaheuristics.
Quantum computer, as the name suggests, principally works on several quantum
physical features. These could be used as an immense alternative to today’s apposite
computers since they possess faster processing capability (even exponentially) than
classical computers. The term quantum computing stems from the synergistic combi-
nation of quantum mechanical principles and classical information theory conjoined
with principles of computer science. Utilization of the basic features of quantum
computing into different evolutionary algorithmic frameworks is foremost part of
this research in soft computing discipline. A number of researchers has coupled the
underlying principles of quantum computing with various metaheuristic structures to
introduce different quantum-inspired algorithmic approaches. The evolution of the
quantum computing paradigm has led to the evolution of time efficient and robust
hybrid metaheuristics by means of conjoining the principles of quantum mechanics
with the conventional metaheuristics, thereby enhancing the real-time performance
of the hybrid metaheuristics.
This volume aims to bring together recent advances and trends in methodologi-
cal approaches, theoretical studies, mathematical and applied techniques related to
hybrid quantum metaheuristics, and their applications to engineering problems. The
scope of the volume in essence is confined into but not bounded on introducing
different novel hybrid quantum metaheuristics for addressing glaring optimization
problems ranging from function optimization, data analysis (both discrete and con-
tinuous), system optimization, and signal processing to a host of scientific and engi-
neering applications. It is also aimed to emphasize the effectiveness of the proposed
approaches over the state-of-the-art existing approaches by means of illustrative ex-
amples and real-life case studies.
This volume comprises nine well-versed chapters on different facets of hy-
brid quantum metaheuristics along with an introductory and concluding chap-
ters. Quantum-inspired metaheuristics can be described as an integrative algorith-
mic structure, which are designed by exploiting the basics of quantum computing
(QC) and metaheuristics. Chapter 1 presents an outline of different categories of
xix
xx Preface
quantum-inspired metaheuristics. A brief summary of entanglement-induced opti-

mization is explored in this chapter. This chapter also provides a concise review of
W-state encoding-based optimization algorithms.
The performance of any machine learning algorithm depends on the proper choice
of parameters of the algorithm and the selected structure of the model. A number of
methods have been proposed to solve the problem of optimal choice of a model.
These include (i) hyperparametric optimization algorithms and (ii) ensemble meth-
ods, in which several machine learning algorithms are used in parallel to collectively
solve the problem. Although none of these algorithms guarantee a successful so-
lution to the problem, sufficient experimental evidence is available to prove their
effectiveness. As far as classifier decision-making is concerned, the voting method is
usually used to collectively combine the decisions of the base classifiers in a machine
learning model. In Chapter 2, the authors explore alternative methods of combining
such as stacking and ensemble selection, and also propose a new quantum-inspired
approach based on metaheuristic hyperparameter tuning algorithms.
Chapter 3 presents the various modern optimization problem-solving techniques
available in literature and their limitations. Convergence speed is the primary prob-
lem of the modern optimization methods, which has been resolved by using quantum
computing. IBM provided quantum computing cloud platform for all the users. IBM
Q experience is considered here for field of study. It has two key features, circuit
composer and QISkit.
Distribution network (DN) acts as a last mile link between the transmission sys-
tem and the end users. DN has more power losses in power system network due to
the low admittance ratio. Many techniques have been implemented in DN to reduce
the losses. Implementation of capacitors with optimal location and capacity, increas-
ing or decreasing the size of the conductors, changing transformer taps, network re-
configuration and optimal allocation of distributed generation (DG) are some of the
techniques implemented in DN to reduce the losses. A novel Multipartite Adaptive
Quantum-inspired Evolutionary Algorithm (MAQiEA) is proposed in Chapter 4 and
used to find the optimal location and capacity of DG. MAQiEA is an updated ver-
sion of Adaptive Quantum-inspired Evolutionary Algorithm (AQiEA), which is an
improvement over QiEA. AQiEA used two Q-bits and entanglement, whereas QiEA
had used a single Q-bit. Two modifications have been made in the rotation strategies
of AQiEA in MAQiEA. The first modification is made on rotation strategy respon-
sible for exploitation, i.e., rotation away from worse in which bi-partite adaptive
crossover operator is augmented with multi-partite operator for better exploitation.
The second modification is made on rotation strategy responsible for exploration,
i.e., rotation toward better, which has been converted to rotation around better as it
has less restriction so provides for better exploration. The effectiveness of MAQiEA
is tested on the IEEE benchmark test bus system.
Chapter 5 introduces a simple and effective approach to identify an optimum
number of clusters of color images at run. In the real-world problems, determin-
ing the appropriate number of clusters from a data set is a challenging task. This
chapter proposes a novel automatic clustering algorithm for identifying the optimal
Preface xxi
number of clusters in color images by using a quantum-inspired framework incorpo-

rated with foraging optimization algorithm. The proposed Quantum-Inspired Manta
Ray Foraging Optimization (QIMRFO) algorithm has been compared with the clas-
sical version of Manta Ray Foraging Optimization (MRFO) algorithm and the well-
known Genetic Algorithm (GA) for of color images. The computational results indi-
cate that the proposed QIMRFO outperforms others quantitatively and qualitatively.
Furthermore, the utilization of the proposed QIMRFO algorithm can be viewed as
an input for solving image segmentation and classification problems. The automatic
stenosis detection in X-ray coronary angiograms is very important in medical anal-
ysis and information systems. Coronary artery diseases are the most common cause
of death in the world over all other cardio vascular diseases.
In Chapter 6, a novel method for coronary stenosis detection using automatic
feature extraction is proposed. In a first stage, a set of 31 features were extracted from
the image dataset. On a second stage, an automatic feature selection process was
performed driven by a Quantum Genetic Algorithm in order to find the most suitable
feature subset, able to train a Support Vector Machine-based classifier. The achieved
results demonstrate the effectiveness of the proposed method by reducing the number
of necessary features for the classification of positive and negative stenosis cases,
keeping at the same time an optimal classification rate in terms of the accuracy and
the Jaccard index metrics.
Coronary artery disease is the leading cause of mortality rate worldwide. This
condition is caused by an abnormal narrowing or occlusion of the artery lumen,
known as atherosclerosis (a specific type of stenosis). Currently, many imaging tech-
niques have been tested; nevertheless, conventional X-ray Coronary Angiography
(XCA) remains the gold standard for coronary artery disease diagnosis, such as
atherosclerosis. Chapter 7 presents a Hybrid Quantum-Convolutional Neural Net-
work to detect coronary artery atherosclerosis in XCA images automatically. It em-
ploys a Quantum Convolutional Layer (QCL) that resembles a typical convolutional
layer. However, unlike the convolution operation where a classical convolutional fil-
ter is applied, a QCL transforms the input using a quantum circuit. This QCL gen-
erates a multichannel image from a single channel XCA that feeds a traditional
Convolutional Neural Network (CNN). The CNN was trained and optimized us-
ing two different architectures previously employed for atherosclerosis detection,
viz., a DenseNet-based and a VGG-based architecture. Additionally, a comparison
of Stochastic Gradient Descent (SGD) and SGD with Momentum (SGDM) were
conducted. These architectures were analyzed and compared using a real dataset
composed of 250 real XCA images (split into 125 images for testing and 125 for
training) regarding their training without the QCL preprocessing and SGDM. The
results showed that the CNNs trained with the outcome of the QCL surpass the base-
line CNNs performance.
Automatic clustering of hyperspectral images is a very strenuous task due to the
presence of a huge number of redundant bands and complexity to process them. In
Chapter 8, two quantum versions of Elephant Herd Optimization algorithm are pro-
posed for this purpose. The use of the binary and ternary quantum logics enhances
xxii Preface
the exploration and exploitation capability of the elephant herd optimization. These
algorithms are compared to their classical counterpart. They are implemented on the
Salinas dataset. The proposed qutrit-based algorithm is found to converge faster and
produce more robust results. The Xie-Beni Index is used as the fitness function. A
few statistical tests like mean, standard deviation, and Kruskal-Wallis test are per-
formed to establish the efficiency of the proposed algorithm. The F score is used
to compare the segmented images using the optimal cluster numbers. The proposed
algorithms are found to perform better in most of the cases.
Salp Swarm Algorithm (SSA) is a recently introduced metaheuristic algorithm
applied for solving benchmark and real-world optimization problems. SSA has good
exploitation ability during the search process. However, its exploration ability is lim-
ited. Chapter 9 attempts to enhance the balance between exploration and exploitation
in SSA by hybridizing with the quantum-inspired framework. The Delta potential-
well model (DPWM) from quantum mechanics is known for improving the conver-
gence and diversity in the population to enhance the exploration ability of SSA. The
proposed hybrid method is tested using well-known complex convex and noncon-
vex multiobjective benchmark problems. A comparative study is conducted between
proposed MQSSA and well-regarded algorithms MSSA and NSGA-II. The experi-
mental results exemplify that the overall performance of MQSSA is competitive as
compared with other approaches.
Chapter 10 proposes a quantum-inspired multiobjective algorithm for automatic
clustering of gray scale images. The well-known Non-dominated Sorting Genetic
Algorithm II (NSGA-II) inspired by the intrinsic principles of quantum mechanical
phenomena is presented here for solving the aforesaid problem by optimizing two
objective functions simultaneously. The proposed algorithm has been compared with
its classical counterpart and the experimental results over six Berkeley gray scale im-
ages certify the efficiency and robustness of the proposed algorithm with reference to
the optimal computational time, mean fitness value, standard deviation, and standard
error. Finally, a statistical superiority t-test has been conducted between these two
algorithms to ascertain the supremacy of the results of the proposed algorithm.
Chapter 11 summarizes the findings reported in the volume with future directions
of research.
The book is intended for researchers, academicians, and practitioners. This will
serve as a readymade material for the researchers and academicians as it covers a
wide range of subject areas belonging to several majors falling under the umbrella
of e-waste management. Additionally, it may be treated as a handbook which will
suffice the needs of policymakers, supply chain managers, and technology Ninjas.
The editors would feel rewarded if the concepts presented in the book come to the
social cause.
August, 2021
Siddhartha Bhattacharyya, Birbhum, India
Elizabeth Behrman, Kansas, USA
Mario Köppen, Fukuoka, Japan
Ivan Cruz-Aceves, Guanajuato, Mexico
Contributors
Dora Elisa Alvarado-Carrillo Debashis De

Department of Computer Science, Maulana Abul Kalam Azad University
Center for Research in Mathematics of Technology, Kolkata, India
(CIMAT), A.C.
Jalisco S/N, Col. Valenciana,
Sourav De
Guanajuato, Mexico
Department of Computer Science &
Juan Gabriel Avina-Cervantes Engineering
Telematics (CA), Engineering Division Cooch Behar Government Engineering
(DICIS), Campus College, Cooch Behar, India
Irapuato-Salamanca
University of Guanajuato, Mexico Alokananda Dey
Department of Computer Science &
Laura-Paulina Badillo-Canchola Engineering
Universidad Tecnológica de León. Blvd. RCC Institute of Information
Universidad Tecnológica No. 225, Technology Kolkata, India
Col. San Carlos, CP. 37670
León, Gto., México Sandip Dey
Department of Computer Science
Andrey Batranin Sukanta Mahavidyalaya, Dhupguri,
School of Mathematics and Information Jalpaiguri West Bengal, India
Engineering
Chongqing University of Education, Tulika Dutta
China Department of Computer Science &
Engineering
Siddhartha Bhattacharyya Assam University Silchar, Assam, India
Rajnagar Mahavidyalaya
Birbhum, India Miguel-Angel Gil-Rios
Universidad Tecnológica de León. Blvd.
A Chatterjee Universidad Tecnológica No. 225,
Department of Computer Science Col. San Carlos, CP. 37670
California State University Dominguez
León, Gto., México
Hills Carson, CA
D K Chaturvedi Sergey Gorbachev

Department of Electrical Engineering School of Mathematics and Information
Faculty of Engineering, Dayalbagh Engineering
Educational Institute (Deemed Chongqing University of Education,
University), Agra, India China
xxiii
xxiv Contributors
Soumyajit Goswami Sanjai Pathak

IBM India Private Limited Amity University, Uttar Pradesh
India Noida, India
Martha Alicia Hernandez-Gonzalez Jan Platos

Unidad Mdica de Alta Especialidad Department of Electrical Engineering &
(UMAE) - Hospital de Computer Science
Especialidades No.1. Centro Mdico VSB Technical University of Ostrava,
Nacional del Bajio, IMSS Czech Republic
Len, Gto., Mexico
Marı́a-del-Carmen Ruiz-Robledo
Maria-Dolores Juarez-Ramirez Universidad Tecnológica de León. Blvd.
Universidad Tecnológica de León. Blvd. Universidad Tecnológica No. 225,
Universidad Tecnológica No. 225, Col. San Carlos, CP. 37670
Col. San Carlos, CP. 37670 León, Gto., México
León, Gto., México
Mayank Sharma
Debanjan Konar Amity University, Uttar Pradesh Noida,
SRM University-AP India
India
Dmytro Shevchuk
Victor Kuzin School of Mathematics and Information
Russian Academy of Engineering Engineering
Russia Chongqing University of Education,
China
Ashish Mani
Amity Innovation & Design Centre H P Singh
Amity School of Engineering & Department of Electrical & Electronics
Technology, Amity University Uttar Engineering
Pradesh, Noida, India Amity School of Engineering &
Technology, Amity University Uttar
G. Manikanta Pradesh, Noida, India
Department of Electrical & Electronics
Engineering Vaclav Snasel
Amity School of Engineering & Department of Electrical Engineering &
Technology, Amity University Uttar Computer Science
Pradesh, Noida, India VSB Technical University of Ostrava,
Czech Republic
Emmanuel Ovalle-Magallanes
Telematics (CA), Engineering Division Sergio Eduardo Solorio-Meza
(DICIS), Campus Universidad Tecnolgica de Mxico
Irapuato-Salamanca (UNITEC) Campus Len
University of Guanajuato, Mexico Len, Gto., Mexico
1 An Introductory Illustration
to Hybrid Quantum-
Inspired Metaheuristics
1.1 INTRODUCTION
Metaheuristic is basically a stochastic algorithmic framework, which may adopt a
predefined set of strategies to build several heuristic algorithms [1][2]. They are
efficient, problem-independent approximate algorithms, which can be successfully
applied to solve various optimization problems. In the 1970s, the researcher devel-
oped a new kind of algorithm, which integrates fundamental heuristics in higher
level frameworks to effectively explore a search space. These techniques are nowa-
days popularly known as metaheuristics. There exists a number of metaheuristics in
the literature, which may include, but not restricted to, Particle Swarm Optimization
(PSO) [3], Ant Colony Optimization (ACO) [4], Simulated Annealing (SA) [5], Tabu
Search (TS) [6][7][8], Differential Evolution (DE) [9] to name a few.
Optimization is an effort of attaining the best possible outcomes under a specified
state of affairs. In many real-life scenarios, such as maintenance, construction, and
other several cases, a concrete decision may need to be taken to reach at certain
goal. Generally, the outcome of such problems is achieving either minimum effort
or maximum benefit. This effort can be typically manifested as a function of a set
of specific variables [10][11]. As a consequence, optimization can be coined as the
process of deriving the criteria, which provides the minimum or value merit of a
function. When the optimization deals with a single objective function, it is called
“single objective optimization.” In case of “multi-objective optimization,” there must
have a systematic approach to be followed to derive the best possible solutions by
using at least two objective functions simultaneously. In some occasions, the solution
set is called pareto-optimal solutions in multi-objective optimization [12][13].
The paper is arranged in the following manner: The outline of different quantum-
inspired metaheuristics is briefly presented in Section 1.2. This section presents the
detailed discussion on metaheuristics of different types. Section 1.3 provides few
important entanglement-induced optimization techniques that are available in the
literature. The W -state encoding of various optimization techniques is presented
and discussed in Section 1.4. Section 1.5 throws light on several quantum system-
based optimization techniques. Different bi-level and multi-level quantum-based
systems have been highlighted in this section. A few popular applications of
quantum-inspired metaheuristics are presented in Section 1.6. The chapter finally
concludes with relevant information in Section 1.7.
DOI: 10.1201/9781003283294-1 1
2 Hybrid Quantum Metaheuristics: Theory and Applications
1.2 QUANTUM-INSPIRED METAHEURISTICS

In essence, the metaheuristics are characterized by high-level strategies to explore
search spaces by utilizing different techniques. The “diversification” and “intensifi-
cation” are two popular terms used to define functioning of metaheuristics. The first
and second terms generally point to the exploration of search space and the search
experience gained from diversification. There is a balance between these two to ex-
ecute metaheuristic application. This helps in speedy identification of regions in the
search space, which possesses high-quality solutions. In addition, it helps to identify
the useful regions of the specified search space in a very small time frame that have
already been explored and also the regions, which cannot furnish any high-quality
solutions.
The metaheuristic algorithms can be classified and described in different ways
based on a specific point of view. This classification is shown in the following points:
1. Nature-inspired metaheuristics vs. non-nature-inspired metaheuristics
2. Memory-based methods vs. memory-less methods
3. Single point-based search vs. population-based search
The metaheuristics as shown in point 1 is basically done by considering the ori-
gins of various algorithms. Representative of nature-inspired algorithms may include
Fuzzy Systems (FS), Artificial Neural Networks (ANN), Evolutionary Algorithm
(EA), and Swarm Optimization (SO), etc. So far, these algorithms have successfully
been applied for solving different real-world problems. Particle Swarm Optimization
(PSO), Ant Colony Optimization (ACO) [9] belongs to SO, whereas Genetic Algo-
rithm (GA) and Differential Evolution (DE) are two examples of EA. Some popular
non-nature-inspired metaheuristics are known to be Tabu Search (TS) and Iterated
Local Search (ILS). The metaheuristics are often classified based on the memory
utilization, i.e., whether they deploy memory or not. Memory-based metaheuristics
as referred to point 2 generally use memory at every iterations and save the search
history. On the contrary, memory-less algorithms perform other tasks, for example,
sometimes they might accomplish a Markov process, and based on the information
gained, the next course of action is determined in the search process. Lastly, meta-
heuristics of point 3 are classified into two types, called a single point-based meta-
heuristics and population-based metaheuristics.
In general, algorithms that use a single solution at any point of time are called
trajectory methods. These methods include such metaheuristics that deal with local
search, like TS, ILS, and variable neighborhood search (VNS). In these metaheuris-
tics, the search procedure expounds a trajectory in the said search space. On the other
hand, other one (population-based metaheuristics) can be expounded as the evolution
of either points or a probability distribution in/over the search space.
The quantum-inspired metaheuristics are developed by considering two differ-
ent subareas of computer science called quantum computing (QC) and Evolutionary
Computing (EC). Unlike “pure” quantum algorithms, such as Grover’s search al-
gorithm [14] or Shor’s factorization algorithm [15], in which the functional quan-
tum computer is necessitated for their efficient execution, the quantum-inspired
An Introductory Illustration to Quantum-Inspired Metaheuristics 3
algorithms exploit the features inspired by theory and principles of quantum mechan-
ical systems, like qubits (quantum bits), superposition of states, etc. It necessitates
these ideas for developing a computing paradigm much speedy than the conven-
tional computing framework. The viable reason for this enhanced computing speed
is achieved by dint of exploiting the inherent parallelism perceived in the qubits,
the basic unit of a quantum computer. This makes QCs to be far more effective in
comparison with their classical counterparts for obtaining factorization of large num-
bers [15] and searching elements from databases [14].
Quantum-Inspired Metaheuristics are comparatively the new area of research of
metaheuristics class. They draw their faces fundamentally from two dissimilar fields,
viz., Metaheuristic and Quantum Computing. Quantum computing is fundamentally
harnessing and utilizing the astounding laws of quantum mechanics for processing
information. A conventional computer (traditional computer) employs long stream
of “bits,” which essentially encode either 0 or 1. On the contrary, a quantum com-
puter employs quantum bits or qubits. A quantum bits is a quantum system, which
basically encodes 0 and 1 into two perceptible quantum states. Owing to the fact that
quantum bits behave quantumly, the happening of “superposition” and “entangle-
ment” can be capitalized for getting better efficiency. A quantum computer is able to
process a huge number of calculations at once. Unlike a classical computer, which
works with 0s and 1s, a quantum computer can use 0s, 1s, and their superposed
form. Hence, using the amazing features of quantum computing, called “superpo-
sition” and “entanglement,” a quantum computer can perform any certain impen-
etrable tasks more efficiently and more quickly in comparison with their classical
counterpart.
1.2.1 LOCAL SEARCH METAHEURISTICS

Nowadays, local search metaheuristics gained very much popularity among other
different approaches. Local search (LS) makes an effort to find good solutions in
an iterative way by changing the current solution. The said changes are generally
called moves. These moves are usually “small” keeping it in mind that the adjacent
solutions are comparatively in close proximity to each other.
The neighborhood of a solution can be described as the set of solutions, which
are achieved due to a single move to the specific solution. The general and most
common approach adopted by these algorithms is that instead of taking the best
improvement solution, they explore a predefined neighborhood to find the best move.
They usually do not explore neighborhood at a whole; rather the first improving move
is espoused by these kinds of algorithms by applying their first improvement strategy.
The move can be varied for different occasions and can be defined accordingly. At
each iteration, a solution is being chosen from its neighborhood to replace the current
solution using a specified rule called search strategy.
The local search metaheuristics are generally found to be computationally effec-
tive since in most cases it takes lesser time to search for an improving move. There
exist numerous search strategies in the literature; out of them, steepest ascent or
steepest descent is probably the most common one. In this strategy, the best move
is found from its neighborhood and used for its later operation. This strategy is of-
ten used in metaheuristics that is called hill-climbers. The other popular strategy is
popularly known as mildest ascent/descent strategy, which usually chooses that par-
ticular solution that improves the present solution by a very little amount. On the
other hand, one important move strategy may be exemplified as the first improving
strategy, where instead of other moves, the first move is chosen that improves the
present solution.
Simulated Annealing (SA) is a very popular metaheuristics, which employs a
move strategy that imitates the annealing procedure of a crystalline solid. It was
first introduced by Kirkpatrick et al. [5] in 1983, which was designed on the basis
of the method as presented by Metropolis et al. [16] in 1953. A concept called “lo-
cal optimum” was introduced to signify the solution, which is better than any other
solutions in the neighborhood. The “global optima” is the best solution to be found
in any optimization problem. When it is found that the current solution falls into a
local optimum, there must have a strategy that the metaheuristic must adopt to “es-
cape” from this situation. These kindx of metaheuristics are basically structures that
depend mainly on iterative improvement for finding good solutions.
Other popular and commonly used strategies that are very frequently espoused to
have a random change, called perturbation to the present solution. Two among them
are popularly known as Iterated Local Search (ILS) and Multi-start Local Search
(MLS), respectively [17].
One more important strategy of this category uses memory structures to record in-
formation of the previous progresses of the search procedures aiming to acquire good
solutions. The commonly used metaheuristics that adopt this strategy are termed as
Tabu Search (TS) [6][7][8] algorithms. Different categories of memory structures can
be used to recollect specific features of the trajectory that the algorithm has taken up
in its search space. A tabu list stores the last solutions that has been encountered
during operation and prohibits these solutions such that they are not visited again
until they are available on the list. Other popular kind of metaheuristic, known as
Guided Local Search (GLS) [18], presents a dissimilar type of memory, known to be
an augmented objective function, which encompasses a penalty factor for each of the
potential element. While it reaches a local optimum, the penalty factor is increased
for each element of the present solution. This allows the search process to escape
from falling in the local optimum.
1.2.2 CONSTRUCTIVE METAHEURISTICS

Constructive metaheuristics represent a discrete class from the perspective of local
search metaheuristics in such a way that they usually contrive solutions from their
basic elements, rather operating on complete solutions. In their way, they start oper-
ating taking an empty set and add a single element in the course of each iteration. It
can be noted that an operation can also be referred to as a move. Like other meta-
heuristics, they also continue their execution until a solution is found out or other
predefined stopping criteria are met.
There exists few popular constructive metaheuristics in the literature, out of

them Ant Colony optimization and Greedy Randomized Adaptive Search Procedure
(GRASP) are the most popular and mostly used metaheuristics. The basics of these
two methods are described below.
Foraging behavior of real ants is the sole inspiration for designing a very popular
and well-admired metaheuristic approach, called Ant Colony Optimization [19][20].
This behavior empowers ants to discover shortest paths between their nest and the
food sources. In their way, ants try to explore the surroundings of their nest at ran-
dom. Once an ant succeeds to locate a food source, thereafter carries some food to
their nest. While they are on the way of returning to nest, the ant spurts a chem-
ical from its body, called pheromone on the ground. More ants traverse the same
path means more amount of pheromone deposited, which in turn attract other ants to
follow the same path for food source. Hence, this chemical (pheromone trails) em-
powers the real ants to communicate with each other indirectly so that they can easily
locate shortest paths to reach at food sources from their nest. This amazing behavior
of real ant colonies is capitalized on artificial ant colonies for solving different op-
timization problems. Ant Colony Optimization (ACO) is a very popular example to
design meta-heuristic algorithms that can effectively be used to solve various Com-
binatorial Optimization (CO) problems. In 1991, Dorigo et al. [19, 20] first presented
an algorithm within this framework. Since then, several algorithms, especially dif-
ferent variants of this basic algorithm, have been presented in the literature. In ACO,
a parameterized probabilistic model, sometimes known to be the pheromone model,
is imposed to define simulation of this chemical. This model comprises a set of pa-
rameters, whose values behave like a memory that can be used to keep track of the
search procedure. Several quantum-inspired ant colony optimization techniques have
been introduced in the recent years. Wang et al. [21] developed a novel evolutionary
optimization technique where qubit was represented in ACO. Dey et al. [22][23][24]
developed few popular quantum-inspired metaheuristics that have been designed for
bi-level/multi-level thresholding of gray scale/color images.
Like ACO, GRASP is also a popular metaheuristic use to solve variant prob-
lems of combinatorial optimization. GRASP constructs a greedy randomized solu-
tion and triees to improve it through a local search in succession within different iter-
ations [25]. The GRASP basically exploits in two phases, called construction phase
and local search phase. The greedy solutions (randomized) are usually generated by
appending elements to the set of solution of the given problem. They are chosen from
the list of elements, which are basically ranked by a selected greedy function. Each
solution is ranked on the basis of its quality which it will attain. The elements having
well-ranked are frequently put down in a restricted candidate list (RCL) for acquiring
variability in the said candidate solutions. The solution in RCL is chosen at random
at the time of building up the solution. Hart and Shogan [26] first introduced this
greedy randomized construction approach in 1987; sometimes this method is called
semi-greedy heuristic.
In the literature, there exist several variations of this algorithm, like the Reactive
GRASP. In Reactive GRASP, the parameter that is used to define the limit of the
RCL during its first phase is self-adjusted based on the quality of the solutions found
earlier [27]. In addition, there also exist several other techniques of this kind; some of
them are employed to speed up in searching in various fields, like cost perturbations,
memorization and learning, and many others [28].
1.2.3 POPULATION-BASED METAHEURISTICS

These metaheuristics deal with a population comprising a set of solutions at each
iteration instead of a single solution. The individuals in the population are updated at
every iteration using certain operators provided by the particular metaheuristics. This
is essential for getting population diversity, aiming to have better outcome at every
iteration. Population-based metaheuristics explore the search space in a consistent,
intrinsic way. The performance of different metaheuristics differs, mainly because of
the way of manipulation of the population.
Population-based metaheuristics are intended to find favorable solutions by iter-
atively choosing and thereafter combining current solutions from a set, known as
population. The paramount members in this class are popularly known as evolution-
ary algorithms as they imitate the features of natural evolution. The term “EA” can be
used to present batch of algorithms that surrounds the extensive range of evolution-
based metaheuristics. Some representative of most investigated population-based
metaheuristics are Genetic Algorithm (GA) [29][11], Evolution Strategies (ES) [30],
Evolutionary Computation (EC) [31], Differential Evolution [9] to name a few. The
population of individuals is updated differently for different population-based meth-
ods. For example, the recombination and mutation operators are used to modify the
population of Evolutionary Computation, and the same is occurred for ant colony op-
timization by the guidance the pheromone trails and other heuristic information. Dif-
ferent population-based metaheuristics are described in the following subsections.
Deterministic population-based algorithms for EA are known to be Path Relinking
(PT) and Scatter Search (SS) [32][33]. SS encodes solutions as real-valued vectors
and thereafter locates new solutions by producing concave or convex linear combina-
tions of the said vectors. PT, on the other hand, introduces the idea of a path between
various high-quality solutions, basically generalizing the idea of linear combination.
Paths generally comprise elementary moves, which may be used by one in local
search metaheuristics. This moves to a path that transmute a single solution (initi-
ating solution) into a corresponding second solution (guiding solution) one at ones.
PT can therefore be appraised a local search heuristic, where the move to be carried
out is selected on the basis of the condition that the move must guide the solution in
close proximity to the guiding solution. The population (also called reference set),
which is used usually used to select initiating and guiding solutions, is also updated
after the generation of the new solutions as stated by deterministic rules in both SS
and PT.
1.2.4 HYBRID METAHEURISTICS

In the literature, metaheuristics have been exhaustively and extensively applied in
solving different global optimization problems [34][35][36]. In the recent years, the
idea of coalescing several metaheuristics has come out, that is popularly known as
hybridization of metaheuristics or, in simple word, hybrid metaheuristics. The hy-
brid metaheuristics are intended to exploit the features of different metaheuristics
and their optimization strategies [37]. The concept of hybridization can play a very
effective role, especially in solving different complex optimization problems. The
selection of right combination of metaheuristics usually instigates in achieving opti-
mum performance. The number of researchers has presented several hybrid systems
in the literature that have been efficiently and successfully applied in solving tradi-
tional optimization problems [38][39][40][41][42]. Two popular hybrid architectures
of this are Multi-agent Metaheuristic Hybridization (MAMH) [38] and Multi-agent
Metaheuristic Architecture (MAGMA) [43].
1.3 ENTANGLEMENT-INDUCED OPTIMIZATION

Quantum entanglement is an important and interesting concept in quantum comput-
ing. It is somehow related to the observed physical experience that happens at the
time of generating, communicating, and sharing spatial proximity of a set of two or
group of particles in such a way that the quantum state of the said particles cannot be
elucidated separately of the state of the others, in spite of the particles being kept in
separation at a large distance. The term “optimization” is referred to the method that
can be used for successfully locating the best solution among the obtained feasible
solutions. There are different kinds of optimization problems (continuous or discrete)
available that can be grouped into a pair of categories based on the variables used to
define the problem.
Of late, several quantum-inspired evolutionary algorithms (QIEAs) have been
introduced in the literature. These algorithms generally have expertise in their
search ability maintaining a healthy balance between exploration and exploitation.
QIEAs can usually be considered as a type of estimation of distribution algorithms
(EDAs) [44][45][46]. Han and Kim first presented this kind of algorithm (QEA) in
2002 [47]. This algorithm has the ability to [47] [48] explore a search space us-
ing very little amount of individuals and exploit efficiently to find the global solu-
tion. A lot of researchers have coalesced QEA with other popular heuristics, such as
cuckoo search algorithm, PSO, gravitational search, immune clonal algorithm, and
tabu search to propose various quantum-inspired algorithms. The principles of QEA
are combined with these quantum-inspired approaches to form several quantum-
inspired algorithms. Despite having certain drawbacks, QIEAs may possess several
appealing features, which may include better search abilities, least computational
costs, and even simple from implementation point of view in comparison with other
existing metaheuristics.
The entanglement induced optimization algorithms are new achievement from
research point of view. Like other quantum-inspired optimization algorithms that
exploit various features of quantum computing in the development process, the en-
tanglement induced optimization algorithms use the quantum entanglement feature
to develop these kinds of algorithms. When dealing with high-dependence problems,
the conventional optimization techniques have been proved to be non-efficient, espe-
cially in locating the global optimum. This problem has been addressed by introduc-
ing a novel meta-heuristic, called the entanglement-enhanced quantum-inspired tabu
search algorithm (Entanglement-QTS) [49].
The quantum-inspired tabu search (QTS) is a popular, effective, simple, and ro-
bust meta-heuristic. Unlike other QIEAs, this novel (QTS) algorithm interestingly
applies both the best and worst solutions. The reason behind this is to guide the indi-
vidual to chase toward finding a better solution and also to go out of the way from a
worse solution. QTS can show its efficiency to speedily reach at the global optimum.
The several applications of QTS algorithm can be found in the literature [50][51][52],
which have encouraging and incomparable searching capability. Entanglement-QTS
uses backbone of the quantum-inspired tabu search and the encouraging feature of
quantum computing, called quantum entanglement.
In comparison with the other QIEAs, the Entanglement-QTS uses qubits, which
are in entangled states. Basically, these entangled states can be demonstrate as an
inflated degree of correlation, render the variables twine together. It signifies a state-
of-the-art thought that can notably perk up to deal with high-dependence and multi-
modal problems. This algorithm is able to find optimal solutions, balance in diversi-
fication and intensification. Entanglement-QTS uses quantum not gate, which helps
to escape several local optimal solutions, strengthen the intensification consequence
by entanglement local search, and speed up the optimization procedure by applying
entangled states.
1.4 W-STATE ENCODING OF OPTIMIZATION ALGORITHMS

The W state is a very new, interesting, and appealing feature of quantum comput-
ing. This can be defined as an entangled quantum state comprising three qubits as
presented below.
1
|Ws ⟩ = √ (|001⟩ + |010⟩ + |100⟩) (1.1)
3
where Ws represents the said quantum entangled state.
This is an astonishing property of quantum computing that is usually used to
represent a certain class of multipartite entanglement. It may be very useful for taking
part for performing different applications in several fields like quantum information
theory. The Ws can be thought as the representative of one non-bi-separable class
comprising three-qubit states. The interesting feature is that it can not be changed
by any means into each other using any of the local quantum functioning [53]. One
interesting characteristic of Ws state is that it is robust as because when one qubit is
lost by any means, the entanglement property among the participating states are not
getting destroyed as a whole; rather, the other two states will be remained entangled
to each other.
The entanglement feature is an encouraging topic in the field of research now

adays. Many researchers have found interest in working in this direction. In opti-
mization purpose, several researchers have widely used quantum entanglement in a
variety of ways [49]. The computation power of Ws state has been elaborately de-
scribed in [54]. Ws state has few appealing application in communication point of
view. A protocol, called quantum secure direct communication (QSDC), has been
introduced by Chen et al. [55] in 2008. The foundation and applications of W state
have been described in [56]. Cruz et al. presented few Ws state-based efficient quan-
tum algorithms in 2019 [57]. Like other encoding schemes, W -state can be encoded
to form population in different quantum behaved applications. The appealing feature
of W -state, in turn, can assist to form efficient algorithms, especially in optimization
perspective.
1.5 QUANTUM SYSTEM-BASED OPTIMIZATION

Optimization problem deals with observing the best possible solutions to any criteria-
based problem. In general, any optimization problem is expressed as either a max-
imization problem or a minimization problem. There are several optimization tech-
niques available in the literature that can be successfully applied in numerous fields
like engineering, economics and mechanics, etc. The potential of quantum comput-
ing may encourage to solve problems, which may not realistically be feasible using
traditional computers. In addition to this, there may have a considerable speed-up
in execution with regard to the popular classical algorithm. The appealing features
of quantum computing might help to build quantum optimization algorithms that are
built to solve optimization problems. There are several algorithms available in the lit-
erature that have been basically designed for dealing with several optimizations with
regard to image processing [23]. The following subsections illustrate the bi-level and
multi-level quantum system-based optimization.
1.5.1 BI-LEVEL QUANTUM SYSTEM-BASED OPTIMIZATION

Many researchers have already designed several quantum behaved system for bi-
level optimization. MirHassani et al. presented a Quantum Binary Particle Swarm
Optimization (QBPSO) algorithm [58] in which an effectual nested strategy was
adopted to deal with competitive facility location problems. This problem is about
to apprehend the almost all of a given market, to make maximum profit. Zhang
et al. [59] presented a novel strategic bidding optimization method using bi-level
programming and swarm intelligence. This proposed method works in two steps.
Firstly, the idea of generalized Nash equilibrium has been used to develop a general
multi-leader-one-follower nonlinear bi-level (MLNB) optimization method. After-
wards, a PSO-based method has been developed for solving the problem expounded
in the MLNB decision model. Dey et al. [60] introduced quantum behaved bi-level
optimization method for image thresholding. The authors proposed two novel al-
gorithms, called quantum-inspired genetic algorithm and quantum-inspired particle
swarm optimization to find optimum threshold value from grey scale images. Zhang
et al. [61] proposed an elite quantum-inspired PSO (EQPSO) algorithm, where an

elite approach is taken for the best particle (global best) to avert premature con-
vergence of the swarm. This algorithm has been employed to solve bi-level multi-
objective programming problem. The working principles of MRLDE have been stud-
ied by Kumar et al. [62]. The authors have developed a recently introduced variant
of DE merged with Otsu method. Chang et al. [63] presented a bi-level semantic
representation analyzing approach that can eliminate the shortcomings of existing
systems. Yan et al. [64] developed a bi-level programming model that can be applied
to optimize the strategy, taken for augmentation of transportation network capacity
within the proposed budget. The lower and upper level problems face some difficul-
ties, which were solved by proposing a quantum-inspired evolutionary algorithm in
this paper. Later, the quantum behaved procedure relating to bi-level image thresh-
olding has been proposed using simulated annealing by Dey et al. [65].
1.5.2 MULTI-LEVEL QUANTUM SYSTEM-BASED OPTIMIZATION

Like bi-level quantum behaved system, there exist many applications in the multi-
level domain. The functionality of a bi-level system can be enhanced to its corre-
sponding multi-level frame by modifying its computational structure. The complex-
ity, in general, may increase for this required modification. Dey et al. [66] presented
two quantum-inspired approaches, namely, Quantum-Inspired Simulated Annealing
and Quantum-Inspired Ant Colony Optimization that deal with multi-level thresh-
olding. A novel evolutionary algorithm has been presented by Tkachuk [67] that can
efficiently handle optimization problems. This algorithm has been built using quan-
tum computations technology. In this algorithm, the notion of many-valued quantum
logic has been incorporated, which in turn made the proposed system more efficient
than other classical quantum genetic algorithm. Dey et al. [22][23][24] proposed sev-
eral quantum-inspired metaheuristics for color image thresholding in multi-level do-
main. A protocol for performing quantum reinforcement learning (QRL) with quan-
tum technologies (QTs) has been presented by Cárdenas et al. [68]. In comparison
with recent QRL, this method does not rely on coherent feedback during its learning
process, which in turn enables its implementation in an extensive diversity of quan-
tum systems. Niemann et al. [69] proposed a strategy that has the ability to verify
the equivalence of two different quantum behaviors irrespective of the dimension of
the specified underlying quantum system. This proposed system can be assimilated
into data-structures like Quantum Multiple-Valued Decision Diagrams (QMDDs),
which enables an effectual verification of the proposed system. Carrasco et al. [70]
presented a paper, where they investigated a fast and sturdy approach to control the
quantum states of a given multi-level quantum system. For this purpose, they used
a dual frequency time-varying potential. Grace et al. [71] presented a formalism to
encode the logical basis of a qubit into multiple physical level’s subspaces. Roy and
Das [72] presented a multi-level quantum system, where optimal control of this sys-
tem has been synthesized. The steering problem associated with this quantum system
has been solved by minimizing energy cost functional of this control.
1.6 APPLICATIONS OF QUANTUM-INSPIRED METAHEURISTICS

Quantum-inspired Particle Swarm Optimization (QPSO) algorithm [73] presented by
Gao et al. with the help of quantum evolutionary theory and PSO. This algorithm is
employed to optimize some benchmark functions and also some cognitive radio spec-
trum allocation problem. A quantum-inspired evolutionary algorithm (QEA), named
quantum-inspired Tabu search (QTS), is presented in [74]. The classical Tabu search
algorithm and the characteristics of quantum computation, such as superposition, are
applied to form this method. This proposed method applied to solve different types
of NP-complete problems. A new metaheuristic algorithm, named the entanglement-
enhanced quantum-inspired tabu search algorithm (Entanglement-QTS) proposed by
Kuo and Chou [75] to take care of combinatorial and numerical improvement issues.
This algorithm depends on the quantum-inspired tabu search (QTS) algorithm in
blend with the quantum entanglement. The authors gave emphasis on the entangled
Q-bits, which can extemporize the treatment of multimodal and high-dependence
problems significantly. Wang et al. [76] applied an improved version of the conven-
tional quantum genetic algorithm to solve different types of optimization problems.
This approach can decide the rotating angle, the self-adaptive rotating angle strat-
egy likewise with the quantum disaster operation, and quantum mutation operation.
A survey and analysis of the quantum-inspired evolutionary algorithm is illustrated
by Xiong et al. [77]. Among different operators for the operation of quantum gate
update in quantum algorithm, Quantum rotation gate (QRG) gave emphasis in this
article. From the start, an itemized investigation on the classification of QRG per-
formed with the assistance of rotation direction and the magnitude of rotation angle
by analyzing and condensing different sorts of QRGs and from that point forward, the
comparing definitions, depictions, and analyzes are delineated [77]. A comparative
study also presented with the help of different kinds of complex function optimiza-
tion problems. The hybrid quantum-inspired firefly and particle swarm optimization
(QIFAPSO) algorithm presented in [78] to solve continuous optimization problems.
In this method, the basic concepts of quantum computing, i.e., superposition states
of Q-bit and quantum measure used to obtain better diversified solutions. The fire-
fly algorithm and PSO algorithm are adapted with the quantum representation for
potential solutions.
Quantum-inspired metaheuristic algorithms are also applied for pattern recogni-
tion and image analysis. Quantum-inspired Particle Swarm Optimization (QiPSO) in
combination with Evolving Spiking Neural Network (ESNN) applied for string pat-
tern recognition [79]. A quantum genetic algorithm (IQGA) with the help of adap-
tive adjustment strategy of the rotation angle and the cooperative learning scheme is
presented for multilevel thresholding-based image segmentation [80]. In this article,
the adaptive adjustment strategy tried to improve the convergence strategy, search
ability and stability, and the cooperative learning improved the search ability in the
high-dimensional solution space. Dey et al. [81] presented a quantum-inspired sperm
whale metaheuristic algorithm to multilevel thresholding-based image segmenta-
tion. The fundamental operators of quantum computing fused with the sperm whale
metaheuristic algorithm. Color MRI image segmented efficiently by the proposed
quantum-inspired genetic algorithm-based FCM [82][83]. Liu et al. [84] presented

a chaotic quantum-behaved PSO based on lateral inhibition (LI-CQPSO) algorithm
to solve the image matching problems. In this approach, the chaos theory, and quan-
tum and lateral inhibition combined to get better result. The problem of local best
of PSO can be solved by chaos, the searching performance, as well as parameter
control of PSO can be handled efficiently by quantum and the edge of the images
can be extracted by the literal inhibition. The fuzzy C-means (FCM) in combination
with four-chain quantum bee colony optimization (FQABC) applied to segment the
images [85]. The four chains quantum encoding method induced in the artificial bee
colony (ABC) optimization algorithm, i.e., QABC, to improve the performance of
ABC algorithm and the improved QABC algorithm is employed to search the opti-
mal initial centers of FCM.
Swarm intelligence also made a notable signature in the field of robotics. Due
to self-autonomy and unexampled levels of distributed intelligence, advancement in
the field of robotics is notable. Robots can be applied in different real-world scenar-
ios, like complex manufacturing, health monitoring, disaster management, complex
logistics, etc. To make it more effective, the swarm intelligence, a significant part
of the computational intelligence, can be merged with the robotics. Combination of
these two fields is known as swarm robotics. Swarm robotics successfully applied
for smart sensing, correspondence and association functionalities blessed to these
small robots, which take into consideration cooperative data detecting, activity and
information deduction from the environment [86]. Osaba et al. [86] illustrated an
overview of the research contribution in the field of swarm robotics, which will end
up being a significant research impetus of the Computational Intelligence arena in
the recent future. Tan and Zheng [87] narrated a detailed survey in the field of swarm
robotics. Basically, inspired from the nature, swarm robotics is a combination of the
vast field of swarm intelligence and robotics. Contreras-Cruz et al. [88] suggested a
hybrid algorithm of artificial bee colony and evolutionary programming to solve the
mobile robot path planning problem.
In the field of optimization engineering, real-time tasks scheduling in different
complex environments is an all-time daunting task for the researchers. Some of the
relevant and prominent research areas of real-time tasks are manufacturing automa-
tion, telecommunication systems, multimedia applications, robotics,control system,
embedded systems, etc. As the time progresses, the applications of real-time tasks
are becoming more complex, time-consuming, and sophisticated. The application
area of quantum-inspired genetic algorithm (QIGA) are found in various domains,
including flow shop scheduling [89][90], power system optimization [91], thermal
unit commitment [92], network design problems [93], etc. The NP-hard combinato-
rial optimization problem with strong engineering backgrounds [94], multiobjective
flow shop scheduling problem (FSSP) [94] solved by the hybrid quantum-inspired
genetic algorithm (HQGA) [94]. Han and Kim [95] proved that the quantum-inspired
evolutionary algorithm (QEA) performed better than the conventional genetic algo-
rithm on combinatorial optimization of the knapsack problem. The termination cri-
terion, a two-phase scheme and Q-gate are changed to improve the presentation of
QEA. Konar et al. [96] proposed a Hybrid Quantum-Inspired Genetic Algorithm

(HQIGA) in multiprocessor environment for real-time scheduling efficiently. Ro-
tation gate used in HQIGA to investigate the variable chromosomes portrayed by
qubits in Hilbert hyperspace. In this approach, random key distribution is utilized to
change over the qubits chromosomes to legitimate timetable arrangements.
Economic load dispatch (ELD), a significant research field in the operation of
thermal power plants, is a non-linear and complex optimization problem. Using ELD,
power generation can be assigned to coordinate the load demand at negligible con-
ceivable expense while fulfilling all the units and system constraints [97]. Due to
different types of non-smooth cost function with equality and inequality constraints
applied in the ELD problem, finding the global optimal solutions is very much diffi-
cult using the traditional approaches. Chaos Quantum-Inspired Particle Swarm Opti-
mization (CQPSO) presented in [97] for ELD problem with better search and higher
convergence ability. The concept of quantum computing as well as the implemen-
tations of self-adaptive probability selection and chaotic sequences mutation in this
approach. Hassan et al. [98] presented quantum-inspired bat algorithm for solving
economic load dispatch problem. Many-objective environmental economic dispatch
(EED) issues can be unraveled by the quantum-inspired particle swarm optimization
(QPSO) technique [99]. A definitive study of research works in solving fathoming
different aspects of Economic Load Dispatch (ELD) problems of power system engi-
neering using various kinds of PSO algorithms narrated in [100]. This article focused
on five main areas of ELD problems [100], such as single objective economic load
dispatch, economic load dispatch with non-conventional sources, dynamic economic
load dispatch, multi-objective environmental/economic dispatch, and economic load
dispatch of microgrids.
Wireless Sensor Network (WSN) [101], an emerging and fast growing technology
in the context of Industry 4.0 [102], has been viewed as one of the predicting tech-
nologies in smart manufacturing. In this modern life, WSNs are applied to monitor
the running equipment in a complex manufacturing environment. WSNs are utilized
not exclusively to give administrations lifetime yet in addition to achieve fast and
high-quality transmission of equipment monitoring data to monitoring centers [103].
The Quantum Ant Colony Multi-Objective Routing (QACMOR), a WSN routing
algorithm, presented in mix with quantum computing and multi-objective fitness
function into the routing research algorithm. It is developed to monitor the manu-
facturing environments. In this method, the node pheromone is represented by the
quantum bits and the pheromone of the search path are updated by rotating the quan-
tum gates. Ullah and Wahid [104] proposed a quantum-inspired genetic algorithm
for topology control in the wireless sensor network. In this method, a linked quan-
tum register that are binary pair of genes has been presented and its local proximity
comprise the highly connected nodes and the energy consumption is very low. A
quantum-inspired genetic algorithm can be proficient enough for grouping of WSNs
to boost the lifetime of WSNs [105].
1.7 CONCLUSION
This chapter presents an outline of the basic theory and concept pertaining to quan-
tum inspired metaheuristics. This chapter throws light on several types of quantum
inspired metaheuristics in details. This chapter also comes up with a bird’s eye view
on different bi-level/multi-level quantum system-based optimization techniques. In
addition to that, several entanglement induced optimization techniques and W -state
encoding of optimization methods have also been discussed. The applications related
to the theme of the topic have been provided that would also certainly bring up to
date the readers.
REFERENCES
1. C. Blum and A. Roli. Metaheuristic in combinatorial optimization: Overview and concep-
tual comparison. Technical report, IRIDIA: Technical Report, 13, 2001.
2. F. Glover and G. A. Kochenberger. Handbook on Metaheuristics. Kluwer Academic Pub-
lishers, 2003.
3. K. Kennedy and R. Eberhart. Particle swarm optimization. In: Proceedings of the IEEE
International Conference on Neural Networks (ICNN95), Perth, Australia, 4:1942–1948,
1995.
4. M. Dorigo, V. Maniezzo, and A. Colorni. The ant system: Optimization by a colony of co-
operating agents. IEEE Transactions on Systems, Man, and Cybernetics, Part B, 26(1):29–
41, 1996.
5. S. Kirkpatrik, C.D. Gelatt, and M.P. Vecchi. Optimization by simulated annealing. Science,
220:671–680, 1995.
6. F. Glover. Tabu search – Part I. ORSA Journal on Computing, 1(3):190–206, 1998.
7. F. Glover. Tabu search – Part II. ORSA Journal on Computing, 2(1):4–32, 1990.
8. F. Glover. Tabu search and adaptive memory programming: Advances, applications and
challenges. In: Barr, Helgason, and Kennington, editors, Interfaces in Computer Science
and Operations Research. Kluwer Academic Publishers, 1996.
9. R. Storn and K. Price. Differential evolution – a simple and efficient heuristic for global
optimization over continuous spaces. Technical report, Technical Report TR-95-012, ICSI.
10. J. L. Cohon. Multiobjective Programming and Planning. Academic Press, New York, 1978.
11. D. E. Goldberg. Genetic Algorithms in Search, Optimization and Machine Learning.
Addison-Wesley Longman Publishing Co., Inc., Boston, MA, 1989.
12. D. A. V. Veldhuizen and G. B. Lamont. Multiobjective evolutionary algorithms: Analyzing
the state-of-the-art. Journal of Evolutionary Computation, 8(2):125–147, 2000.
13. K. Deb. Multi-objective optimization using evolutionary algorithms. Wiley, Chichester,
UK, 1989.
14. L. K. Grover. Quantum computers can search rapidly by using almost any transformation.
Physical Review Letters, 80(19):4329–4332, 1998.
15. P. W. Shor. Polynomial-time algorithms for prime factorization and discrete logarithms on
a quantum computer. SIAM Journal of Computing, 26(5):1484–1509, 1997.
16. N. Metropolis, A. W. Rosenbluth, M.N. Rosenbluth, A. H. Teller, and E. Teller. Equa-
tion of state calculations by fast computing machines. The Journal of Chemical Physics,
21(6):1087–1092, 1953.
17. H. Lourenco, O. Martin, and T. St utzle. Iterated local search. Handbook on Metaheuris-
tics, 2003.
18. C. Voudouris and E. Tsang. Guided local search and its application to the traveling sales-
man problem. European Journal of Operational Research, 113(2):469–499, 1999.
19. M. Dorigo, V. Maniezzo, and A. Colorni. The ant system: An autocatalytic optimizing
process. Technical Report Technical Report TR91-016, Politecnico di Milano, Italy, 1991.
20. A. Colorni, M. Dorigo, and V. Maniezzo. Distributed optimization by ant colonies. In:
Proceedings of ECAL91 European Conference on Artificial Life, pages 131–142, Elsevier,
Amsterdam, The Netherlands, 1991.
21. L. Wang, Q. Niu, and M. R Fei. A novel quantum ant colony optimization algorithm and
its application to fault diagnosis. Transactions of the Institute of Measurement and Control,
30(3–4):313–329, 2008.
22. S. Dey, S. Bhattacharyya, and U. Maulik. Quantum inspired meta-heuristic algorithms for
multi-level thresholding for true colour images. In: Proceeding of 2013 Annual IEEE India
Conference (INDICON), Mumbai, India, 2013.
23. S. Dey, S. Bhattacharyya, and U. Maulik. Efficient quantum inspired metaheuristics for
multi-level true colour image thresholding. Applied Soft Computing, 56:472–513, 2017.
24. S. Dey, S. Bhattacharyya, and U. Maulik. New quantum inspired meta-heuristic techniques
for multi-level colour image thresholding. Applied Soft Computing, 46:677–702, 2016.
25. T. A. Feo and M. G. C. Resende. Greedy randomized adaptive search procedures. Journal
of Global Optimization, 6(2):109–133, 1995.
26. J. P. Hart and A. W. Shogan. Semi-greedy heuristics: An empirical study. Operations Re-
search Letters, 6(3):107–114, 1987.
27. M. Prais and C. C. Ribeiro. Reactive grasp: An application to a matrix decomposition
problem in TDMA traffic assignment. INFORMS Journal on Computing, 12(3):164–176,
2000.
28. M. G. C. Resende and C. C. Ribeiro. Greedy Randomized Adaptive Search Procedures.
Handbook on Metaheuristics, Springer, 2003.
29. J. Holland. Adaptation in neural artificial systems. University of Michigan, Ann Arbor, MI,
1975.
30. H.-G. Beyer and H.-P. Schwefel. Evolution strategies: A comprehensive introduction. Jour-
nal Natural Computing, 1(1):3–52, 2002.
31. D. B. Fogel. Evolutionary Computation. IEEE Press, Piscataway, NJ, 1995.
32. F. Glover, M. Laguna, and R. Marti. Fundamentals of scatter search and path relinking.
Control and Cybernetics, 39(3):653–684, 2000.
33. F. Glover, M. Laguna, and R. Marti. Scatter search and path relinking: Advances and ap-
plications. Handbook of metaheuristics, 2003.
34. I. Salman, O. Ucan, O. Bayat, and K. Shaker. Impact of metaheuristic iteration on artificial
neural network structure in medical data. Processes, 6(57), 2018.
35. P. Bertolazzi, G. Felici, P. Festa, G. Fiscon, and E. Weitschek. Integer programming mod-
els for feature selection: New extensions and a randomized solution algorithm. European
Journal of Operational Research, 250:389–399, 2016.
36. P. Festa and M. G. C. Resende. Basic components and enhancements. Telecommunication
Systems, 46:253–271, 2011.
37. C. Blum, J. Puchinger, G.R. Raidl, and A. Roli. Hybrid metaheuristics in combinatorial
optimization: A survey. Applied Soft Computing, 11:4135–4151, 2011.
38. G. Souza, E. Goldbarg, M. Goldbarg, and A. Canuto. A multiagent approach for meta-
heuristics hybridization applied to the traveling salesman problem. In: Proceedings of the
2012 Brazilian Symposium on Neural Networks, Curitiba, Parana, Brazil, 20–25 October
2012, pages 208–213, 2012.
39. Z. Tiejun, T. Yihong, and X. Lining. A multi-agent approach for solving traveling salesman
problem. Wuhan University Journal of Natural Sciences, 11:1104–1108, 2006.
40. X.-F. Xie and J. Liu. Multiagent optimization system for solving the traveling salesman
problem (TSP). IEEE Transactions on Systems, Man, and Cybernetics, Part B, 39:489–
502, 2009.
41. F. Fernandes, S. Souza, M. Silva, H. Borges, and F. Ribeiro. A multiagent architecture for
solving combinatorial optimization problems through metaheuristics. In: Proceedings of
the IEEE International Conference on Systems, Man and Cybernetics, San Antonio, TX,
USA, 11–14 October 2009, pages 3071–3076, 2009.
42. R. Malek. An agent-based hyper-heuristic approach to combinatorial optimization prob-
lems. In: Proceedings of the IEEE International Conference on Intelligent Computing and
Intelligent Systems (ICIS), Xiamen, China, 29–31 October 2010, pages 428–434, 2010.
43. M. Milano and A. Role. MAGMA: A multiagent architecture for metaheuristics. IEEE
Transactions on Systems, Man, and Cybernetics, Part B, 34:925–941, 2004.
44. G. X. Zhang. Quantum-inspired evolutionary algorithms: A survey and empirical study.
Journal of Heuristics, 17(3):303–351, 2011.
45. G. X. Zhang. Time-frequency atom decomposition with quantum-inspired evolutionary
algorithms. Circuits, Systems, and Signal Processing, 29(2):209–233, 2010.
46. M. D. Platel, S. Schliebs, and N. Kasabov. Quantum-inspired evolutionary algorithm: A
multimodel EDA. IEEE Transactions on Evolutionary Computation, 13(6):1218–1232,
2009.
47. K. H. Han and J.-H. Kim. Quantum-inspired evolutionary algorithm for a class of com-
binatorial optimization. IEEE Transactions on Evolutionary Computation, 6(6):580–593,
2002.
48. K. H. Han and J.-H. Kim. Quantum-inspired evolutionary algorithms with a new termina-
tion criterion, h gate, and two-phase scheme. IEEE Transactions on Evolutionary Compu-
tation, 8(2):156– 169, 2004.
49. S. Y. Kuo and Y. H. Chou. Entanglement-enhanced quantum-inspired Tabu search algo-
rithm for function optimization. IEEE Access, 5, 2017.
50. H.-P. Chiang, Y.-H. Chou, C.-H. Chiu, S.-Y. Kuo, and Y.- M. Huang. A quantum-inspired
Tabu search algorithm for solving combinatorial optimization problems. Soft Computing,
18(9):1771–1781, 2014.
51. Y.-H. Chou, S.-Y. Kuo, C.-Y. Chen, and H.-C. Chao. A rule-based dynamic decision-
making stock trading system based on quantum-inspired Tabu search algorithm. IEEE
Access, 2:883–896, 2014.
52. Y.-H. Chou, S.-Y. Kuo, C. Kuo, and Y.-C. Tsai. Intelligent stock trading system based on
QTS algorithm in Japan’s stock market. In: Proceedings of the IEEE International Confer-
ence on Systems, Man and Cybernetics, San Antonio, TX, pages 997–982, 2013.
53. W. Dur, G. Vidal, and J. I. Cirac. Three qubits can be entangled in two inequivalent ways.
Physical Review A, 62(6), 2000.
54. E. DHondt and P. Panangaden. The computational power of the W and GHZ states. Quan-
tum Information & Computation, 6(2), 2005.
55. X. B. Chen, Q. Y. Wen, F. Z. Guo, Y. Sun, G. Xu, and F. C. Zhu. Controlled quantum
secure direct communication with W state. International Journal of Quantum Information,
6(4):899–906, 2008.
56. M. M. Cunha, A. Fonseca, and E. O. Silva. Tripartite entanglement: Foundations and ap-
plications. Universe, 5(209), 2019.
57. D. Cruz, R. Fournier, F. Gremion, A. Jeannerot, K Komagata, T. Tosic, J. Thiesbrummel,
C. L. Chan, N. Macris, M. A. Dupertuis, and J. G. Clement. Efficient quantum algorithms
for GHZ and W states, and implementation on the IBM quantum computer. Advanced
Quantum Technologies, 2(5–6), 2019.
58. S. A. MirHassani, S. Raeisi, and A. Rahmani. Quantum binary particle swarm
optimization-based algorithm for solving a class of bi-level competitive facility location
problems. Optimization Methods and Software, 30(4):756–768, 2015.
59. G. Zhang, G. Zhang, Y. Gao, and J. Lu. Competitive strategic bidding optimization in
electricity markets using bi-level programming and swarm technique. IEEE Transactions
on Industrial Electronics, 58(6):2138–2146, 2011.
60. S. Dey, S. Bhattacharyya, and U. Maulik. Quantum inspired genetic algorithm and parti-
cle swarm optimization using chaotic map model based interference for gray level image
thresholding. Swarm and Evolutionary Computation, 15:38–57, 2014.
61. T. Zhang, T. Hu, J. W. Chen, Z. Wan, and X. Guo. Solving bi-level multiobjective pro-
gramming problem by elite quantum behaved particle swarm optimization. Abstract and
Applied Analysis, 2012, 2012.
62. S. Kumar, P. Kumar, T. K. Sharma, and M. Pant. Bi-level thresholding using PSO, artificial
bee colony and MRLDE embedded with Otsu method. Memetic Computing, 5:323–334,
2013.
63. X. Chang, Z. Ma, Y. Yang, Z. Zeng, and A. G. Hauptmann. Bi-level semantic representation
analysis for multimedia event detection. Memetic Computing, 27(5):1180–1197, 2017.
64. X. Yan, N. Lv, Z. Liu, and K. Xu. Quantum-inspired evolutionary algorithm for transporta-
tion network design optimization. In: Proceeding of 2008 Second International Conference
on Genetic and Evolutionary Computing, Hubei, China, 2008.
65. S. Dey, S. Bhattacharyya, and U. Maulik. Quantum-inspired multi-objective simulated an-
nealing for bi-level image thresholding. Quantum Inspired Computational Intelligence, Re-
search and Applications, 2017.
66. S. Dey, I. Saha, U. Maulik, and S. Bhattacharyya. New quantum inspired meta-heuristic
methods for multi-level thresholding. In: Proceeding of 2013 International Conference
on Advances in Computing, Communications and Informatics (ICACCI), Mysore, India,
2013.
67. V. Tkachuk. Quantum genetic algorithm on multilevel quantum systems. Mathematical
Problems in Engineering, 2018:1–12, 2018.
68. F. A. Cardenas-Lopez, L. Lamata, J. C. Retamal, and E. Solano. Multiqubit and multi-
level quantum reinforcement learning with quantum technologies. PLoS ONE, 13(7):1–12,
2018.
69. P. Niemann, R. Wille, and R. Drechsler. Equivalence checking in multi-level quantum
systems. International Conference on Reversible Computation, LNCS, 8507:201–215,
2014.
70. S. Carrasco, J. Rogan, and J. A. Valdivia. Speeding up maximum population transfer in
periodically driven multi-level quantum systems. Scientific Reports, 9, 2019.
71. M. Grace, C. Brif, H. Rabitz, I. Walmsley, R. Kosut, and D. Lidar. Encoding a qubit into
multilevel subspaces. New Journal of Physics, 8, 2006.
72. B. C. Roy and P. K. Das. Optimal control of multi-level quantum system with energy cost
functional. International Journal of Control, 80(8):1299–1306, 2007.
73. J. Cao H. Gao and M. Diao. A simple quantum-inspired particle swarm optimization and
its application. Information Technology Journal, 10(12):2315–2321, 2011.
74. H.-P. Chiang, Y.-H. Chou, C.-H. Chiu, S.-Y. Kuo, and Y.-M. Huang. A quantum-inspired
Tabu search algorithm for solving combinatorial optimization problems. Soft Computing,
18(9):1771–1781, September 2014.
75. S. Kuo and Y. Chou. Entanglement-enhanced quantum-inspired Tabu search algorithm for
function optimization. IEEE Access, 5:13236–13252, 2017.
76. H. Wang, J. Liu, J. Zhi, and C. Fu. The improvement of quantum genetic algorithm
and its application on function optimization. Mathematical Problems in Engineering,
2013(730749):1–10, 2013.
77. H. Xiong, Z. Wu, H. Fan, G. Li, and G. Jiang. Quantum rotation gate in quantum-inspired
evolutionary algorithm: A review, analysis and comparison study. Swarm and Evolutionary
Computation, 42:43–57, 2018.
78. D. Zouache, F. Nouioua, and A. Moussaoui. Quantum-inspired firefly algorithm with parti-
cle swarm optimization for discrete optimization problems. Soft Computing, 20(7):2781–
2799, July 2016.
79. H. N. Abdull Hamed, N. Kasabov, Z. Michlovsky, and S. M. Shamsuddin. String pattern
recognition using evolving spiking neural networks and quantum inspired particle swarm
optimization. In: Neural Information Processing. Lecture Notes in Computer Science, vol.
5864, pages 611–619. Springer, Berlin, Heidelberg, 2009.
80. J. Zhang, H. Li, Z. Tang, Q. Lu, X. Zheng, and J. Zhou. An improved quantum-inspired
genetic algorithm for image multilevel thresholding segmentation. Mathematical Problems
in Engineering, 2014(295402):1–12, 2014.
81. S. Dey, S. De, D. Ghosh, D. Konar, S. Bhattacharyya, and J. Platos. A novel quantum
inspired sperm whale metaheuristic for image thresholding. In: 2019 Second International
Conference on Advanced Computational and Communication Paradigms (ICACCP), pages
1–7, Feb 2019.
82. S. Das, S. De, and S. Bhattacharyya. True color image segmentation using quantum-
induced modified-genetic-algorithm-based FCM algorithm. In: Quantum-Inspired Intelli-
gent Systems for Multimedia Data Analysis, pages 55–94. Research Essentials Collection.
IGI Global, 2018.
83. S. Das, S. De, S. Bhattacharyya, and A. E. Hassanien. Color MRI image segmentation us-
ing quantum-inspired modified genetic algorithm-based FCM. In: Recent Trends in Signal
and Image Processing. Advances in Intelligent Systems and Computing, vol. 727, pages
151–164. Springer, Singapore, 2019.
84. F. Liu, H. Duan, and Y. Deng. A chaotic quantum-behaved particle swarm optimization
based on lateral inhibition for image matching. Optik, 123(12):1955–1960, 2012.
85. Y. Feng, H. Yin, H. Lu, L. Cao, and J. Bai. FCM-based quantum artificial bee colony
algorithm for image segmentation. In: Proceedings of the 10th International Conference
on Internet Multimedia Computing and Service, pages 6:1–6:7, 2018.
86. E. Osaba, J. Del Ser, A. Iglesias, and X.-S. Yang. Soft computing for swarm robotics: New
trends and applications. Journal of Computational Science, 39, 2020.
87. Y. Tan and Z. Zheng. Research advance in swarm robotics. Defence Technology 9:18–39,
2013.
88. M. Contreras-Cruz, V. Ayala, and U. Hernandez-Belmonte. Mobile robot path planning us-
ing artificial bee colony and evolutionary programming. Applied Soft Computing, 30:319–
328, 2015.
89. J. Gu, X. Gu, and B. Jiao. A quantum genetic based scheduling algorithm for stochas-
tic flow shop scheduling problem with random breakdown. IFAC Proceedings Volumes,
41(2):63–68, 2008.
90. J. Gu, X. Gu, and M. Gub. A novel parallel quantum genetic algorithm for stochastic
job shop scheduling. Journal of Mathematical Analysis and Applications, 355(1):63–81,
2009.
91. A. K. Al-Othman, F. S. Al-Fares, and K. M. EL-Nagger. Power system security constrained

economic dispatch using real coded quantum inspired evolution algorithm. International
Journal of Nuclear and Quantum Engineering, 1(5):4–10, 2007.
92. Y.-W. Jeong, J.-B. Park, J.-R. Shin, and K. Y. Lee. A thermal unit commitment approach
using an improved quantum evolutionary algorithm. Electric Power Components and Sys-
tems, 37(7):770–786, 2009.
93. H. Xing, X. Liu, X. Jin, L. Bai, and Y. Ji. A multi-granularity evolution based quantum
genetic algorithm for QoS multicast routing problem in WDM networks. Computer Com-
munications, 32(2):386–393, 2009.
94. B. Li and L. Wang. A hybrid quantum-inspired genetic algorithm for multiobjective flow
shop scheduling. IEEE Transactions on Systems, Man, and Cybernetics, Part B (Cybernet-
ics), 37(3):576–591, June 2007.
95. K.-H. Han and J.-H. Kim. Quantum-inspired evolutionary algorithms with a new termina-
tion criterion, h/sub /spl epsi// gate, and two-phase scheme. IEEE Transactions on Evolu-
tionary Computation, 8(2):156–169, April 2004.
96. D. Konar, S. Bhattacharyya, K. Sharma, S. Sharma, and S. R. Pradhan. An improved hybrid
quantum-inspired genetic algorithm (HQIGA) for scheduling of real-time task in multipro-
cessor system. Applied Soft Computing, 53:296–307, 2017.
97. K. Meng, H. G. Wang, Z. Dong, and K. P. Wong. Quantum-inspired particle swarm opti-
mization for valve-point economic load dispatch. IEEE Transactions on Power Systems,
25(1):215–222, Feb 2010.
98. H. Tehzeeb ul Hassan, M. U. Asghar, M. Z. Zamir, and H. M. A. Faiz. Economic load dis-
patch using novel bat algorithm with quantum and mechanical behaviour. In: 2017 Inter-
national Symposium on Wireless Systems and Networks (ISWSN), pages 1–6, Nov 2017.
99. F. Mahdi, P. Vasant, M. Abdullah-Al-Wadud, J. Watada, and V. Kallimani. A quantum-
inspired particle swarm optimization approach for environmental/economic power dis-
patch problem using cubic criterion function. International Transactions on Electrical En-
ergy Systems, page e2497, 12 2017.
100. M. N. Alam. State-of-the-art economic load dispatch of power systems using particle
swarm optimization. CoRR, abs/1812.11610, 2018.
101. C. Li, X. Xie, Y. Huang, H. Wang, and C. Niu. Distributed data mining based on deep
neural network for wireless sensor network. International Journal of Distributed Sensor
Networks, 11(7):1–7, 2015.
102. Xiaomin Li, D. Li, J. Wan, A. V. Vasilakos, C.-F. Lai, and S. Wang. A review of industrial
wireless networks in the context of industry 4.0. Wireless Networks, 23(1):23–41, 2017.
103. F. Li, M. Liu, and G. Xu. A quantum ant colony multi-objective routing algorithm in WSN
and its application in a manufacturing environment. Sensors, 19(15), 2019.
104. S. Ullah and M. Wahid. Topology control of wireless sensor network using quantum in-
spired genetic algorithm. International Journal of Swarm Intelligence and Evolutionary
Computation, 04, 08 2015.
105. M. Rathee and S. Kumar. Quantum inspired genetic algorithm for energy efficient clus-
tering in wireless sensor networks. In: 2016 IEEE 1st International Conference on Power
Electronics, Intelligent Control and Energy Systems (ICPEICES), pages 1–6, 07 2016.
2 AApproach
Quantum-Inspired
to Collective
Combine Basic Classifiers
in an Ensemble Bagging
2.1 INTRODUCTION
Machine learning algorithms are used to solve classification problems (supervised
learning), clustering (unsupervised learning), reinforcement learning, etc. Learning
consists in using the available data to adjust the variable parameters of the algorithm
in order to fit the algorithm to the specifics of the problem being solved. Hyperpara-
metric optimization (the problem of choosing a model) is the problem of finding a
set of hyperparameters of a learning algorithm, at which the highest efficiency of
the algorithm is achieved when solving a specific problem. The choice of a suit-
able method and its tuning have the greatest impact on the success of the solution
of the task at hand. The performance of machine learning algorithms depends on the
parameters of the algorithm and the chosen structure of the model. To solve the prob-
lem of optimal choice of a model, a number of methods have been proposed, which
can be divided into 2 groups: hyperparametric optimization algorithms and ensemble
methods, in which several machine learning algorithms are used in parallel to col-
lectively solve the problem. Although none of the proposed algorithms is a panacea
and does not guarantee a successful solution to the problem, experimental evidence
of their effectiveness was obtained [1][2]. To collectively combine the decisions of
the base classifiers, the voting method is usually used. In this chapter, we will look at
alternative methods of combining such as stacking and ensemble selection, and also
propose a new quantum-inspired approach based on metaheuristic hyperparameter
tuning algorithms.
2.2 BAGGING METHOD

The Bootstrap aggregating method was proposed by Leo Breiman to improve clas-
sification efficiency by combining classifiers built on random subsamples of a given
training sample.
Let there be a training sample X of size n. The bagging procedure generates L
new training samples Xi , each of size n, by uniform selection with substitution from
the original sample. Thus, the elements of the Xi samples may contain duplicates.
The probability of being sampled (a subsample of size L from the main sample “with
return”) for sufficiently large samples is approximately 0.632, i.e. about 36.8% of
DOI: 10.1201/9781003283294-2 21
Figure 2.1: Bagging with final ensemble combiner (Adapted from Gorodetsky, V.
& Serebryakov, S. (2006). Methods and algorithms of the collective recognition: a
survey. SPIIRAS Proceedings, 3(1), 139-171)
examples from the training sample may be unclaimed. This type of selection is called
Bootstrap. The L obtained samples are used to train L models, which in turn are com-
bined into a team. In regression problems, the outputs of the classifiers are averaged;
in classification problems, the voting method is usually used.
Bagging leads to improvements in unstable algorithms such as artificial neural
networks and classification and regression trees (CARTs). Due to the use of bagging
in some works, an improvement in pattern recognition was noted [3][4]. Thus, bag-
ging is useful in the case of different classifiers and instability, when small changes
in the initial sample lead to significant changes in the classification [5].
As seen in Figure 2.1, the final ensemble combiner can use different ways to
combine the classifier decisions, including voting, stacking and ensemble selection.
Existing classification algorithms can be combined into several groups according to
the principle of their operation:
1. Classifiers based on the similarity of objects.

2. Algorithms for statistical classification.
3. Classifiers based on the separability of classes in the attribute space.
4. Logical classification algorithms.
5. Neural networks.
2.3 CLASSIFIERS BASED ON SIMILARITY OF OBJECTS

The algorithms of this group are based on the compactness hypothesis, which as-
sumes that objects of the same class are most often similar to each other, and objects
A Quantum-Inspired Approach to Collective Combine Basic Classifiers 23
of different classes are different. The similarity of objects in an n-dimensional feature

space can be calculated using various metrics:
1. Euclidean distance: s
n
dE (x, y) = ∑(xi − yi )2 (2.1)
i
2. Manhattan distance:
n
dMan (x, y) = ∑ |xi − yi |2 (2.2)
i
3. Chebyshev distance:
dCh (x, y) = maxi (|xi − yi |) (2.3)
4. Minkowski distance of the pth order:
n 1
dMink (x, y) = (∑ |xi − yi | p ) p (2.4)
i
The most famous algorithms in this group are the following:
1. The method of constructing standards: Based on the training sample, reference

points are calculated, which are the centers of objects of each class in the attribute
space, according to the formula:
1
nk ∑
Eik = (xi : f (x̄) = k); i = 1, 2, . . . m; k = 1, 2, . . . K (2.5)
where, Ek is the standard of the kth class; nk is the number of objects in the sample
of the kth class; m is the dimension of the feature space and K is the number of
classes.
According to a certain metric, the distance from the classified object to each of
the standards is calculated, the object is attributed to the class, the distance to the
standard of which is minimal.
2. Method of k nearest neighbors [6]: For this algorithm, the data must be repre-
sented as a matrix of distances between the sampled objects, calculated according
to a certain metric. Classes of k objects closest to the classified object are consid-
ered. The object belongs to the class that is most often found among its neighbors.
When solving practical problems with this method, it is very important to choose
the correct metric for calculating the distance between objects, as well as the
value of the parameter k. If the value of the parameter is too small (for example,
k = 1), the algorithm becomes susceptible to the negative influence of outliers; if
the value of k is too high, too many neighboring objects are included in the cal-
culations, which can also negatively affect the quality of the classification, since
among the neighbors of the classified object there may be many objects of another
class.
3. The method of potential functions [7]: This method is based on the physical prin-
ciple of the potential of the electric field of a charged particle. The distance from
the classified object to each object of the training sample is calculated. The de-
cision rule is constructed as in the method of nearest neighbors, the difference is
that the sample object has some measure of importance (“charge”) relative to the
classified object.
2.4 STATISTICAL CLASSIFICATION ALGORITHMS

Statistical classification algorithms are based on estimating the distribution density
of classes over a sample. Depending on the estimation method, algorithms are dis-
tinguished with parametric and nonparametric density estimation, as well as density
estimation as a mixture of parametric distributions. Well-known algorithms of this
group:
1. Naive Bayesian classifier [8]: This method is based on the assumption that the
features that describe the objects of the sample are statistically independent. This
assumption greatly facilitates the problem of estimating the distribution density,
since instead of the n-dimensional density, it is necessary to estimate n one-
dimensional densities. Densities can be estimated both parametrically and non-
parametrically. According to the Bayes rule, the posterior probabilities of each
of the K classes are found, provided that the attribute x of the classified object is
measured:
f (x|i) × P(i)
P(i|x) = m ; i = 1, 2, . . . K (2.6)
∑ j f (x| j) × P( j)
where, f (x|i) is the assessment of the conditional distribution density of the at-
tribute x for the ith class and P(i) is an estimate of the prior probability of the
class.
The decision rule for the classified object x is as follows:
i∗ = arg maxi P(i|x); i = 1, 2, . . . K (2.7)
2. Parzen window method: This method uses nonparametric estimation of the den-
sity [9] of the distribution of classes for the available sample, therefore, it does
not put forward hypotheses about the structure of the distribution density func-
tion. The decision rule for classifying object x is as follows:
n
d(x, x j )
i∗ = arg maxi λi ∑ [y j = i]K( ); i = 1, 2, . . . K (2.8)
j=1 h
where, λi is the price of the correct answer for class i; n is the sample size; y j
is the class of the jth object; K(t) is the nuclear function; d(x, x j ) is the distance
between the classified object x and the object x j and h is the window width.
3. EM-algorithm (expectation-maximization): The EM algorithm [10] estimates the
density as a mixture of parametric distributions. In this algorithm, two stages are
iteratively performed: the estimation stage, in which the expected value of the
likelihood function is calculated, and the maximization stage, in which the pa-
rameters of the likelihood function are calculated that maximize it.
2.5 CLASSIFIERS BASED ON CLASS SEPARABILITY IN ATTRIBUTE

SPACE
These algorithms build a dividing surface in the feature space, dividing objects into
disjoint classes. The most famous methods of this group are as follows:
1. Fisher’s linear discriminant [11], also known as linear discriminant analysis, is ap-
plicable if the sample satisfies the following hypotheses: the classes are normally
distributed and the class covariance matrices are equal. Fisher’s linear discrimi-
nant is a simplification of the quadratic discriminant. In the case of two classes
in two-dimensional space, the dividing surface constructed using this method will
be a straight line. In the case of a larger number of classes, the dividing surface
will be piecewise linear.
2. Logistic regression [12]. For the case of 2 classes, a linear classification algorithm
is constructed with a decision rule of the form:
m
log reg(x, w) = sign( ∑ w j x j − w0 ) = sign(x, w) (2.9)
j=1
where, w j is the weight of the jth feature; w0 is the decision threshold; w is the
vector of weights and (x, w) is the scalar multiplication of the weights vector and
features of the object.
The problem of training the logistic regression algorithm is to find the optimal
vector of weights w that minimizes the loss function of the form:
n
1
L(w) = ∑ ln(1 + ) (2.10)
i=1 eyi (x,w)
3. The support vector machine [13] is one of the most popular supervised learning
methods for several reasons:
(i) The fastest method for finding the decision rule.
(ii) Reduces to solving a quadratic programming problem in a convex domain,
which always has a unique solution.
(iii) Finds the dividing surface of the classes with a dividing bar of maximum
width, which contributes to more confident classification.
In the case of two classes and a linearly separable sample, the decision rule of the
SVM algorithm takes the form:
yi ((w, xi ) + b) ≥ 1; i = 1, 2, . . . n (2.11)
where, yi is the class label of the ith sample object; w is the vector of weight
coefficients and n is the sample size.
The vector w is sought in the form:

n
w = ∑ λi yi xi (2.12)
i=1
Into this sum with nonzero coefficients λi includes only those objects of the se-
lection that lie on the dividing surface. These objects are called support vectors.
In the case of a linearly inseparable sample, the feature space Rn is transferred to
a space of higher dimension H using function ϕ (x), in which the sample becomes
linearly divisible. In this case, the decision rule is sought in the form:
yi ((w, ϕ (xi )) + b) ≥ 1; i = 1, 2, . . . n (2.13)

n
w = ∑ λi yi ϕ (xi ) (2.14)
i=1
There is also a multiclass support vector machine, which is reduced to dividing the
problem into several binary classification problems according to the “one against
all” scheme.
2.6 LOGICAL CLASSIFICATION ALGORITHMS

The principle of operation of logical classification algorithms is based on the con-
struction of hierarchical compositions of simple rules. The most prominent repre-
sentative of this group of algorithms is a decision tree [14]. The structure of the tree
consists of the so-called branches and leaves, in the branches signs are written on
which the objective function depends, and in the leaves - the values of the objective
function. Decision trees are suitable for both classification problems and regression
problems. In the case of the classification problem, the class to which the classifi-
cation object belongs is recorded in the leaves of the tree. In the case of regression
problems - a real number, The tree construction scheme looks like this:
1. The next feature fi(x) is selected.

2. The values of the selected feature are divided into several subgroups, thereby
dividing the feature space into several subspaces.
3. The procedure is recursively repeated until one of the stopping criteria is reached
- the maximum tree depth, the number of leaves, or the maximum size of each
leaf.
Depending on the principle by which the next feature for splitting is selected, and
how it is split, there are several variants of this algorithm:
1. Algorithm ID3 - in it, the next feature is selected according to the criterion of
information gain.
2. Algorithm C4.5 [15] is an improved version of ID3, the choice of a feature ac-
cording to the criterion of normalized increment of information.
3. Algorithm CART (classification and regression tree) [16].
In practice, in order to avoid the effect of overfitting after building a decision tree,
some of its branches are truncated to maintain better generalizing ability, this proce-
dure is called pruning.
2.7 NEURAL NETWORKS

A group of classification algorithms called neural networks [17] is based on the prin-
ciple of linking a large number of simple elements built in the form of an optimal
structure. Artificial neurons, mathematical constructs inspired by the study of nerve
cells in the human brain, act as simple elements. The concept of a neural network
was formalized by McCulloch and Pitts in 1943, and the first learning algorithm was
proposed by Hebb in 1949. In 1958, Rosenblatt invented the perceptron, the first
neural network used for pattern recognition. In those years, there was a high interest
in neural networks in scientific circles, but after Minsky in 1969 published a formal
proof of the limitations of the perceptron in recognizing complex, variable images,
interest in neural networks in general, and to the perceptron in particular, it dropped
sharply. A little later, in 1974, an error backpropagation algorithm was invented, and
in 1982 neural networks with feedback were proposed, the training of which was car-
ried out on the basis of minimizing the so-called network energy function. In 1986,
interest in the direction of neural networks was renewed after the backpropagation
algorithm was significantly improved. Later, this algorithm was successfully applied
to train a new generation of deep neural networks.
2.8 METHODS OF COMBINING BASIC CLASSIFIERS

As seen in 9.9, the final ensemble combiner can use different ways to combine the
classifier decisions, including voting, stacking and ensemble selection. Those com-
bination methods are briefly introduced as follows.
2.8.1 VOTING
Let a classifier ensemble consists of L base classifiers in the set D =
{D1 , D2 , . . . , DL }, and any object x ∈ Rn is assigned to one of the c possible classes
Ω = {ω1 , ω2 , . . . , ωc }.
For x to be classified, L classifiers output a matrix M = [mi, j ]; i = 1, 2, . . . , L; j =
1, 2, . . . , c.
1. Majority voting rule: Suppose mi, j ∈ {0, 1}, where mi, j = 1 if Di predicts x in
class ω j and mi, j =0, otherwise. x is assigned to ωk if
L L
c
∑ mi,k = max
j=1
∑ mi, j (2.15)
i=1 i=1
2. Average of probabilities rule: Suppose mi, j ∈ {0, 1}, where mi, j is the degree of
support that classifier Di gives to the hypothesis that x comes from class ω j , de-
noted as mi, j = PDi (ωi |x). x is assigned to ωk if
1 L c 1
L
∑
L i=1
mi,k = max ∑ mi, j
j=1 L i=1
(2.16)
2.8.2 STACKING
Stacking constructs a set of heterogonous or homogeneous base classifiers, and the
outputs of base classifiers are used to train metaclassifier which produces single out-
put as final classification result. The task of the metaclassifier is to correct any mis-
takes made by base classifiers and minimize the generalization error. Any classifica-
tion algorithm can be used to train base classifier or metaclassifier. The procedure of
stacking algorithm is as follows:
1. Step 1: split a dataset into three disjoint subsets: the training set, the validation
set, and the testing set;
2. Step 2: train a set of base classifiers on the training set;
3. Step 3: apply those base classifiers to classify the validation set;
4. Step 4: using the outputs of base classifiers from Step 3 as the features, along with
the true class label, to train the metaclassifier;
5. Step 5: test the metaclassifier on the testing set to evaluate the performance of
Stacking ensemble.
2.8.3 ENSEMBLE SELECTION

Most ensemble algorithms combine all base classifiers to construct an ensemble,
but it is difficult to determine that what base classifier should be included in the
ensemble, how many base classifiers are needed, and what ensemble strategy should
be used to combine base classifiers.
Combining a portion of base classifiers instead of using all of them to construct
an ensemble may be better. Such an ensemble is called as ensemble selection, which
can achieve strong generalization performance with small size of base classifiers. As
seen in Figure 2.2, the ensemble selection is built in three steps.
1. Step 1: A set of heterogonous or homogeneous base classifiers are trained for the
same task.
2. Step 2: The chosen algorithm is employed to compute the weight of base classi-
fiers, and a subset of base classifiers whose weight is bigger than a preset threshold
is combined to construct an ensemble.
3. Step 3: The output layer calculates the final degree of membership Y of jth exam-
ple x j to the class base on Majority voting rule (soft vote) or Average of probabil-
ities rule.
Figure 2.2: Ensemble selection (Modified from Gorbachev, S., Arkhipov, A., Gor-
bacheva, N., Bhattacharyya, S., Cao, J. & Kale, S. (2021). Study and Developing
of Diversity Generation Methods in Heretogeneous Ensemble Models. International
Journal of Distributed Computing and Technology, 7(1), 816)
(k)
The outputs Y j of the penultimate layer (Figure 2.2) for each ith class calculate
the degree of membership of the jth example x j of training sample to the class as
a weighted average linear combination of the normalized outputs of each classifier
(neuroexpert):
1 L (1) (1) 1 L (c) (c)
Y j = ∑ νi yi j , . . .Y j = ∑ νi yi j
(1) (c)
(2.17)
L i=1 L i=1
The weights of each classifier (neuroexpert) are calculated based on the number of
errors they made using the following methods:
(i) Fisher linear discriminant
(ii) logistic regression
(iii) Single-layer perceptron
(iv) SVM support vector machine, which is most significant in terms of maximizing
the separation ability between classes and in terms of reliability
(v) the <<naive>> Bayesian classifier (the most popular of the simple ones)
(vi) heuristic algorithms.
2.8.3.1 The <<naive>> Bayesian Classifier

(1) (1) (c) (C)
Let y1 (x j ), . . . yL (x j ), . . . y1 (x j ), . . . yL (x j ) be independent random variables.
(k)
Then the weighted voting coefficients νi (i = 1, 2, . . . L) for each kth class (k =
1, 2, . . . c) can be calculated using the formula:
1 − pi
νi = ln ; i = 1, 2, . . . L (2.18)
pi
where pi is the probability of errors of the ith classifier on the training sample. The
fewer errors it makes, the greater its weight. As a probability estimate, we can take
the error rate γi or γi + N1 , so that the denominator does not vanish, i.e.:
1 − γi
νi = ln ; i = 1, 2, . . . L (2.19)
γi + N1
If any neuroexpert makes more than half of the mistakes, then his weight is νi < 0.
Such an unreliable classifier is not taken into account in the meta-network, putting
νi = 0.
2.8.4 QUANTUM-INSPIRED METAHEURISTICS METHOD

This section provides an overview of population-based quantum inspired metaheuris-
tics. A group of researchers have intended to design a variety of quantum inspired
algorithms when no quantum computers were available in the literature. There exist
some appealing features in quantum computers that may provide remarkable out-
comes when executing quantum inspired algorithms in classical computers. Nowa-
days, first generation of quantum computers are at hand, different scientists through-
out the globe are working hard to assess the quantum efficacy in a variety of fields.
A large variety of research works have already been introduced since the invention
of first version of quantum behaved of metaheuristics. Metaheuristics are stochastic
kind of algorithms, in which, based on some predefined policies, numerous heuris-
tic algorithms are constructed [18]. Several problem-independent, efficient, popular
approximate metaheuristics are available in the literature that can efficiently deal
with different optimization problems. Metaheuristics are referred to as the high-level
algorithmic structure that come up with a set of predefined strategies to introduce
heuristic optimization techniques [19].
An approximation technique is a process that provides approximate results to any
specific problem. An approximation technique runs in polynomial time depending
on the size of input variable, rather than exponential time to search for optimal solu-
tion (or near optimal solution). Execution time is a crucial metric for developing an
approximation technique. This metric is highly influenced in positive direction with
the advancement of computer architecture. Nowadays, algorithms are designed and
modified in such a way that they can perform parallel processing. The multicore mi-
croprocessor is new generation hardware device which are frequently and efficiently
used for executing these new versions of algorithms. Apart from that GPU cards
which can accommodate plenty of processing units and CUDA (a specialized soft-
ware) are also very useful for the same purpose [20]. In the recent times, a number
of renowned companies have successfully designed powerful quantum computers
which are more efficient than classical computers in all respects. These quantum
computers can efficiently be used in different fields such as machine learning, simu-
lations, optimization to name a few. Some of the popular companies which have suc-
cessfully developed quantum computers are Google [21], IBM [22], Intel [23] and
D-Wave Systems [24] to name a few. In the literature, some popular metaheuristics
can be listed as simulated annealing (SA) [25], particle swarm optimization (PSO)
[26], differential evolution (DE) [27], and ant colony optimization (ACO) [28] to
name a few.
In 1996, Narayanan and Moore used the thought and features of quantum mechan-
ics to develop efficient metaheuristics [22]. The authors used a quantum-inspired
crossover to find optimal solution for the traveling salesman problem (TSP). A set
of guidelines has been introduced as an effort to characterize a method for designing
and developing quantum algorithms. The theory and feathers of quantum computing
have been used by different researchers several times to design a number of quantum
inspired algorithms. One pioneer algorithms of this category is popularly known as
Genetic Quantum Algorithm (GQA) [30]. In 2019, Montiel et al. presented a popular
quantum-inspired algorithm, called quantum-inspired Acromyrmex evolutionary al-
gorithm (QIAEA) [31]. The authors carefully observed the colony habits of the Atta
and Acromyrmex in their daily life. This fact motivated them to develop QIAEA.
Han et al. [32] developed a programmed version (parallel) of QGA, called paral-
lel QGA (PQGA). PQGA has been applied on knapsack problem for optimization.
PQGA has been compared with QGA to judge its efficacy, where PQGA outper-
formed other. In 2012, Li et al. [33] presented a watershed-based quantum evolu-
tionary algorithm for texture image clustering and SAR segmentation. Wang et al.
[34] introduced a QEA and particle swarm optimization method (PSO) based quan-
tum swarm evolutionary algorithm (QSwE). A variety of Q-bit expression form and
an improved version of PSO have been introduced in QSwE for updating quantum
angle. Zouache et al. [35] introduced PSO based quantum-inspired firefly algorithm
(QIFAPSO). The communal habits of the firefly, swarm and the concept of quantum
computing have been combined in a single skeleton to design QIFAPSO.
Apart from that, several quantum based metaheuristics are available in the litera-
ture. Some of them are quantum-inspired evolutionary algorithm based on p-system
(QEPS) [36] developed by Zhang et al., quantum-inspired DE and PSO algorithm
(QDEPSO) [37] proposed by Zouache and Moussaoui etc. Chaos Quantum-Inspired
Particle Swarm Optimization (CQPSO) was developed in [38] to handle Economic
load dispatch (ELD) problem. Hassan et al. [39] proposed quantum inspired bat al-
gorithm to deal with various economic load dispatch problem. In the wireless sen-
sor network, Ullah and Wahid [40] designed a quantum inspired genetic algorithm
framework for topology control. Several quantum inspired algorithms have been in-
troduced in the literature so far for handling bi-level optimization problem. Zhang
et al. [41] used bi-level programming in collaboration with swarm intelligence to
develop a strategic bidding optimization algorithm. Dey et al. [42] presented quan-
tum inspired bi-level optimization algorithms using GA and PSO for gray-level im-
age thresholding. The quantum inspired multi-objective based simulated annealing
has been designed by Dey et al. [3] for bi-level image thresholding. The compu-
tational capability of bi-level system has been enhanced to the multi-level frame by
altering its basic structure. Dey et al. [2] introduced Quantum inspired particle swarm
optimization and quantum inspired differential evolution for multi-level colour image
thresholding. Tkachuk [45] used quantum technological approach to develop quan-
tum inspired evolutionary algorithm. Later, Dey et al. [2][4][47] designed a number
of quantum inspired metaheuristics in multi-level and colour domain. Cardenas et
al. [48] designed a protocol to perform quantum reinforcement learning (QRL) and
quantum technologies (QT).
Quantum inspired metaheuristics have been widely used in pattern recognition.
Dey et al. [49] introduced a quantum inspired sperm whale algorithm for multi-level
thresholding. The basic operators of quantum computing have been fused with the
sperm whale algorithm. Dutta et al. [50] introduced a novel metaheuristic, called the
Border Collie Optimization. The motivation behind designing of this algorithm is
mimicking the sheep herding behaviour of Border Collie dogs.
Unlike, single objective optimization (SOO), multi-objective optimization
(MOO) handles more than one objective function at a time. Kim et al. [51] designed
a QEA based quantum-inspired multi-objective evolutionary algorithm (QMEA) for
solving the 0-1 knapsack problem. Moghadam et al. [52] first introduced a quantum
version of gravitational search algorithm (GSA), called quantum-behaved gravita-
tional search algorithm (QGSA).
Later, in 2015, Chakraborti et al. [53] introduced a modified version of QGSA,
called modified binary quantum-behaved gravitational search algorithm with differ-
ential mutation (MBQGSA-DM). Like QGSA, this algorithm also used differential
mutation strategy. Li and Wang [54] proposed a hybrid quantum-inspired genetic
algorithm (HQGA). This algorithm has been designed to efficiently deal with multi-
objective based combinatorial optimization problem, called low shop scheduling
problem (FSSP). A novel algorithm, called Quantum Ant Colony Multi-Objective
Routing (QACMOR) has been developed to deal with WSN routing problem. In this
algorithm, the concepts of quantum computing and multi-objective function have
been utilized.
2.9 CONCLUSION
With the development of machine learning theory and the accumulation of practical
experience of using various algorithms, it became clear that there is no ideal classifi-
cation method that would be better than all others for all sizes of the training sample,
for any percentage of noise in data, for any complexity of the boundaries of dividing
objects into classes etc. Therefore, at present, ensemble classification methods that
combine many different classifiers trained on different data samples. One of the most
accurate and fast parallelization methods available today is bagging, which turns
out to be useful in the case of heterogeneous classifiers and instability, when small
changes in the initial sample lead to significant changes in the classification. To in-
crease the speed of combining decisions of basic classifiers, a new quantum-inspired
method of collective decision-making based on metaheuristic quantum algorithms is
proposed. The development of ensemble methods in high-speed online learning is
expected in the future.
REFERENCES
1. Friedman, J. & Greedy, H. (2001). Function approximation: a gradient boosting machine.
Annals of statistics, 29(5), 1189-1232.
2. Zhou, ZH. & Wu, J. & Tang, W. (2002). Ensembling neural networks: many could be better
than all. Artificial intelligence, 137(1), 239-263.
3. Sahu, A. & Runger, G. & Apley, D. (2011). Image denoising with a multi-phase kernel
principal component approach and an ensemble version. IEEE applied imagery pattern
recognition workshop, 1-7.
4. Shinde, A. & Sahu, A. & Apley, D. & Runger, G. (2014). Preimages for variation patterns
from kernel PCA and bagging. IIE Transactions, 46(5), 429-456.
5. Buhlmann, P. & Hothorn, T. (2007). Boosting algorithms: Regularization, prediction and
model fitting. Statistical Science, 477-505.
6. Arya, S. & Mount, D. & Netanyahu, N. & Silverman, R. & Wu, A. (1998). An optimal
algorithm for approximate nearest neighbor searching fixed dimensions. Journal of the
ACM, 45(6), 891-923.
7. Aizerman, M. & Bravermann, E. and Rosonoer, L. (1970). The Potential Function Method
in Machine Learning Theory. Moscow: Science.
8. Beletskaya, S. & Asanov, Yu. & Povalyaev, A. & Gaganov, A.V. (2015). Research of the
effectiveness of genetic algorithms for multicriteria optimization. Voronezh State Technical
University Bulletin, 11(1), 1-4.
9. Epanechnikov, V. (1969). Nonparametric estimation of multidimensional probability den-
sity. Probability theory and its applications, 14(1), 156-161.
10. Dempster, A. & Laird, N. & Rubin, D. (1977). Maximum likelihood from incomplete data
via the EM algorithm. Journal of the Royal Statistical Society. Series B (Methodological),
39(1), 1-38.
11. Scholkopft, B. & Mullert, K. & Fisher, R. (1999). Discriminant analysis with kernels.
Neural Networks for Signal Processing, 1(1), 41-48.
12. Hosmer Jr, DW. & Lemeshow, S. & Sturdivant, R. (2013). Applied logistic regression.
New- York: John Wiley & Sons.
13. Cortes, C. & Vapnik, VN. (1995). Support-vector networks. Machine Learning, 20(3),
273297.
14. Kamiski, B. & Jakubczyk, M. & Szufel, P. (2017). A framework for sensitivity analysis of
decision trees. Central European Journal of Operations Research, 26(1), 135159.
15. Quinlan, JR. (2014). C4.5: programs for machine learning. Amsterdam: Elsevier.
16. Breiman, L. & Friedman, J. & Stone, CJ. & Olshen, RA. (1984). Classification and regres-
sion trees. Monterey: Wadsworth & Brooks.
17. Rutkovskaya, D. & Rutkovsky, L. & Pilinsky, M. (2013). Neural networks, genetic algo-
rithms and fuzzy systems. Moscow: Hotline-Telecom.
18. Glover, F. & Kochenberger. G.A. (2003). Handbook on Metaheuristics. Kluwer Academic
Publishers.
19. Veldhuizen, DAV. & Lamont, GB. (2000). Multiobjective evolutionary algorithms: An-
alyzing the state-of-the-art. Journal of Evolutionary Computation, The MIT Press, 8(2),
125147.
20. Fabris, F. & Krohling, RA. (2012). A co-evolutionary differential evolution algorithm for
solving min-max optimization problems implemented on gpu using c-cuda. Expert Syst.
Appl., 39(6), 1032410333.
21. Google llc. (2019). quantum, google ai. https://ai.google/research/teams/applied-
science/quantum-ai/, 2019. Accessed: May. 18, 2021, [Online].
22. IBM. (2019). ibm q, quantum computing. https://www.research.ibm.com/ibm-q/, 2019.
Accessed: May. 4, 2021. [Online].
23. Intel corporation. (2019). 2018 ces: Intel advances quantum and neuromorphic com-
puting research. https://newsroom.intel.com/news/intel-advances-quantumneuromorphic-
computing-research/,2019. Accessed: May. 2, 2021, [Online].
24. D-wave systems. (2019). dwave systems. https://www.dwavesys.com/home/, 2019. Ac-
cessed: May. 15, 2021, [Online].
25. Kirkpatrik, S. & Gelatt, CD. & Vecchi. MP. (1995). Optimization by simulated annealing.
Science, 220, 671680.
26. Kennedy, K. & Eberhart, R. (1995). Particle swarm optimization. in: Proceedings of
the IEEE International Conference on Neural Networks (ICNN95), Perth, Australia, 4,
19421948.
27. Storn, R. & Price, K. (1995). Differential evolution a simple and efficient heuristic for
global optimization over continuous spaces. Technical Report TR-95-012, ICSI.
28. Dorigo, M. & Maniezzo, V. & Colorni, A. (1996). The ant system: optimization by a colony
of cooperating agents. IEEE Trans. Syst. Man & Cybernet. Part B, 26(1), 2941.
29. Narayanan, A. & Moore, M. (1996). Quantum-inspired genetic algorithms. in: Proceedings
of IEEE Int. Conf. Evol. Comput., 6166.
30. Han, KH. & Kim, JH. (2000). Genetic quantum algorithm and its application to combina-
torial optimization problem. in: Proceedings of Congr. Evol. Comput.(CEC), 2, 13541360.
31. Montiel, O. & Rubio, Y. & Olvera, C. & Rivera, A. (2019). Quantum inspired acromyrmex
evolutionary algorithm. Nat. Sci. Rep., 9(12181), 169176.
32. 32. Han, KH. & Park, KH. & Lee, CH. & Kim, JH. (2001). Parallel quantum inspired
genetic algorithm for combinatorial optimization problem. in: Proceedings of Congr. Evol.
Comput., 2, 14221429.
33. Li, Y. & Shi, H. & Jiao, L. & Liu. R. (2012). Quantum evolutionary clustering algorithm
based on watershed applied to sar image segmentation. Neurocomputing, 87(10), 9098.
34. Wang, Y. & Feng, XY. & Huang, YX. & Pu, DB. & Zhou, WG. & Liang, YC. & Zhou,
CG. (2007). A novel quantum swarm evolutionary algorithm and its applications. Neuro-
computing, 70(4), 633640.
35. Zouache, D. & Nouioua, F. & Moussaoui, A. (2016). Quantum inspired firefly algorithm
with particle swarm optimization for discrete optimization problems. Soft Computing,
20(7), 27812799.
36. Zhang, G. & Gheorghe, M. & Wu. C. (2008). A quantum-inspired evolutionary algorithm
based on p systems for knapsack problem. Fundam. Inform., 87(1), 93116.
37. Zouache, D. & Moussaoui, V. (2015). Quantum-inspired differential evolution with particle
swarm optimization for knapsack problem. J. Inf. Sci. Eng., 31, 17791795.
38. Meng, K. & Wang, HG. & Dong, Z. & Wong, KP. (2010). Quantum inspired particle swarm
optimization for valve-point economic load dispatch. IEEE Transactions on Power Sys-
tems, 25(1), 215222.
39. Tehzeeb ul Hassan, H. & Asghar, MU. & Zamir, MZ. & Faiz, HMA. (2017). Economic
load dispatch using novel bat algorithm with quantum and mechanical behaviour. In: Pro-
ceedings of 2017 International Symposium on Wireless Systems and Networks (ISWSN),
16.
40. Ullah, S. & and Wahid, & M. (2015). Topology control of wireless sensor network us-
ing quantum inspired genetic algorithm. International Journal of Swarm Intelligence and
Evolutionary Computation, 04, 08.
41. Zhang, G. & Zhang, G. & Gao, Y. & Lu, J. (2011). Competitive strategic bidding op-
timization in electricity markets using bilevel programming and swarm technique. IEEE
Transactions on Industrial Electronics, 58(6), 21382146.
42. Dey, S. & Bhattacharyya, S. & Maulik, U. (2014). Quantum inspired genetic algorithm
and particle swarm optimization using chaotic map model based interference for gray level
image thresholding. Swarm and Evolutionary Computation, 15, 3857.
43. Dey, S. & Bhattacharyya, S. & Maulik, U. (2017). Quantum-inspired multi-objective sim-
ulated annealing for bilevel image thresholding. Quantum Inspired Computational Intelli-
gence, Research and Applications.
44. Dey, S. & Bhattacharyya, S. & Maulik, U. (2013). Quantum inspired meta-heuristic algo-
rithms for multi-level thresholding for true colour images. In: Proceeding of 2013 Annual
IEEE India Conference (INDICON), Mumbai, India.
45. Tkachuk. V. (2018). Quantum genetic algorithm on multilevel quantum systems. Mathe-
matical Problems in Engineering, 112.
46. Dey, S. & Bhattacharyya, S. & Maulik, U. (2016). New quantum inspired meta-heuristic
techniques for multi-level colour image thresholding. Applied Soft Computing, 46,
677702.
47. 47. Dey, S. & Bhattacharyya, S. & Maulik, U. (2017). Efficient quantum inspired meta-
heuristics for multi-level true colour image thresholding. Applied Soft Computing, 56,
472513.
48. Ctardenas-Ltopez, FA. & Lamata, L. & Retamal, JC. & Solano, E. (2018). Multiqubit and
multilevel quantum reinforcement learning with quantum technologies. PLoS ONE, 13(7),
112.
49. Dey, S. & De, S. & Ghosh, D. & Konar, D. & Bhattacharyya, S. & Platos, J. (2019). A novel
quantum inspired sperm whale metaheuristic for image thresholding. In: Proceedings of
2019 Second International Conference on Advanced Computational and Communication
Paradigms (ICACCP), 17.
50. Dutta, T. & Bhattacharyya, S. & Dey, S. & Platos, S. (2020). Border Collie Optimization.
IEEE Access, 8, 109177-109197.
51. Kim, Y. & Kim, JH. & Han, KH. (2006). Quantum-inspired multiobjective evolutionary
algorithm for multiobjective 0/1 knapsack problems. In: Proceedings of IEEE Int. Conf.
Evol. Comput., 26012606.
52. Moghadam, MS. & Nezamabadi-Pour, H. & Farsangi, MM. (2005). A quantum behaved
gravitational search algorithm. Intell. Inf. Manage., 2012(4), 390395.
53. Chakraborti, T. & Chatterjee, A. & Halder, T. & Konar. A. (2015). Automated emotion
recognition employing a novel modified binary quantum-behaved gravitational search al-
gorithm with differential mutation. Expert Syst., 32(4), 522530.
54. Li, B. & Wang, L. (2007). A hybrid quantum-inspired genetic algorithm for multiobjective
flow shop scheduling. IEEE Trans. Syst., Man, Cybern. B, Cybern., 37(3), 576591.
3 Function Optimization
Using IBM Q
3.1 INTRODUCTION
An optimization problem is the searching technique to find out the best solution
from all the possible solutions. When the practical optimization problem in any dis-
cipline is represented by a mathematical function, it is called objective function.
Real-world optimization problems are typically having several objective functions.
There are many ways available in literature for resolving the practical optimization
problems. In this chapter, the extensively used optimization techniques are described
briefly. The main focus has been given to solve the optimization problems using
IBM Q. After introduction in the first section of this chapter, a brief overview to
single-objective and multi-objective optimization with their difficulties have been
discussed.
Section 3.3 has been reserved to discuss about modern techniques of resolving
optimization problems. Genetic algorithms, simulated annealing, particle swarm op-
timization, differential evolution, ant colony optimization, bee-colony optimization,
harmony search algorithm, bat-algorithm, Cuckoo search, neural network-based op-
timization, fuzzy optimization, etc. are available in literature, which are considered
as modern methods of optimization problem solving. All the afore-mentioned meth-
ods are described in short with their consequences.
When the complexity of optimization problems and amount of data involved rise,
more efficient ways of solving optimization problems are needed. The power of
quantum computing can be used for solving problems which are not practically feasi-
ble on classical computers, or suggest a considerable speed up with respect to the best
known classical algorithm. In Section 3.4, the author has enlightened on the basics
of quantum computing and the quantum algorithms used for solving optimization
problems.
IBM provides an experimental cloud-enabled quantum computing user-interface
platform, known as IBM Q, for the students, researchers, and general science en-
thusiasts. It allows users to run established algorithms and experiments, work with
quantum bits (qubits), etc. The different features of IBM Q have been reflected in
Section 3.5.
The circuit composer in IBM Q is a tool that allows users to visually learn how to
create quantum circuits. In Section 3.6, the step-by-step approach has been presented
to build a sample quantum application using circuit composer.
Qiskit (Quantum Information Software Kit) in IBM Q is an open-source quantum
computing framework, which enables developers and researchers to conduct quan-
tum explorations using Python scripts. Section 3.7 has been utilized to showcase a
DOI: 10.1201/9781003283294-3 37
sample Qiskit project with its different components like Qiskit Terra, Qiskit Aqua,
Qiskit Ignis, and Qiskit Aer.
In Section 3.8, the author has enlightened on the application of quantum comput-
ing, where IBM Q can be utilized seamlessly. Out of those applications, objective
function optimization using IBM Q has been broadly illustrated in this section. Port-
folio optimization, risk analysis, and Monte-Carlo-like applications are considered
as few examples of optimization.
The last section of this book chapter is dedicated for drawing a conclusionary
communication in terms of the application of IBM Q in objective function optimiza-
tion.
3.2 FUNCTION OPTIMIZATION

The main reason to rely on any model in a decision-making process is to provide
a quantitative assessment of the effects of management decisions on the system be-
ing considered. A model also provides a fairly objective assessment as opposed to
subjective opinions of system behavior. Thus, models should be used in support of
decision-making. Optimization is just another type of modeling and the same ap-
plies to optimization models. Optimization tools should be used for supporting deci-
sions rather than for making decisions, that is, should not substitute decision-making
process.
3.2.1 DIFFICULTIES IN OPTIMIZATION METHODS

Real-world optimization problems often involve a number of characteristics, which
make them difficult to solve up to a required level of satisfaction. The characteristics
are mentioned below [1]
1. Existence of mixed type of variables (such as Boolean, discrete, integer, and real)
2. Existence of non-linear constraints
3. Existence of multiple conflicting objectives
4. Existence of multiple optimum (local and global) solutions
5. Existence of stochasticity and uncertainty in describing the optimization problem
Most real-world search and optimization problems involve multiple objectives. A

solution that is optimum (maximum or minimum) with respect to one objective may
not be optimum with respect to another objective. So some compromise must be
required. This is more difficult to implement multi-objective optimization problems
(MOOPs) than single objective optimization problems (SOOPs).
3.2.2 DEFINITION OF MULTI-OBJECTIVE OPTIMIZATION PROBLEM (MOOP)

A multi-objective optimization problem (MOOP) has a number of objective func-
tions, which are to be minimized or maximized. Multi-objective optimization
Function Optimization Using IBM Q 39
Figure 3.1: Decision variable space and the corresponding objective space.
problem can be described in its general form:
Minimize/Maximize fm (x), m = 1, 2, . . . , M;
Sub ject to g j (x) >= 0, j = 1, 2, . . . , J;
(3.1)
hk (x) = 0, k = 1, 2, . . . , K;
(L) (U)
xi <= xi <= xi , i = 1, 2, . . . , n;
As indicated in Figure 3.1, a solution “x” is a vector of “n” decision variables:

x = (x1 , x2 , . . . , xn )T . The last set of constraints is called variable bounds. These
bounds constitute a decision variable space, D. The terms “g j (x)” and “hk (x)” are
called constraint functions. In this problem, there are ‘J’ inequality (greater-than-
equal-to type) and “K” equality constraints. A solution “x” that does not satisfy all of
the (J + K) constraints and all of the variable bounds is called an infeasible solution.
On the other hand, if any solution “x” satisfies all constraints and variable bounds
is known as feasible solution. So, in the presence of constraints, the entire decision
variable space D may not be feasible. The set of all feasible solutions is called the
feasible region, S. For each solution “x” in the decision variable space, there exists a
corresponding point in another space, which is called objective space, Z. The map-
ping takes place between an n-dimensional solution vector and an M-dimensional
objective vector[1].
3.2.3 DIFFERENCES BETWEEN SOOPS AND MOOPS

The main goal of single-objective optimization is to find the best solution, which
corresponds to the minimum or maximum value of a single objective function that
lumps all different objectives into one. This type of optimization is a useful tool,
which should provide decision makers with insights into the nature of the problem,
but usually cannot provide a set of alternative solutions that trade different objectives
against each other. On the contrary, in a multi-objective optimization with conflicting
objectives, there is no single optimal solution. The interaction among different ob-
jectives gives rise to a set of compromised solutions, largely known as the trade-off,
non-dominated, noninferior, or Pareto-optimal solutions.
1. In a single-objective optimization, there is only one goal, the search for an opti-
mum solution. In multi-objective optimization, there are two goals, progressing
toward the Pareto-optimal front and maintaining diversity among the solutions in
the Pareto-optimal front.
2. In a single-objective optimization, there is only one search space, the decision
space. But in multi-objective optimization, there are two search spaces, decision
space and objective space.
3. Single-objective optimization is the degenerate case of multi-objective optimiza-
tion. In many cases, multi-objective optimization can be converted into single-
objective optimization.
3.3 MODERN OPTIMIZATION PROBLEM-SOLVING TECHNIQUES

Heuristic and metaheuristic are the two common terms while considering modern
optimization techniques to solve the optimization problems. Heuristic means to find
or discover certain value or solution by trial and error. Heuristic algorithms can attain
to the optimum solutions within a given finite time, but it never assures its perfor-
mance in achieving the best solution. To achieve more better and accurate solutions,
meta-heuristic algorithms have been developed by making some improvements over
heuristic algorithms. The modern metaheuristic algorithms work on the basis of cer-
tain randomization of the search process. In the subsequent subsections, a number of
modern optimization problem-solving techniques have been discussed briefly with
their limitations. These methods operate by considering the issues such as impre-
cision, partial truth, uncertainty, and approximation, which are often present in few
real-world problems.
3.3.1 GENETIC ALGORITHM

Genetic Algorithm (GA), which mimics the natural process of Evolution and Dar-
win’s principle of “Survival of the Fittest”, was invented by John Holland in the
1960s at the University of Michigan. According to [7], GAs are particularly suitable
for solving complex optimization problems like wire routing, scheduling, adaptive
control, game playing, cognitive modeling, transportation problems, traveling sales-
man problems, optimal control problems, database query optimization, etc.
A GA comprises a set of individual elements (the population) and a set of bio-

logically inspired operators defined over the population itself. According to evolu-
tionary theories, only the most suited elements in a population are likely to survive
and generate offspring, thus transmitting their biological heredity to new genera-
tions. In computing terms, a genetic algorithm maps a problem onto a set of strings,
each string representing a potential solution. A simple GA cycle has four stages:
population of strings, evaluation of each string, selection of best string, and genetic
operators to create the new population of strings.
The biggest limitation of GA is that it cannot guarantee optimality. The solution
quality also deteriorates with the increase of problem size.
3.3.2 SIMULATED ANNEALING

Simulated Annealing (SA) uses a thermodynamic evolution process to search min-
imum energy states [3]. The primary objective of simulated annealing is to find the
global minimum of a function that characterizes large and complex systems. It pro-
vides a powerful tool for solving non-convex optimization problems. Simulated an-
nealing starts from a random initial solution. It then proceeds by generating new
solution and accepting/rejecting them probabilistically. As the search proceeds, the
temperature cools down and the process converges to a global minima. The tempera-
ture needs to be reduced at a slow and controlled rate to ensure proper solidification
with a low energy crystalline state that corresponds to the best required result.
If the cost function is expensive to compute, the algorithm becomes very slow. For
problems where the energy landscape is smooth, or there are few local minima, SA
is overkill. The method cannot tell whether it has found an optimal solution. Some
other method is required to do this.
3.3.3 PARTICLE SWARM OPTIMIZATION

As per [8], Particle Swarm Optimization (PSO) algorithm was inspired from the
behavior of swarm of insects, such as ants, bees, wasps, or a school of fish or a
flock of birds. The algorithm operates through the behavioral intelligence of each
solution/particle or insect in a swarm/ function.
Ant Colony Optimization (ACO) is based on the cooperative behavior of real ant
colonies, that is, their characteristics of finding the shortest path from their nest to a
food source.
Bee Colony Optimization (BCO) works mainly on the basis of foraging behavior
of the bees. The natures of bee, that is, forage which means hunt or explore, are taken
consideration for this optimization algorithm.
Like GA, PSO algorithms do also not guarantee about the optimum solution. Al-
though ACO and BCO have higher probability and efficiency in finding the global
optima, it is difficult to define their initial design parameters and not suitable for
large-scale problems.
3.3.4 BAT ALGORITHM

Bats emit a loud sound pulse and listen for the echo that reflects from adjacent ob-
jects. Depending on this echo voice, they avoid the obstructions and treasure their
root while searching their target. Based on certain properties of their pulses emitted
and the time for detecting the echo by their ears, the bats are able to detect the dis-
tance, orientation, type of victims, etc. This optimization algorithm [5], utilizes the
echolocation nature of bats.
Bat algorithm converges very quickly at the early stage and there after conver-
gence rate slows down.
3.3.5 CUCKOO SEARCH ALGORITHM

Cuckoo Search (CS) algorithm [6] is inspired by the obligate brood parasitism of
some cuckoo species by laying their eggs in the nests of other host birds of different
species. Female cuckoos can lay their eggs very similar (in color and pattern of eggs)
to the host so that the host can take responsibility of those eggs.
CS has its own limitation while performing for multi-objective problems and dis-
crete problems.
3.3.6 FUZZY SYSTEM

Most natural language is fuzzy, which involves vagueness and imprecision. A fuzzy
logic proposition is a linguistic statement relating some perception without clearly
defined boundaries. In other words, all truths in fuzzy logic are partial or approx-
imate. Linguistic statements that tend to express subjective ideas and that can be
interpreted slightly differently by various individuals typically involve fuzzy propo-
sitions. Fuzzy propositions are assigned to fuzzy sets. The objective and constraint
functions are characterized by the membership functions in a fuzzy system, where
membership in a classical subset, of a classical crisp sets of objects.
Fuzzy logic-based control systems, i.e. Fuzzy Logic Controller (FLC), can han-
dle many real-world problems that cannot be efficiently managed by conventional
(crisp) control systems. Fuzzy control systems theory uses the routine fuzzification
fuzzy operation defuzzification [7]. The purpose of designing and applying FLCs is
to tackle vague, ill-described, and complex processes that can barely be handled by
classical systems theory, classical control techniques, and classical two-valued logic.
The majority of FLCs are knowledge-based systems in that either their fuzzy models
or their fuzzy logic controllers are described by fuzzy IF-THEN rules. Rules are es-
tablished based on experts’ knowledge about the systems, controllers, performance,
etc.
Fuzzy system is also having the limitations. The main concern is stability of the
system. Determining the fuzzy rules and membership functions are not very easy
task to achieve.
3.3.7 NEURAL NETWORK BASED OPTIMIZATION

An Artificial Neural Network (ANN) is an information processing system that has
several performance features in common with biological neurons. An ANN is char-
acterized by (a) its design pattern between the neurons, called its architecture, (b)
its technique of determining the weights on the connections, called its learning al-
gorithm, and (c) its activation functions. ANN with its different architectures and
learning mechanism is suitable for solving optimization problems.
However, a few limitations of this technique are unexpected behavior of the net-
work, determination of the proper network structure, hardware dependency, etc.
In many cases, good results have been achieved by combining the above-
mentioned modern optimization problem-solving methods. Neuro-fuzzy architecture
is a very good example in which the good properties of both neural network and
fuzzy logic are attempted to bring together. Neuro-fuzzy systems are mostly fuzzy
rule-based systems, in which different techniques of neural networks are used for
rule introduction and calibration. Fuzzy logic may also be employed to improve the
performance of optimization methods used with neural networks [8].
3.4 QUANTUM COMPUTING AND OPTIMIZATION ALGORITHMS

The common limitation of the above-mentioned modern optimization problem-
solving techniques is convergence speed. This problem has been resolved by us-
ing quantum computing, which is much faster than the traditional computing. This
section is dedicated for a brief overview of quantum computing and optimization
algorithms.
3.4.1 QUANTUM COMPUTING

All computing systems rely on a fundamental ability to store and manipulate infor-
mation. Current computers manipulate individual bits, which store information as
binary 0 and 1 states. Quantum computers leverage quantum mechanical phenomena
to manipulate information. To do this, they rely on quantum bits, or qubits, depicted
in IBMs quantum computing reference website. Three quantum mechanical proper-
ties (superposition, entanglement, and interference) are used in quantum computing
to manipulate the state of a qubit.
The fundamental building block, i.e. Quantum bit or Qubit, of quantum computer
has been represented by Bloch sphere in Figure 3.2. A qubit can be at 1 state or 0
state or in any superposition of the two. The state of a qubit can be represented as
Ψ⟩ = α 0⟩ + β 1⟩ (3.2)
where α and β are complex numbers and |α |2 + |β |2 = 1
Quantum logic gates, operating on a number of qubits, are the building blocks of
quantum circuits. There are many types of quantum gates, like H (Hadamard) gate,
CX (Controlled-X) gate, ID (Identity) gate, U3 gate, U2 gate, U1 gate, Rx gate, Ry
gate, Rz gate, X gate, Y gate, Z gate, S gate, Sdg gate T gate, Tdg gate, cH gate, cY
Figure 3.2: Bloch sphere.
gate, cZ gate, cRz gate, cU1 gate, cU3 gate, ccX gate, SWAP gate, etc. Even custom
gates can also be created for using in quantum circuits.
Digital quantum computers use quantum logic gates to do computation. A quan-
tum computer consists of the below-mentioned blocks or chambers, IBMs quantum
computing reference website [9]
(a) Qubit Signal Amplifier

(b) Input Microwave Lines
(c) Superconducting Coaxial Lines
(d) Cryogenic Isolators
(e) Quantum Amplifiers
(f) Cryoperm Shield
(g) Mixing Chamber
3.4.2 OPTIMIZATION USING QUANTUM COMPUTING

The limitations of the individual modern optimization problem-solving techniques
highlighted in Section 3.3, have been tried to minimize by using quantum comput-
ing. In general, logistics or supply chain network infrastructure, air traffic control
work scheduling, and financial services are few optimization examples, which can
be resolved by quantum computing.
In [10], a novel Genetic Quantum Algorithm (GQA) is introduced to resolve
the optimization problem. GQA has an outstanding capability of global search due
to its diversity produced by the probabilistic representation. Well-known Knapsack
problem has been used to discuss the performance of GQA. Further improvement to
this algorithm had been made using parallelism feature.
Quantum annealing algorithm such as Quantum Processing Unit (QPU) is ap-
plicable for solving binary optimization problems [11]. In [12], quantum annealer
has been utilized to optimize the traffic flow, as mentioned by [13]. The QPU is de-
signed to solve Quadratic Unconstrained Binary Optimization (QUBO) problems,
where each qubit represents a variable and couplers between qubits represent the
costs associated with qubit pairs. Quantum annealing algorithm has been used in
[14] to resolve the Nurse Scheduling Problem (NSP), which arises when searching
the optimal schedule for a set of available nurses to create a rotating roaster.
Quantum Adiabatic Algorithm (QAA) [15] has been used on a quantum computer
for finding the global minima of a classical cost function. Performance have been
measured in [16] by generating over 200,000 instances of MAX 2-SAT on 20 qubits.
Few real-life optimization examples, which can be resolved by quantum comput-
ing are
a. Telecommunications companies upgrading their network infrastructure

b. Healthcare firms optimizing patient treatments
c. Governments improving air traffic control
d. Consumer products and retail companies tailoring marketing offers
e. Financial services firms enhancing their risk optimization
f. Organizations developing employee work schedules
g. Universities scheduling classes
3.5 FEATURES OF IBM Q EXPERIENCE

The IBM Q Experience is an online platform available in Cloud, which gives users
access to a set of IBM’s prototype quantum processors. It was launched by IBM in
May 2016. IBM Q is an industry first initiative to build universal quantum comput-
ers for business, engineering, and science. This effort includes advancing the entire
quantum computing technology stack and exploring applications to make quantum
broadly usable and accessible. The snapshots of IBM Q have been demonstrated in
Figure 3.3. IBM Q is applicable to solve the most challenging problems in chemistry,
optimization, machine learning, finance, etc. Many self-explanatory documentations
are provided in IBMs website on quantum computing.
Official website - https://quantum-computing.ibm.com/
3.6 CIRCUIT COMPOSER IBM Q

The circuit composer is a tool in IBM Q that allows users to visually create quantum
circuits. Figure 3.4 is the visual representation of circuit composer in IBM Q.
Figure 3.5 shows a sample circuit diagram in visual form and in internal coding
form for 5 quantum registers. The circuit can be exported as .qasm file, which can be
again modified and imported. Figure 3.6 shows the result dashboard after running
the circuit.
Figure 3.3: Snapshot of IBM Q.
Circuit composer of IBM Q is being widely used to simulate the quantum circuits
visually and test those very easily. A quantum calculator (addition, subtraction, mul-
tiplication, and division) is demonstrated and simulated using circuit composer [17].
The algorithm for obtaining maximum and minimum of any mathematical model
has been presented in [18]. Circuit composer has been used as a simulator to find
out the minimum from Titanic passengers age. In [19], controlled square root of Z
gate has been constructed and tested using circuit composer. An exhaustive survey
Figure 3.4: Circuit composer.
on 20 different quantum algorithms is available in literature [20]. All the algorithms

are simulated and tested using circuit composer.
3.7 QISKIT IN IBM Q

QISKit [21] is an open-source quantum computing software development kit for
leveraging today’s quantum processors in research, education, and business. It is a
Python-based software library that can be used to create quantum computing pro-
grams, compile, and execute them on one of several backends [22]. So, QISKit can
be installed on top of Python using the below command pip install qiskit This is also
available in IBM Q Experience with the Qiskit Notebooks as shown in Figure 3.7.
3.7.1 CREATING 5-QUBIT CIRCUIT WITH THE HELP OF QISKIT IN IBM Q

Once the New Notebook button is clicked, a Jupyter notebook gets opened with
QISKit installed into it. After writing the notebook completely, it can be saved as
.ipynb. This notebook can be downloaded as different formats as shown in Figure
3.8.
A sample QISKit code has been written to create a quantum circuit using a
Hadamard gate H on qubit 0, a Controlled-X gate on control qubit 0 and target qubit
1 and measure on each qubit to visualize the result. In Figure 3.9, a sample circuit
has been drawn using QISKit programming and Figure 3.10 shows simulation and
plotting the result.
(a)
(b)
Figure 3.5: (a) Diagrammatic representation, (b) Pseudo-code representation.
3.7.2 TESTING THE CIRCUIT USING IBM QUANTUM COMPUTER

Up to this point, QISKit codes were running on Jupyter notebook locally. Now, this
data has been sent to IBM Q machine ibm˙16˙melbourne, which has been shown in
Figure 3.11. It is treated as a job, which has different statuses like
a. Job is validating
b. Job is successfully queued
c. Job has successfully run
Once job has successfully run, the results can be plotted using plot˙histogram
function.
Figure 3.6: Result dashboard.
Figure 3.7: QISKit notebooks.

Figure 3.8: Working with QISKit notebook.
Figure 3.9: Sample 5 qubit circuit using QISKit.

Figure 3.10: Simulating and plotting the result.
From the two plots provided in Figure 3.10 and Figure 3.11, the differences are
very clear. When the result on Jupyter notebook is run, probability of occurrences
is ideal and caught at 2 significant qubits 00000 and 00011. The probability of oc-
currences in other qubits is ignored. When the result is sent to the real quantum
computer, that is, IBM Q for analysis, it provides probability of occurrences in many
qubits by considering the minimal errors too. By improving the technology of real
quantum computer, the scientists are working out to diminish these insignificant er-
rors.
print(result) = ’00011’: 513, ’00000’: 511
print(result˙ibmq) = ’00101’: 1, ’01011’: 5, ’00010’: 36, ’10010’: 2, ’01010’:
1, ’10000’: 20, ’00011’: 311, ’00111’: 2, ’00001’: 54, ’00000’: 571, ’01000’: 4,
’10001’: 1, ’11111’: 1, ’11011’: 1, ’00100’: 2, ’10011’: 11, ’01001’: 1
3.8 OPTIMIZATION USING IBM Q

One of the application areas of IBM Q is to resolve the most challenging optimiza-
tion problems. Quantum computers offer the best or optimal solution among varying
weighted selections more competently than classical computers. Qiskit Aqua [23]
is a tool, which allows researchers to experiment with optimization applications. It
contains a library of cross-domain quantum algorithms upon which applications for
near-term quantum computing can be built. The Aqua algorithms run on top of Qiskit
Terra, which builds, compiles, and executes quantum circuits on simulators or real
quantum devices.
Figure 3.11: Creating a job for IBM Q and visualizing the result.
Travelling Salesman Problem (TSP), considered as NP-hard problem, is an impor-

tant category of optimization problems that is mostly encountered in various areas
of science and engineering. In [24], Circuit Composer of IBM Q has been used to
simulate the TSP by considering four cities.
Vehicle Routing Problem (VRP), another NP-hard problem, comprises of search-
ing the smallest distance for delivering goods from a depot to many geographically
distributed locations with numerous journeys. VRP has been considered to solve us-
ing IBM Q Experience and presented in [25].
The classical Tabu Search is a metaheuristic that explores search spaces and con-
ducts a local heuristic search procedure to explore the solution space beyond local
optimum using a Tabu list with forbidden moves. In [26], the Quantum Tabu Search
(QTS) algorithm has been proposed based on Knapsack problem approach. The pro-
posed QTS has been simulated and tested by using backend IBM Q 16 Rueschlikon
(16-qubits) and IBM Q 5 Yorktown (5-qubits) simulator.
The number of gates and the number of levels used in IBM Q have to be optimized
to obtain the better result. This improves the reliability of the output state, reduces
the effect of noise, and increases the accuracy of the quantum computation. In [27],
circuits implemented in IBMs QX2 and QX4 have been utilized with maximum 16
qubits to establish experimental violation of Mermin inequality.
Variational Quantum Eigensolver (VQE) [28] is a hybrid between classical and
quantum computing. A classical computer controls the preparation of a quantum
state using few experimental parameters; then a quantum computer prepares that state
and calculates its properties. Optimization problems can be resolved by quantum
computer using VQE. In [29], VQE in IBM Q has been applied to solve the MaxCut
NP-complete binary optimization problem with 5 qubits.
Grover’s Adaptive Search (GAS) [30] is a fast quantum mechanical algorithm
for combinatorial optimization problems, which can resolve an O(N/2) optimization
problem into O(N) steps. In [31], a modified version of Grover’s search algorithm
with fewer gates, optimized number of iterations and improved performance has been
presented. To establish the upgraded and optimized quantum search, set search and
array search algorithms have been implemented using IBM Q. Grover optimizer can
be easily applied to solve the QUBO problem [32].
In July 2020, IBM came up with Qiskit Optimization Module [33]. This is now in
an initial stage by keeping in mind the goal of providing a super-optimized solution
to the users within few milliseconds for any input problem. This module will act
as a black box with the combination of quantum and classical resources. Users do
not need to have the knowledge of quantum theory and mechanics. According to the
documentation, it empowers easy and efficient modeling of optimization problems
for developers and optimization experts without quantum expertise. IBMs DOcplex
(Decision Optimization CPLEX) [34] modeling for python is used to develop the
Qiskit Optimization Module.
3.9 CONCLUSION
As per definition, optimization is the technique of finding an alternative with the
most cost-effective or highest achievable performance under the given constraints,
by maximizing desired factors and minimizing undesired ones. Since many years, lot
of efforts have been incorporated to resolve the optimization problems, specifically
the NP-hard and multi-objective problems. It becomes slight easy after the invention
of quantum computer, which is very costly and sensitive. IBM makes it available
freely for the common people by introducing IBM Q Experience online platform.
In this chapter, efforts have been made to discuss the two features, circuit com-
poser and QISkit, available in IBM Q. A 5-qubit circuit has been created on Jupyter
notebook using QISKit software development kit and the same has been tested using
real IBM quantum computer. The chapter has been concluded by discussing on the
techniques to resolve optimization problems using IBM Q.
ACKNOWLEDGMENTS
This work had been accomplished on IBM Cloud to leverage its IBM Quantum ex-
perience platform. The author is grateful to IBM for providing access to its IBM Q
environment.
REFERENCES
1. Kalyanmoy Deb. Optimization for engineering design: Algorithms and examples. PHI
Learning Pvt. Ltd., 2012.
2. David E. Goldberg and John Henry Holland. Genetic algorithms and machine learning.
Machine Learning, 3: 95–99, 1988.
3. Scott Kirkpatrick, C Daniel Gelatt, and Mario P Vecchi. Optimization by simulated an-
nealing. Science, 220(4598):671–680, 1983.
4. James Kennedy and Russell Eberhart. Particle swarm optimization. In Proceedings of
ICNN’95-International Conference on Neural Networks, volume 4, pages 1942–1948.
IEEE, 1995.
5. Xin-She Yang. Bat algorithm for multi-objective optimisation. International Journal of
Bio-Inspired Computation, 3(5):267–274, 2011.
6. Xin-She Yang and Suash Deb. Cuckoo search via Levy flights. In 2009 World Congress on
Nature & Biologically Inspired Computing (NaBIC), pages 210–214. IEEE, 2009.
7. Guanrong Chen, Trung Tat Pham, and N.M. Boustany. Introduction to fuzzy sets, fuzzy
logic, and fuzzy control systems. Applied Mechanics Reviews, 54(6):B102–B103, 2001.
8. Detlef Nauck, Frank Klawonn, and Rudolf Kruse. Foundations of neuro-fuzzy systems.
John Wiley & Sons, Inc., 1997.
9. https://www.ibm.com/quantum-computing/.
10. Kuk-Hyun Han and Jong-Hwan Kim. Genetic quantum algorithm and its application to
combinatorial optimization problem. In Proceedings of the 2000 Congress on Evolutionary
Computation. CEC00 (Cat. No. 00TH8512), volume 2, pages 1354–1360. IEEE, 2000.
11. Kuk-Hyun Han, Kui-Hong Park, Ci-Ho Lee, and Jong-Hwan Kim. Parallel quantum-
inspired genetic algorithm for combinatorial optimization problem. In Proceedings of the
2001 Congress on Evolutionary Computation (IEEE Cat. No. 01TH8546), volume 2, pages
1422–1429. IEEE, 2001.
12. Florian Neukart, Gabriele Compostella, Christian Seidel, David Von Dollen, Sheir Yarkoni,
and B. Parney. Traffic flow optimization using a quantum annealer. Frontiers in ICT, 4:29,
2017.
13. Mark W Johnson, Mohammad HS Amin, Suzanne Gildert, Trevor Lanting, Firas Hamze,
Neil Dickson, Richard Harris, Andrew J Berkley, Jan Johansson, Paul Bunyk, et al. Quan-
tum annealing with manufactured spins. Nature, 473(7346):194–198, 2011.
14. Kazuki Ikeda, Yuma Nakamura, and Travis S Humble. Application of quantum annealing
to nurse scheduling problem. Scientific Reports, 9(1):1–10, 2019.
15. Edward Farhi, Je rey Goldstone, Sam Gutmann, and Michael Sipser. Quantum computation
by adiabatic evolution. arXiv preprint quant-ph/0001106, 2000.
16. Elizabeth Crosson, Edward Farhi, Cedric Yen-Yu Lin, Han-Hsuan Lin, and Peter Shor.
Different strategies for optimization using the quantum adiabatic algorithm. arXiv preprint
arXiv:1401.7320, 2014.
17. Prathamesh P Ratnaparkhi and K. Bikash. Demonstration of a quantum calculator on IBM
quantum experience platform, DOI: 10.13140/RG.2.2.12661.63209 (2018)
18. Yanhu Chen, Shijie Wei, Xiong Gao, Cen Wang, Jian Wu, and Hongxiang Guo. An opti-
mized quantum maximum or minimum searching algorithm and its circuits. arXiv preprint
arXiv:1908.07943, 2019.
19. Petar Nikolov and Vassil Galabov. Experimental realization of controlled square root of
z gate using ibm’s cloud quantum experience platform. arXiv preprint arXiv:1806.02575,
2018.
20. J. Abhijith, Adetokunbo Adedoyin, John Ambrosiano, Petr Anisimov, Andreas artschi,
William Casper, Gopinath Chennupati, Carleton Corin, Hristo Djidjev, David Gunter, et
al. Quantum algorithm implementations for beginners. arXiv e-prints, pages arXiv:1804,
2018.
21. https://qiskit.org/.
22. Robert Wille, Rod Van Meter, and Yehuda Naveh. Ibm’s qiskit tool chain: Working with
and developing for real quantum computers. In 2019 Design, Automation & Test in Europe
Conference & Exhibition (DATE), pages 1234–1240. IEEE, 2019.
23. https://github.com/Qiskit/qiskit-aqua.
24. Karthik Srinivasan, Saipriya Satyajit, Bikash K Behera, and Prasanta K Panigrahi. Efficient
quantum algorithm for solving travelling salesman problem: An IBM quantum experience.
arXiv preprint arXiv:1805.10928, 2018.
25. Abhimanyu Nowbagh and K. Bikash. A quantum approach for solving vehicle routing
problem: An IBM quantum experience.
26. Carla Silva, Ines Dutra, and Marcus S Dahlem. Driven tabu search: A quantum inherent
optimisation. arXiv preprint arXiv:1808.08429, 2018.
27. Mitali Sisodia, Abhishek Shukla, Alexandre AA de Almeida, Gerhard W. Dueck, and Anir-
ban Pathak. Circuit optimization for IBM processors: A way to get higher fidelity and
higher values of nonclassicality witnesses. arXiv preprint arXiv:1812.11602, 2018.
28. Alberto Peruzzo, Jarrod McClean, Peter Shadbolt, Man-Hong Yung, Xiao-Qi Zhou, Peter
J Love, Alan Aspuru-Guzik, and Jeremy L O’brien. A variational eigenvalue solver on a
photonic quantum processor. Nature Communications, 5(1):1–7, 2014.
29. Nikolaj Moll, Panagiotis Barkoutsos, Lev S Bishop, Jerry M Chow, Andrew Cross, Daniel
J Egger, Stefan Filipp, Andreas Fuhrer, Jay M Gambetta, Marc Ganzhorn, et al. Quantum
optimization using variational algorithms on near-term quantum devices. Quantum Science
and Technology, 3(3):030503, 2018.
30. Lov K Grover. A fast quantum mechanical algorithm for database search. In Proceedings
of the Twenty-Eighth Annual ACM Symposium on Theory of Computing, pages 212–219,
1996.
31. Austin Gilliam, Marco Pistoia, and Constantin Gonciulea. Optimizing quantum search us-
ing a generalized version of Grover’s algorithm. arXiv preprint arXiv:2005.06468, 2020.
32. https://qiskit.org/documentation/tutorials/optimization/index.html.
33. https://www.ibm.com/blogs/research/2020/07/quantum-optim-module/.
34. https://developer.ibm.com/docloud/documentation/optimization-modeling/modeling-for-
python/.
4 Multipartite Adaptive
Quantum-Inspired
Evolutionary Algorithm to
Reduce Power Losses of a
Radial Distribution
Network
4.1 INTRODUCTION
Over the last few decades, some new approximation algorithms have emerged with
the aim to explore the search space, which is commonly known as metaheuristics.
Generally, metaheuristic is defined as an iterative process which guides the heuristic
method by combining the intelligent concepts for exploiting and exploring the search
space. If a heuristic optimization algorithm is expressed in a metaheuristic frame-
work with different intelligent concepts to explore the search space is also referred
as metaheuristic. Glover introduced the term metaheuristic in 1986 by combing the
Greek prefix Meta with heuristic. Heuristic means to find or discover, whereas the
suffix Meta means beyond or higher level solution [1]. Metaheuristic method allows
the local search operators to escape from local optima by generating new initial so-
lutions or allowing worsening moves for the local search in an intelligent way. High
quality solutions are produced in metaheuristics by introducing a bias with various
forms [2].
Metaheuristic methods have demonstrated to the scientific community that they
are often feasible, alternative, and superior to more traditional methods such as dy-
namic programming and branch and bound, etc. In comparison with traditional meth-
ods, metaheuristics are often providing a better trade-off between computing time
and solution quality for large and complicated problems. Metaheuristic methods are
often more flexible than traditional methods in two different ways. Firstly, these are
adapted to fit for most real-life optimization problems in terms of computational time
and solution quality, which can vary greatly across different situations. Secondly,
they do not offer any demands on formulation of optimization problem. Metaheuris-
tic methods are implemented by several commercial vendors in their software as
primary optimization engine.
Metaheuristic algorithms attempt to find the best feasible solution of an optimiza-
tion problem out of all possible solutions. Series of operations are performed on the
DOI: 10.1201/9781003283294-4 57
optimization problem to search for the better solution. Local search methods and
population based methods are normally employed by metaheuristics to obtain the
feasible solutions. Local search methods use iterative process to find the optimal
solutions [3]. Population-based methods find the optimal solution by iteratively se-
lecting and then combining existing solutions from a set, usually called as popula-
tion. The most important member of this class, which mimics the principle of natural
evolution, is evolutionary algorithm (EA). In EA, selection operator generally gives
direction to the search process by using Darwinian principle. The solution for next
iteration is generated through variation operators like crossover by recombining the
solutions from current iteration. Local heuristics and mutation operator are used to
improve the exploration and exploitation, that is, escaping from local minima and
increasing the convergence rate. EAs are popular due to their ease of implementation
and employed for solving difficult and complex optimization problems. However,
EA often suffers from some limitations.
Quantum-inspired Evolutionary Algorithm (QiEA) is used to overcome these lim-
itation, which use probabilistic representation along with some concepts and opera-
tions of quantum computing [4]. It uses a single qubit with small population size and
is governed by principle of quantum mechanics [5]. Q-gates are used in QiEA as a
variation operator to drive the individuals in the population towards better solution.
In recent times, Adaptive Quantum-inspired Evolutionary Algorithm (AQiEA) [6] is
applied on various engineering optimization problems with a measurement opera-
tor, which is a modified version of QiEA. AQiEA uses two sets of qubits, whereas
QiEA uses a single set of qubit. Recently, AQiEA is applied on optimization prob-
lem of Distributed Generator (DG) [7]-[8], Network Reconfiguration [9]-[11], ce-
ramic grinding [12], Cost analysis of DG and Capacitor [13], Siting and sizing of
Capacitors [14], and simultaneous implementation of both DG and capacitors [15].
In this chapter, we are proposing Multipartite Adaptive Quantum-inspired Evolution-
ary algorithm , which is an updated version of AQiEA. MAQiEA improves on both
exploration and exploitation ability of AQiEA by introducing changes in Rotation
towards Better Strategy and Rotation away from Worse Strategy. In MAQiEA, the
Rotation toward Better Strategy of AQiEA is converted into Rotation Around Better
Strategy as it offers more un-restricted exploration as compared to the previous strat-
egy, which allowed exploration in only improving direction. Similarly, the Rotation
away from Worse Strategy in AQiEA involved only two individuals, one was Best
Individual and the other was sequential selected Individual and was primarily used
for exploitation purpose, i.e., searching around the Best individual with the help of
other individuals in the population, so it was bipartite. However, recent algorithms
like Grey Wolf Optimizer (GWO) [16], Symbiotic Organism Search [17] and Salp
Swarm Algorithm (SSA) [18] etc. are shown to be good exploiters and they tend to
use multiple individuals based variation operators rather than bipartite variation op-
erators, therefore, it was decided to augment the exploitation strategy, i.e., Rotation
away from worse by converting it into Multipartite Adaptive Variation operator from
the current bipartite version in AQiEA.
Multipartite Adaptive QIEA to Reduce Power Losses 59
Figure 4.1: Types of DG Technologies.
Distributed Generator: Optimal location and capacity of DG play a key role to

minimize the power losses in DN. In general, DG is defined as the small scale power
generating source which is placed very nearer to the load centers i.e., supplies elec-
tric power near to point of consumption, which reduce the transmission cost and
enhance the percentage of power loss reduction in the system [19]. Definition of DG
differs from one agency to another [20]. DG is a power generating source used near
load centers, which is not connected to high voltage transmission system directly.
In comparison with conventional power plants, DG has modular and small size, low
installation time with low investment. DGs are operated as power backup supply
devices. Implementing the DGs into distribution system has great impact on protec-
tion, operation and stability of the system. Figure 4.1 shows the different types of
DG technologies, which comprise of Renewable as well as non-renewable energy
resources. Figure 4.2 shows the diagrammatic representation of DG capacities. The
renewable DG technologies are PV Solar, Geo Thermal, Bio-Mass and Wind turbine,
whereas non-renewable technologies are Gas-Turbines, Micro-turbine, reciprocating
engines and fuel cells etc. Integrating DG into distribution network have both posi-
tive and negative impacts. Positive impact is improvement in percentage power loss
reduction, enhancement in voltage profile, improvement in reliability and loadability.
Whereas negative impact indicates increment in the short circuit current and reversed
power flow. These impacts may vary depending on the optimal placement, sizing and
type of DG. Inappropriate location and capacity of DG leads the system to increase
in power loss and reduction in voltage profile. Distribution network has dynamic
load structure which is not fixed for certain length of time. Industrial, commercial
and residential loads are mainly dependent on voltage, which varies with time. How-
ever most of the work done in DN to reduce the losses with DG implementation has
considered only constant power load (CP) model, which is independent of voltage
and doesn’t vary with time. In this study, an investigation has been performed with
Figure 4.2: DG Capacities.
voltage dependent load model (VDLM) to find the power losses incurred in the sys-
tem. In this study constant current (CC) and constant impedance (CZ) which varies
linearly and square of the voltage are considered. The industrial load (IL), commer-
cial load (CL) and residential load (RL) power requirements vary exponentially with
the terminal voltage. In addition to VDLM, CP load model is also used in the study.
A class of mixed load (ML) is also investigated to find the power losses incurred in
the system i.e., combination of all load including both VDLM and CP load model.
4.2 LITERATURE REVIEW

Research on optimization of DG in DN is becoming more and more popular due to
its ease of implementation and ecofriendly technology. Developed and under devel-
oping countries are focusing more on renewable energy sources to meet the required
load demand with both dispatchable and un-dispatchable DGs. Optimal placement
and capacity of DG is one of the most complex optimization problems of electrical
power systems. Over the past few decades, optimal location and sizing of DG has
become an interesting and challenging area of research to reduce the power losses.
Many authors have tried to solve this important optimization problem with three dif-
ferent techniques. In general, these techniques are classified as Analytical, Numerical
and Metaheuristic methods. Analytical techniques requires less computational time
and easy to implement on smaller bus systems [21]. However, for larger and com-
plex test bus system, its computational complexity is very high. Similarly, numerical
methods are difficult to implement on larger test bus systems when new constraints
are added [21]. Many techniques have been implemented in DN for optimal location
and capacity of DG to reduce the losses. In comparison with analytical techniques,
metaheuristic techniques have more ability to provide significant solution. Thus more
and more metaheuristic techniques are applied to solve this important optimization
problem.
Vijay babu et al. [22] used an analytical approach with real power loss expression
to find the optimal location and capacity of DG. An investigation has been performed
with four different scenarios to minimize the losses with different types of DGs. All
class of mix DGs viz., Type-I (which injects only active power), Type-II (both injects
both active and reactive powers), Type-III (only reactive power), is also used in this
study as fourth scenario. Prakash and Kathod [23] presented an analytical technique
to reduce losses with implementation of single DG. In this study, optimization of DG
reduces the magnitude of active and reactive power components. A hybrid technique
i.e., combination of both analytical and metaheuristic technique (genetic algorithm)
is used in ref [24], to determine the location and capacity of DG. An investigation
has been performed on DGs mode of operation, two different scenarios are used
with different power factors. Ahmed et al. [25] used a linearized model to estimate
the optimal capacity of DG with graph flow and Kalman filter. Optimal size of DG is
obtained with a two stage method, where graph flow is used to create a linear model
and Kalman filter is used to find the optimal size of DG. Mahmoud et al. [26] pre-
sented an analytical approach with integration of DG in distribution system to reduce
losses. Multiple DGs are installed in the system with optimal power factors and four
different scenarios are used to maximize the percentage power loss reduction with
different types of DGs.
Khoa et al. [27] proposed an optimization technique known as one rank cuckoo
search algorithm (ORCSA). ORCSA is used to solve the combinatorial optimization
problem of DG by finding its optimal location and capacity with different power
factors. Multiple objective optimizations are used in the study, Similar to the above,
Sultana et al. [16] studied the effect of DG allocation on distribution system with
grey wolf optimization algorithm with an objective to reduce the losses. Yifei et al.
[28] studied the impact of DG in DN with loss sensitivity factor (LSF), which is used
to determine the optimal allocation of DG. Mistry [29] used two different optimiza-
tion techniques to reduce the losses in a DN with implementation of multiple DGs.
Carvalho and Niraldo [30] used an optimization technique known as Ant Colony Op-
timization for optimization DG with same objective as mentioned above. Vizhiy and
Santhi [31] presented a multiobjective optimization problem with DG to reduce the
losses. Biogeography based optimization (BBO) is used to find optimal placement
and capacity of DG. Ali et al. [32] used evolutionary algorithm technique called as
Ant Lion Optimization (ALO), which is used to find the optimal allocation of DG.
Proposed algorithm is based on behaviour of hunting lion ants, a multiobjective ap-
proach is used to reduce the losses with improvement in voltage profile.
Mahajan and Vadhera [33] used particle swarm optimization technique with an
objective to minimize the losses in DN by finding the ideal location of DG with op-
timal size. An investigation has been performed to find the improvement in voltage
profile with multiple weight factors. Snigdha and Panigrahi [34] studied the effect
of DG in distribution system with multiobjective differential evolution algorithm to
maximize the benefits of DG owners and utilities by minimizing the power losses
with different scenarios. Chaotic symbiotic organism search algorithm with multi
objective optimization problem is used in ref [35] to minimize the losses in a DN by
allocating the DG at optimal location with capacity. Benchmark test bus systems are
also used to show the effectiveness of CSOS. Power loss reduction index is consid-
ered as primary objective to reduce the losses with implementation of DG, Selective
Particle Swarm Optimization (SPSO) determines the optimal placement and capacity
of DG [36]. Sarfaraz et al. [37] used same algorithm to find the optimal location and
capacity of DG with same objective. Multiple DGs are used with small rating DGs
in [36], whereas ref [37] uses a single DG with high operating power (size). Power
loss reduction with multiple DGs is high in comparison with single DG.
Devang and Ritesh [38] studied the effect of DG with different power factors i.e,
lagging power factor, 0.8 power factor and unity power factor on distribution system
at various locations. In this study, two DGs are employed with small size to minimize
the losses. Jamian et al. [39] used Gravitational search algorithm to reduce the losses
in the DN by placing the DG at its optimal location with optimal capacity. Zhang and
Bo [40] studied the impact of DG in radial DN on power loss minimization. Simu-
lation results conclude that, DG should be placed at the end of the line when it is
operating with low rating and at the middle of the line when it is operating with high
power rating. DG should be placed nearer to substation when it has high operating
power greater than the load demand. An optimization technique known as Genetic
Algorithm is used to find the placement and capacity of DG with an objective to
minimize the power losses [41]. Ang et al. [42] used a new metaheuristic technique
viz., sine cosine algorithm to find the optimal placement and capacity of single and
multiple DGs in DN with an objective to improve the voltage profile in the net-
work and maximize the percent power loss reduction in the system. Simultaneous
implementation of DG and Capacitors are also used by some authors to reduce
the power losses. In such studies, DG injects only active power with unity power
factor into the system, whereas Capacitor injects only reactive power with zero power
factor into the system, injection of both powers into the system results in high re-
duction in power losses as compared with independent implementation. However,
independent implementation of DG has high reduction in power losses in compari-
son with independent implementation of Capacitors. Simultaneous implementation
of both DG and capacitor induces high investment cost, maintenance and operation
cost. In our study, only independent implementation of DG is considered.
It has been observed from the above literature that only constant power load is
used which doesn’t vary with time. In distribution network, consumers use different
load models with different ratings. Majority of load used in distribution network is
dependent on voltage, however CP load model is independent of voltage. If the opti-
mal placement and capacity of DG obtained with CP load model is used in practical
distribution network, which induces more power losses into the system due to the
improper location and sizing of DG. Some have used different load models other
than CP load. Roy et al. [43] studied the impact of DG in a DN on different load
models with voltage profile. Initial investigation has been performed by analyzing
the impact of static load (CP load) on DN. Dynamic analysis shows that composite
load model has high voltage dips and CZ load model has low voltage dips. Oscar
et al. [44] used a novel approach to minimize the power losses by varying the load
Figure 4.3: Different Load Models.
model i.e., twenty four load model is used for optimization of DG by keeping the
voltages within the limits. Divya and Srinivasan [45] studied the effect of DG un-
der fault conditions with a simple radial distribution system. Voltage Stability Index
(VSI) is used to find the optimal allocation of DG, whereas its optimal capacity is
obtained with Particle Swarm Optimization. Aashish et al. [46] studied the effect of
practical load models with integration of DG in DN. Several performance indices are
developed as multi-objective function. Genetic algorithm and Particle Swarm opti-
mization are also used to determine the optimal location and size of DG. Das et al.
[47] investigated the effect of DG on a VDLM i.e., residential time varying load to
reduce the power losses. Sensitivity index-based method is used to find the optimal
placement of DG with variation in load whereas, Genetic Algorithm determines the
optimal capacity of DG.
In this study, an investigation has been performed to study the effect of DG with
variation in load. Different types of load models, which are dependent on exponen-
tial characteristics of node voltage, are shown in Figure 4.3. Optimal placement
and sizing of DG is a nondifferentiable combinatorial complex optimization prob-
lem. Multi-partite Adaptive Quantum-inspired Evolutionary Algorithm (MAQiEA)
is used to find the optimal location and capacity of DG for VDLM. MAQiEA is the
updated version of AQiEA. In AQiEA, three rotation strategies are used to converge
the population towards global optima, whereas MAQiEA uses probabilistic rotation
around better and multi-partite rotation away from worse rotation strategies which
provides for relatively better exploration and exploitation. MAQiEA has high ro-
bustness and has better exploitation and exploration of search space in comparison
with AQiEA as shown by test results.
4.3 PROBLEM FORMULATION

DN is normally designed to supply the power to all consumers. But the power de-
mand is increasing day by day due to the vast increase in industrial and human needs.
In order to cater the required load demand, distributed generators are used as alter-
native solution. It is a well-known fact that the load at distribution system changes
continuously. In this work, an investigation has been performed to study the effect
of DG on VDLM other than CP load. IL, RL, CL, CC and CZ are used in this study
[48]-[51]. These are expressed as follows:

Vi (m) µcp Vi (m) µi Vi (m) µr
PL (m) = PL0 (m) S1 + T1 +U1
Vo (m) Vo (m) Vo (m)
µc µci (4.1)
Vi (m) Vi (m) Vi (m) µcc
+ PL0 (m) V1 +W1 + X1

Vi (m) γcp Vi (m) γi Vi (m) γr
QL (m) = QL0 (m) S2 + T2 +U2
γc γci (4.2)
Vi (m) Vi (m) Vi (m) γcc
+ QL0 (m) V2 +W2 + X2
For Load Type I: Constant Power Load: S1 = 1 and T1 = U1 = V1 = W1 =
X1 = 0
S2 = 1 and T2 = U2 = V2 = W2 = X2 = 0
For Load Type II: Industrial Load: T1 = 1 and S1 = U1 = V1 = W1 = X1 = 0
T2 = 1 and S2 = U2 = V2 = W2 = X2 = 0
For Load Type III: Residential Load: U1 = 1 and S1 = T1 = V1 = W1 = X1 = 0
U2 = 1 and S2 = T2 = V2 = W2 = X2 = 0
For Load Type IV: Commercial Load: V1 = 1 and S1 = T1 = U1 = W1 = X1 = 0
V2 = 1 and S2 = T2 = U2 = W2 = X2 = 0
For Load Type V: Constant Impedance Load: W1 = 1 and S1 = T1 = U1 =
V1 = X1 = 0
W2 = 1 and S2 = T2 = U2 = V2 = X2 = 0
For Load Type VI: Constant Current Load: X1 = 1 and S1 = T1 = U1 = V1 =
W1 = 0
X2 = 1 and S2 = T2 = U2 = V2 = W2 = 0
Where PL (m) and QL (m) are the active and reactive power loads at mth bus. Table 4.1
shows the active and reactive power components of voltage dependent load model
[17], [46]. The objective of the study is to reduce the power losses with optimal
location and capacity of DG. Power losses are expressed as follows:
F = min(Ploss )
Nb
Min(Ploss ) = ∑ Im2 ∗ Rm . (4.3)
m=1
Voltage Stability Index: Distribution network has complex structure, poor voltage
regulation and high power losses are observed in the system at the end nodes. In
Table 4.1
Exponent Values of Different Voltage Dependent Loads
Load Type Exponent
Constant Power Load µcp γcp
0 0
Industrial Load µi γi
0.18 6.0
Residential Load µr γr
0.92 4.04
Commercial Load µc γc
1.51 3.40
Constant Impedance Load µci γci
1 1
Constant Current Load µcc γcc
2 2
order to maintain the system voltage within acceptable limits, voltage stability index
is considered.
Min(V SIk+1 ) = Vk4 − 4(Pk+1 Xk − Qk+1 Rk )2 − 4(Pk+1 Rk − Qk+1 Xk )Vk2 (4.4)
Power loss minimization in distribution network with optimal placement and sizing
of DG is an interesting and challenging area of research. Optimal location and ca-
pacity of DG minimizes power losses and improves the voltage profile in the system.
Power injected by at a particular bus, m is given as follows.
PL (m) = PL (m) − PDG (m) (4.5)
Constraints for power loss minimization with DG are given as follows:

a) Operation of Distributed Generator: DG is used to minimize the power losses
with optimal location and capacity, inappropriate sizing of DG induces high power
losses in the system. It is necessary that optimal size of DG has to be in specified
limits.
min max
PDG,i < PDG,i < PDG,i (4.6)
b) Power injection for the system: DGs are placed nearer to load centers, if DG
induces high power into the system i.e., greater than total power demand and losses,
bi directional power flow may damage the system equipment. It is necessary that,
total power injected by DG always less than demand and losses in the system.
n
∑ PDG,i < PDemand + Ploss (4.7)
m=1
c) Power injection: Total power injected by substation and Distributed generator has
to meet power demand at load centers including losses.
n
PSubstation + ∑ PDG,i < PDemand + Ploss (4.8)
m=1
d) Voltage limit: Optimal location and capacity of DG not only reduces the power
losses but also improves the voltage profile in the system. After installing DG at
optimal location with capacity, voltage has to be in permissible limits.
V min < V < V max (4.9)
4.4 POWER FLOW

In a power system network, ac and dc power flows from generating station to the
utilities at load end through different branches and busses. The flow of power i.e.,
active or reactive in a system is called load flow or power flow. Load flow or power
flow study is generally a mathematical approach used to determine the load under
steady state condition, active or reactive power flow through different branches, gen-
erators, total active and reactive power losses, individual power losses obtained at
each branch and bus voltages. Load flow analysis is used to determine the steady
state operation of network. It is widely used by power system professional during
planning stages of distribution system or adding an additional network to existing
one. Traditional power flow methods such as Gauss-Seidel, fast decoupled and New-
ton Raphson methods are used to find the power losses, voltage profile of the net-
work. These methods are mainly designed and used for transmission systems and
unsuitable for distribution networks. Distribution system has complex structure with
large number of branches and nodes. Distribution networks are normally designed as
mesh structures but in normal operating conditions they are operated in radial struc-
ture. The load at distribution network generally poses unbalanced operation with
unbalanced distribution loads. Traditional power flow methods require set of equa-
tions whose size is equal to number of buses. If the traditional power flow methods
are used in distribution network, for large test bus systems the computational time
is too long. In addition, Y-bus formation in traditional power flow methods creates
waste of memory storage. Hence, traditional power flow methods are inefficient for
distribution networks. Distribution system has high R/X ratio, which causes ill con-
ditioning for traditional power flow methods. Hence, there is need for a load flow
which uses topological characteristics of distribution network to determine the to-
tal power losses and bus voltage in the network. Jen-Hao Teng [52] has proposed a
load flow which uses topological characteristic of distribution network. Based on the
topological characteristics of the network, Bus Injection to Branch Current (BIBC)
matrix and Branch Current to Bus Voltage (BCBV) matrix are normally developed.
This method generally has three important steps which are given below.
Equivalent current injection
Formation of BIBC Matrix
Formation of BCBC Matrix
Equivalent Current Injection: Equivalent current injection of the distribution sys-

tem at bus k, the complex power Sk for t th iteration is given as
Sk = Pk + jQk (4.10)
Ikt = Ikr (Vkt ) + jIki (Vkt ) (4.11)

Formation of BIBC Matrix: By applying Kirchhoffs Current Law (KCL) to the sim-
ple radial distribution network shown in Figure 4.4, the equivalent current injection
obtained is
B2 = I3 + I4 B4 = I5 + I6 + I7 B5 = I6 + I7 + I8 B7 = I8 (4.12)
Bus Injection to Branch Current (BIBC) matrix is obtained as follows
Figure 4.4: Radial Distribution System.
    
B1 1 1 1 1 1 1 1 I2
B2  0 1 1 0 0 0 0  
    I3 
B3  0 0 1 0 0 0 0 
 
   I4 
B4  = 0 0 0 1 1 1 1 I5 
 
   
B5  0 0 0 0 1 1 1  
    I6 
B6  0 0 0 0 0 1 0 I7 
B7 0 0 0 0 0 0 1 I8
The relation between equivalent current injection and branch currents are expressed
as
B = BIBC I (4.13)
Algorithm for formulation of BIBC Matrix:
Step I: If a distribution network has u branch section and v bus section, a null matrix
with dimension u*(v-1) is created.
Figure 4.5: Diagrammatic representation of BIBC matrix.
Step II: If a line section (BL ) is located between two buses bus m and bus n, copy the
column of mth bus of BIBC matrix to the nth column of BIBC matrix and a value +1
is added at a position of l-row, n-bus column. If a line section (B5 ) is located between
two buses bus 6 and bus 8, copy the 6th column of BIBC matrix to 8th column of
BIBC matrix and +1 is added at a 5-row, 8-column of BIBC matrix.
Step III: The above process is repeated for all branch line sections in the distribution
network.
The above process of copying the columns from one column to other column and
addition of value +1 is shown in Figure 4.5.
Formation of BCBV Matrix:
Branch-current and bus voltage relationship (BCBV) is obtained by applying KVL
to the simple radial distribution network shown in Figure 4.6.
V2 = V1 − B1 Z12V5 = V2 − B4 Z25V6 = V5 − B5 Z56 (4.14)
Using above equation the bus voltage at 6th bus is written as
V6 = V1 − B1 Z12 − B4 Z25 − B5 Z56 (4.15)
From the above equation, it has been observed that the bus voltage can be expressed
as function of substation voltage, branch current and line parameters. Branch Current
to Bus Voltage (BCBV) matrix is obtained as follows.

      
V1 V2 Z12 0 0 0 0 0 0 B1
V1  V3  Z12 Z23 0 0 0 0 0  B2 
      
V1  V4  Z12 Z23 Z34 0 0 0 0   
      B3 
V1  − V5  = Z12 0 0 Z25 0 0 0   
      B4 
V1  V6  Z12 0 0 Z25 Z56 0 0   
      B5 
V1  V7  Z12 0 0 Z25 Z56 Z67 0  B6 
V1 V8 Z12 0 0 Z25 Z56 0 Z68 B7
Therefore branch currents and bus voltages are expressed as

∆V = BCBV B (4.16)
Algorithm for formulation of BCBV Matrix:

Step I: If a distribution network has u branch section and v bus section, a null matrix
with dimension (v-1)*u is created.
Step II:If a line section (BL ) is located between two buses bus m and bus n, copy the
row of mth bus of BCBV matrix to the nth row of BCBV matrix and fill the line sec-
tion impedance (Zmn ) to the position of the n-th row to l-th column. If a line section
(B5 ) is located between two buses bus 6 and bus 8, copy the 6th row of BCBV ma-
trix to 8th row of BCBV matrix and an impedance of value Z68 is added at a 8-row,
6-column of BCBV matrix.
Step III:The above process is repeated for all branch line sections in the distribution
network.
The above process of copying the rows from one row to other row and addition of
value ′ Zmn
′ is shown in Figure 4.6. Algorithm for Distribution Load Flow: Distri-
Figure 4.6: Diagrammatic representation of BCBV matrix.
bution load flow is used to find the total power losses induced in the system, bus
voltages and individual power losses for each branch. Step-by-step process of direct
load flow is given as follows.
Step I: Read the initial line data and load data of the distribution network.
Step II:Bus Injection to Branch Current matrix is formulated based on data.

Step III:Branch Current to Bus Voltage matrix is formulated based on data.
Step IV:Direct load flow matrix is created. The relationship between bus current in-
jection and bus voltage is given by
∆V =[BCBV][BIBC][I]
=[DLF][I]
Step V:Initialize the iteration count t=0
Step VI:Increment in iteration count t=t+1
Step VII:Update the voltage by solving the equation iteratively. By solving the below
equations iteratively, the solution of the direct load flow is obtained.
Ikt = Ikr (Vkt ) + jIki (Vkt )
[∆V (t + 1)] = [DLF][I t ]
[V (t + 1)]=[V 0 ] [∆V (t + 1)]
If tolerance exists or converges print results else go to step (6).
Flow chart is representation of direct load flow is shown in Figure 4.7.
4.5 ALGORITHM
In recent times, metaheuristics are mainly used to solve several optimization prob-
lems. Metaheuristic methods have demonstrated the scientific community that they
are often feasible, alternative, and superior to more traditional methods such as dy-
namic programming and branch and bound, etc. In comparison with traditional meth-
ods, metaheuristics are often providing a better trade-off between computing time
and solution quality for large and complicated problems. In EA, individuals will
compete with one another. The fittest individual in the population will move for-
ward to the next generation. It acts as a parent and it will again compete with child
in the next generation, the fittest among them will move forward to further gener-
ation. This process repeats until convergence exists. EA often suffers from some
major limitations i.e., stagnation, sensitivity to the choice of parameter, premature
and Slow convergence. QiEA overcomes the above limitations by creating a good
balance between exploration and exploitation. QiEA is designed by integrating prin-
ciples of Quantum mechanics viz., measurement, entanglement, superposition and
interference into current framework of EA. It is proposed to solve difficult combi-
natorial and non-differentiable optimization problems. In this study, A Multi-partite
Adaptive Quantum-inspired Evolutionary Algorithm (MAQiEA) is used to solve non
linear large-scale optimization problem. AQiEA [9] is different from QiEA, AQiEA
uses two Q-bits per solution vector, whereas Quantum-inspired Evolutionary Algo-
rithm uses a single Q-bit. In Adaptive Quantum-inspired Evolutionary Algorithm,
the smallest information element in a quantum computer is a quantum-bit (qubit)
analogous to classical bits. The basis states are represented in Hilbert space by a
vector as |0⟩ and |1⟩. The qubit can be represented by vector |C⟩ and it is defined as
|C⟩ = A|1⟩ + B|0⟩ (4.17)

Figure 4.7: Flow Chart of Direct Load Flow.
Where A and B are complex numbers which specifies the probability amplitudes
associated with states |1⟩ and |0⟩ respectively and should satisfy the condition.
| A |2 + | B |2 = 1 (4.18)
Where | A |2 and | B |2 specify the probability of qubit to be in state 0 and 1.
The proposed Multi-partite Adaptive Quantum-inspired Evolutionary Algorithm
employs two qubits, first qubit is used to store the solution vector of design vari-
ables and the second qubit is used to store the scaled and ranked objective function
value [8]. The classical implementation of Entanglement principles are mathemati-
cally represented as follows.
|C2i (t)⟩ = f1 |C1i (t)⟩ (4.19)
|C1i (t + 1)⟩ = f2 (|C2i (t)⟩, |C1i (t)⟩, |C1 j (t)⟩) (4.20)
Where |C2i is ith vector of second qubit, |C1i is ith solution vector of first qubit
and |C1 j is jth solution vector of first qubit, t is iteration number, f1 and f2 are the
functions through which both the qubits are classically entangled. The second qubit
is used as feedback in parameter/ tuning free adaptive quantum-inspired rotation
crossover operator. A1i is the probability amplitude of the scaled value of ith variable
in the ith qubit. The variables are scaled between upper and lower limits, the limits
are taken as zero and one. The qubits are stored in quantum register. Number of
variables is equal to number of qubits per quantum register Qi . The structure of Qi is
shown below:
Q1,i = [A1,i,1 , A1,i,2 .....A1,i,n ]
..................
Q1,m = [A1,m,1 , A1,m,2 , .....A1,m,n ]
The second set of qubit in quantum register Qi+1 is used to store the scaled and
ranked objective function value of corresponding solution vector in Qi . The fittest
vector for objective function value is assigned 1, whereas the worst vector for objec-
tive function value is assigned 0 of second qubit set. The remaining solution vectors
for objective function value of second qubit is also ranked in the range of zero and
one.
If nv represents the number of variables used to solve the optimization problem
of DG for optimal placement and capacity of DG with total population np.
   
Q1 X11 ... ... Y1nv
   
... ...
... ...
... ...
... ...
... ...
 =  (4.21)
   
Qnp Xnp1 ... ... Ynpnv
Minimization of power losses with optimal location and sizing of DG is considered
as main objective. The solution vector for solving the above mentioned objective is
represented in Figure 4.8 as follows: The solution vector for simultaneous placement
Figure 4.8: Solution vector representation for DG with optimal location and sizes.
and sizing of DG to reduce the power losses is given as follows.

Q=
 
LDG1,1 LDG1,2 ... LDG1,np SDG1,1 SDG1,2 ... SDG1,np
 LDG2,1 LDG2,2 ... LDG2,np SDG2,1 SDG2,2 ... SDG2,np 
 
 
...
...
...
...
 ... ... ... ... 

L LDGnp−1,2 ... LDGnp−1,np SDGnp−1,1 SDGnp−1,2 ... S 
DGnp−1,1 DGnp−1,np
LDGnp,1 LDGnp,2 ... LDGnp,np SDGnp,1 SDGnp,2 ... SDGnp,np
(4.22)
Three rotation strategies have been applied to converge the population adaptively
towards global optima.
Rotation towards the Best Strategy (R-I): All the solution vectors in the popula-
tion are rotated towards the best solution vector. It is expected that better candidate
solution will be found for all other vectors by rotating the remaining solution vectors
towards the best solution vector.
Rotation Around the Better Strategy (R-IIA): This strategy is primarily used for
exploration purpose, that is two individuals are randomly selected and the search
takes place around the better individual. The direction and the magnitude of search
region is determined by the relative fitness represented in second set of qubits and
relative position of the two individuals stored in the first set of qubits.
Multi Parent Rotation away from worse (R-IIIM): This is inspired from multi par-
enting strategy, which has been previously used by some metaheuristics such as Grey
Wolf Optimization [16] and Symbiotic Organism Search [17]. These metaheuristics
are known for their exploitation, so R-IIIM now employs the best individual, a se-
quentially selected individual, a randomly selected individual and the worst individ-
ual in the population.
Flow chart of proposed algorithm with direct load flow is shown in Figure 4.9.
Pseudo code of the Multipartite Adaptive Quantum-inspired Evolutionary Algo-
rithm is shown below:
Pseudo Code:
Initialization
N p =Numberof Quantum Registers i.e., Quantum-inspired Registers Q1
for i=1: N p
Q1 (i)
=rand (0, 1) ;
Do
Measurement Operator
for i=1: N p
if rand(0,1)¡(Q1 (i))2
Qm (i)= (Q1 (i))2
else
Qm (i)= (1-(Q1 (i))2 )
Fitness calculation

for i=1: N p
var(i) = Back transform(Qm (i))
fitness function(i) = DFL PF(var(i))
Assign Q2 using fitness level of solution vector of Q1
Apply Adaptive Quantum based crossover operator using Q1 and Q2 to gener-
ate Q1c
Elitist selection between Q1c and Q1
While (!termination criteria)
Description:
1. Population size, number of variables and maximum number of iterations are ini-
tially assigned for quantum register i.e., the Quantum-inspired register Q1 is initial-
ized randomly.
Figure 4.9: Flow chart for the proposed algorithm with Load Flow.
Figure 4.10: Single line diagram of test bus system.
2. A new measured value string (Qm ) is generated by implementing measurement

operation on the qubit string.
3. Quantum register Qm is then back transformed into solution vector, var, i.e. pheno-
type and is used to compute the fitness through the load flow by Jen-Hao Teng [52]
explained in Section Power Flow.
4. Second qubit QR2 stores scaled and ranked for solution vector within value [0,
1]. The worst solution vector and fittest solution vector in the second qubit is given
the value 0 and 1. The remaining solution vectors in the second qubit are arranged
between 0 to 1.
5. Three rotation strategies R-I, R-IIA, R-IIIM are applied on Q1 in Adaptive Quan-
tum Crossover variation operator using Q2 to generate children population Q1c .
6. By applying tournament selection between Q1 and Q1c , the corresponding win-
ning individuals amongst Q1 and Q1c will move to the next generation.
7. Maximum number of iterations is used as the termination criterion.
4.6 RESULTS AND DISCUSSION

The performance of the MAQiEA is validated on IEEE standard benchmark test bus
system i.e., 69-bus system to reduce the power losses in distribution network. Under
planning stage, distribution network is designed as meshed or looped structure. In
normal operating conditions the network is operated in radial structure. Figure 4.10
shows the single line diagram of test bus system. All the buses in the benchmark test
Table 4.2
Initial Load Data of the Test Bus System
Particular Value (69 Bus System)
Total Active Power Demand (MW) 3.81
Total reactive Power Demand (MVAr) 2.694
Buses 1 - 69
Sectionalizing switches 1 - 68
Tie line switches 69 - 73
Maximum power rating of DG (MW) 1
bus system are considered as candidate nodes for optimal location and capacity of
multiple DGs except the substation node. All the parameters in the system i.e., line
data and load data of the benchmark test bus system is converted into per units (p.u)
for calculation purpose. Experimental results on test bus systems are carried out on
MATLAB environment hosted on a Intel Core TM i3 CPU computing machine with
4 GB RAM capacity @1.80GHz. MAQiEA has shown better performance as com-
pared with other algorithms available in the literature. The initial data of the medium
voltage test bus system viz., line data and load data are shown in Table 4.2, the to-
tal real and reactive power demand of the system i.e., active and reactive loads on
the network is 3801kW and 2694kVAr. The 69 bus system mainly consists of sixty
nine buses, which has 5 tie line switches and 68 sectionalizing switches which are
numbered as 1 to 68 and 69 to 73 respectively. Active and reactive power losses
of constant power load model with normally open switches 69, 70, 71, 72 and 73
i.e., without opening any tie line switches and without DG implementation is 224.94
kW and 102.12 kVAr with minimum voltage 0.9092p.u respectively. The parameters
used for testing of MAQiEA is given in Table 4.3. First of all, testing has been per-
formed to validate the changes made in the original AQiEA [8] to arrive at MAQiEA.
Thereby, testing was done for validating design decision for changing Rotation To-
wards Better Strategy in AQiEA [7] to Rotation Around Better Strategy (TR-II).
Further, testing was also performed for validating the design decision for changing
Bipartite Rotation away from Worse Strategy in AQiEA [8] to Multipartite Rotation
away from Worse Strategy in MAQiEA. Constant power load model is considered to
validate the results.
Tests for validating design decision for Rotation Around Better Strategy (R-IIA):
The R-II in AQiEA was generating a random variable between [0, 1] as is done
in majority of EAs and so it was termed as a Rotation towards Better strategy. How-
ever, In R-IIA, a random variable is generated between [-1, 1], and is termed as
Rotation Around Better Strategy in MAQiEA. We have arrived at [-1, 1], after thor-
ough investigations, which have been performed to find the power losses incurred in
the system with different search spaces other than 0 and 1. Seven different search
intervals have been investigated to arrive at the best performing one. In Case I, the
Table 4.3
Parameters Different State-of-Art-techniques
GA Population Size=50; Number of Generations=100; Muta-
tion probability= 0.02; Crossover probability= 0.8
PSO Population Size=50; Acceleration factor C1 =C2 =2; Inertia
weights Wmax =0.9, Wmin =0.4
ALO Number of Agents N=50; Itermax =200
GSA Number of Agents N=50; Itermax =200
MAQiEA Population Size=50; Itermax =200
Table 4.4
Wilcoxon Signed Rank Test on MAQiEA with Different Cases
Comparison R+ R− PValue Hypothesis
Case-I Vs Case-II 0 465 0 H0 : Case-I ≥ Case-II
H1 : Case-I ≤ Case-II
Case-I Vs Case-III 26 439 0 H0 : Case-I ≥ Case-III
H1 : Case-I ≤ Case-III
Case-I Vs Case-IV 26 439 0 H0 : Case-I ≥ Case-IV
H1 : Case-I ≤ Case-IV
Case-I Vs Case-V 0 465 0 H0 : Case-I ≥ Case-V
H1 : Case-I ≤ Case-V
Case-I Vs Case-VI 71 394 0.0004 H0 : Case-I ≥ Case-VI
H1 : Case-I ≤ Case-VI
Case-I Vs Case-VII 0 465 07 H0 : Case-I ≥ Case-VII
H1 : Case-I ≤ Case-VII
search interval limits are [-1, 1], for Case II the limits are [0, 1], similarly for other
remaining the search intervals as Case III to Case VII are [-0.5, 1], [-1, 0.5], [-0.25,
1], [-0.25, 0.25] and [0.25, 0.5], respectively. The proposed algorithm is tested with
different search intervals to minimize the fitness function. It was observed [-1 1], has
minimum fitness value in comparison with others. Wilcoxon signed rank test is per-
formed between Case I and other Cases to test the statistical significance of Cass I.
In Wilcoxon signed rank test two hypothesis are created, Null Hypothesis and Alter-
nate Hypothesis i.e., H0 and H1 , respectively and significance level, α is 0.05, which
is compared with the PValue to arrive at a conclusion [53]-[54] shown in Table 4.4.
Table 4.4 shows, the pair wise comparison of Case-I with other Cases. Case-I shows
significant improvement over other Cases with level of significance α =0.05. It has
been observed from tabulated results that null hypothesis is rejected based on PValue
i.e., level of significance, hence Case I is the best performing search interval.
Table 4.5
Performance Analysis of MAQiEA with Other Alternatives
AQiEA MAQiEA1 MAQiEA2 MAQiEA
St. Dev 0.6573 0.0908 0.0984 0.0821
Average 73.4776 71.9906 71.9624 71.8834
Tests for validating design decision for Multipartite Rotation away from Worse
Strategy: In MAQiEA, a novel rotation strategy i.e., Multi Parent Rotation away
from worse is used, R-IIIM, which now employs the best individual, a sequentially
selected individual, a randomly selected individual and the worst individual in the
population. Whereas R-III, in AQiEA, had used the best individual, and a sequen-
tially selected individual. In order to show the effectiveness of MAQiEA, two algo-
rithms other than AQiEA, is incorporated in the study viz., MAQiEA1 and MAQiEA2 .
MAQiEA1 uses tripartite multi parent R-III strategy with best individual, a sequen-
tially selected individual, and a randomly selected individual from the population,
whereas MAQiEA2 uses tripartite multi parent strategy with best individual, a se-
quentially selected individual, and the worst individual in the population. Each algo-
rithm is analysed with thirty independent runs. Based on the performance of theses
runs minimum power loss, maximum power loss, average power loss and standard
deviation are calculated. In addition, Wilcoxon signed rank test is also used to vali-
date the results statistically. Null hypothesis and alternate hypothesis are represented
as H0 and H1 for each test case. Table 4.5 shows the power losses i.e., minimum,
maximum and average power losses after thirty independent runs. Table 4.6 shows,
the pair wise comparison of MAQiEA and other algorithms. It has been observed
from tabulated results that null hypothesis is rejected based on PValue i.e., level of
significance. That is MAQiEA is the best performing version on the basis of Average
and Standard deviation as well as on the basis of Wilcoxon Signed Rank test. For
constant impedance load model, total active and reactive power losses induced in the
system before installing DG is 188.6958kW, 86.5806kVAr with minimum 0.9174p.u.
Similarly, for constant current load model, real and reactive power losses obtained
for base case i.e., without implementation of DG is 158.878kW and 73.767kVAr
with minimum VSI 0.7305p.u. For industrial load model, the load used in the test
bus system is totally dependent on industrial load. Similarly for residential and com-
mercial load models also load used in the system is purely dependent on residential
and commercial loads. In case of constant current load and constant impedance load
model the total load varies with square of the voltage and linearly with voltage re-
spectively. Power losses obtained for practical load models without implementing
DG for industrial, residential and commercial load is171.4316kW, 164.9382kW and
157.0083kW, respectively. A class of mix load is also considered in the study, which
incorporates all voltage-dependent load models including constant power load which
is independent of voltage. The total active and reactive power losses obtained before
implementing DG is 208.8178kW and 95.0525kVAr respectively.
Table 4.6
Wilcoxon Signed Rank Test on MAQiEA with Other Algorithms
Comparison R+ R− PValue Hypothesis
AQiEA Vs MAQiEA1 464 1 0 H0 : AQiEA ≤ MAQiEA1
H1 : AQiEA ≥ MAQiEA1
AQiEA Vs MAQiEA2 464 1 0 H0 : AQiEA ≤ MAQiEA2
H1 : AQiEA ≥ MAQiEA2
AQiEA Vs MAQiEA 464 1 0 H0 : AQiEA ≤ MAQiEA
H1 : AQiEA ≥ MAQiEA
MAQiEA1 Vs MAQiEA 438 27 0 H0 : MAQiEA1 ≤ MAQiEA
H1 : MAQiEA1 ≥ MAQiEA
MAQiEA2 Vs MAQiEA 366 99 0.03 H0 : MAQiEA2 ≤ MAQiEA
H1 : MAQiEA2 ≥ MAQiEA
The parameters used in the algorithms are given in Table 4.3. For constant power
load model, tabulated results in Tables 4.7, 4.8 and 4.9 demonstrate that, MAQiEA
has high reduction in power losses of 71.7497kW and 35.9981kVAr with minimum
voltage 0.979 by implementing DG at optimal location 60, 61 and 17 with optimal
sizes 1MW, 774kW and 509kW respectively. Whereas, GA has minimum reduction
in power losses as compared with all other algorithms. GA has power loss reduction
of 89.737kW and 43.2099kVAr with minimum voltage 0.9093 at optimal location 61,
12, and 23 with capacities 1MW, 520kW and 500kW respectively. ALO has power
loss reduction of 78.1456kW and 38.5394kVAr with optimal location 68, 62 and
63 and capacity of 361kW, 933kW and 935kW. Placing DG at optimal location with
optimal capacity not only reduces the power loss but also improves the voltage profile
of the system. MAQiEA has better improvement in voltage profile in comparison
with GA, PSO, GSA and ALO.
For constant impedance load, proposed algorithm has high reduction in power
loss of 61.3645kW and 31.4816kVAr with implementation of DG at 61, 16 and 60
with capacities of 639kW, 486kW and 1MW respectively. Except the proposed al-
gorithm, ALO has high power loss reduction of 65.3053kW and 32.8232kVAr with
location 61, 68 and 62 and sizing of 1MW, 410kW and 750kW in comparison with
other algorithms. MAQiEA has high improvement in VSI in comparison with GA,
GSA, PSO and ALO. GA has better improvement in voltage profile and VSI in com-
parison with PSO and GSA, however in this case also GA minimum loss reduction
in power loss of 77.8474kW and 38.6822kVAr with DG location 65,55 and 61 and
with capacity of 976kW, 940kW and 467kW respectively.
Similarly for constant current load, ALO has high reduction in power loss ex-
cept MAQiEA. MAQiEA has maximum power loss reduction in comparison with
other algorithms. The overall active and reactive power loss obtained after imple-
menting DG with ALO is 59.1839kW and 30.0164kVAr at locations 64, 60 and 68
with optimal capacities 736kW, 691kW and 900kW respectively. Whereas, PSO and
80
Table 4.7
Comparative Analysis of MAQiEA with Other Algorithms.
Base Case GA PSO GSA ALO MAQiEA
Location ... 61, 12, 23 68, 57, 61 60, 63, 56 68, 62, 63 60, 61, 17
Size (MW) ... 1, 0.52, 0.5 0.87, 1, 0.88 0.7964, 0.9312, 0.3607, 0.9326, 1, 0.7733,
Hybrid Quantum Metaheuristics: Theory and Applications

0.9657 0.9351 0.5084
Constant Power Load Ploss (kW ) 224.9244 89.737 83.2238 84.2603 78.1456 71.7497
Qloss (kW ) 102.1108 43.2099 39.8034 41.1438 38.5394 35.9981
Vmin (p.u) 0.9093 0.9541 0.9714 0.9731 0.9745 0.979
VSI(p.u) 0.6827 0.8277 0.8895 0.8968 0.9016 0.9177
Pload (MW ) 3.8019 1.7819 1.0519 1.1086 1.5735 1.5202
Location ... 65, 55, 61 68, 11, 61 15, 60, 63 61, 68, 62 61, 16, 60
Size (MW) ... 0.9751, 0.9394, 0.4766, 0.6815, 0.7108, 0.6714, 1, 0.4109, 0.6386, 0.4855,
0.4667 1 0.5972 0.7493 1
Constant Impedance Ploss (kW ) 188.6958 77.8474 73.3395 68.6522 65.3053 61.3645
Load
Qloss (kW ) 86.5806 38.6822 35.9269 34.5953 32.8232 31.4816
Vmin (p.u) 0.9174 0.9736 0.9624 0.9682 0.977 0.9813
VSI(p.u) 0.7076 0.8986 0.8569 0.8765 0.9109 0.9263
Pload (MW ) 3.619 1.2378 1.4609 1.6316 1.4588 1.4584
Location ... 27, 64, 49 61, 50, 68 60, 55, 13 64, 60, 68 16, 60, 63
Size (MW) ... 0.7165, 0.9724, 1, 0.3386, 0.9561, 0.8975, 0.7356, 0.6909, 0.5324, 1,
0.7993 0.4249 0.1449 0.9003 0.4384
Multipartite Adaptive QIEA to Reduce Power Losses
Table 4.8
Comparative Analysis of MAQiEA with Other Algorithms (Contd...)
Location ... 61, 12, 23 68, 57, 61 60, 63, 56 68, 62, 63 60, 61, 17
Size (MW) ... 1, 0.52, 0.5 0.87, 1, 0.88 0.7964, 0.9312, 0.3607, 0.9326, 1, 0.7733,
0.9657 0.9351 0.5084
Constant Current Ploss (kW ) 158.8787 71.1685 63.5002 66.9969 59.1839 52.6348
Load
Qloss (kW ) 73.767 31.9649 30.1927 33.6443 30.0164 27.6494
Vmin (p.u) 0.9247 0.9678 0.9647 0.9673 0.9826 0.9832
VSI(p.u) 0.7305 0.8694 0.8655 0.8748 0.924 0.9264
Pload (MW ) 3.45 0.9618 1.6865 1.4515 1.1232 1.4584
Location ... 18, 60, 63 8, 13, 62 62, 51, 17 58, 64, 68 16, 61, 60
Size (MW) ... 0.9861, 0.6573, 0.9559, 0.7189, 1, 0.7399, 0.9444, 0.8343, 0.514, 0.7138,
0.5394 1 0.3653 0.7527 1
Industrial Load Ploss (kW ) 171.4316 49.24 43.8302 44.6244 39.7631 30.7966
Qloss (kW ) 79.0819 25.3316 23.1395 23.9029 21.4549 18.2978
Vmin (p.u) 0.9197 0.9692 0.9681 0.9648 0.9838 0.9864
VSI(p.u) 0.7135 0.8782 0.8739 0.862 0.9286 0.943
Pload (MW ) 3.7679 1.5851 1.0931 1.6627 1.2365 1.5401
Location ... 62, 67, 10 67, 57, 63 62, 57, 63 68, 63, 62 60, 16, 61
Size (MW) ... 0.9479, 0.9535, 0.9984, 0.7229, 0.8763, 0.6872, 0.5084, 0.7894, 1, 0.4909,
0.5762 0.8594 0.7251 1 0.583
81
82
Table 4.9
Comparative Analysis of MAQiEA with Other Algorithms (Contd...)
Location ... 61, 12, 23 68, 57, 61 60, 63, 56 68, 62, 63 60, 61, 17
Size (MW) ... 1, 0.52, 0.5 0.87, 1, 0.88 0.7964, 0.9312, 0.3607, 0.9326, 1, 0.7733,
0.9657 0.9351 0.5084
Residential Load Ploss (kW ) 164.9382 56.1564 48.5161 54.8719 45.0331 39.6854
Hybrid Quantum Metaheuristics: Theory and Applications

Qloss (kW ) 76.3458 27.9655 25.0357 27.7732 24.1137 22.1611
Vmin (p.u) 0.9218 0.9674 0.978 0.9746 0.9803 0.9834
VSI(p.u) 0.721 0.8729 0.9091 0.902 0.9233 0.9332
Pload (MW ) 3.6331 1.1555 1.0524 1.3445 1.3353 1.5592
Location ... 65, 59, 22 63, 51, 23 11, 8, 62 62, 68, 63 61, 16, 60
Size (MW) ... 1, 0.8524, 1, 0.9492, 0.9291, 0.7764, 0.6022, 0.8156, 0.5774, 0.4854,
0.6827 0.6685 0.8329 0.8462 0.9666
Commercial Load Ploss (kW ) 157.0083 58.0609 53.1034 56.9336 47.7983 43.1541
Qloss (kW ) 72.9508 30.0867 27.2928 28.6663 25.1554 23.6074
Vmin (p.u) 0.9242 0.9961 0.9726 0.9662 0.984 0.9842
VSI(p.u) 0.7289 0.9615 0.8892 0.8697 0.9312 0.9367
Pload (MW ) 3.5312 0.9961 0.9135 0.9662 1.2672 1.5018
Location ... 62, 56, 61 68, 62, 18 62, 66, 61 63, 26, 61 61, 16, 60
Size (MW) ... 0.8412, 0.6739, 0.7077, 0.7578, 0.8043, 0.8651, 0.811, 0.3389, 0.7482, 0.5317,
0.9415 0.6394 0.7429 1 1
Mixed Load Ploss (kW ) 208.8178 71.3398 66.3952 65.7329 64.6133 62.3229
Qloss (kW ) 95.0525 35.2727 33.5744 32.821 32.7734 31.7996
Vmin (p.u) 0.9121 0.9736 0.9727 0.9758 0.9826 0.9808
VSI(p.u) 0.6913 0.8986 0.8898 0.9054 0.9281 0.9242
Pload (MW ) 3.7415 1.2849 1.6366 1.3292 1.5916 1.4616
GSA produce active power loss of 63.5kW and 66.99kW. GA has minimum power
loss reduction of 71.1685kW and 31.9649kVAr with locations 27, 64 and 49 with
optimal capacity 717kW, 972kW and 800kW. Maximum percentage power loss re-
duction is obtained with MAQiEA as 52.6348kW and 27.6494kVAr with voltage
profile improvement of 0.9832 p.u. MAQiEA has maximum improvement in VSI in
comparison with all other algorithms, minimum VSI of MAQiEA is 0.9264p.u, ALO
has minimum VSI of 0.924p.u followed by GSA with 0.8748p.u, followed by GA
with 0.8694p.u and PSO with minimum VSI of 0.8655p.u. Improvement in voltage
profile is observed, after installing DG, proposed algorithm has maximum improve-
ment in voltage profile of 0.9832p.u and ALO has improvement in voltage profile of
0.9826p.u.
For industrial load model, the performance of MAQiEA is better in all instances
i.e., real power loss, reactive power loss, voltage profile and VSI. The tabulated re-
sults demonstrate that MAQiEA has high reduction in power loss in comparison with
other algorithms. Overall real and reactive power losses incurred in the system after
implementing DG is 30.7966kW and 18.2978kVAr at location 16, 61, and 60 with
optimal capacity 514kW, 714kW and 1MW with minimum VSI and voltage pro-
file of 0.943p.u and 0.9864p.u respectively. ALO, GSA, PSO and GA has minimum
power loss of 39.7631kW, 44.62kW, 43.83kW and 49.24kW. Minimum improve-
ment in voltage profile is observed in GA of 0.9692p.u, whereas GSA has minimum
VSI of 0.862p.u.
Similarly for residential load and commercial load, the power losses obtained af-
ter installing DG with MAQiEA is 39.6854kW and 43.1541kW. GA has power loss
reduction of 56.1564kW and 58.0609kVAr. For residential load, optimal location and
capacity of DG with MAQiEA is 60, 16, 61 and 1MW, 491kW and 583kW. For com-
mercial load, optimal placement and sizing of DG with the proposed algorithm is 61,
16, 60 and 578kW, 485kW and 967kW respectively. MAQiEA has maximum reduc-
tion in power loss as compared with other algorithm for residential and commercial
load models. For residential load model, ALO has power loss reduction of 45.033kW
and 24.113kVAr with locations 68, 63, 62 and 508kW, 789kW and 1MW respec-
tively. Similarly for commercial load model, ALO has active and reactive power
loss reduction of 47.7983kW and 25.1554kVAr with optimal placement and sizing
of DG 62, 68, 63 and 602kW, 816kW and 846kW with minimum voltage profile of
0.984kW and VSI of 0.9312p.u. For mixed load model, integration of all voltage
dependent load models is considered. The load at every bus is given in appendix.
MAQiEA has maximum improvement in voltage profile and VSI of 0.9808p.u and
0.9242p.u. Whereas other algorithms such as ALO, GSA, PSO and GA has mini-
mum voltage profile of 0.9823p.u, 0.9758p.u, 0.9727p.u, and 0.9736p.u. Minimum
VSI of ALO, GSA, PSO and GA are 0.9281p.u, 0.9054p.u, 0.8898p.u, and 0.8986p.u
respectively. Minimum power losses are obtained with MAQiEA of 62.3229kW and
31.7996kVAr with location 61, 16, 60 and capacities of 748kW, 532kW and 1MW
respectively.
Optimal location and capacity of DG not only reduces the power losses but also
improves the voltage profile. Improvement in voltage profile for constant power
Figure 4.11: Voltage profile improvement for test bus system.
load, constant impedance load, constant current load, industrial load, residential
load, commercial load and mix load models are shown in Figure 4.11 with all al-
gorithms including base case. Similarly for improvement in VSI for constant power
load, constant impedance load, constant current load, industrial load, residential load,
Figure 4.12: Voltage Stability Index for test bus system
commercial load and mix load models are shown in Figure 4.12 with all algorithms
including base case.
Figure 4.13 shows the overall power loss comparison of MAQiEA with ALO,
GSA, PSO, GA and base case. Proposed algorithm has high reduction in power loss
for all load models including constant power load, Figure 4.14 shows the overall im-
provement in voltage profile of MAQiEA with all different load models. Figure 4.15
shows the overall improvement in VSI of MAQiEA with all different load models.
Figure 4.13: Comparison of power loss with MAQiEA with other algorithms for all
load models.
Figure 4.14: Voltage Profile improvement of MAQiEA for all Load Models.
Discussions:
Distribution network has major power loss percentage in power system network as
compared with generation and transmission system. Distributed generators are nor-
mally employed in DN to reduce the power losses. Majority of researchers have im-
plemented DG in DN with a CP load model. However, it is well known that load at
DN varies from time to time. In this study, an investigation has been performed with
DG on different load models to reduce the losses. Voltage-dependent load are used
in this study. CP load model is generally independent of voltage, which doesn’t vary
with voltage. Most of the loads used by consumers at load centers are dependents on
Figure 4.15: Voltage Stability Index of MAQiEA for all Load Models.
voltage. So it is necessary to know the effect of DG on VDLM. CC load and CZ load

models varies linearly and square of the voltage. In addition, some practical loads
which are majorly used in distribution network are also considered viz., RL, IL and
CL models. A class of ML is considered which combines all the voltage dependent
loads. Seven different cases are used to study the effect of DG. If a CP load model is
used, the total load on the test bus system is independent of voltage. Total load used
in the distribution network didn’t dependent on voltage. If a CZ load model is used,
the total load on the test bus system linearly varies with voltage. Total load used in
the test bus linearly varies with voltage. Similarly, in case of CC load model, total
load on the test bus system varies with square of the voltage. In case of practical
loads, if an IL is used, total load used by the test bus system is only industrial load.
Similarly, for CL and RL the total load on the test bus system is commercial load and
residential load. Combination of all load models including CP load model is used in
ML model. Tables 4.10 and 4.11 shows the data of benchmark test bus system.
Multiple DGs are used in this study to reduce the power losses. Multiple DGs have
high reduction in power loss as compared with single DG implementation. In case
of single DG, one DG with high power rating is used to reduce the losses, whereas
multiple DG uses small power rating in comparison with single DG rating. Optimal
placement and sizing of DG is complex combinatorial optimization problem. Numer-
ical, analytical and meta-heuristic algorithms are used to reduce the losses. For large
test bus system, metaheuristics has high reduction in power losses and low computa-
tion time in comparison with analytical and numerical methods. Many metaheuristic
techniques are implemented for this important optimization problem. Normally, evo-
lutionary algorithms suffer from slow convergence, stagnation of choice of param-
eter and premature convergence. An evolutionary algorithm which overcomes the
limitations in EA is QiEA. In this study, a new and efficient quantum-inspired evolu-
tionary algorithm is used to find the optimal location and size of DG, which is termed
as Multi-partite Adaptive Quantum inspired Evolutionary Algorithm. MAQiEA uses
probabilistic approach with Q-bits. MAQiEA is an updated version of AQiEA, which
was an improvement over QiEA. MAQiEA didnt require any additional operator to
avoid premature convergence. QiEA uses single Q-bit whereas AQiEA used two Q-
bits. Q-gates are used in QiEA to move the system towards convergence whereas
MAQiEA used a Multi-partite Adaptive Crossover operator for better convergence.
AQiEA has used three rotation strategies to move the search towards better solutions,
which were bipartite Rotation towards Best, Rotation towards Better and Rotation
away from Worse. Whereas MAQiEA uses improved probabilistic rotation around
better & multi-partite rotation away from worse rotation strategies which provides
for relatively better exploration and exploitation, in addition to Rotation towards
Best Strategy in AQiEA. MAQiEA has high robustness and has better exploitation
and exploration of search space in comparison with AQiEA as shown by test results.
Wilcoxon signed rank test has been used to arrive at best design of MAQiEA amongst
various alternatives. It has been observed from tabulated results that for voltage de-
pendent load models that location of DG is fixed for all load models except constant
power and constant current load. The load on the system is varying exponentially
with voltage level at the node in case of IL, CL and RL. The optimal location of DG
for practical loads is fixed for MAQiEA but other algorithms have different optimal
location for different load models. The robustness of the proposed algorithm is very
high. The results of simulated experiments in the tables demonstrate that MAQiEA is
performing better in comparison with other algorithms (GA, GSA, PSO and ALO).
4.7 CONCLUSIONS
Minimization of power loss in a DN is one of the challenging areas of research for
the distribution utilities. In recent times, power losses are reduced by implementing
DGs into distribution network. However, majority of research has been done on this
important optimization problem with CP load model. Majority of consumers at load
center uses VDLMs such as CZ, CC, IL, RL and CL, whereas CP load model is inde-
pendent of voltage. If the optimal placement and capacity of DG with CP load model
is used on practical distribution system, it induces high power losses and poor voltage
regulation in the system. In this study, an investigation has been performed to reduce
the losses in the distribution system with DG for different VDLMs. Optimal location
and capacity of DG is a difficult non differentiable, non linear, complex combinato-
rial optimization problem. A Multipartite Adaptive Quantum-inspired Evolutionary
Algorithm is proposed for optimal location and sizing of DG. MAQiEA uses prob-
abilistic approach with Q-bits, and is an updated version of AQiEA that has intro-
duced two Q-bits per solution vector and entanglement inspired adaptive crossover
operator. MAQiEA has introduced a Multipartite Adaptive Crossover operator as a
variation operator for better convergence. The effectiveness of MAQiEA is tested on
standard IEEE benchmark test bus system. Tabulated result shows the effectiveness
of the proposed algorithm as compared with other algorithms.
REFERENCES
1. Sorensen, K., Sevaux, M., and Glover, F. (2017). A history of metaheuristics. arXiv
preprint arXiv:1704.00853.
2. Blum, C., and Roli, A. (2008). Hybrid metaheuristics: An introduction. In Hybrid Meta-
heuristics (pp. 1–30). Springer, Berlin, Heidelberg.
3. Singh, B., and Sharma, J. (2017). A review on distributed generation planning. Renew-
able and Sustainable Energy Reviews, 76, 529–544.
4. Han, K. H., and Kim, J. H. (2002). Quantum-inspired evolutionary algorithm for a class
of combinatorial optimization. IEEE Transactions on Evolutionary Computation, 6(6),
580–593.
5. Zhang, G. (2011). Quantum-inspired evolutionary algorithms: A survey and empirical
study. Journal of Heuristics, 17(3), 303–351.
6. Mani, A., and Patvardhan, C. (2009, May). A novel hybrid constraint handling technique
for evolutionary optimization. In 2009 IEEE Congress on Evolutionary Computation
(pp. 2577–2583). IEEE.
7. Manikanta, G., Mani, A., Singh, H. P., and Chaturvedi, D. K. (2016, September).
Placing distributed generators in distribution system using adaptive quantum inspired
evolutionary algorithm. In 2016 Second International Conference on Research in
Computational Intelligence and Communication Networks (ICRCICN) (pp. 157–162).
IEEE.
8. Manikanta, G., Mani, A., Singh, H. P., and Chaturvedi, D. K. (2019). Adaptive quantum-
inspired evolutionary algorithm for optimizing power losses by dynamic load allocation
on distributed generators. SJEE, 16(3), 325–357.
9. Manikanta, G., Mani, A., Singh, H. P., and Chaturvedi, D. K. (2019). Distribution Net-
work Reconfiguration using Adaptive quantum-inspired evolutionary algorithm. Inter-
national Conference on Recent innovation in Electrical Electronics and Communication
Engineering (ICRIEECE-2018) at School of Electrical Engineering, Kalinga Institute of
Industrial Technology (KIIT), Bhubaneswar, India
10. Manikanta, G., Mani, A., Singh, H. P., and Chaturvedi, D. K. (2018, December). Mini-
mization of Power Losses in Distribution System with Variation in Loads Using Adap-
tive Quantum inspired Evolutionary Algorithm. In 2018 4th International Conference on
Computing Communication and Automation (ICCCA) (pp. 1–6). IEEE.
11. Manikanta, G., Mani, A., Singh, H. P., and Chaturvedi, D. K. (2018, October). Distri-
bution Network Reconfiguration with Different Load Models using Adaptive Quantum
inspired Evolutionary Algorithm. In 2018 International Conference on Sustainable En-
ergy, Electronics, and Computing Systems (SEEMS) (pp. 1–7). IEEE.
12. Mani, A., and Patvardhan, C. (2012). An improved model of ceramic grinding process
and its optimization by adaptive Quantum inspired evolutionary algorithm. International
Journal of Simulations: Systems Science and Technology, 11(6), 76–85.
13. Manikanta, G., Mani, A., Singh, H. P., and Chaturvedi, D. K. (2017). DG and Capacitor
Placement in Distribution system considering Cost and Benefits using AQiEA, National
System Conference, DEI, Agra.
14. Manikanta, G., Mani, A., Singh, H. P., and Chaturvedi, D. K. (2016, November). Sitting
and sizing of capacitors in distribution system using adaptive quantum inspired evo-
lutionary algorithm. In 2016 7th India International Conference on Power Electronics
(IICPE) (pp. 1–6). IEEE.
15. Manikanta, G., Mani, A., Singh, H. P., and Chaturvedi, D. K. (2019). Simultaneous
placement and sizing of DG and capacitor to minimize the power losses in radial
distribution network. In Soft computing: Theories and applications (pp. 605–618).

Springer, Singapore.
16. Sultana, U., Khairuddin, A., Mokhtar, A. S., Qazi, S. H., and Sultana, B. (2017). An
optimization approach for minimizing energy losses of distribution systems based on
distributed generation placement. Jurnal Teknologi, 79(4), 87–96.
17. Manikanta, G., Mani, A., Singh, H. P., and Chaturvedi, D. K. (2017, November). Min-
imization of power losses in distribution system using symbiotic organism search al-
gorithm. In 2017 IEEE PES Asia-Pacific Power and Energy Engineering Conference
(APPEEC) (pp. 1–6). IEEE.
18. Seyedali M., Gandomi A., Mirjalili S., Saremi S., Faris H., and Mirjalili S. (2017) Salp
Swarm Algorithm: A bio-inspired optimizer for engineering design problems. Advances
in Engineering Software, 114:163–191.
19. Lopes, J. P., Hatziargyriou, N., Mutale, J., Djapic, P., and Jenkins, N. (2007). Integrating
distributed generation into electric power systems: A review of drivers, challenges and
opportunities. Electric Power Systems Research, 77(9), 1189–1203.
20. Ackermann, T., Andersson, G., and Sder, L. (2001). Distributed generation: A definition.
Electric Power Systems Research, 57(3), 195–204.
21. Viral, R., and Khatod, D. K. (2012). Optimal planning of distributed generation systems
in distribution system: A review. Renewable and Sustainable Energy Reviews, 16(7),
5146–5165.
22. Babu, P. V., Singh, S., and Singh, S. P. (2017, July). Distributed generators allocation in
distribution system. In 2017 IEEE Power & Energy Society General Meeting (pp. 1–5).
IEEE.
23. Prakash, P., and Khatod, D. K. (2016, July). An analytical approach for optimal sizing
and placement of distributed generation in radial distribution systems. In 2016 IEEE 1st
International Conference on Power Electronics, Intelligent Control and Energy Systems
(ICPEICES) (pp. 1–5). IEEE.
24. Vatani, M., Alkaran, D. S., Sanjari, M. J., and Gharehpetian, G. B. (2016). Multiple dis-
tributed generation units allocation in distribution network for loss reduction based on a
combination of analytical and genetic algorithm methods. IET Generation, Transmission
and Distribution, 10(1), 66–72.
25. Al Ameri, A., Nichita, C., Abbood, H., and Al Atabi, A. (2015, March). Fast Estimation
Method for Selection of Optimal Distributed Generation Size Using Kalman Filter and
Graph Theory. In 2015 17th UKSim-AMSS International Conference on Modelling and
Simulation (UKSim) (pp. 420–425). IEEE.
26. Mahmoud, Karar, Yorino, Naoto and Ahmed, Abdella. (2015). Power loss minimiza-
tion in distribution systems using multiple distributed generations. IEEJ Transactions on
Electrical and Electronic Engineering, 10: 521–526, 2015. doi:10.1002/tee.22115.
27. Khoa, T. H., Nallagownden, P., Baharudin, Z., and Dieu, V. N. (2017, November). One
rank cuckoo search algorithm for optimal placement of multiple distributed generators in
distribution networks. In TENCON 2017-2017 IEEE Region 10 Conference (pp. 1715–
1720). IEEE.
28. Guo, Y., Gao, H., Wang, J., Wu, Z., and Han, C. (2015, November). Analysis of dis-
tributed generation effect on system losses in distribution network. In 2015 5th Interna-
tional Conference on Electric Utility Deregulation and Restructuring and Power Tech-
nologies (DRPT) (pp. 1998–2002). IEEE.
29. Mistry, K. (2016, March). MSFL based determination of optimal size and location of
distributed generation in radial distribution system. In 2016 International Conference on
Electrical, Electronics, and Optimization Techniques (ICEEOT) (pp. 530–535). IEEE.

30. de Carvalho, T. L. A., and Ferreira, N. R. (2018, May). Optimal allocation of distributed
generation using ant colony optimization in electrical distribution system. In 2018 Sim-
posio Brasileiro de Sistemas Eletricos (SBSE) (pp. 1–6). IEEE.
31. Vizhiy, S. A., and Santhi, R. K. (2016, March). Biogeography based optimal placement
of distributed generation units in distribution networks: Optimal placement of distributed
generation units. In 2016 International Conference on Electrical, Electronics, and Opti-
mization Techniques (ICEEOT) (pp. 2245–2250). IEEE.
32. Ali, E. S., Abd Elazim, S. M., and Abdelaziz, A. Y. (2017). Ant Lion Optimization Al-
gorithm for optimal location and sizing of renewable distributed generations. Renewable
Energy, 101, 1311–1324, 2017.
33. Mahajan, S., and S. Vadhera, S. (2016, March). Optimal sizing and deploying of dis-
tributed generation unit using a modified multiobjective Particle Swarm Optimization.
In 2016 IEEE 6th International Conference on Power Systems (ICPS) (pp. 1–6). IEEE.
34. Saha, S., and Mukherjee, V. (2016). Optimal placement and sizing of DGs in RDS us-
ing chaos embedded SOS algorithm. IET Generation, Transmission and Distribution,
10(14), 3671–3680.
35. Behera, S. R., and Panigrahi, B. K. (2019). A multi objective approach for placement of
multiple DGs in the radial distribution system. International Journal of Machine Learn-
ing and Cybernetics, 10(8), 2027–2041.
36. Saini, S., and Kaur, G. (2016, December). Real power loss reduction in distribution
network through Distributed Generation integration by implementing SPSO. In 2016
International Conference on Electrical Power and Energy Systems (ICEPES) (pp. 35–
40). IEEE.
37. Bansal, A., and Singh, S. (2016). Optimal allocation and sizing of distributed generation
for power loss reduction.
38. Mer, D. K., and Patel, R. R. (2016, March). The concept of distributed generation and
the effects of its placement in distribution network. In 2016 International Conference on
Electrical, Electronics, and Optimization Techniques (ICEEOT) (pp. 3965–3969). IEEE.
39. Jamian, J. J., Mustafa, M. W., Mokhlis, H., Baharudin, M. A., and Abdilahi, A.
M. (2014). Gravitational search algorithm for optimal distributed generation opera-
tion in autonomous network. Arabian Journal for Science and Engineering, 39(10),
7183–7188.
40. Zhang, J., and Bo, Z. Q. (2010, September). Research of the impact of distribution gener-
ation on distribution network loss. In 45th International Universities Power Engineering
Conference UPEC2010 (pp. 1–4). IEEE, 2010.
41. Hasibuan, A., Masri, S., and Othman, W. A. F. W. B. (2018, February). Effect of dis-
tributed generation installation on power loss using genetic algorithm method. In IOP
Conference Series: Materials Science and Engineering (Vol. 308, No. 1, p. 012034). IOP
Publishing.
42. Ang, S., Leeton, U., Chayakulkeeree, K., and Kulworawanichpong, T. (2018). Sine co-
sine algorithm for optimal placement and sizing of distributed generation in radial dis-
tribution network. GMSARN International Journal, 12, 202–212.
43. Roy, N. K., Hossain, M. J., and Pota, H. R. (2011, September). Effects of load modeling
in power distribution system with distributed wind generation. In AUPEC 2011 (pp.
1–6). IEEE.
44. Angarita, O. F. B., Leborgne, R. C., Gazzana, D. D. S., and Bortolosso, C. (2015, Oc-
tober). Power loss and voltage variation in distribution systems with optimal allocation
of distributed generation. In 2015 IEEE PES Innovative Smart Grid Technologies Latin
America (ISGT LATAM) (pp. 214–218). IEEE.
45. Divya, K., and Srinivasan, S. (2016, January). Optimal siting and sizing of DG in radial
distribution system and identifying fault location in distribution system integrated with
distributed generation. In 2016 3rd International Conference on Advanced Computing
and Communication Systems (ICACCS) (Vol. 1, pp. 1–7). IEEE.
46. Bohre A. K., Agnihotri G. (2016). Optimal sizing and sitting of DG with load mod-
els using soft computing techniques in practical distribution system. IET Generation,
Transmission & Distribution, 10(11), 2606–2621.
47. Das, S., Das, D., and Patra, A. (2016, July). Distribution network reconfiguration using
distributed generation unit considering variations of load. In 2016 IEEE 1st International
Conference on Power Electronics, Intelligent Control and Energy Systems (ICPEICES)
(pp. 1–5). IEEE.
48. Price, W. W., Casper, S. G., Nwankpa, C. O., Bradish, R. W., Chiang, H. D., Concordia,
C., ... and Wu, G. (1995). Bibliography on load models for power flow and dynamic
performance simulation. IEEE Power Engineering Review, 15(2), 70.
49. Price, W. W., Taylor, C. W., and Rogers, G. J. (1995). Standard load models for
power flow and dynamic performance simulation. IEEE Transactions on Power Sys-
tems, 10(CONF-940702-), 1302–1313.
50. Concordia, C., and Ihara, S. (1982). Load representation in power system stability stud-
ies. IEEE Transactions on Power Apparatus and Systems, (4), 969–977.
51. Price, W. W., Chiang, H. D., Clark, H. K., Concordia, C., Lee, D. C., Hsu, J. C., ... and
Vaahedi, E. (1993). Load representation for dynamic performance analysis. IEEE Trans-
actions on Power Systems (Institute of Electrical and Electronics Engineers);(United
States), 8(2).
52. Teng, J. H. (2003). A direct approach for distribution system load flow solutions. IEEE
Transactions on Power Delivery, 18(3), 882–887.
53. Garca, S., Molina, D., Lozano, M., and Herrera, F. (2009). A study on the use of non-
parametric tests for analyzing the evolutionary algorithms behaviour: A case study on
the CEC2005 special session on real parameter optimization. Journal of Heuristics,
15(6), 617–644.
54. Derrac, J., Garca, S., Molina, D., and Herrera, F. (2011). A practical tutorial on the use of
nonparametric statistical tests as a methodology for comparing evolutionary and swarm
intelligence algorithms. Swarm and Evolutionary Computation, 1(1), 3–18.
4.8 PARAMETERS OF IEEE BENCHMARK TEST BUS SYSTEM

The detailed parameters of IEEE benchmark test bus system are listed in the Ta-
bles 4.10 and 4.11.
Table 4.10
Line Parameters for IEEE 69 Bus Radial Distribution System
Branch No From Bus To Bus R(ohms) X(ohms)
1 1 2 0.0005 0.00112
2 2 3 0.0005 0.00112
3 3 4 0.0015 0.0036
4 4 5 0.0251 0.0294
5 5 6 0.366 0.1864
6 6 7 0.381 0.1941
7 7 8 0.0922 0.047
8 8 9 0.0493 0.0251
9 9 10 0.819 0.2707
10 10 11 0.1872 0.0619
11 11 12 0.7114 0.2351
12 12 13 1.03 0.34
13 13 14 1.044 0.345
14 14 15 1.058 0.3496
15 15 16 0.1966 0.065
16 16 17 0.3744 0.1238
17 17 18 0.0047 0.0016
18 18 19 0.3276 0.1083
19 19 20 0.2106 0.069
20 20 21 0.3416 0.1129
21 21 22 0.014 0.0046
22 22 23 0.1591 0.0526
23 23 24 0.3463 0.1145
24 24 25 0.7488 0.2475
25 25 26 0.3089 0.1021
26 26 27 0.1732 0.0572
27 3 28 0.0044 0.0108
28 28 29 0.064 0.1565
29 29 30 0.3978 0.1315
30 30 31 0.0702 0.0232
31 31 32 0.351 0.116
32 32 33 0.839 0.2816
33 33 34 1.708 0.5646
34 34 35 1.474 0.4873
35 3 36 0.0044 0.0108
36 36 37 0.064 0.1565
37 37 38 0.1053 0.123
38 38 39 0.0304 0.0355
39 39 40 0.0018 0.0021
40 40 41 0.7283 0.8509
41 41 42 0.31 0.3623
42 42 43 0.041 0.0478
43 43 44 0.0092 0.0116
44 44 45 0.1089 0.1373
45 45 46 0.0009 0.0012
46 4 47 0.0034 0.0084
47 47 48 0.0851 0.2083
48 48 49 0.2898 0.7091
49 49 50 0.0822 0.2011
50 8 51 0.0928 0.0473
51 51 52 0.3319 0.1114
52 9 53 0.174 0.0886
53 53 54 0.203 0.1034
54 54 55 0.2842 0.1447
55 55 56 0.2813 0.1433
56 56 57 1.59 0.5337
57 57 58 0.7837 0.263
58 58 59 0.3042 0.1006
59 59 60 0.3861 0.1172
60 60 61 0.5075 0.2585
61 61 62 0.0974 0.0496
62 62 63 0.145 0.0738
63 63 64 0.7105 0.3619
64 64 65 1.041 0.5302
65 11 66 0.2012 0.0611
66 66 67 0.0047 0.0014
67 12 68 0.7394 0.2444
68 68 69 0.0047 0.0016
Table 4.11
Load Parameters for IEEE 69 Bus Radial Distribution System
Bus No Active Load (kW) Reactive Load (kVAr) Load Model
1 0 0 Substation
2 0 0 Constant Impedance Load
3 0 0 Constant Current Load
5 2.6 2.2 Constant power Load
6 40.4 30 Residential Load
7 75 54 Industrial Load
8 30 22 Constant power Load
11 145 104 Commercial Load
13 8 5.5 Commercial Load
14 0 0 Residential Load
15 45.5 30 Constant Current Load
19 1 0.6 Constant power Load
21 5 3.5 Constant Impedance Load
28 26 18.6 Residential Load
33 19.5 14 Commercial Load
35 26 18.55 Constant Impedance Load
36 26 18.55 Constant Current Load
40 1.2 1 Commercial Load
42 6 4.3 Constant Current Load
44 39.22 26.3 Industrial Load
45 39.22 26.3 Commercial Load
48 384.7 274.5 Residential Load
49 384.7 274.5 Constant power Load
50 40.5 28.3 Residential Load
51 3.6 2.7 Constant Impedance Load
52 4.35 3.5 Constant Impedance Load
53 26.4 19 Constant Impedance Load
54 24 17.2 Industrial Load
5 Quantum-Inspired Manta
Ray Foraging Optimization
Algorithm for
Automatic Clustering of
Color Images
5.1 INTRODUCTION
Clustering or cluster analysis can be defined as a process of discovering the underly-
ing structure of a data set by partitioning the entire data into two or more groups. In
this process, similar data points are kept together in the same group and dissimilar
data points are kept separate.
The extensive use of clustering can be applied in the field of data mining, engi-
neering, economics, sociology, biology, and physics [1][2]. In order to deal with the
clustering problems, several approaches have been proposed so far, which includes
hierarchical clustering, non-hierarchical clustering, fuzzy clustering, artificial neural
network-based clustering and evolutionary approach-based clustering [1, 2] to name
a few. In the literature, a variety of clustering methods have been developed so far
since the past few years. The prerequisite for most of the existing methods is that they
must have the knowledge of apposite number of clusters beforehand. In most occa-
sions, the dataset may suffer from insufficient and inappropriate knowledge about the
data, that makes the functioning of clustering algorithm, a challenging and tedious
task. To cope up with this limitation, few automatic clustering techniques have been
already developed by several researchers [8][9][10][11].
In the recent years, metaheuristic algorithms have been considered a good choice
for solving several kinds of optimization problems. They are able to provide an ap-
propriate solution for different simple and complex optimization problems within
a short time frame. Some popular metaheuristic algorithms may include Genetic
Algorithm [7], Particle Swarm Optimization [8], Differential Evaluation [9], etc.
Though, the nature-inspired metaheuristic algorithms are capable to solve a prob-
lem very quickly, still they may suffer from premature convergence. In order to
handle this situation, several new approaches can be adopted efficiently and effec-
tively. These approaches may include introducing new parameters in the existing
algorithm, hybridizing more than one algorithm or even incorporating the features
of quantum computing into an existing algorithm. In this regard, quantum-inspired
metaheuristic algorithms have achieved a remarkable efficiency with reference to
DOI: 10.1201/9781003283294-5 95
fast computation without getting stuck in the local optima [10][11][28][13][14][15]

[34][17].
In this paper, an effort has been made to introduce a new quantum-inspired algo-
rithm by coupling the features of quantum computing and a newly developed meta-
heuristic algorithm, called Manta Ray Foraging Optimization [18] algorithm.
The main features of the proposed work are stated below.
A new computational methodology has been developed for automatic clustering
of color image by incorporating the features of quantum computing into a meta-
heuristic, called Manta Ray Foraging Optimization [18] algorithm. This algorithm
is able to identify the optimal number of cluster effectively and efficiently without
getting stuck in local optima.
The exploration and exploitation strategies in the search space have been achieved
by using quantum rotation gate and Pauli-X gate.
The implementation process of the proposed algorithm is simple.
The parameters of the proposed algorithm have been tunes by the Sobol’s sensi-
tivity analysis [19][20][21] to enhance its efficacy.
The rest of the paper has been organized as follows: A brief review has been pre-
sented in Section 5.2. Section 5.3 presents a brief idea about quantum computing.
Section 5.4 describes the validity measurement process of the clustering algorithms.
A brief overview of the Manta Ray Foraging Optimization algorithm has been pre-
sented in Section 5.5. Section 5.6 elaborates the working principle of the proposed
method. The experimental details and analysis of results have been presented in Sec-
tion 5.7. Finally, Section 5.8 presents a brief conclusion with possible future scope.
5.2 LITERATURE REVIEW

Clustering algorithms are basically categorized into two types, viz., partitional clus-
tering and hierarchical clustering [22][1][2]. In case of partitional clustering, the en-
tire dataset is decomposed into a set of disjoint clusters. Few examples of partitional
clustering methods are the K-means algorithm, its variations [23][24], and the expec-
tation maximization (EM) algorithm [25] to name a few. In contrast, a tree structure
is generated during the hierarchical clustering, in which, each cluster represents the
partition of the data set. Few representative examples of hierarchical clustering are
single-linkage, complete-linkage and average-linkage methods [1][26][27] to name
a few. An extensive survey of various clustering methods is presented in [2].
Nowadays, any kind of simple or complex optimization problems can be eas-
ily solved with the help of nature-inspired metaheuristics algorithms. The nature-
inspired metaheuristics algorithms basically accumulate the social behavior of some
creatures, viz., birds, fish, dolphins, ants, lions, wolves, honey bees, fireflies, crow,
bat, etc. In order to solve any kinds of complex problems, they use their intelli-
gence to learn from their surroundings and act accordingly. Some researchers have
been motivated to design some nature-inspired meta-heuristic algorithms to solve
various kinds of complex optimization problems by assimilating the intelligent be-
havior of the social creatures. Examples of some efficient meta-heuristic algorithms
Quantum-Inspired Manta Ray Foraging Optimization for Automatic Clustering 97
may include Genetic algorithm [7], Differential Evaluation [9], Particle Swarm Op-
timization [8], Ant Colony Optimization [28], Bat Optimization [29], Bacterial For-
aging algorithm [30], Firefly algorithm [31], Cuckoo Search [32], and Crow Search
algorithm [33][34] to name a few. By identifying the problem of automatic clus-
tering, several nature-inspired metaheuristics algorithms have been developed so
far, which are available in the literature [8][9][10][11]. The nature-inspired meta-
heuristic algorithms are capable to solve any kind of complex optimization problems
within a short time frame; in addition, they are also capable to provide an optimal or
near optimal solution. In spite of these capabilities, they may suffer from premature
convergence. In order to overcome this situation, the quantum-inspired framework
has been combined with several meta-heuristic approaches. These quantum-inspired
meta-heuristic algorithms can be efficiently and effectively applied to solve various
types of optimization problems, viz., task scheduling on distributed systems [23],
combinational optimization problems [24][17], multi-cast routing problems in wire-
less mesh networks [37], multi-level thresholding problems [38][39], image anal-
ysis [15], mathematical function optimization [40][41], and automatic clustering
[10][11][28][13][14][34] to name a few.
5.3 FUNDAMENTALS OF QUANTUM COMPUTING

In quantum computing, the smallest processing element can be referred as quantum
bit or qubit, which is analogous to binary bit in classical computer. In a classical
computer, a single register can store only one state of information, such as zero (0)
or one(1) at any point of time, but for quantum computer, a single register is capable
to store the states like |0⟩ and |1⟩ and their superposition state |Ψ⟩ at a time. The
superposition state, |Ψ⟩ is basically a combination of the states |0⟩ and |1⟩, which
can be represented as follows:
|Ψ⟩ = α |0⟩ + β |1⟩ (5.1)
where α and β represent the probability amplitudes of the corresponding states, re-
spectively. The following normalization condition should be satisfied by them.
| α |2 + | β |2 = 1 (5.2)
Here, for the purpose of quantum measurement, the superposition state |Ψ⟩ is col-
lapsed either to |0⟩ or to |1⟩ by satisfying the following equation.

|0⟩ if | α |2 > | β |2 ,
|Ψ⟩ = (5.3)
|1⟩ Otherwise
The processing capabilities of quantum-based algorithms depend upon the quantum

logic gates, which are basically congregation of several hardware devices. Quantum
gates are reversible in nature. These gates can be used in different ways for devel-
oping quantum-based algorithms. Some popular quantum gates are Hadamard gate,
Phase Shift gate, Controlled gate, Toffoli gate, Fredkin gate, Rotation gate, Pauli
gate, etc [39]. In this paper, two very useful gates, viz., Rotation gate and Pauli-X
gate [28][13] have been incorporated with the classical MRFO to develop proposed
QIMRFO.
5.3.1 ROTATION GATE

The superposition state, |Ψ⟩ as presented in Equation (5.1), can also be written as
follows:
|Ψ⟩ = cos θ |0⟩ + sin θ |1⟩ (5.4)
where the values of cos2 θ and sin2 θ represent the probability
amplitudes of the
cos θ
states |0⟩ and |1⟩, respectively, and in this case, represents a single qubit.
sin θ
Mathematically, the quantum rotation gate can be represented by the following ma-
trix.
cos δ − sin δ
U(δ ) = (5.5)
sin δ cos δ

cos θ
where δ is the small rotational angle that can be applied to to produce new
" ′ # sin θ
cos θ
qubit , ′ . The process to update quantum states can be accomplished by using
sin θ
the quantum rotation gate as given by
" ′ #
cos θ cos θ cos δ − sin δ cos θ cos(θ + δ )
′ = U(δ ) = = (5.6)
sin θ sin θ sin δ cos δ sin θ sin(θ + δ )
5.3.2 PAULI-X GATE

The Pauli-X gate is a quantum equivalent to NOT gate, which acts on a single qubit
to invert its value. This can be defined as follows:

0 1
X= (5.7)
1 0
Mathematically, one qubit state is converted to other by using Pauli-X gate as follows:

α 0 1 α β
X |Ψ⟩ = X = = (5.8)
β 1 0 β α
The computational details of the Quantum Rotation gate and the Pauli-X gate have
been elaborately described in [28][13].
5.4 VALIDITY MEASUREMENT OF CLUSTERING

In this paper, Pakhira Bandyopadhyay Maulik (PBM) index [42] has been used as
a cluster validity index to identify the appropriate number of clusters in the input
image. This index was proposed by Pakhira et al in 2004. The maximum value of
PBM index indicates optimal result.
Mathematically, PBM index can be defined as follows:
2
1
PBM = × f racE0 EK × DK (5.9)
K
where K represents the number of clusters. The parameter E0 is constant that can be
defined as E0 = ∑P∈DS ∥P −V ∥ where V represents the center of the patterns P ∈ DS.
The parameter EK can be defined as EK = ∑Ki=1 ∑Nj=1 Ui j Pj −Vi , where N is
the number of data points in the data set, U(DS) = [Ui j ]K×N is the partition matrix
of the data points. Vi represents the center of the ith cluster. DK signifies the cluster
separation measure, which can be defined as DK = maxKi,j=1 Vi −V j . The details of
this cluster validity index are available in [42].
5.5 OVERVIEW OF MANTA RAY FORAGING OPTIMIZATION ALGO-

RITHM
Manta Ray Foraging Optimization algorithm [18] has been proposed by Weiguo
Zhao et al in 2020, which was inspired by the foraging behaviors of manta rays.
Manta rays generally use three unique foraging strategies, viz., chain foraging, cy-
clone foraging, and somersault foraging. In MRFO, the population has been per-
turbed by applying these three foraging strategies depending upon some constrains.
These strategies have been elaborately described in [18]. The position of the ith manta
ray having dimension d at a time stamp t can be represented as Pid (t), where i ∈ N.
In chain foraging, Manta rays swim towards the position of plankton by making
a foraging chain. During this time, they not only accept the first move towards the
food but also they follow their preceding one. The mathematical model used for chain
foraging is represented as follows:
 d d (t) − Pd (t)) + α .(Pd (t) − Pd (t)),

 P (t) + r.(Pbest
 i i best i
d i=1
Pi (t + 1) = d (t) − Pd (t)) + α .(Pd (t) − Pd (t)), (5.10)

 Pd (t) + r.(Pi−1
 i i best i
i = 2, 3, ..., N
p
where α = 2.r. |log(r)| and r ∈ [0, 1] is a random number.
In cyclone foraging strategy, whenever a shoal of manta rays identifies a patch of
plankton in deep water, they make a long chain for foraging and swim towards the
food source with a spiral motion. The mathematical model used for cyclone foraging
is represented as follows:
 d d (t) − Pd (t)) + β .(Pd (t) − Pd (t)),

 P + r.(Pbest
 best i best i
d i=1
Pi (t + 1) = d (t) − Pd (t)) + β .(Pd (t) − Pd (t)), (5.11)

 Pd + r.(Pi−1
 best i best i
i = 2, 3, ..., N
T −t+1
where β = 2.er1 ( T ) .sin(2π r1 ), r1 is a random number selected between [0, 1] and
T represents the maximum number of iteration.
The mathematical model used for this strategy can also be represented as follows:
 d d (t) − Pd (t)) + β .(Pd (t) − Pd (t)),

P + r.(Prand
 rand i rand i
d i=1
Pi (t + 1) = d (t) − Pd (t)) + β .(Pd (t) − Pd (t)), (5.12)

 Pd + r.(Pi−1
 rand i rand i
i = 2, 3, ..., N
It can be noted that, based on some predefined condition, either Equation (5.11) or
Equation (5.12) is used in cyclone foraging strategy.
In somersault foraging strategy, each manta ray tries to update its position to the
best position found so far by swimming around the food source. It then somersaults
to new position. The mathematical model used for somersault foraging is represented
as follows.
d
Pid (t + 1) = Pid (t) + S.(r2.Pbest − r3.Pid (t)), i = 1, 2, ..., N (5.13)
where the somersault range of manta rays is decided by using somersault factor,
S = 2. r2 and r3 are two random numbers selected from [0, 1].
5.6 PROPOSED METHODOLOGY

In this paper, a novel algorithm has been proposed for automatic clustering of color
images, named Quantum-Inspired Manta Ray Foraging. In this algorithm, the fea-
tures of quantum gates, viz., Quantum Rotation and Pauli-X gate, have been used to
enhance its computational capability.
In the proposed QIMRFO, the quantum encoding mechanism has been performed
to encode each of the manta ray of the population for identifying the active cluster
centroids from the original population. The basic steps of QIMRFO algorithm are
presented as follows:
1. Initially, a population POP[N×L] with N number of manta rays is created using

pixel intensities of a color image. The pixel intensities are chosen randomly from
a color image and thereafter these are normalized between 0 to 1 to form the
elements of POP[N×L] . In POP[N×L] , each manta ray of length L is considered as a
solution of the problem and L is chosen as the square root of the highest intensity
value among all the components of pixel intensity (red, green, and blue).
2. The POP[N×L] is encoded by a quantum state population QSP[N×L] , which contains
the quantum states either |0⟩ or |1⟩ depending upon the probability amplitudes of
the qubits. The probability amplitudes αi j and βi j of a qubit are produced by
randomly generated values of θ ∈ [0, 2π ], which yields the formula as mention
in Equation (5.4). The following condition is used to detect the state of a single
qubit which belongs to QSP[N×L] .
If | αi j |2 < | βi j |2 then Ψi j = |0⟩, otherwise, Ψi j = |1⟩
where, i = 1, 2, ..., N and j = 1, 2, ..., L.
3. The cluster centroids are identified from the original population POP[N×L] with
the help of quantum state population QSP[N×L] . In QSP[N×L] , the values of |Ψ⟩
that are identified as |1⟩, indicate active cluster centroids. Then, fitness FTN of
each individual manta ray is computed by using Equation (5.9).
4. Quantum rotation gate is used to generate the new quantum state population
QSNP[N×L] by using Equation (5.6). The same set of operations as specified in
Step 3 are performed again to compute new fitness values NFTN of each individ-
ual manta ray by identifying the cluster centroids which belong to POP[N×L] with
the help of QSNP[N×L] .
5. Thereafter, the quantum state population QSP[N×L] and the fitness FTN of the
manta rays are updated based on the fitness values of FTN and NFTN .
6. The basic steps of MRFO are performed on the original population POP[N×L] to
perturb it.
a. The position of the ith manta ray at a time stamp t is updated by using the
following criteria.
If rand < 0.5, then
t
i. If Tmax < rand then, Equation (5.12) is used
ii. Otherwise, Equation (5.11) is used
Else Equation (5.10) is used.
b. The fitness FTN of all manta rays is computed by executing Step 3 and
thereafter, the best individual among them is identified.
c. The positions of all manta rays at time stamp t are updated by Equation
(5.13). Thereafter, Step 7 is performed.
7. If the fitness values are not improved, then quantum Pauli-X gate is used (by
using Equation (10.24)) based on a predefined mutation probability to achieve
diversity in POP[N×L] and afterward, the Step 3 is performed; otherwise Step 8 is
performed.
8. Steps 5 to 9 are repeated for a predefined number of times or until the stopping
criteria is met.
9. Finally, the best fitness value and the corresponding number of cluster centroids
are reported as the optimal results.
In the proposed QIMRFO, a single manta ray may have more than one fitness val-
ues due to the use of quantum rotation gate operation. As the quantum states of that
individual are changed due to the operation of quantum rotation gate, there is a pos-
sibility that the selection of active cluster centroids will be changed. The execution is
always carried out only with the best fitness values of the manta rays, which produces
the exploration in the search space. Similarly, the exploitation in the search space is
accomplished by the use of Pauli-X gate as it acts like a mutation operation which
enables the quantum states to flip their values from 0 to 1 and vice versa, depending
upon a predefined mutation probability. So, by incorporating the features of quan-
tum gates, the search space can be diversified and it always yields better solution.
The flowchart of the proposed algorithm is presented in Figure 5.1.
Figure 5.1: Flowchart of QIMRFO algorithm for automatic clustering of color im-
ages.
5.7 EXPERIMENTAL RESULTS AND ANALYSIS

This paper presents a Quantum-Inspired Manta Ray Foraging Optimization algo-
rithm to identify the optimum number of clusters in color images on the run. The
proposed algorithm has been compared with two different algorithms, viz., MRFO
and GA to judge its supremacy. The proposed algorithm uses the quantum me-
chanical principles, viz., the quantum rotation gate and Pauli-X gate strategy in its
structure. The experimental process and the analysis of the results are presented in
the following subsections.
5.7.1 DEVELOPMENTAL ENTERTAINMENT

The experiments have been conducted on four Berkeley [31] and four publicly avail-
able real-life color images [44][45]. As a developmental environment, Windows 10
operating system with Intel(R) Core (TM) i5-8250U processor, 8.00GB RAM and
1.60GHz CPU speed, has been used. The Python language has been chosen for im-
plementing the proposed algorithm.
5.7.2 DATASET USED

In this paper, eight color images, four are Berkeley images [31] and four publicly
available real-life images [44][45], have been chosen for the experimental purpose.
The test images are presented in Figure 5.2, in which (a)-(d) are Berkley images [31]
of size 481 × 321 and (e)-(h) are real life images [44][45] of sizes 510 × 383, 512 ×
480, 768 × 512 and 512 × 480, respectively.
5.7.3 CLUSTERED IMAGES

The clustered images corresponding to the test images are presented in Figure 5.3
and the threshold values obtained by executing the proposed algorithm have been
presented in Table 5.4.
5.7.4 SENSITIVITY ANALYSIS OF QIMRFO

The performance of the proposed algorithm has been evaluated by executing it sev-
eral times with different parameter settings. As the convergence of any algorithm
depends upon the proper tuning of parameters, the proposed algorithm has also been
gone through Sobol’s sensitivity analysis [19][20][21] for its proper tuning. The ex-
periments for MRFO and GA have been conducted with their original parameter
values as mentioned in their original works [8][18]. The Python’s SALib library has
been used to accomplish the sensitivity analysis for identifying the impact of relative
contribution of each input parameters toward generating the output.
The conditional variances are evaluated to measure the performance of the anal-
ysis. During the evaluation, the first-order index (IA j ) is used to measure the direct
contribution of each input factor to the output variance, which is represented by the
following equation:
V (EX[Q | Pj ])
IA j = (5.14)
V (Q)
where Pj represents the jth input parameter and Q represents the output of the model.
V represents the variance, and the expectation of Q conditional probability on a par-
ticular value of Pj is computed by EX[Q | Pj ].
Figure 5.2: Data Sets : (a)#22093, (b)#163014, (c)#102061, (d)#159045, (e)#pool,

(f)#flower, (g)#tulips and (d)#fruits [31][44][45].
Figure 5.3: Clustered Images of (a)#22093, (b)#163014, (c)#102061, (d)#159045,

(e)#pool, (f)#flower, (g)#tulips and (d)#fruits [31][44][45].
Higher values of IA j as mentioned in Equation (5.14) signify better effectiveness

of the output parameter. The measurement of total indices IT j is used to show the in-
fluence of the jth parameter on the output, which can be mathematically represented
by
EX[V (Q | Pj )]
IT j = (5.15)
V (Q)
The output is influenced by the corresponding non-zero value of IT j .
In this paper, Sobol’s sensitivity analysis test [19][20][21] has been used to tune
four parameters, viz., population size, maximum number of iterations, somersault
factor and the mutation probability of the proposed algorithm. The results of the
sensitivity analysis tests are presented in Table 5.1. The other parameters, viz., α , β ,
small rotation angle δ as used in QIMRFO are basically random in nature; hence no
tuning is required for them. The length of each manta ray has not been tuned since it
has been assigned the same value for all the participating algorithms.
After tuning the parameters of QIMRFO, the best possible combination of those
parameters is reported in Table 5.2. This table also presents the original settings
of parameters of MRFO and GA, which have been mentioned in their respected
Table 5.1
Experimental Results of Sensitivity Analysis [19][20][21] for QIMRFO
Parameters Range 1st Order Effect Total Effect
Population 10 -1.0714 0.0999
20 -0.9827 0.1678
30 0.7925 0.5478
40 0.3803 0.5320
50 0.2999 0.8432
Maximum Iteration 100 -0.4746 1.0764
500 0.0632 1.1200
1000 0.8305 1.3829
1500 0.8211 0.9851
2000 0.5319 0.8733
Somersault Factor 1.5 0.0461 1.7926
1.6 0.1297 1.1091
1.7 0.4793 0.8737
1.8 0.4305 1.2862
1.9 0.3922 1.7701
2.0 0.3774 1.6529
Mutation Probability 0.1 -0.6052 0.0542
0.05 -0.4899 0.0833
0.03 -0.1701 0.0798
0.01 0.6915 0.0916
0.005 -0.0762 1.2940
0.003 0.2824 1.9963
0.001 0.5350 0.0374
Table 5.2
Settings of Parameters for QIMRFO, MRFO and GA
Parameters QIMRFO MRFO GA
Maximum Iteration : MAX I 1000 1000 1000
Population Size: N 30 50 50
Somersault Factor : SF 1.7 2 -
Crossover Probability : CP - - 0.85
Mutation Probability : MP 0.01 - 0.001
Small Rotation Angle : δ [-1.0, 1.0] - -
works [8][18]. Each algorithm has been executed for 40 times using these differ-
ent settings of parameters and the average values among different runs have been
considered for reporting purpose.
5.7.5 ANALYSIS OF EXPERIMENTAL RESULTS

In this paper, a novel Quantum-Inspired Manta Ray Foraging Optimization algorithm
has been introduced for automatic clustering of color images. The proposed algo-
rithm has been implemented on Python environment. The experiments have been
conducted on four Berkeley images [31] and four publicly available real-life color
images [44][45] having different dimensions. All the test images are presented in
Figure 5.2.
The efficiency of proposed algorithm has been compared with its classical coun-
terpart and the well-known genetic algorithm in different aspects. Comparisons have
been made with reference to the mean fitness value, standard deviation of fitness,
standard error of fitness, and computational time. The experimental results of all the
algorithms have been presented in Tables 5.3–5.6. The optimal number of cluster
(ηc ), mean fitness value (µ ), standard deviation (σ ), standard error (ε ), and compu-
tational time (τ ) for all the algorithms are recorded in Table 5.3. The results prove
the efficacy of QIMRFO over MRFO and GA. The optimal number of cluster (ηc )
and the threshold values of three color components, viz., red, green, and blue, of the
proposed algorithm are presented in Table 5.4.
A statistical test, called t-test has also been conducted between the proposed al-
gorithm and each of the comparable algorithm separately. The t-test [38] has been
conducted with 95% confidence level to find p-values. If p < 0.05, then null hypothe-
sis is rejected and the corresponding alternative hypothesis is accepted. The results of
statistical t-test are presented in Table 10.4. While performing t-test between QIM-
RFO and MRFO, five results are extremely significant and two are quite significant
and rest of one is not significant. The test between QIMRFO and GA shows that all
eight results are extremely significant.
Finally, with the help of Friedman test [47][48], the average ranking has been
accomplished for all the algorithms. The test results are presented in Table 5.6. The
average rank of QIMRFO, MRFO, and GA has been found to be 1.2, 2.0 and 2.75,
Table 5.3
Number of Cluster (ηc ), Mean (µ ), Standard Deviation (σ ), Standard Error
(ε ), Computational Time (τ ) in Second of QIMRFO, MRFO, and GA
Data Sets Algorithms ηc µ σ ε τ
#22093 QIMRFO 5 3.369005 0.150678 0.008699 65.97
MRFO 5 2.378611 0.294815 0.037547 126.06
GA 5 2.049843 0.349721 0.060214 162.19
#163014 QIMRFO 4 0.089985 0.001435 0.001015 73.24
MRFO 3 0.077392 0.039875 0.028196 152.54
GA 4 0.058110 0.047322 0.065918 177.45
#102061 QIMRFO 6 1.836728 0.173652 0.000462 200.11
MRFO 6 1.969722 0.329838 0.007926 134.26
GA 4 1.373659 0.759327 0.009992 276.08
#159045 QIMRFO 5 2.060309 0.000439 0.000273 89.29
MRFO 5 1.687840 0.011284 0.000206 111.13
GA 5 1.405063 0.069531 0.009842 151.24
#pool QIMRFO 4 22.10002 0.059366 0.000117 112.36
MRFO 4 15.36671 0.624183 0.002393 147.10
GA 4 10.73810 0.547716 0.007555 163,44
#flower QIMRFO 4 0.015439 0.002204 0.000833 297.28
MRFO 4 0.009497 0.042827 0.006027 297.01
GA 5 0.006394 0.007843 0.005196 312.14
#tulips QIMRFO 5 0.418231 0.000438 0.002458 356.29
MRFO 5 0.292812 0.000126 0.000972 382.15
GA 5 0.248261 0.004701 0.067391 430.38
#fruits QIMRFO 4 2.575446 0.000278 0.000197 86.45
MRFO 4 1.528733 0.023288 0.016467 191.62
GA 4 1.793581 0.019843 0.008652 222.47
respectively. The test values prove that QIMRFO is the best performing algorithm
among all of the competitive algorithms.
The convergence curves of QIMRFO, MRFO, and GA are shown in Figure 5.4.
Each figure shows that QIMRFO converges faster than others. Hence, the superior-
ity of the proposed algorithm has been visually and quantitatively established using
different measures. The population diversity curves using the quantum rotation gate
and the Pauli-X gate have been presented in Figures 5.5 and 5.6, respectively.
5.8 CONCLUSION AND FUTURE SCOPE

In this paper, a Quantum-Inspired Manta Ray Foraging Optimization algorithm has
been presented for automatic clustering of color images. A novel quantum rotation
gate and Pauli-X gate strategy have been introduced to achieve the exploration and
Table 5.4
Number of Cluster (ηc ) with its Corresponding Threshold Values of the Clus-
tered Images
Data Sets ηc Color Component Threshold Value
#22093 5 R [30, 86, 100, 135, 232]
G [51, 92, 143, 175, 247]
B [36, 75, 140, 169, 188]
#163014 4 R [40, 103, 140, 189]
G [42, 125, 150, 176]
B [20, 75, 100, 153]
#102061 6 R [47, 80, 91, 132, 145, 240]
G [62, 100, 155, 210, 229, 241]
B [51, 74, 135, 212, 240, 250]
#159045 5 R [39, 75, 112, 147, 180]
G [30, 61, 100, 143, 175]
B [11, 25, 63, 95, 140]
#pool 4 R [15, 25, 140, 200]
G [43, 61, 90, 192]
B [10, 53, 142, 191]
#flower 4 R [50, 85, 158, 200]
G [25, 60, 106, 153]
B [10, 42, 147, 196]
#tulips 5 R [25, 61, 83, 212, 229]
G [30, 100, 112, 146, 200]
B [31, 62, 80, 150, 191]
#fruits 4 R [37, 77, 148, 162]
G [10, 62, 100, 146]
B [25, 36, 70, 98]
exploitation in the search space for identifying the optimal results. Several tests have
been conducted among the competitive algorithms to judge the effectiveness of the
proposed algorithm. The experimental results prove that QIMRFO outperforms oth-
ers in all aspects.
In future, the functionality of the proposed algorithm can be expanded in such a
way that it can efficiently handle the high dimensional data sets. Various quantum
gate strategies can also be used to develop new algorithms for solving different types
of optimization problems.
Table 5.5
Results of t-Test Between QIMRFO vs. MRFO, and QIMRFO vs. GA
Data Sets p - value Significance
#22093 QIMRFO & MRFO <0.0001 Extremely Significant
QIMRFO & GA <0.0001 Extremely Significant
#163014 QIMRFO & MRFO 0.04571 Significant
#102061 QIMRFO & MRFO 0.0268 Significant
QIMRFO & GA 0.0003 Extremely Significant
#159045 QIMRFO & MRFO <0.0001 Extremely Significant
#pool QIMRFO & MRFO <0.0001 Extremely Significant
#flower QIMRFO & MRFO 0.3835 Not Significant
#tulips QIMRFO & MRFO <0.0001 Extremely Significant
#fruits QIMRFO & MRFO <0.0001 Extremely Significant
Table 5.6
Results of Friedman Test [47][48] for QIMRFO, MRFO and GA
Data Sets QIMRFO MRFO GA
#22093 3.9218 (1) 2.7263 (2) 2.4497 (3)
#163014 0.0921 (1.5) 0.0921 (1.5) 0.0765 (3)
#102061 1.9728 (2) 2.1097 (1) 1.7383 (3)
#159045 2.0999 (1) 1.6982 (2.5) 1.6982 (2.5)
#pool 22.8993 (1) 15.9609 (2) 14.9761 (3)
#flower 0.0171 (1.5) 0.0171 (1.5) 0.0162 (3)
#tulips 0.4314 (1) 0.2983 (2.5) 0.2983 (2.5)
#fruits 2.5833 (1) 1.5934 (3) 1.8217 (2)
Average Rank : 1.25 2.0 2.75
ACKNOWLEDGMENT
This work was supported by the AICTE sponsored RPS project on Automatic Clus-
tering of Satellite Imagery using Quantum-Inspired Metaheuristics vide F.No 8-
42/RIFD/RPS/Policy-1/2017-18.
Figure 5.4: Convergence curves of QIMRFO, MRFO, and GA for test images.
Figure 5.5: Population Diversity using Quantum Rotation Gate of test images
[31][44][45].
Figure 5.6: Population Diversity Increased by using Pauli-X Gate of test images
[31][44][45].
REFERENCES
1. A. K. Jain and R. C. Dubes. Algorithms for Clustering Data. Prentice-Hall, Inc., USA,
1988.
2. A. K. Jain, M. N. Murty, and P. J. Flynn. Data clustering: A review. ACM Computing
Surveys, 31(3):264–323, 1999.
3. S. Bandyopadhyay and U. Maulik. Genetic clustering for automatic evolution of clusters
and application to image classification. Pattern Recognition, 35(6):1197–1208, 2002.
4. A. E. Ezugwu. Nature-inspired metaheuristic techniques for automatic clustering: A survey
and performance study. SN Applied Sciences, 2, 2020.
5. A. Jose-Garca and W. Gomez-Flores. Automatic clustering using nature-inspired meta-
heuristics: A survey. Applied Soft Computing, 41:192–213, 2016.
6. S. Das, A. Abraham, and A. Konar. Automatic clustering using an improved differential
evolution algorithm. IEEE Transactions on Systems, Man, and Cybernetics – Part A: Sys-
tems and Humans, 38(1):218–237, 2008.
7. John H. Holland. Adaptation in Natural and Artificial Systems: An Introductory Analysis
with Applications to Biology, Control and Artificial Intelligence. MIT Press, Cambridge,
MA, 1992.
8. J. Kennedy and R. Eberhart. Particle swarm optimization. In Proc. IEEE International Con-
ference on Neural Networks, Perth, Australia, pp. 1942–1948, 1995.
9. R. Storn and K. Price. Differential evolution – a simple and efficient heuristic for global
optimization over continuous spaces. Journal of Global Optimization, 11:341–359, 1997.
10. S. Bhattacharyya, V. Snasel, A. Dey, S. Dey, and D. Konar. Quantum spider monkey opti-
mization (QSMO) algorithm for automatic gray-scale image clustering. International Con-
ference on Advances in Computing, Communications and Informatics (ICACCI 2018), pp.
1869–1874, 2018.
11. A. Dey, S. Bhattacharyya, S. Dey, J. Platos, and V. Snasel. Quantum-inspired bat optimiza-
tion algorithm for automatic clustering of grayscale images, vol. 922, pp. 89–101. Springer,
Singapore, 2019.
12. A. Dey, S. Dey, S. Bhattacharyya, J. Platos, and V. Snasel. Novel quantum-inspired ap-
proaches for automatic clustering of gray level images using particle swarm optimization,
spider monkey optimization and ageist spider monkey optimization algorithms. Applied
Soft Computing, 88(106040), 2020.
13. A. Dey, S. Dey, S. Bhattacharyya, J. Platos, and V.S. Snasel. Quantum-Inspired Automatic
Clustering Algorithms: A comparative study of Genetic Algorithm and Bat Algorithm, pp.
89–114. De Gruyter, 2020.
14. A. Dey, S. Dey, S. Bhattacharyya, V. Snasel, and A.E. Hassanien. Simulated Annealing
Based Quantum-Inspired Automatic Clustering Technique, pp. 73–81. Cairo, 2018.
15. S. Dey, S. Bhattacharyya, and M. Ujjwal. Quantum Behaved Swarm Intelligent Techniques
for Image Analysis: A Detailed Survey, pp. 1–39. GI Global, Hershey, USA, 2015.
16. S. Dey and U. Bhattacharyya and S. Maulik. Quantum-Inspired Automatic Clustering
Technique Using Ant Colony Optimization Algorithm, 2018.
17. K.H. Han and J.H. Kim. Quantum-inspired evolutionary algorithm for a class of combina-
torial optimization. IEEE Transactions on Evolutionary Computation, 6(6):580–593, 2002.
18. W. Zhao, Z. Zhang, and L. Wang. Manta ray foraging optimization: An effective bio-
inspired optimizer for engineering applications. Engineering Applications of Artificial In-
telligence, 87:103300, 2020.
19. A. Saltelli, P. Annoni, I. Azzini, F. Campolongo, M. Ratto, and S. Tarantola. Variance
based sensitivity analysis of model output. design and estimator for the total sensitivity
index. Computer Physics Communications, 181(2):259–270, 2010.
20. A. Saltelli and I.M. Sobol. Sensitivity analysis for nonlinear mathematical models: Numer-
ical experience. Matematicheskoe Modelirovanie, 7(11):16–28, 1995.
21. I.M. Sobol. Global sensitivity indices for nonlinear mathematical models and their Monte
Carlo estimates. Mathematics and Computers in Simulation, 55(1–3):271–280, 2001.
22. H. Frigui and R. Krishnapuram. A robust competitive clustering algorithm with applica-
tions in computer vision. IEEE Transactions on Pattern Analysis and Machine Intelligence,
21(5):450–465, 1999.
23. P.S. Bradley and U.M. Fayyad. Refining initial points for k-means clustering. In Proceed-
ings of the Fifteenth International Conference on Machine Learning, pp. 91–99. Morgan
Kaufmann Publishers Inc., 1998.
24. D. Pelleg and A. Moore. X-means: Extending k-means with efficient estimation of the
number of clusters. In Proceedings of the 17th International Conference on Machine Learn-
ing, pp. 727–734. Morgan Kaufmann, 2000.
25. X.L. Meng and D.V. Dyk. The em algorithm an old folksong sung to a fast new tune.
Journal of the Royal Statistical Society. Series B (Methodological), 59(3):511–567, 1997.
26. F. Murtagh. A survey of recent advances in hierarchical clustering algorithms. The Com-
puter Journal, 26(4):354–359, 1983.
27. F.J. Rohlf. 12 single-link clustering algorithms. In Classification Pattern Recognition and
Reduction of Dimensionality, vol. 2 of Handbook of Statistics, pp. 267–284. Elsevier,
1982.
28. M. Dorigo, M. Birattari, and T. Stutzle. Ant colony optimization, vol. 1. IEEE, 2006.
29. X.S. Yang. A new metaheuristic bat-inspired algorithm. In Nature Inspired Cooperative
Strategies for Optimization (NICSO 2010), pp. 65–74. Springer, 2010.
30. K.M. Passino. Biomimicry of bacterial foraging for distributed optimization and control.
IEEE Control Systems Magazine, 22(3):52–67, 2002.
31. X.-S. Yang. Firefly algorithm, stochastic test functions and design optimisation. Interna-
tional Journal of Bio-inspired Computation, 2(2):78–84, 2010.
32. X.-S. Yang and S. Deb. Cuckoo search via levy flights. pp. 210–214, 2010.
33. A Askarzadeh. A novel metaheuristic method for solving constrained engineering opti-
mization problems: Crow search algorithm. Computers & Structures, 169:1–12, 2016.
34. Z.A. Babak, B.H. Omid, and X. Chu. Crow Search Algorithm (CSA), pp. 143–149.
Springer Singapore, Singapore, 2018.
35. T. Gandhi, Nitin, and T. Alam. Quantum genetic algorithm with rotation angle refinement
for dependent task scheduling on distributed systems. In 2017 Tenth International Confer-
ence on Contemporary Computing (IC3), pp. 1–5. IEEE, Aug 2017.
36. H.P. Chiang, Y.H. Chou, C.H. Chiu, S.Y. Kuo, and Y.M. Huang. A quantum-inspired
tabu search algorithm for solving combinatorial optimization problems. Soft Computing,
18:1771–1781, 2013.
37. M. Mahseur, A. Ramdane-Cherif, D. Acheli, and Y. Meraihi. A quantum-inspired bi-
nary firefly algorithm for qos multicast routing. International Journal of Metaheuristics,
6(4):309, 2017.
38. S. Dey, S. Bhattacharyya, and U. Maulik. Efficient quantum inspired metaheuristics for
multi-level true color image thresholding. Applied Soft Computing, 56:472–513, 2017.
39. S. Dey, I. Saha, S. Bhattacharyya, and U. Maulik. Multilevel thresholding using quantum-
inspired metaheuristics. Knowledge-Based System, 67:373–400, 2014.
40. S.S. Tirumala. A quantum-inspired evolutionary algorithm using Gaussian distribution-
based quantization. Arabian Journal for Science and Engineering, 43:471–482, 2018.
41. Y.-J. Yang, S.Y. Kuo, F.-J. Lin, I.-I. Liu, and Y.-H. Chou. Improved quantum-inspired tabu
search algorithm for solving function optimization problem. 2013 IEEE International Con-
ference on Systems, Man, and Cybernetics, pp. 823–828, 2013.
42. M. Pakhira, S. Bandyopadhyay, and M. Ujjwal. Validity index for crisp and fuzzy clusters.
Pattern Recognition, 37:487–501, 2004.
43. Berkley images. www2.eecs.berkeley.edu/Research/ Projects/CS/vision/bsds/BSDS300/
html/dataset/ images.html. Accessed on 15/01/2020.
44. Real life images. www.hlevkin.com/06testimages.htm. Accessed on 15/01/2020.
45. Real life images. https://homepages.cae.wisc.edu/ ece533/images/. Accessed on
23/05/2020.
46. B. Flury. A First Course in Multivariate Statistics. Springer Texts in Statistics
47. M. Friedman. The use of ranks to avoid the assumption of normality implicit in the analysis
of variance. Journal of the American Statistical Association, 32(200):675–701, 1937.
48. M. Friedman. A comparison of alternative tests of significance for the problem of m rank-
ings. Annals of Mathematical Statistics, 11(1):86– 92, 1940.
6 Automatic Feature
Selection for Coronary
Stenosis Detection in X-Ray
Angiograms Using
Quantum Genetic
Algorithm
6.1 INTRODUCTION
The automatic coronary stenosis detection problem is a challenging task since it in-
volves detailed analysis over X-Ray Coronary Angiograms (XCA) in the form of
gray-level digital images. XCA remains as the gold-standard imaging technique for
medical diagnosis of arterial diseases, including stenosis and other related condi-
tions. In this procedure, a liquid dye, such as fluorescein, is injected using a thin
catheter inserted into an access point to the bloodstream (usually in arm or groin).
The dye reveals an arterial structure that can be easily seen on X-ray images and
allows to cardiologists to detect narrowed or blocked areas through the coronary
arteries.
In clinical practice, the stenosis cases detection process is performed by cardiolo-
gists. To detect the possible stenosis cases, the specialist performs a visual scan over
an X-ray coronary angiogram (XCA) image which can be printed over a physical
media or as gray-level digital image. Until the process, the specialist labels the an-
giogram over different regions where a stenosis case is present according with his
expertise and knowledge. Figure 6.1 illustrates an angiogram and their respective
stenosis regions labeled by a specialist. However, given the limited access to such
delicate clinical expertise, the variability of diagnoses among specialists has allowed
that the automatic Computer-Aided Diagnosis (CAD) systems play a vital role in
cardiology to assist the detection of coronary artery stenosis.
The problem of automated stenosis cases detection in XCA has been addressed
from different approaches in literature. For example, Kishore and Jayanthi [1] make
use of a fixed-size window (patch) that is manually selected from an previously en-
hanced image and after, an adaptive threshold algorithm is applied to keep only ves-
sel pixels. With the segmented image, a vessel width is calculated by adding the
intensity values from the left to the right edge. In this approach, there is no need to a
DOI: 10.1201/9781003283294-6 117

Figure 6.1: X-ray coronary angiogram. From left to right: Original angiogram, steno-
sis diseases labeled by a cardiologist and, the zoom of the stenosis labels.
skeletonization process of the arteries, using only the vessel width to determine the
grade of a stenosis case into the selected window. Saad [2] was able to detect stenosis
cases by using vessel skeletons. This approach requires a previous image segmenta-
tion process in order to extract vessel pixels and after, a skeletonization procedure
is performed to extract only the center lines corresponding to the vessels. With the
vessel center lines (skeleton), the length of the orthogonal line is computed using a
fixed-size window that is moving over the image in order to obtain a vessel-width
measure that is compared with a fixed value to determine if a stenosis case exists or
not in that region of the image. Sameed et al. [3] make use of the Hessian matrix to
enhance vessel pixels and determine candidate stenosis regions, by identifying nar-
rowed vessel areas. Wan et al. [4] carried out the vessel diameter estimation using
a smoothed vessel centerline curve from the candidate stenosis regions detected by
the Hessian-approach. Both, from the artery lumen (vessel diameter), allow deter-
mining the stenosis measurement and final classification. Cervantes-Sanchez et al.
[5] proposed a method for computing the vessel width along the arteries by applying
the second-order derivatives directly over the enhanced images, where the cases of
stenosis were labeled as local minima through the vessel width; in this approach,
no additional skeletonization or vessel diameter estimation was needed. Posteriorly,
Cruz-Aceves et al. [6] used a Bayes classifier over a handcrafted 3D feature vector
that was obtained from the results of potential cases of coronary stenosis identi-
fied previously by a second-order derivative operator. Major disadvantage for this
method relies in the need of a predefined threshold or fixed value in the form of
a vessel width, narrowest measure, etc., which a current computed value, must be
compared or contrasted against to determine if it is greater, equal or lower than that
fixed threshold, resulting in a problem by itself the determination of that fixed value.
Convolutional Neural Networks (CNNs) have emerged to overcome the disadvan-
tages of methods where a-priori fixed value (or a set of them) must be established
to perform a classification. They have been successfully used to solve diverse kinds
Automatic Feature Selection for Coronary Stenosis Detection in X-Ray Angiograms 119
of problems involving image processing in medical imaging, such as image segmen-

tation, enhancement, classification, and location of specific interest regions. Instead
of dealing with pre-defined features, CNN’s input are raw images, which are seg-
mented although the convolutional layers by applying a set of specific filters in order
to keep only useful information according to the expected results during the train-
ing stage. One of the most reliable CNN architectures in medical imaging is the
U-Net [7] proposed by Harouni in 2018. U-Net is composed of 18 convolutional lay-
ers plus the output layer. Figure 6.2 illustrates the U-Net architecture. In addition,
Figure 6.2: U-Net architecture.
CNNs have been applied successfully for the coronary stenosis detection problem.
Antczak et al. [8] tested several CNNs architectures using a natural image dataset
plus 10000 instances of synthetically generated vessels patches, which includes pos-
itive and negative cases. An additional strength of CNNs is the possibility to pass
knowledge from a pre-trained CNN to a new one. This process is called Transfer
Learning [9]. However, success of knowledge transfer depends on dissimilarity be-
tween the source domain (where the CNN has been trained) and the target domain
(where the knowledge is transferred) [10].
Major disadvantages of CNNs are related with choosing the right CNN archi-
tecture and the large amount of instances required to achieve a correct training and
classification, despite the risk to fall in an over-fitting state due mainly to an unbal-
anced dataset (where the number of positive and negative cases are significantly dif-
ferent). For instance, in the automated coronary stenosis detection problem, the num-
ber of positive cases are significantly smaller than negative cases. To overcome this
disadvantage, strategies such as data augmentation [11] and synthetic data genera-
tion [12] are applied. However, those strategies entails the inconvenient of manually
select a representative and significant dataset from which the new augmented data
will be generated or a correct methodology or model to generate synthetic cases
closed to the real ones.
In this research is proposed a novel method based on the automated feature se-
lection for detection of coronary stenosis cases in X-ray coronary angiograms. The
method was tested using an image database with ≈ 2800 instances. In addition, the
proposed method was compared with other five strategies taken from the state-of-
the-art, which includes machine learning and deep learning techniques. To measure
the effectiveness of the proposed and compared methods, the Accuracy and the Jac-
card’s Index metrics were used, achieving a rate of 0.92 and 0.85, for the accuracy
and Jaccard index, respectively. The obtained results probed the method effectiveness
for the automated stenosis cases detection problem in X-ray coronary angiograms by
selecting only a subset of features and keeping at same time, an optimal classification
rate.
6.2 BACKGROUND
6.2.1 FEATURE EXTRACTION
In Image Processing, feature extraction term refers to the multiple metrics that can
be computed or extracted from an image, a region, or a single pixel from it. Since
single-pixel metrics does not offer valuable information by itself, the most common
approach to extract significant information from an image are the windows (patches)
or the entire image. The feature extraction process can be performed in an automated
way, for instance, by using a CNN. However, almost all CNN-based procedures ob-
fuscate the extracted features because the complexity inherent to the process or the
feature itself. This kind of strategies leads to produce a classification result only.
On the other hand, specific or manually extracted features can be analyzed in detail.
However, this approach has the disadvantage that features must be collected by the
researcher in a non-fully automated way most of the time.
From an overall point of view, extracted features from an image can be classified
in three categories: intensity, texture, and morphology. Each of these groups will
be described on next sections and how they are related with the studied problem.
It is important to mention that, since the gray-level scale pixel intensity is the most
common approach for X-ray coronary angiogram digital imaging, all features will
be related to work with gray-level images.
6.2.1.1 Pixel Intensity-based Features

Pixel intensity features are the most single metrics that can be computed from an im-
age or region of it. They are extracted computing classical statistical measurements
from the intensities (values) of the pixels such as minimum (min), maximum (max),
mean, median, and standard deviation. Figure 6.3 illustrates two images with their
respective intensity values and their corresponding intensity-based feature values.
Min: 127.00 Min: 65.00

Max: 127.00 Max: 137.00
Mean: 127.00 Mean: 113.91
Median: 127.00 Median: 117.00
Std. Dev.: 0.00 Std. Dev.: 15.49
Figure 6.3: Intensity-based features for two images of 16x16 in the first row. Second
row contains the corresponding pixel intensities for each of them. Third row contains
the corresponding intensity-basic features.
6.2.1.2 Texture Features

Texture features provide information about the rugosity level present in an image.
Normally, texture features are related with sudden intensity variations in a region of
the image. Intensity changes are useful to describe interest objects or regions inside
the image. Most common approach to extract texture related features are the Haral-
ick texture features [13], which are computed using the Gray-Level Co-Occurrence
Matrix or GLCM. The GLCM is a matrix in which rows and columns are formed
by the pixel intensities taken from the entire image or a region from it. The GLCM
has the same number of rows and columns. The matrix contents are computed based
on the frequency of a pixel with intensity i occurring in a specific spatial relationship
to a pixel with intensity j. This operation is expressed by the following equation:

G−1G−1
P(i, j, d, θ ) = ∑ ∑ P(i, j), (6.1)
i=0 j=0
where P(i, j, d, θ ) is the frequency which two pixels with intensities i and j at a
distance d and an angle θ occurs. G is the number of gray levels used.
6.2.1.3 Morphologic Features

Morphologic features are related with shape and the inherent properties of the inter-
est objects or regions inside the image. For problems related to artery study, morpho-
logic features provide information about the vessel’s shape, the proportion of vessel
pixels against non-vessels pixels and others which will be described later.
Typically, morphologic features require a previous image enhancement or seg-
mentation process in order to highlight or keep only the interested objects or regions,
while the rest of the image is treated as background. Most of the coronary X-ray
angiograms contain noise and narrowed artery regions. For this reason, an effective
segmentation procedure must be applied to the raw image in order to be able to ex-
tract morphologic features from vessels.
It is important to mention that some features are computed from the vessel skele-
ton. For this reason, a second processing task must be performed after the enhance-
ment or segmentation procedure in order to extract the vessel’s center lines that are
known as the artery skeleton. The artery skeleton provides information about the
vessels such as their tortuosity, number of branches, their segment length, etc. Fig-
ure 6.4 illustrates a raw X-ray coronary angiogram image and their respective vessel
segmentation and skeleton extraction.
Figure 6.4: Sample of a raw X-ray coronary angiogram image and their correspond-
ing vessel segmentation and skeleton extraction, from left to right.
Additional texture and morphologic features can be extracted by combining in-

formation from the raw image and their corresponding vessels and skeleton. For
instance, vessel texture measures can be extracted with the segmented vessel infor-
mation from the raw image.
Figure 6.5: The same above X-ray angiogram with their respective vessel segmenta-
tion and skeleton overlapped in order to illustrate a combined feature extraction.
For this study, the Frangi [14] method was used in order to segment vessels from
the angiographies because the optimal results are obtained. The Frangi method uses
the eigenvalues obtained from a Hessian matrix. The Hessian matrix is computed
from the second-order derivative of the original image. It is calculated by convolving
a Gaussian kernel at different orientations with the original image as follows:

x2 + y2
G(x, y) = − exp − ∥y∥ < L/2, (6.2)
2σ 2
where σ is the spread of the Gaussian profile and L is the length of the vessel seg-
ment. The resultant Hessian matrix is expressed as follows:

Hxx Hxy
H= , (6.3)
Hyx Hyy
where Hxx , Hxy , Hyx , and Hyy are the directional second-order partial derivatives of
the image.
The segmentation function defined by Frangi for 2-D vessel detection is as follows:

0 if λ2 > 0,
f (x) = R2b 2 (6.4)
exp − 2α 2 1 − exp 2Sβ 2 elsewhere.
The α parameter is used with Rb to control the shape discrimination. The β parameter
is used by S2 for noise elimination. Rb and S2 are calculated as follows:
|λ1 |
Rb = , (6.5)
|λ2 |
q
S2 = λ12 + λ22 , (6.6)
where λ1 and λ2 are the eigenvalues of Hessian matrix.
The Frangi method response is a gray-scale image with the vessel pixels en-
hanced. In order to fully eliminate non-vessel pixels (background and noise), a bi-
narization of the Frangi response must be applied. In this research, the Otsu [15]
method was used because the threshold value is calculated automatically based on
the image pixels by computing the weighted sum of variance of the two classes,
expressed as follows:
σω2 (t) = ω0 (t)σ02 (t) + ω1 (t)σ12 (t), (6.7)
where ω0 and ω1 weights are the probabilities of the two classes separated by a
threshold t, and σ02 and σ12 are the statistical variances of ω0 and ω1 , respectively.
After the vessel segmentation procedure was applied, the Medial Axis Transform
technique was used to extract the vessel skeletons, which makes use of the Voronoi
method, expressed as follows:
Rk = {x ∈ X|d(x, Pk ) ≤ d(x, Pj ) for all j ̸= k}, (6.8)
where Rk is the Voronoi region associated with the site Pk (a tuple of nonempty
subsets in the space X), which contains the set with all points in X whose distance to
Pk is not greater than their distance to the other sites Pj . j is any index different from
k. d(x, Pk ) is a closeness measure from point x to point Pk . The Euclidean distance is
commonly used as a closeness measure and it is defined as follows:
q
D(p1 , p2 ) = (x2 − x1 )2 + (y2 − y1 )2 (6.9)
where D(p1 , p2 ) is the distance between two points p1 and p2 defined by coordinates
(x1 , y1 ) and (x2 , y2 ), respectively, in a 2-D plane.
Figure 6.6 illustrates five coronary angiograms with their respective segmentation
response applying the Frangi method and the corresponding results after the binariza-
tion and skeletonization tasks were performed. In Section A.1 is presented a Matlab
implementation to extract vessel segments. The code function makes use of other
auxiliary functions presented in subsequent sections in the appendix. For instance,
in Section A.2 is presented the code that finds and returns the positions related with
the pixels of a segment in order to calculate their morphology on a posterior process.
The code in Section A.3 is useful to extract a window from a 2-D matrix. Finally,
the code presented in Section A.4 finds and returns the first row position of a matrix
where a specific row vector is located. Those methods are useful to extract distinct
features related with vessel’s morphology.
6.2.2 FEATURE SELECTION

Feature selection refers to select only those features from a larger set that are use-
ful to solve some specific problem. Since a significant amount of features can be
Figure 6.6: Image vessel detection. In first row, original angiograms are presented.
Second row shows the Frangi segmentation response. In third row, the Otsu method
response is illustrated. Last row contains the vessel skeletons by applying the Medial
Axis Transform method.
extracted, the challenge is to find the best feature subset that is able to classify coro-
nary stenosis and non-stenosis cases correctly. The complexity associated to solve
the posed problem relies in the large number of combinations of features to achieve
the best classification rate. The problem complexity can be computed as 2n possibili-
ties to combine features, where n is the number of features. Figure 6.7 illustrates two
examples of feature selection.
6.2.3 SUPPORT VECTOR MACHINES

Support Vector Machines (SVMs) were designed primary to be linear classifiers by
applying concepts of vector calculus in physics [16][17]. Major strength of SVMs
relies in their capacity to project data to high dimensional orders than they are rep-
resented initially and, where a linear separation can be performed correctly [18]. In
order to perform a linear classification, SVMs select a subset of data points that are
able to influence the position and orientation of the hyperplane, which can be used
to perform the linear classification. It is important to be mentioned that SVMs are
Feature Set
Two distinct feature selection subsets
Figure 6.7: Feature selection example.
supervised learning models. This means that a training data set and its correspond-
ing label set are required to perform the SMV training. In order to project the data
represented in a space χ to a higher dimensional space F , the SVM makes use of the
Mercer kernel operator. For a given training data x1 , ..., xn , that are vectors in some
space χ ⊆ Rd , the support vectors can be considered as a set of classifiers expressed
as follows [19]:
n
f (x) = ∑ αi K(xi , x) . (6.10)
i=1
When K satisfies the Mercer condition [20], it can be expressed as follows:
K(u, v) = Φ(u) · Φ(v), (6.11)
where Φ : χ → F and “·” denotes an inner product. With this assumption, f can be
rewritten as follows:
f (x) = w · Φ(x),
n
(6.12)
w = ∑ αi Φ(xi ).
i=1
6.2.4 QUANTUM GENETIC ALGORITHM

As described before in Section ??, feature selection is a complicated task when the
number of features is considerably large, since the total number of different combi-
nations of features can be computed as 2n . For this reason, by performing random or
manual trials will have low probabilities to find the optimal combination of features
able to achieve an accurate classification rate and minimizing at same time, the num-
ber of involved features. In order to overcome this problem, the use of metaheuristic
search strategies is a suitable way to explore. In this research, the Quantum Genetic
Algorithm (QGA) was employed due to the advantages that it has over other search
strategies.
QGA is the product of the combination of quantum computation and genetic al-
gorithms, and it is a new evolutionary algorithm of probability [21]. In 1996, QGA
was first proposed by Narayanan and Moore [22]. It was successfully used to solve
the TSP problem [23]. QGA can be applied to solve problems where the conven-
tional genetic algorithm has been applied. In the state of the art literature a wide
of QGAs implementations are described to solve optimization problems success-

fully [24][25][26][27][28], such as the personnel scheduling problem [29], dynamic
economic dispatch problem [30], multi-sensor image registration [31], cryptanaly-
sis [32], and others.
Quantum computing is based on the evolution concept from quantum mechanics,
which establishes that the evolution of an isolated quantum system from an initial
state to another final state is governed by the Schrödinger equation [33], expressed
as follows:
∂
iℏ = |ψ (t)⟩ = H(t) |ψ (t)⟩ , (6.13)
∂t
√
where i is the imaginary number −1 and ℏ is a constant term named “reduced
Planck constant” and ℏ = h/2π . The |ψ (t)⟩ vector depends on time and describes the
state of a quantum system (or state) at time t. In addition, it can be noted that quantum
mechanics equations are expressed using proper notation symbols to describe vector
and matrix operations [34]. In 1930, the theoretical physicist Paul Dirac published the
book The Principles of Quantum Mechanics introducing the ket notation to describe
a column vector as follows:  
v1
v2 
 
|V ⟩ =  .  (6.14)
 .. 
vi
and bra notation to denote a row vector as following:

⟨M| = w1 w2 . . . wi (6.15)
therefore, the product of bra and ket vectors can be expressed using the notation of
Dirac as follows:
 
v1

v2 
⟨M|V ⟩ = w1 w2 . . . wi  .  = w1 v1 + w2 v2 + . . . + wi vi (6.16)
 .. 
vi
The products of ket and bra vectors are expressed as follows:

   
v1 v1 w1 v1 w2 v1 w... v1 wi
v2  v2 w1 v2 w2 v2 w... v2 wi 
   
|V ⟩ ⟨M| =  .  w1 w2 . . . wi =  . .. .. ..  (6.17)
 ..   .. . . . 
vi vi w1 vi w2 vi w... vi wi
When translating previous physics concepts to computing field, the term qubit is used
to represent the minimal unit of information that stores |0⟩ and ⟨1| states. Applying
this principle to the QGA, an initial quantum population Q(0) is composed of a set of
quantum individuals. Each individual will be composed of a quantum chromosome
i that is defined as a string of j qubits representing a quantum system |ψ ⟩i with 2 j

simultaneous states, expressed as following:
j
α1 α2 α3 . . . α j
→ |ψ ⟩i = ∑ ci |ψk ⟩ , (6.18)
β1 β2 β3 . . . β j i k=1
being the gene j, the qubit represented by a vector expressed as follows:

αj
→ ψj (6.19)
βj
and the quantum population expressed as following:

α1 α2 α3 . . . α j
β1 β2 β3 . . . β j 1

α1 α2 α3 . . . α j
β1 β2 β3 . . . β j 2
(6.20)
..
.

α1 α2 α3 . . . α j
β1 β2 β3 . . . β j i
In QGA the initial population is generated by setting the value for all qubits in the
chromosomes by representing the quantum superposition of all states with equal
probability [33]. The initialization can be computed using the Hadamard matrix by
the vector |0⟩ as follows:

1 1 1 1 1 1
H. |0⟩ = √ =√ , (6.21)
2 1 −1 0 2 1
obtaining a superposition vector. After the product was computed, a phase angle
θ (0, π2 ] is randomly obtained, being the argument of the trigonometric functions or
elements in the rotation matrix U(t), expressed as follows:

cos(θ ) −sin(θ )
U(t) = (6.22)
sin(θ ) cos(θ )
Finally, to conclude the initialization step, the product of the rotation matrix by the
superposition vector must be performed, obtaining a pair of amplitudes (α , β ), which
is defined as the state of j qubit as follows:

αj cos(θ ) −sin(θ ) 1 1
= √ (6.23)
βj sin(θ ) cos(θ ) 2 1
In order to obtain a discrete population of chromosomes, an observation must be
performed as follows:
(
p(α ) ≤ |α 2j |, x j = 0(basisstate |0⟩)
(6.24)
p(α ) > |α 2j |, x j = 1(basisstate |1⟩)
Such as in classic GA, on each iteration, crossover and mutation of population

is performed in order to produce new individuals. These operations are known as
Quantum Crossover Gate and Quantum Mutation Gate. The mutation occurs in two
different manners: inversion and insertion. To perform the inversion-based mutation,
an inter-qubit mutation of the jth qubit is computed by swapping the amplitudes with
the quantum Pauli X gate as follows:

0 1
U(t) = , (6.25)
1 0
which results in ! t
β t+1
j 0 1 αj
=
α t+1
j 1 0 β tj
The quantum mutation using the insertion gate is performed by the permutation or
swapping between two qubits chosen randomly. For instance, given the following
quantum chromosome with first and third qubits chosen randomly:

α1 α2 α3 . . . α j
,
β1 β2 β3 . . . β j
the new mutated chromosome applying the insertion strategy will be as follows:

α3 α2 α1 . . . α j
β3 β2 β1 . . . β j
The quantum crossover is performed similar to classic GA by selecting a randomly

cut point and exchanging chromosomal segments. For instance, suppose that two
chromosomes m and n expressed as follows:

α1 α2 α3 . . . α j
β1 β2 β3 . . . β j m
′
α1 α2′ α3′ . . . α ′j
β1′ β2′ β3′ . . . β j′ n
were selected for a crossover operation with a randomly selected point between first
and second positions. The resultant recombined chromosomes are expressed as fol-
lows:

α1 α2′ α3′ . . . α ′j
β1 β2′ β3′ . . . β j′ m∗
′
α1 α2 α3 . . . α j
β1′ β2 β3′ . . . β j n∗
Using the previous concepts, the QGA steps are defined and compared with its clas-
sical corresponding GA in Table 6.1.
Table 6.1
QGA Steps
Step Quantum GA Classic GA
1 Initialize quantum population Q0 Generate initial population P0
2 Make P(0) by measuring each indi- Evaluate P(0)
vidual Q(0) → P(0)
3 while (not termination condition) while (not termination condition)
do do
4 begin begin
5 t ← t +1 t ← t +1
6 Perform Quantum Crossover Perform Crossover
7 Perform Quantum Mutation Perform Mutation
8 Measure Q(t) → P(t) Evaluate population P(t)
9 end end
6.3 PROPOSED METHOD

This research is focused in the use of a QGA to select a subset of features that are
suitable to perform the classification of stenosis and non-stenosis cases keeping or
improving at same time, the accuracy rate that is obtained by using the original fea-
ture set. The method starts with a feature extraction task in order to obtain an ini-
tial set of features from a bank of image patches. In this first step, distinct types
of features are extracted such as basic-intensity measures, texture-based and shape
features. By performing this step, an initial set of 31 features are extracted from the
image bank. Features 1 to 14 correspond to Haralick texture features described as
follows:
1. Angular second moment (energy)

2. Contrast
3. Correlation
4. Variance
5. Inverse difference moment (homogeneity)
6. Sum average
7. Sum variance
8. Sum entropy
9. Entropy
10. Difference variance
11. Difference entropy
12. Information measure of correlation 1
13. Information measure of correlation 2
14. Maximum correlation coefficient
Features 15 to 18 correspond to basic intensity measurements in the image.

15. Minimum intensity value

16. Maximum intensity value
17. Mean intensity of the image
18. Standard deviation of the intensities in the image
The remaining features are related with the vessel morphology. As mentioned before,
a previous vessel enhancement and skeletonization tasks are required to keep only
information related with vessel pixels. The Frangi method was applied in order to
perform a vessel enhancement over the original images. Later, a binarization of each
Frangi response was performed by applying the Otsu method in order to discriminate
non-vessel pixels. Finally, the Medial Axis Transform procedure was applied in order
to extract vessel skeletons. Shape-based features are described as follows:
19. The number of pixels corresponding to vessels in the patch

20. The number vessel segments in the patch
21. Vessel density. The proportion of vessel-pixels present in the patch
22. Tortuosity 1. The tortuosity of each segment was calculated using the true length
(measured with the chain code) divided by the Euclidean length. The mean tortu-
osity was calculated from all the segments within the sub window
23. Vessel Length
24. Number of bifurcation points. The number of bifurcation points removed within
the sub window when creating segments.
25. Gray level coefficient of variation. The ratio of the standard deviation to the mean
of the grey level of all segment pixels within the sub window.
26. Gradient mean. The mean gradient magnitude along all segment pixels within
the sub window calculated using the Sobel gradient operator applied on the pre-
processed image.
27. Gradient coefficient of variation. The ratio of the standard deviation to the mean
of the gradient of all segment pixels within the sub window.
28. Mean vessel width. Skeletonization correlates to vessel center lines. The distance
from the segment pixel to the closest boundary point of the vessel using the vessel
map prior to skeletonization. This gives the half-width at that point, which is then
multiplied by 2 to achieve the full vessel width. The mean is calculated for all
segment pixels within the sub-window
29. The minimum standard deviation of the vessels length, based on the vessels
present in the patch
30. The maximum standard deviation of the vessels length, based on the vessels
31. The mean of the standard deviations of the vessels length, based on the vessels
It is important to mention that a vessel segment is referred to the skeletonized vessel

segments. After a feature set is defined, a QGA is used with an SVM in order to
search for an optimal feature selection. The proposed method procedure is described
in detail in Figure 6.8.
1. Input: Image bank of 2. Feature Extraction

coronary angiogram Extract features and generate a
patches in grayscale format feature bank.
3. Feature Selection
Perform a Feature Selection using QGA with SVM.
3.1. Generate Initial Quantum Population
3.2. Evaluate population using SVM training
3.3. Perform Quantum Crossover and Mutation tasks
Repeat until max iterations are reached
3.4. Select Best solution found based on training effectiveness
Figure 6.8: Feature selection example.
After the feature selection task is finished, the SVM can be trained using only
the feature subset found previously by the QGA. In order to probe the effective-
ness, a testing set is used. In addition, the classification results are measured us-
ing two different metrics in order to assess the achieved results. As a first instance,
the True-Positive (TP), True-Negative (TN), False-Negative (FN), and False-Positive
(FP) fractions are used to obtain the Accuracy (Acc) metric. This metrics are com-
puted as follows:
TP
T PR = , (6.26)
T P + FN
TN
T NR = , (6.27)
T N + FP
FP
FPR = , (6.28)
FP + T N
TP+TN
Acc = , (6.29)
T P + T N + FP + FN
where T PR is the True-Positive Rate, T NR is the True-Negative Rate, FPR is the

False-Positive Rate, T P is the number of positive instances that were labeled as pos-
itive by the classifier, T N is the number of negative instances labeled as negative by
the classifier, FP is the number of negative instances labeled as positive by the clas-
sifier, and FN is the number of positive instances labeled as negative by the classifier.
As a second accuracy metric, the Jaccard index is used. The Jaccard index is a
measure that determines how similar are two sets of elements. It is useful to measure
the efficiency of a classifier by comparing the obtained results versus those expected.
It is computed as follows:
A∩B
J(A, B) = , (6.30)
A∪B
where A and B are the expected and response sets, respectively.
6.4 EXPERIMENT DETAILS

For the experimentation, an image bank of coronary patches was formed. The images
correspond to the natural Antczak [8] dataset, which is formed by 1394 negative and
122 positive instances. However, in order to avoid an overfitting of the SVM, 1272
positive instances were taken from the Antczak synthetic dataset. With the synthetic
positive-stenosis patches addendum, the dataset was equilibrated with 1394 positive
cases and 1394 negative cases. Each patch is of size 32 × 32 pixels and they are
defined in gray-scale. Figure 6.9 illustrates a set of samples with positive and negative
cases.
(a) Positive instances (b) Negative instances
Figure 6.9: 40 instances of coronary X-ray angiogram patches taken from the
Antczak [8] image dataset. (a) Positive stenosis cases. (b) Negative stenosis cases.
For the training and the feature selection process, 1670 instances were taken ran-
domly from the image set in a proportion of 50−50% for positive and negative cases,
respectively. The remaining instances were used for testing purposes.
As mentioned before, an SVM was used to perform the classification of positive
and negative cases. It was configured to use a polynomial kernel of order 6, a ker-
nel scale of 6.5, and a kernel offset of 0.1. All SVM parameters were established
manually as a trade-off between time and accuracy. Additionally, in order to increase

the train rate of the SVM, the training dataset was partitioned into 10 groups with
167 instances in order to perform a cross-validation with k = 10.
All experiments were performed using the Matlab 2018Rb platform using an Intel
i3 PC with 4GB of RAM.
6.5 RESULTS
After conducting the experiment, significant results were obtained related with the
feature selection process and the classification accuracy achieved. In order to assess
the feature selection results, a statistical analysis was performed considering the best
result achieved by the QGA on each trial from a set of 30 trials. In its best global
solution, the QGA was able to find an optimal feature subset of 20 instead of the ini-
tial set containing 31 features. Additionally, it is interesting to notice the frequency
which each feature was present in the best solution achieved by QGA over all trials,
since it provides information about the influence of that particular feature on the clas-
sification result. Table 6.2 describes the frequency which each feature was present in
the best solution achieved by the QGA in each trial.
Table 6.2
Feature Selection Frequency over All QGA Trials
Feature Frequency
Contrast 0.55
Correlation 0.64
Sum Variance 0.73
Sum Entropy 0.45
Entropy 0.45
Difference Entropy 0.45
Min 1.00
Max 1.00
Mean 0.64
Std.Dev. 0.91
Number of Vessel Pixels 0.91
Number of Vessel Segments 0.73
Vessel Density 0.91
Vessel Length 0.91
Number of Bifurcation Points 0.64
Grey Level Coefficient of Variation 1.00
Gradient Mean 0.73
Gradient Coefficient of Variation 0.91
Mean Vessel Width 0.91
Vessel. Std. Dev. Max 0.64
Figure 6.10 illustrates the frequency for all set of 31 features. Based on the results
1
Frequency
0.5
0 Feature
Name
Figure 6.10: Frequency in which each feature was present in best solution achieved
by QGA over all trials. x-axis describes each feature name. y-axis represents the
frequency of each feature, where 1 means that the feature appears in the solution
achieved by the QGA in all trials (the 100%).
described in Table 6.2 and the chart presented in Figure 6.10, the importance or effect
of each individual feature in the classification task and how the combination of them
can lead to an optimal classification rate can be observed. It is remarkable that some
features appear on all solutions achieved by the QGA, meaning that those features
are really important in the classification process. In addition, it can be contrasted the
results described in Table 6.2 with Figure 6.10 and how almost a half of selected
features are above of the mean frequency, which was 0.72.
The performance results are presented in Table 6.3. The proposed method is com-
pared with other four classification methods in the literature, such as Feed Forward
(FFNN) and Back Propagation (BPNN) Neural Networks, UNET and CNN-16C
Convolutional Neural Networks.
Based on results presented in Table 6.3, the classification process was performed
with the highest rates in terms of the Accuracy and Jaccard Index metrics by
the SVM-based classification method. However, the proposed method was able to
achieve closest performance results to the SVM by using only a subset of 20 features
instead of the total feature set, which contains a total of 31 features. It is important
to notice that the performance of non-linear classification methods such as FFNN
and BPNN decreases when a feature selection task is applied to them. This can
be attributed to the loss of information associated with the elimination of features
which are relevant for a non-linear classification task. On the other hand, deep learn-
ing methods such as UNET and CNN-16C Convolutional Neural Networks were
able to achieve the closest results to the best rate after the proposed method. In this
Table 6.3
Accuracy Rate and Jaccard Index Comparison for the Proposed Method and
5 additional Classification Methods from the State of the Art
Number of Jaccard
Method Features
Accuracy Index
20 0.70 0.54
FFNN
31 0.72 0.61
20 0.69 0.57
BPNN
31 0.71 0.00
SVM 31 0.92 0.86
UNET [7] – 0.76 0.72
CNN-16C [8] – 0.86 0.74
Proposed Method 20 0.92 0.85
context, it is important to mention that CNNs were applied using only the same
training and testing set such as the rest of techniques and no data-augmentation or
transfer-learning processes were applied in order to measure the effectiveness of all
techniques under the same conditions. In Figure 6.11, a subset of instances cor-
responding to the True-Positive, True-Negative, False-Positive, and False-Negative
fractions are illustrated.
(a) True-Positive instances (b) True-Negative instances
(c) False-Positive instances (d) False-Negative instances
Figure 6.11: Samples of the true-positive, true-negative, false-positive, and false-

negative, fractions of the accuracy metric. Each fraction contains 10 instances.
6.6 CONCLUSION
In this research, a method is proposed for the stenosis classification problem, which
makes use of a Quantum Genetic Algorithm (QGA) to perform an automatic feature
selection in order to select only those features that has a strong influence on a Support
Vector Machine-based (SVM) classifier. The QGA performs a search over the space
formed by the feature set in order to find an optimal combination of features and
at the same time, keeping or decreasing the loss rate in the training stage. Initially,
a set of 31 features were extracted from an image database which contains ≈ 2780
instances of X-ray coronary angiogram patches. The image database is balanced in
terms of the positive and negative stenosis cases. After the feature selection process
ends, a subset of 20 features was able to keep the classification rate in terms of
the Accuracy metric and the Jaccard index, compared with the original set with 31
features. By using only 20 features, the accuracy rate and Jaccard index were 0.92
and 0.85, respectively, which are very similar to those obtained using the full set of
31 features. In addition, the reduction of features has effect on the time required to
perform an exhaustive feature extraction of new angiograms, since the required time
for extract the 31 features was 0.94 seconds, versus 0.62 seconds that are required to
extract the 20 selected features, considering a window that will perform a scan over
an angiogram to detect regions with possible stenosis cases. Based on the results
obtained in this study, it can be concluded that the proposed method can be applied
in clinical practice to assists cardiologists in the evaluation and finding of possible
stenosis cases in X-ray coronary angiograms.
ACKNOWLEDGMENT
The present research has been supported by the Universidad Tecnológica de León.
REFERENCES
1. A. Kishore and V. Jayanthi. Automatic stenosis grading system for diagnosing coronary
artery disease using coronary angiogram. International Journal of Biomedical Engineering
and Technology, 31(3):260–277, 2018.
2. I. Saad. Segmentation of coronary artery images and detection of atherosclerosis. Journal
of Engineering and Applied Sciences, 13:7381–7387, 2018.
3. S. Sameh, M. Azim, and A. AbdelRaouf. Narrowed coronary artery detection and clas-
sification using angiographic scans. In 2017 12th International Conference on Computer
Engineering and Systems (ICCES), pages 73–79, 2017.
4. T. Wan, H. Feng, C. Tong, D. Li, and Z. Qin. Automated identification and grading of
coronary artery stenoses with x-ray angiography. Computer Methods and Programs in
Biomedicine, 167:13–22, 2018.
5. F. Cervantes-Sanchez, I. Cruz-Aceves, and A. Hernandez-Aguirre. Automatic detection of
coronary artery stenosis in x-ray angiograms using Gaussian filters and genetic algorithms.
AIP Conference Proceedings, 1747, 2016.
6. I. Cruz-Aceves, F. Cervantes-Sanchez, and A. Hernandez-Aguirre. Automatic detection
of coronary artery stenosis using Bayesian classification and Gaussian filters based on
differential evolution. Hybrid Intelligence for Image Analysis and Understanding, pages
369–390, 2017.
7. A. Harouni, A. Karargyris, M. Negahdar, D. Beymer, and T. Syeda-Mahmood. Universal
multi-modal deep network for classification and segmentation of medical images. In 2018
IEEE 15th International Symposium on Biomedical Imaging (ISBI 2018), pages 872–876,
2018.
8. Karol Antczak and ukasz Liberadzki. Stenosis detection with deep convolutional neural
networks. MATEC Web of Conferences, 210:04001, 2018.
9. Shin Hoo-Chang, Holger Roth, Mingchen Gao, Le Lu, Ziyue Xu, Isabella Nogues, Jianhua
Yao, Daniel Mollura, and Ronald-M. Summers. Deep convolutional neural networks for
computer-aided detection: CNN architectures, dataset characteristics and transfer learning.
IEEE Transactions on Medical Imaging, 35(5):1285–1298, 2016.
10. Hossein Azizpour, Ali Razavian, Josephine Sullivan, Atsuto Maki, and Stefan Carlsson.
From generic to specific deep representations for visual recognition. In 2015 IEEE Con-
ference on Computer Vision and Pattern Recognition Workshops (CVPRW), pp. 36–45,
2015.
11. David-A. van Dyk and Xiao-Li Meng. The art of data augmentation. Journal of Computa-
tional and Graphical Statistics, 10(1):1–50, 2001.
12. Jessamyn Dahmen and Diane Cook. Synsys: A synthetic data generation system for health-
care applications. Sensors, 19(5): 2019.
13. R. Haralick, K. Shanmugam, and I. Dinstein. Textural features for image classification.
IEEE Transactions on Systems, Man, and Cybernetics, 3(6):610–621, 1973.
14. A. Frangi, W. Niessen, K. Vincken, and M. Viergever. Multiscale vessel enhancement filter-
ing. Medical Image Computing and Computer-Assisted Intervention (MICCAI98), pages
130–137, 1998.
15. Otsu Nobuyuki. A threshold selection method from gray-level histograms. IEEE Transac-
tions on Systems, Man and Cybernetics, 9(1):62–66, 1979.
16. Corinna Cortes and Vladimir Vapnik. Support-vector networks. Machine Learning,
20:273–297, 1995.
17. Nello Cristianini, John Shawe-Taylor, et al. An introduction to support vector machines
and other kernel-based learning methods. Cambridge University Press, 2000.
18. William-S. Noble. What is a support vector machine? Nature Biotechnology, 24:1565–
1567, 2006.
19. Simon Tong and Daphne Koller. Support vector machine active learning with applications
to text classification. Journal of Machine Learning Research, pages 45–66, 2001.
20. Christopher-J.C. Burges. A tutorial on support vector machines for pattern recognition.
Data Mining and Knowledge Discovery, 2:121–167, 1998.
21. F. Shi, H. Wang, L. Yu, and F. Hu. Analyze of 30 Cases of MATLAB Intelligent Algo-
rithms. Beihang University Press, 2010.
22. A. Narayanan and M. Moore. Quantum-inspired genetic algorithms. In Proceedings of
IEEE International Conference on Evolutionary Computation, pages 61–66, 1996.
23. Davendra Donald. Traveling salesman problem, theory and applications. Intech, Rijeka,
2011.
24. Kuk-Hyun Han and Jong-Hwan Kim. Quantum-inspired evolutionary algorithms with a
new termination criterion, h/sub /spl epsi// gate, and two-phase scheme. IEEE Transactions
on Evolutionary Computation, 8(2):156–169, 2004.
25. Zhifeng Zhang and Hongjian Qu. A new real-coded quantum evolutionary algorithm. In
Proceedings of the 8th WSEAS International Conference on Applied Computer and Ap-
plied Computational Science, pages 426–429, 2009.
26. Gexiang Zhang. Quantum-inspired evolutionary algorithms: A survey and empirical study.
Journal of Heuristics, 17:303–351, 2011.
27. Utpal Roy, Sudarshan Roy, and Susmita Nayek. Optimization with quantum genetic algo-
rithm. International Journal of Computer Applications, 102(16):1–7, 2014.
28. Ying Sun and Xiong Hegen. Function optimization based on quantum genetic algorithm.
Research Journal of Applied Sciences, Engineering and Technology, 7(1):144–149, 2014.
29. Wang Huaixiao, Li Ling, Liu Jianyong,Wang Yong, and Fu Chengqun. Improved quan-
tum genetic algorithm in application of scheduling engineering personnel. Abstract and
Applied Analysis, 2014:1–10, 2014.
30. Lee Jia-Chu, Lin Whei-Min, Liao Gwo-Ching, and Tsao Ta-Peng. Quantum genetic algo-
rithm for dynamic economic dispatch with valve-point effects and including wind power
system. International Journal of Electrical Power & Energy Systems, 33(2):189–197, 2011.
31. Ying Sun and Xiong Hegen. A novel quantum-inspired evolutionary algorithm for mul-
tisensor image registration. The International Arab Journal of Information Technology,
3(1):9–15, 2006.
32. Hu Wei. Cryptanalysis of tea using quantum-inspired genetic algorithms. Journal of Soft-
ware Engineering and Applications, 3:50–57, 2010.
33. Rafael Lahoz-Beltra. Quantum genetic algorithms for computer scientists. Computers,
5(4):243–249, 2012.
34. Leonard Susskind and Art Firedman. Quantum Mechanics: The Theoretical Minimum.
Penguin Books: London, UK, 2015.
Quantum Preprocessing
7 for Deep Convolutional
Neural Networks in
Atherosclerosis Detection
7.1 INTRODUCTION
Atherosclerosis is a specific type of stenosis, i.e., narrowing or occlusion of the artery
lumen, that occurs due to the accumulation of some substances as cholesterol on
the coronary arteries’ inner walls. An opportune diagnosis and atherosclerosis treat-
ments are essential since it represents the leading cause of Coronary Artery Dis-
ease. According to the World Health Organization (WHO), this heart condition has
a high mortality rate worldwide, with 17.9 million estimated deaths every year [1].
Atherosclerosis detection consists of arteries visual inspection through a screening
test, either non-invasively employing computed tomography (Coronary Computed
Tomography Angiography, CCTA) or with the regular procedure consisting of insert-
ing a catheter through the groin or arm into the coronary arteries (X-ray Coronary
Angiography, XCA). In both cases, a contrast medium is injected to guide or locate
the arteries. Nonetheless, XCA remains the gold standard used by specialists since
it offers enough resolution for diagnosis. Besides, the patients can receive treatment
during the same session. For instance, during the visual examination of the XCA
images, the physician may identify regions with stenosis as shown in Figure 7.1.
Figure 7.1: X-ray coronary angiography image. Stenosis and non-stenosis samples
regions marked in green and red, respectively.
DOI: 10.1201/9781003283294-7 141

However, given the limited access of specialists and the time consumption for
diagnoses, it has allowed the Computer-Aided Diagnosis (CAD) systems to play
a vital role in cardiology. CAD systems have been a significant field of research
during the last few decades, developed to improve and support the medical diagnosis
process. CAD uses Machine Learning (ML) methods to analyze imaging or non-
imaging (i.e., clinical profile) patient data to assess patients’ conditions.
In computer vision, one of the breakthroughs occurred when Krizhevsky et al.
[2] presented Alexnet, a Deep Convolutional Neural Network (DCNN) that won the
2012 ImageNet Large Scale Visual Recognition Challenge (ILSVRC) [3]. New Deep
Learning (DL) algorithms have been proposed to adapt CNN architectures to chal-
lenging medical imaging problems. Despite significant research problems on the
medical imaging domain using DL algorithms, it often suffers two significant dif-
ficulties in practice: 1) limited amount of labeled data; 2) mislabeling labels. New
hybrid models have emerged to build more efficient ML algorithms. The most rele-
vant is Quantum Machine Learning (QML) [4], which relies on quantum computing
emulated from physical properties. Such improvement can be accomplished using
quantum mechanical systems.
In this work, a Hybrid Quantum-Convolutional Neural Network for atheroscle-
rosis detection in XCA images is presented. This approach considers an automatic
preprocessing step, implemented through a Quantum Convolutional Layer (QCL),
whose behavior corresponds to apply a quantum circuits as an image filter. The pre-
possessing quantum layer receives as input an XCA patch of one single-channel.
Moreover, it generates a multichannel image, which feeds a classical CNN to perform
the atherosclerosis detection. The presented hybrid method showed that employing
a QCL as an XCA image preprocessing improves the network performance against
raw-XCA images, usually feeding a CNN. Five different evaluation metrics were
used to measure the performance of the proposed method. Besides, two different
optimization techniques were compared: Stochastic Gradient Descent and Stochas-
tic Gradient Descent with Momentum. Additionally, two different CNN architec-
tures previously introduced for atherosclerosis detection were studied. The employed
dataset consists of 250 real XCA images, where 125 images are used for training and
the remaining 125 for testing.
The remainder of this document is organized as follows: Firstly, in Section 7.2,
the related work is addressed. Section 7.3 describes briefly the concepts related to
Quantum computation and CNNs. Section 7.4 presents the introduced methodology
for atherosclerosis detection. In Section 7.5, the experimental and numerical results
are carefully detailed and discussed. Finally, the conclusions are stated in Section 7.6.
7.2 RELATED WORK

Some methods for the automatic detection of atherosclerosis have been proposed in
the literature. Most of them have based their strategies on blood vessel extraction.
Sui et al. [5] proposed performing vessel segmentation through a method including
a multi-scale Hessian matrix-based enhancement and a dual-stage region growing
method. The segmentation was used as input of an arterial tree’s skeleton algorithm
Quantum Preprocessing for DCNN in Atherosclerosis Detection 143
to compute the arterial diameter of each skeleton point. Finally, the detection of
general or severe obstruction was assigned based on the ratio given by the segment
minimum diameter over its average diameter for each segment.
Wu et al. [6] used a U-Net architecture to segment the vascular structure of the
XCA sequence taking advantage of the binary output to calculate a contrast-filling
degree for each frame. Next, a percentage of sequences passing to the detection step
was selected, where a Deconvolutional Single-Shot Multibox Detector may generate
the location information. Jevitha et al. [7] combined a contour activation algorithm
with a Frangi filter to extract the left main coronary artery. Next, bifurcation points
were automatically detected using a template kernel. Later, measurements were taken
on the artery bifurcation angles. The result was used to discriminate between stenotic
and healthy arteries. These methods building the atherosclerosis detection and clas-
sification ruled around blood vessel segmentation try to emulate physicians’ proce-
dures when diagnosing patients. This last approach has the advantage that the work-
flow often has a direct interpretation. Regardless, the main task results (atheroscle-
rosis detection) are conditioned by the segmentation algorithm performance, which,
in many cases, struggles to detect narrowed regions.
Recently, proposals have emerged that do not consider vessels’ extraction as a
necessary step. Instead, end-to-end systems have been built, taking advantage of the
rapid increase in computational power and the superior performance that algorithms
based on deep learning have shown.
Au et al. [8] introduced a patch-based CNN that automatically characterizes
and analyzes coronary stenosis in such a context. The network was based on a
DenseNet [9], where skip connections between convolutional layers were added.
Likewise, Antczak and Liberadzki [10] developed a patch-based CNN based on a
VGG architecture [11], adding dropout layers after the convolution’s layers and re-
moving the pooling layers. This approach employed an artificial dataset to overcome
the problem of a limited amount of training data. Both networks were trained from
scratch using a Sigmoid activation function in the last dense layer to compute the
patch probability of belonging to the stenosis class.
Cong et al. [12] presented another interesting solution. The InceptionV3 [13] net-
work was trained from scratch to select a subset of candidate frames (like what was
done in [6]), considering the contrast filling degree and other image quality measures,
such as well-defined vessel borders. A second step used Transfer Learning for steno-
sis detection with an InceptionV3 network pre-trained with the ImageNet database.
Fine-tuning was performed using a strategy called redundancy training that included
pre-classified redundant frames in the training dataset.
Ovalle et al. [14] presented a method for successfully detecting coronary artery
stenosis in XCA images, evaluating three pre-trained state-of-the-art architectures
(VGG16 [11], ResNet50 [15], and InceptionV3 [13]) via Transfer Learning. Such
a method incorporates a network-cut approach where only a sub-set of layers was
considered. Layers between the cut layer and custom classifier are discarded. During
the fine-tuning step, an artificial dataset was exploited. Moreover, the fine-tuning
was substantially improved using a sub-set of real XCA images. This approach
outperformed the results obtained using only the fine-tuning process during Transfer
Learning.
Figure 7.2 shows an overview of the previously published methods evaluated us-
ing XCA images. End-to-end methods have shown worthy potential for detecting
Figure 7.2: Typical workflow of the previously published algorithms for coronary
artery stenosis detection in XCA images.
and classifying atherosclerosis using independent vascular structure extraction tech-

niques. Besides, innovative visualization techniques have been efficiently proposed
to interpret and justify the decision rules.
However, one of the most significant limitations of the previous approach is that
a long training time is required to get accurate results. The number of model pa-
rameters stands in the order of thousands to millions. New hybrid models have
then emerged to build more efficient machine learning algorithms. The Quantum
Convolutional Layer or Quanvolutional Layer is an outstanding example of this ap-
proach [16]. It improves the performance of the CNN through quantum circuits rep-
resenting classical filters. Quantum circuits can transform selected local subsections
of the input data to extract relevant features, like traditional convolutional transfor-
mations whose filters are convolved into every local subsection of the input. Sleeman
et al. [17] introduced a hybrid method combining a classical DCNN autoencoder
with a quantum annealing Restricted Boltzmann Machine (RBM), applied to image
generation. This hybrid autoencoder approach appeared an advantage for RBM re-
sults relative to using a classical RBM implementation into the MNIST (Modified
National Institute of Standards and Technology) and MNIST Fashion dataset classi-
fication problem.
Furthermore, some relevant classification works related to the biomedical images
domain have emerged. Iyer et al. [18] used a modified version of the Variational
Classifier previously introduced by Bergholm et al. in [19] for the binary clas-
sification of pigmented skin lesions. Images’ dimensionality is transformed from
128 × 128 × 3 to a 4-dimensional vector via an autoencoder cleverly being imple-
mented with a CNN. Next, each vector was encoded in quantum amplitudes and sent
to the 2-qubit classifier to obtain the predicted expected values of the image, being
either a melanoma or melanocytic nevi.
On the other hand, Bisarya et al. [20] proposed a quantum convolutional neural
network (QCNN) for cancer detection in breast cell data by exploiting the Wisconsin
dataset. Two cases were theoretically considered: the former uses numerical data
associated with the size of the cells and their texture, which were processed using
an architecture of two QCLs with only ten parameters, each one assigned to a qubit;
the latter uses gray-scale magnetic resonance images, processed in sets of 4 × 4-
pixel arrays, encoded in a 256-qubits system. Regardless, only the first approach was
implemented due to the limited availability of resources.
It is noteworthy that according to the authors’ best knowledge, there are currently
no works that use quantum convolutional networks to detect atherosclerosis or other
types of stenosis. Therefore, in this study, a Hybrid Quantum-Convolutional Neural
Network for atherosclerosis detection in XCA images, employing a QCL, is pre-
sented. The QCL successfully performs a preprocessing over the XCA images to
generate a multichannel image feeding a classical CNN and improving the detection
performance against the raw-XCA images.
7.3 MATHEMATICAL FOUNDATIONS

A summary of essential and associated topics of Quantum Computation and CNNs
related to the proposed study is presented in the following subsections. The reader
may refer to Nielsen and Chuang [21] and De Wolf [22] for in-depth details about
quantum computation.
7.3.1 QUANTUM COMPUTING

The most straightforward computational unit is the bit, which can hold two different
states, denoted as 0 and 1. On the other hand, the qubit is the basic unit in Quantum
Computation [21]. Like a classical bit, a qubit also has two states, the ground state |0⟩
and the excited state |1⟩ (both written in Dirac notation). In contrast to classical bits
whose value is a scalar, that is, zero or one, a qubit can be an arbitrary superposition
of two basic states |0⟩ and |1⟩. Such a superposition might be interpreted as a linear
combination of two basic vectors, weighted by complex amplitudes. This property is
commonly known as superposition. Hence, these states can be manipulated through
quantum circuits. Finally, a well-defined value can be obtained by the measurement
of some (or all) of the qubits’ output by the quantum circuit.
7.3.1.1 Qubit States

The qubit states are stochastic in nature due to the possibility of holding the super-
position states. The two basic states are given as

1 0
|0⟩ = , |1⟩ = . (7.1)
0 1
The superposition state |ψ ⟩ is a linear combination of the basic states denoted by
|ψ ⟩ = α0 |0⟩ + α1 |1⟩ , (7.2)
where α0 and α1 are two complex numbers fulfilling the fundamental property given
by |α0 |2 + |α1 |2 = 1. Thus, each amplitude αi can be represented with the unit imag-
inary circle using the polar representation for a complex number. It can be shown
that |ψ ⟩ can be written as

ϑ ϑ
|ψ ⟩ = cos |0⟩ + eiφ sin |1⟩ , (7.3)
2 2
where ϑ and φ correspond to two angles in spherical coordinates. In such a way, a
quantum state is visualized as a unit vector in the 3-D space (on the Bloch sphere).
The two basic states |0⟩ and |1⟩ are then represented by ϑ = 0 and ϑ = ϕ , pointing
in z and −z directions, respectively, as illustrated in Figure 7.3.
(a) Ground state |0⟩. (b) Excited state |1⟩.
Figure 7.3: Illustration of the two basic states on the Bloch Sphere. The graph was
plotted using the Johansson et al. library [23].
Next, a couple of examples are included to clarify some basic properties of quan-
tum computing.
Example 1. Consider a 2-qubit system whose state is described by the state vector
√1 [1, 0, 0, 1]⊺ = √1 |0⟩ + √1 |3⟩ = √1 |(00)2 ⟩ + √1 |(11)2 ⟩. This representation is a
2 2 2 2 2
valid quantum state since | √12 |2 + | √12 |2 = 1, where (X)2 is a binary representation
reading from left to right of the vector state positions.
A quantum register or a quantum system of n-qubits can be any linear combination

such that
|ψ ⟩n = α0 |0⟩ + α1 |1⟩ + · · · +α2n−1 |2n−1⟩ ,
2n−1 (7.4)
∑ |αi |2 = 1.
i=0
The entanglement is another essential property (unique to quantum computing). It

means that a system of n-qubits cannot be written as the tensor product of two sub-
systems, e.g., a bipartite pure state containing the subsystems A and B, is entangled
if it cannot be expressed as
|ψ ⟩AB = |φ ⟩A ⊗ |ϕ ⟩B , (7.5)
where (⊗) is the tensor product operator.
Example 2. For a 2-qubit system, whose state is given by

1 1
√ (|(00)2 ⟩ + |(01)2 ⟩) = |0⟩ ⊗ √ (|0⟩ + |1⟩)
2 2
is separable since it can be represented as the tensor product of two subsystems
|φ ⟩A = |0⟩ and |ϕ ⟩B = √12 (|0⟩ + |1⟩). On the other hand, the system
1
√ (|(00)2 ⟩ + |(11)2 ⟩) ̸= |φ ⟩A ⊗ |ϕ ⟩B ,
2
is entangled. Such a 2-qubit system is also known as an EPR-pair [24].
7.3.1.2 Qubit Operations

Quantum operations (or quantum gates) can manipulate the state of a quantum sys-
tem such that
ψ ′ = U |ψ ⟩ , (7.6)
where U ∈ CN×N . Hence, it must be a unitary transformation given as
U∗ U = UU∗ = I, (7.7)
where U∗ is the conjugate transpose of U. A unitary operator acting on a small num-

ber of qubits (say, at most 3) is often called a gate, in analogy to the classical logic
gates like AND, OR, and NOT. The simplest single-qubit gates consist of only two
well-known gates: the Hadamard H gate and the T gate (or π8 gate). The Hadamard
gate allows taking a qubit from a definite computational basis state into a two states’
superposition. The Hadamard matrix H is defined by

1 1 1
H= √ . (7.8)
2 1 −1
Meanwhile, the T-gate performs a rotation π4 around the z-axis and is defined as
follows
1 0
T= π . (7.9)
0 e4i
These two operations can be composed to approximate unitary transformations
on a single qubit, such as the Pauli-X, Y, and Z gates (σx , σy , and σz ) used to rotate
the superposition along x-, y-, or z-axis. The Pauli gates are defined by

0 1
σx = = HT4 H, (7.10)
1 0

0 −i
σy = = T2 HT4 HT6 , (7.11)
i 0

1 0
σz = = T4 . (7.12)
0 −1
The Pauli matrices are involutory, that is, a matrix that is its own inverse, such that

1 0
σx = σy = σz = −iσx σy σz =
2 2 2
= I. (7.13)
0 1
It can be shown that by a given angle γ and a Pauli matrix σa , a = [x, y, z]
exp(i γσa ) = cos(γ )I + i sin(γ )σa . (7.14)
Therefore, the rotations operator, which rotate the unit Bloch vector by an angle γ
around a specific axis, is given by
γ
γ γ
Rx (γ ) = e−i 2 σx = cos I − i sin σx , (7.15)
2 2
γ γ γ
Ry (γ ) = e−i 2 σy = cos I − i sin σy , (7.16)
2 2
γ γ γ
Rz (γ ) = e−i 2 σz = cos I − i sin σz . (7.17)
2 2
Moreover, the rotation operators can be expanded as

cos (γ /2) −i sin (γ /2)
Rx (γ ) = , (7.18)
−i sin (γ /2) cos (γ /2)

cos (γ /2) − sin (γ /2)
Ry (γ ) = , (7.19)
sin (γ /2) cos (γ /2)
" γ #
e− 2 i 0
Rz (γ ) = γ . (7.20)
0 e2i
A sequence of quantum operations applied to the state of a given system is de-

scribed by quantum circuits. Quantum circuits can be structured or random. In the
first case, a structured quantum circuit is a sequence of well-defined quantum gates.
On the other hand, the set of quantum gates are randomly selected. A circuit has
a depth d equals the total number of layers conforming to the circuit. A layer is a
circuit whose gates act on disjoint qubits, that is, a layer has a depth of one.
Example 3. For instance, let us consider the following quantum circuit that rotates
a qubit in the x-axis and afterward around the y-axis. First, a qubit in the ground
state |0⟩ = [1, 0]⊺ is rotated through the angle γ1 around the x-axis applying the gate
Rx (γ1 ), and eventually, the y-axis by the angle γ2 via the gate Ry (γ2 ). After these
operations, the qubit is now in the state
|ψ ⟩ = Ry (γ2 )Rx (γ1 ) |0⟩ .
7.3.1.3 Qubit Measurements

A measurement must extract the information of a qubit, yielding a numerical result.
If a quantum state |ψ ⟩ is measured, then only one classical state | j⟩ can be seen
with probability |α j |2 , which corresponds to the amplitude α j . This is known as
Born’s rule. Accordingly, observing a quantum state induces a measurement in the
computational basis state probabilities. However, other kinds of measurements can
be computed from quantum systems, such as the expectation of an observable and
its variance. Observables are Hermitian matrix M such that their eigenvalues are
real numbers. The Pauli gates (σx , σy , and σz ) are examples of observables, with
eigenvalues ±1. Ergo, the expected value of an observable M on a state |ψ ⟩ is
⟨M⟩ψ = ⟨ψ |M|ψ ⟩ . (7.21)
Example 4. Let us consider the state |ψ ⟩ = Ry (γ2 )Rx (γ1 ) |0⟩, applying a mea-
surement over the Pauli-Z observable, the expectation value is given by ⟨σz ⟩ψ =
⟨ψ |σz |ψ ⟩ . Depending on the circuit parameters γ1 and γ2 , the output is yielded in
the range [−1, 1]. For instance, if γ1 = 0.5 and γ2 = 1.0 then ⟨σz ⟩ψ = 0.47416.
7.3.2 CONVOLUTIONAL NEURAL NETWORKS

Convolutional Neural Networks (CNNs) consist of a finite number of convolutional
layers that can accurately extract intricate visual features such as edges, corners, in-
terest points, among others, through a set of input images [2]. It can be applied to
many practical problems such as detection, recognition, and classification. CNNs are
a specific type of neural networks composed of the following layers: convolution
layer, pooling layer, and fully connected layer. The convolution layer and the pool-
ing layer can be fine-tuned concerning the hyper-parameters described in the next
sections. Figure 7.4 illustrates the typical architecture of a CNN.
Figure 7.4: Typical architecture of a Convolutional Neural Network. The design com-
prises one convolutional layer with a kernel size of 3 × 3, followed by a pooling layer
with a window size of 2 × 2. Finally, a fully connected layer is involved into the
model used to optimize the objective functions.
7.3.2.1 Convolutional Layer

The convolutional layer uses filters to perform a set of convolution operations across
the input image, concerning diverse filter dimensions. These layers can be viewed as
2-D kernels (matrices) where their hyper-parameters include the kernel size F and
stride S. The stride S denotes the number of pixels by which the kernel moves after
each operation. Hence, padding is introduced to enlarge the input such that the output
has the same height/width dimension as the input. The resulting outcome O is called
a feature map or activation map. Thus, a convolutional layer consists of K kernels
trained to detect local features. Each feature map Ok is defined as follows:

Ok = f bk + ∑ wk [c] ∗ I[c] , (7.22)
c
where f is an activation function (e.g., Sigmoid), bk is the k-bias, I[c] is the image at
the c-channel, wk [c] is the k-kernel for such a channel, and ∗ denotes the convolution
operator.
Example 5. Consider a 5 × 5 input image represented as a matrix; also consider

another 3 × 3 matrix representing a kernel. Here, the kernel will convolve over each
pixel position of the input image to finally get a feature map, as shown in Figure 7.5.
In this convolution layer, the hyper-parameters F = 3 × 3, S = 1, and zero padding
were used.
7.3.2.2 Pooling Layer

The pooling layer, also called the subsampling or downsampling layer, typically ap-
plied after a convolutional layer, reduces each feature map’s spatial size but retains
the most valuable information. Pooling layers reduce the number of parameters and
computations in the network and control overfitting [25]. Diverse types of pooling
0 0 0 0 0 0 0 1 1 -1
0 0 1 1 1 0 0 0 1 0 0 2 3 2 0
0 0 0 1 1 1 0 0 -1 -1 0 -1 0 1 0
0 0 0 0 1 1 0 0 0 0 1 -1
Kernel
0 0 0 0 1 1 0
0 2 3 1 0
0 0 0 2 1 2 0
0 0 2 0 0
0 0 0 0 0 0 0
Input Output
Figure 7.5: Convolution operation configured with zero-padding, a stride of 1 pixel,

and a kernel size of 3 × 3.
can be applied, the max and average pooling are the most used. In max-pooling, the
maximum value is taken from a certain spatial neighborhood (for example, a 2 × 2
window). Instead of taking the largest element, the average-pooling selects the aver-
age value of the neighborhood.
Example 6. Given a feature map of size 5 × 5 produced by a convolutional layer, a

pooling layer is defined with a spatial neighborhood of 2 × 2. The pooling procedure
is illustrated in Figure 7.6, where the max and average pooling are shown in detail.
2 3 0
2 3 0
0 2 3 2 0 0 2 0
0 -1 0 1 0
Max-Pooling
0 0 0 1 -1
0 2 3 1 0
0 0 2 0 0 0.25 1.5 0
Input 0.5 1.25 -0.25
0 0.5 0
Average-Pooling
Figure 7.6: Pooling operations with a neighborhood of 2 × 2 and a stride of 2.
7.3.2.3 Fully Connected Layer

The fully connected layer takes the result of the convolution/pooling process to be
employed for classification. The previous convolutional or pooling layer output is
flattened into a single vector, each indexed value is representing a probability that
a specific feature belongs to a label. Fully connected layers are usually found at
the end of the CNN architectures; each neuron in the past layer is associated with
each neuron on the next layer. In summary, the fully connected step consists of three
layers:
Fully connected input layer: It makes the last features map into a single one-
dimensional vector, commonly used in the transition from the last convolu-
tional/pooling layer to the fully connected layer.
First fully connected layer: A set of weights are learned to predict the correct label
from one-dimensional feature representation.
Fully connected output layer: It retrieves the final probabilities for each class.
7.3.2.4 Activation Functions

Activation functions are mathematical gates that determine the neuron’s output trans-
mitted to the next layer. Some activation functions normalize each neuron output into
a range of [1, 0] or [−1, 1]. Modern neural networks use non-linear activation func-
tions to learn a complex representation between the network’s inputs and outputs,
as shown in Figure 7.7. The suitable choice of the activation function depends on
𝑥!
𝑤!
𝑥"
∑ 𝑓 𝑓 𝑏 + & 𝑥! 𝑤!
. 𝑤"
!
.
.
𝑥# 𝑤#
Inputs Neuron Output
Figure 7.7: The basic process of an activation function. The sum of all weighted
inputs (w⊺ x) and the bias b is passed through a non-linear activation function f to
generate the neuron output.
the nature of the problem to solve. The most widely used activation functions are
the Sigmoid, Hyperbolic Tangent, Rectified Linear Unit (ReLU), Leaky ReLU, and
SoftMax. The Sigmoid function returns a value close to zero for small values in the
argument and close to 1 for large argument values,
1
Sigmoid(x) = . (7.23)
1 + e−x
As an alternative to the Sigmoid function, the Hyperbolic Tangent function could
be used as an activation function producing outputs in the range of [−1, 1] and is
formally defined as
ex − e−x
tanh(x) = . (7.24)
ex + e−x
The Rectified Linear Unit (ReLU) activation function returns the element-wise max-
imum between 0 and the input value as follows:
ReLU(x) = max (0, x). (7.25)
Leaky ReLU is a variation of the ReLU function, which has a small positive value
when the input is not active
Leaky-ReLU(x) = max (δ x, x). (7.26)
Lastly, the SoftMax function converts a real input vector into a vector of probabilities.
Therefore, the elements of the output vector must sum up to 1. The SoftMax function
applied on the vector x is computed as
ex
SoftMax(x) = N
. (7.27)
∑ exi
i=1
7.4 PROPOSED METHOD

In the presented work, a Quantum Convolutional Layer (QCL) was employed as
a new type of image preprocessing method. From a given XCA image, the QCL
generates a multichannel image resembling the feature maps generated by a typical
Convolutional Layer. Next, the outcome of the QCL feeds a traditional CNN focused
on detecting atherosclerosis. A general framework of the proposed method is shown
in Figure 7.8 and discussed in this section.
Figure 7.8: Outline of the proposed methodology. The input XCA patches pass
through a QCL to generate a 9-channel image to feed the traditional CNN to de-
tect atherosclerosis.
7.4.1 QUANTUM CONVOLUTIONAL LAYER

In traditional image processing methods, there are several techniques to enhance the
image quality to accelerate and improve the training process in machine learning
algorithms. Moreover, new promising hybrid DL methods have emerged. The QCL
is an example of this approach. A QCL produces an advanced set of feature maps
when applied to an input image. However, unlike the classical convolution operation
where each filter produces a feature map, a quantum convolution employs a quantum
circuit as a filter, and it can generate from each circuit multiple feature maps. A
QCL requires some hyper-parameters to be tuned: the number of filters and their
respective input size. Besides, for each filter, an encoding function is necessary to
initialize the quantum state. Finally, the feature maps are generated from a given
decoding function (an observable). It is noteworthy that the filter size defines the
number of qubits required to operate the quantum circuit. In this study, a single 3 × 3
filter size was chosen. Therefore, the circuits dispose of nine qubits and compute nine
feature maps. The proposed methodology is described in the following procedure and
illustrated in Figure 7.10.
Figure 7.9: 9-qubit Random Quantum Circuit employed in the Quantum Convolu-
tional Layer. The circuit has the Pauli-X, Pauli-Y, and Pauli-Z rotations randomly
distributed and controlled-CNOT gates as imprimitives (2-qubit entangled gates).
1. Ground state initialization: The input image sub-regions u were obtained by a

sliding window of 3 × 3 pixels and a stride of 1 pixel to initialize the quantum
ground state as |xn ⟩ = Ry (un ). Notice that the ground state is parameterized as
Figure 7.10: Design of the Quantum Convolutional Layer consisting of three main
stages: ground-state initialization (encoding), Random Quantum Circuit (RQC) defi-
nition and output state computation, and the measurement (decoding). From a sliding
window of 3 × 3, the QCL generates nine expectation values mapped into nine dif-
ferent channels from data.
rotation over the y-axis as encoding function, with an angle γ = un , where un is

the intensity of the pixel n in the image neighborhood u.
2. Random quantum circuit definition: A random quantum circuit, associated with a
unitary U, is generated. Hence, the circuit comprises layers of randomly chosen
single-qubit rotations and 2-qubit entangling gates. The 2-qubit gates and the ro-
tations are randomly distributed in the circuit. Specifically, a 9-qubit circuit with
two layers is generated, as shown in Figure 7.9.
3. Output state computation: Each image neighborhood u is processed by the quan-
tum circuit, whereby |x′ ⟩ = U |x⟩
4. Measurement: The system is measure using a Pauli-Z gate, observable over the
respective output state, such that ⟨σz ⟩x′ = ⟨x′ |σz |x′ ⟩. Each nine expectation values
are mapped into nine channels of a single output pixel.
7.4.2 NETWORK ARCHITECTURE

Two CNN architectures were tested to accurately detect atherosclerosis. For each
CNN, the detection performance was carried out using the images generated by
the QCL and the original XCA images. Table 7.1 presents a detailed description
of these two architectures. The first network introduced by Au et al. [8] is based
on a DenseNet architecture [9]. On the other hand, the second network proposed by
Antczak and Liberadzki [10] is based on a VGG architecture [11] but changing the
kernel size and depth as well as removing the pooling layers.
The CNN weights were trained from scratch, with a random initialization using
the Glorot (also known as Xavier) algorithm [26]. The fundamental idea is to ini-
tialize each layer weight with a small Gaussian value characterized by a zero mean
and variance based on the number of columns (n) in the previous layer wk−1 . In
consequence, the weights are scaled by the inverse of the square root of n as follows:

1 1
w ∼ N −√ , √
k
. (7.28)
n n
Table 7.1
CNN Architectures for Atherosclerosis Detection
Network Architecture Layer Description
64-Conv(3 × 3) Convolutional layer with 64 kernels, each with a size of 3 × 3
pixels
pixels
pixels
I) Au et al. [8] Conc(1, 3) Concatenation of the outputs of the first and third convolu-
tional layer
pixels
pixels
Conc(1, 6) Concatenation of the outputs of the first and sixth convolu-
tional layer
GMP Global Max Pooling layer
1-Dense Dense layer with one neuron and a Sigmoid activation func-
tion
8-Conv(7 × 7) Convolutional layer with eight kernels, each with a size of
7 × 7 pixels
Dropout(0.5) Dropout layer with a rate of 0.5
7 × 7 pixels
II) Antczak and 8-Conv(7 × 7) Convolutional layer with eight kernels, each with a size of
Liberadzki [10] 7 × 7 pixels
7 × 7 pixels
16-Dense Dense layer with 16 neurons and a Sigmoid activation func-
tion
1-Dense Dense layer with one neuron and a Sigmoid activation func-
tion
Let yn be the class indicator variable for the n-th input patch, which operates as
(
0, if non-atherosclerosis,
yn = (7.29)
1, if atherosclerosis.
Furthermore, ŷ corresponds to the estimated probability by the CNN that the input
patch has atherosclerosis. Lastly, the optimization process is achieved by minimizing
the binary cross-entropy loss function defined as
1 N
min J(y, ŷ) = −
w
∑ [yn log(ŷn )+(1−yn )log(1−ŷn )] ,
N n=1
(7.30)
where N represents the batch-size used during the minimization. For artherosclerosis
detection, the loss function can be rewritten as
(
log(ŷn ), if yn = 1,
J(yn , ŷn ) = − (7.31)
log(1 − ŷn ), if yn = 0,
where y = 1 means that the class C1 = Ci is a positive case of atherosclerosis. The

gradient computed with respect to each predicted label ŷn will only depend on the
loss given by its binary problem.
The objective function from (7.31) is minimized through the training step by com-
puting its gradient regarding the current weights’ values. In most neural networks,
the Stochastic Gradient Descent (SGD) optimization [27] is employed to find an opti-
mal solution. SGD computes the gradient of the parameters using only a few training
samples (batch size). For SGD optimization, the gradient is updated by
wk+1 = wk − α ∇wk J(y, ŷ), (7.32)
where α is the learning rate. SGD can lead to slow convergence, particularly after the
initial steep gains. Some methods have been incorporated to overcome such incon-
venience. Thus, the momentum is a method that integrates the past gradients in each
dimension. In SGD with momentum [28] (SGDM), the gradient at every dimension
is incorporated to gain velocity where the parameters have a consistent gradient. The
momentum update is given by
vk+1 = η vk + α ∇wk J(y, ŷ),

(7.33)
wk+1 =wk − vk+1
where v is the velocity term, starting with the initial value zero, and η is the momen-
tum coefficient that controls the fraction of the previous velocity term to be consid-
ered in the current update step. The evaluated networks were optimized using SGD
and SGDM with a learning rate α = 0.01 and momentum η = 0.8, respectively, with
a batch size N = 100 during 1, 000 epochs.
7.4.3 EVALUATION METRICS

A set of binary classification metrics were considered to evaluate the network’s per-
formance in the detection of atherosclerosis. These metrics include the Accuracy,
Sensitivity, Specificity, Precision, and F-Measure.
Accuracy refers to the ratio of correct predictions over the number of instances
evaluated,
t p + tn
Accuracy = , (7.34)
t p + f p + tn + f n
where t p refers to the true positive cases of atherosclerosis detection, tn refers to

the true negative cases, f p represents the number of instances incorrectly classified
as positive (false positives), and fn represents the number of instances incorrectly
classified as negative (false negatives).
Sensitivity measures the fraction of positive instances correctly classified, which
is given by
tp
Sensitivity = . (7.35)
t p + fn
Specificity measures the proportion of negatives samples correctly classified as
follows
tn
Specificity = . (7.36)
tn + f p
Precision is used to measure the positive patterns correctly classified from the
total predicted patterns in a positive class. It is calculated as
tp
Precision = . (7.37)
tp + f p
F-Measure represents the harmonic mean between recall (sensitivity) and preci-
sion values. A general F-Measure is described as follows
Precision × Recall
Fβ -Measure = (1 + β 2 ) × . (7.38)
β 2 × Precision + Recall
In particular, if β = 1, a balanced Fβ -Measure is obtained; consequently, such ex-

pression is acknowledged as the F1 -Measure.
7.5 RESULTS AND DISCUSSIONS

The atherosclerosis detection was evaluated using two different approaches to train
CNN. First, a quantum convolutional layer performs a preprocessing over the XCA
images. On the other hand, CNN is trained using the original XCA images. The
detection results for each strategy are presented. Moreover, a comparative analysis
of the proposed approach using SGD and SGDM is included in this section. The
analysis concerned the CNN architectures introduced by Au et al. [8] and Antczak
and Liberadzki [10].
For the computational experiments, Python 3.6, Keras 2.3.1, TensorFlow 2.2.0,
and the PennyLane library [19] were used and tested on a Cloud Platform, including
CPU and GPU processors. The CPU included an Intel (R) Xeon (R), 12 GB of RAM,
and a 2.00 GHz dual processor. The GPU was powered by a Tesla P4 consisting of
2560 CUDA cores and 8 GB VRAM.
7.5.1 DATASET OF CORONARY STENOSIS

Antczak and Liberadzki [10] introduced a dataset for stenosis detection, publicly
available on the authors GitHub site [29]. The dataset consists of 250 patches of
size 32 × 32 pixels obtained from XCA images. In numerical experiments, the pixel
values were normalized into [0, 1].
The patches were obtained, as seen in Figure 7.11, following the next steps:
1. Input images are downsampled from 256 × 256 to 128 × 128 pixels.
2. A sliding window of size 32 × 32 was moved to obtain a set of patches from each
image.
3. Each patch is labeled as non-stenosis, stenosis, respectively.
The sliding window produces multiple overlapping matches; thus, multiple patches
may be classified as positive cases, even if there is only one stenosis on the image.
Input XCA Output patches
Downsampling Windowing
Figure 7.11: General outline of XCA patches generation. First, the XCA image is
sub-sampled; next, a sliding window generates the output patches.
7.5.2 QUANTUM PREPROCESSING

The quantum convolutional layer (QCL) acts as an image preprocessing. Given an
XCA patch of 32 × 32 for one channel, the QCL outcome is a patch of 32 × 32 with
nine channels. The input pixel values are in the range of [0, 1], and the output is in
the range [−1, 1].
Figure 7.12 shows the effect of the quantum convolution layer on a batch of sam-
ples from XCA images. The QCL results suggest that this layer generates different
images where the contrast is changing in each channel. In other cases, in the fifth
channel (sixth row), the QCL outcome is similar to extract the image texture (e.g.,
Local Binary Pattern [30]).
Figure 7.12: Feature maps generated by the Quantum Convolutional Layer (QCL).
First row: raw XCA images. The following nine rows show the generated images by
the QCL.
7.5.3 TRAINING RESULTS

A k-Fold cross-validation procedure was carried out during the numerical exper-
iments. In k-Fold cross-validation, a parameter k decides in how many folds the
dataset is divided. Thus, each fold gets the chance to appear in the training set k − 1
times. The number of folds chosen was k = 5, as illustrated in Figure 7.13.
The dataset was divided into two subsets: training and testing of 125 images each.
The testing subset was excluded during the cross-validation process. In this approach,
four folds (75% of the training subset) were used to train the CNN, and the remain-
ing fold (25% of the training subset) was saved for validation. This process was
executed five times, selecting the best model for validation loss in each iteration. The
best model was updated if the validation loss is improved over the epochs; this avoids
overfitting. Finally, each iteration’s atherosclerosis detection performance was evalu-

ated using the excluded initial subset (testing). Furthermore, the mean and maximum
coefficients were obtained, as well as the standard deviation for these coefficients.
Dataset (250 images)
Training set (125 images) Test set (125 images)
Fold 1 Fold 2 Fold 3 Fold 4 Fold 5
Figure 7.13: k-Fold cross-validation procedure. The dataset was split into training
and testing sets. Next, the training set is divided into folds to train CNN. The fold
marked as gray was taken as validation, while the folds in light green as training data.
The learning curve can measure the model’s performance during the training and
validation steps in terms of accuracy and loss. In Figure 7.14, losses and accuracy
curves are plotted concerning the number of epochs used to train the model.
Figure 7.15 shows that the accuracy (using SGDM) increases while the loss de-
creases. The Q-CNN-A (CNN proposed by Au et al. [8] with the quantum convolu-
tional layer) achieves the best accuracy in training and validation. Those results are
followed by the Q-CNN-B (CNN proposed by Antczak and Liberadzki [10] using the
quantum convolutional layer). In terms of validation loss, the models start to overfit
during the training process.
On the other hand, the accuracy curves in Figure 7.16 also shown that the CNN
trained with the outcome of the QCL reach the highest accuracy (in particular, Q-
CNN-A) by using SGD as the optimizer. This same network architecture but using
only the raw (normalized) XCA for training, followed in accuracy performance.
In contrast to the loss curves obtained by SGDM, the curves in Figure 7.17 suggest
that the validation loss decreases and has a small gap with the training loss during
the training process.
7.5.4 DETECTION RESULTS

In this research, the approach analysis was carried out using the generated XCA
patches by the QCL and the original (normalized) XCA patches. Additionally, the
Figure 7.14: Training and validation accuracy curves of the CNN models employing
SGD with Momentum concerning the number of epochs (1000) in the training phase.
Top: Training accuracy curves, Bottom: Validation accuracy curves.
detection results were compared using SGD and the innovative variant using mo-
mentum (SGDM). Table 7.2 presents the performance rates of atherosclerosis detec-
tion for each CNN studied. It shows that the use of momentum provides improved
performance for both CNN architectures. The best performance for each metric was
highlighted.
Furthermore, the network performance surpassed the raw-XCA patches version
when the network is trained using the QCL preprocessing, reaching an accuracy of
83.36% against 79.84% for the architecture employed by Au et al. [8].
For the second architecture (Antczak and Liberadzki [10]), the accuracy is im-
proved from 69.44% to 80.80% when the network is fed with the QCL output im-
ages. In general, an improvement for precision, sensitivity, specificity, and F1 score
was achieved using the QCL.
Figure 7.15: Training and validation loss curves of the CNN models employing SGD
with Momentum concerning the number of epochs (1000) in the training phase. Top:
Training loss curves, Bottom: Validation loss curves.
Figure 7.18 shows a prediction sample: true-positive, true-negative, false-positive,

and false-negative cases. Beneath each case, the generated multichannel image by
the QCL was presented. Therefore, in the case of correct detection (true-positive and
true-negative), the vessel pixels were enhanced in each channel of the QCL image
outcome.
Meanwhile, misclassification cases were also found. Thus, in the false-positive
case, the network detects as atherosclerosis a patch with no vessel pixels, contain-
ing a central region with noise that looks like a vessel portion. In the false-negative
samples, the atherosclerosis region contains a cut in the artery, which causes that
vessel pixels to have background intensities both in the original image and in the
multichannel image generated by the QCL.
Figure 7.16: Training and validation accuracy curves of the CNN models employ-
ing SGD, and concerning as 1000 the number of epochs in the training phase. Top:
Training accuracy curves, Bottom: Validation accuracy curves.
7.6 CONCLUDING REMARKS

In this study, a Hybrid Quantum-Convolutional Neural Network method for
atherosclerosis detection in XCA images was introduced. The proposed scheme
includes a Quantum Convolutional Layer used as a preprocessing to improve the
atherosclerosis detection performance of typical CNNs. The numerical experiments,
based on two distinct CNN architectures (DenseNet-based and VGG-based archi-
tectures) and trying on two different training algorithms: Stochastic Gradient De-
scent and Stochastic Gradient Descent with Momentum, have demonstrated that
using a Quantum Convolutional Layer on a limited XCA dataset performs effi-
ciently for atherosclerosis detection. The introduced hybrid methodology improved
Figure 7.17: Training and validation loss curves of the CNN models employing SGD
concerning the number of epochs (1000) in the training phase. Top: Training loss
curves, Bottom: Validation loss curves.
the five evaluation metrics concerning the classical CNN architectures trained with
the normalized-raw XCA images.
Additionally, the numerical results are permitted to assert that momentum during
the optimization step improves the detection performance. The atherosclerosis de-
tection with the DenseNet-based CNN employing the Quantum Convolutional Layer
reached the best accuracy, with an 83.36%, while in precision, sensitivity, specificity,
and F1-score achieved an 81.52%, 86.67%, 83.99%, and 80.00%, respectively.
A future direction of this work will incorporate this quantum layer during the
optimization procedure. Finally, in agreement with numerical results, the proposed
method has shown the potential of a preprocessing Quantum Convolutional Layer to
generate discriminative feature maps (a multichannel image) to feed a classical CNN
to detect atherosclerosis in XCA images.
Table 7.2
Network Detection Results.
Model Accuracy Precision Sensitivity F1 -Score Specificity
83.36% 81.52% 86.67% 83.99% 80.00%
QCL-SGDM-A
(±2.11%) (±2.13%) (±2.77%) (±2.07%) (±2.62%)
80.80% 78.55% 85.08% 81.64% 76.45%
QCL-SGDM-B
(±3.16%) (±2.16%) (±5.37%) (±3.41%) (±2.19%)
82.88% 78.46% 91.11% 84.29% 74.52%
QCL-SGD-A
(±1.65%) (±1.94%) (±1.90%) (±1.42%) (±2.96%)
60.64% 59.09% 71.75% 64.47% 49.35%
QCL-SGD-B
(±1.55%) (±1.16%) (±9.81%) (±3.72%) (±8.20%)
79.84% 75.98% 87.62% 81.35% 71.94%
SGDM-A
(±4.28%) (±3.15%) (±6.06%) (±4.26%) (±3.32%)
69.44% 67.80% 75.87% 71.13% 62.90%
SGDM-B
(±1.99%) (±2.28%) (±10.06%) (±4.17%) (±7.36%)
67.36% 64.96% 76.83% 69.90% 57.74%
SGD-A
(±2.55%) (±0.86%) (±12.20%) (±5.12%) (±7.59%)
60.80% 58.55% 78.10% 65.67% 43.23%
SGD-B
(±2.09%) (±1.42%) (±19.40%) (±7.36%) (±16.11%)
The detection performance for each neural network (A:Au et al. [8], B: Antczak and Liberadzki[10]) was carried out
employing the Quantum Convolutional Layer (QCL) preprocessing and the original XCA. The networks were also opti-
mized using Stochastic Gradient Descent (SGD) and SGD with Momentum (SGDM).
Figure 7.18: Detection result sample. Each row shows one case of True-Positive
(TP), True-Negative (TN), False-Positive (FP), and False-Negative (FN). From left
to right: one raw XCA images and the nine-channel QCL outcome image.
CONFLICTS OF INTEREST
The authors declare that there is no conflict of interest.
ACKNOWLEDGEMENTS
This research was supported by the Engineering Division of the Campus Irapuato-
Guanajuato, grant NUA 147347; and the Mexican Council of Science and Technol-
ogy CONACyT, Doctoral Studies Grants no. 626154 and 626155.
ETHICAL APPROVAL
All procedures performed in studies involving human participants were in accor-
dance with the ethical standards of the institutional and/or national research commit-
tee and with the 1964 Helsinki Declaration and its later amendments or comparable
ethical standards. For this type of study, formal consent is not required.
REFERENCES
1. World Health Organization. Cardiovascular Diseases (CVDs). https://www.who.int/news-
room/fact-sheets/detail/cardiovascular-diseases-(cvds), Aug 2020.
2. Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep
convolutional neural networks. In Advances in Neural Information Processing Systems,
pages 1097–1105, 2012.
3. Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhi-
heng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large
scale visual recognition challenge. International Journal of Computer Vision, 115(3):211–
252, 2015.
4. Jacob Biamonte, Peter Wittek, Nicola Pancotti, Patrick Rebentrost, Nathan Wiebe, and
Seth Lloyd. Quantum machine learning. Nature, 549(7671):195– 202, 2017.
5. Chenxin Sui, Zhuang Fu, Zeyu Fu, Yao Wang, Yu Zhuang, Rongli Xie, Yanna Zhao, Jun
Zhang, and Jian Fei. A Novel method for vessel segmentation and automatic diagnosis of
vascular stenosis. In 2019 IEEE International Conference on Robotics and Biomimetics
(ROBIO), pages 918–923. IEEE, 2019.
6. WeiWu, Jingyang Zhang, Hongzhi Xie, Yu Zhao, Shuyang Zhang, and Lixu Gu. Auto-
matic detection of coronary artery stenosis by convolutional neural network with temporal
constraint. Computers in Biology and Medicine, 118:103657, 2020.
7. S. Jevitha, M. Dhanalakshmi, and Pradeep G Nayar. Analysis of left main coronary bifur-
cation angle to detect stenosis. In International Conference on Intelligent Systems Design
and Applications, pages 627–639. Springer, 2018.
8. Benjamin Au, Uri Shaham, Sanket Dhruva, Georgios Bouras, Ecaterina Cristea, Andreas
Coppi, Fred Warner, Shu-Xia Li, and Harlan Krumholz. Automated characterization of
stenosis in invasive coronary angiography images with convolutional neural networks.
arXiv preprint arXiv:1807.10597, 2018.
9. Gao Huang, Zhuang Liu, Laurens Van Der Maaten, and Kilian Q Weinberger. Densely
connected convolutional networks. In Proceedings of the IEEE Conference on Computer
Vision and Pattern Recognition, pages 4700–4708, 2017.
10. Karol Antczak and Lukasz Liberadzki. Stenosis detection with deep convolutional neural
networks. In MATEC Web of Conferences, volume 210, page 04001. EDP Sciences, 2018.
11. Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale
image recognition. arXiv preprint arXiv:1409.1556, 2014.
12. Chao Cong, Yoko Kato, Henrique Doria Vasconcellos, Joao Lima, and Bharath Venkatesh.
Automated stenosis detection and classification in X-ray angiography using deep neu-
ral network. In 2019 IEEE International Conference on Bioinformatics and Biomedicine
(BIBM), pages 1301–1308. IEEE, 2019.
13. Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir
Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper
with convolutions. In Proceedings of the IEEE Conference on Computer Vision and Pat-
tern Recognition, pages 1–9, 2015
14. Emmanuel Ovalle-Magallanes, Juan Gabriel Avina-Cervantes, Ivan Cruz–Aceves, and Jose
Ruiz-Pinales. Transfer learning for stenosis detection in X-ray coronary angiography.
Mathematics, 8(9):1510, 2020.
15. Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for
image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern
Recognition, pages 770–778, 2016.
16. Maxwell Henderson, Samriddhi Shakya, Shashindra Pradhan, and Tristan Cook. Quan-
volutional neural networks: Powering image recognition with quantum circuits. Quantum
Machine Intelligence, 2(1):1–9, 2020.
17. Jennifer Sleeman, John Dorband, and Milton Halem. A hybrid quantum enabled RBM ad-
vantage: convolutional autoencoders for quantum image compression and generative learn-
ing. In Quantum Information Science, Sensing, and Computation XII, volume 11391, page
113910B. International Society for Optics and Photonics, 2020.
18. Vijayasri Iyer, Bhargava Ganti, A.M. Hima Vyshnavi, P.K. Krishnan Namboori, and Sriram
Iyer. Hybrid quantum computing based early detection of skin cancer. Journal of Interdis-
ciplinary Mathematics, 23(2):347–355, 2020.
19. Ville Bergholm, Josh Izaac, Maria Schuld, Christian Gogolin, Carsten Blank, Keri McK-
iernan, and Nathan Killoran. Pennylane: Automatic differentiation of hybrid quantum-
classical computations. arXiv preprint arXiv:1811.04968, 2018.
20. Aradh Bisarya, Shubham Kumar, Walid El Maouaki, Sabyasachi Mukhopadhyay, Bikash
K Behera, Prasanta K Panigrahi, et al. Breast Cancer Detection Using Quantum Convolu-
tional Neural Networks: A Demonstration on a Quantum Computer. medRxiv, 2020.
21. Michael A. Nielsen and Isaac L. Chuang. Quantum Computation and Quantum Informa-
tion: 10th Anniversary Edition. Cambridge University Press, 2010.
22. Ronald De Wolf. Quantum computing: Lecture notes. arXiv preprint arXiv:1907.09415,
2019.
23. J. Robert Johansson, Paul D Nation, and Franco Nori. QuTiP: An open-source Python
framework for the dynamics of open quantum systems. Computer Physics Communica-
tions, 183(8):1760–1772, 2012.
24. Albert Einstein, Boris Podolsky, and Nathan Rosen. Can quantum-mechanical description
of physical reality be considered complete? Physical Review, 47(10):777, 1935.
25. Dominik Scherer, Andreas Muller, and Sven Behnke. Evaluation of pooling operations in
convolutional architectures for object recognition. In International Conference on Artificial
Neural Networks, pages 92–101. Springer, 2010.
26. Xavier Glorot and Yoshua Bengio. Understanding the difficulty of training deep feedfor-
ward neural networks. In Proceedings of the thirteenth international conference on artificial
intelligence and statistics, pages 249–256, 2010.
27. Herbert Robbins and Sutton Monro. A stochastic approximation method. The Annals of
Mathematical Statistics, pages 400–407, 1951.
28. Ning Qian. On the momentum term in gradient descent learning algorithms. Neural Net-
works, 12(1):145–151, 1999.
29. Karol Antczak and Lukasz Liberadzki. Deep Stenosis Detection Dataset.
https://github.com/KarolAntczak/DeepStenosisDetection, Aug 2020.
30. Timo Ojala, Matti Pietikainen, and David Harwood. A comparative study of texture mea-
sures with classification based on featured distributions. Pattern Recognition, 29(1):51–59,
1996.
8 Multilevel Quantum
Elephant Herd Algorithm
for Automatic Clustering of
Hyperspectral Images
Automatic clustering of hyperspectral images is a very strenuous task due to the
presence of huge number of redundant bands and complexity to process them. In
this work, two quantum versions of Elephant Herd Optimization algorithm are pro-
posed for this purpose. The binary and ternary quantum logics used enhances the
exploration and exploitation capability of the elephant herd optimization. These al-
gorithms are compared to their classical counterpart. They are implemented on the
Salinas dataset. The proposed qutrit based algorithm is found to converge faster and
produce more robust results. The Xie-Beni Index is used as the fitness function. A
few statistical tests like mean, standard deviation and Kruskal Wallis test are per-
formed to establish the efficiency of the proposed algorithm. The F score is used
to compare the segmented images using the optimal cluster numbers. The proposed
algorithms are found to perform better in most of the cases.
8.1 INTRODUCTION
Hyperspectral image (HSI) processing has caught the attention of many researchers
in the past decade. The development of powerful spectral cameras provided re-
searchers with the tools to easily acquire them. HSI can provide extensive and metic-
ulous information about the object or area which are captured by means of the spec-
tral cameras [5]. The reflection and absorption capabilities of different materials
present on the Earth’s surface are unique. These spectral informations are used by
HSI to recognize them individually [41]. HSI is extensively used in various fields
like environmental studies [45], military applications [25] and medical fields [36].
The number of bands in HSI varies from 10 to around 400. The main problem with
hyperspectral images is the presence of redundant and correlated information, lead-
ing to Hughes phenomena [21]. The rich spectral information increases the comput-
ing time in processing HSI. Hence, various dimensionality reduction techniques are
widely researched [11].
To extract useful information from HSI, different methods like classification, clus-
tering, and unmixing are used [5]. Clustering is a very beneficial method in HSI anal-
ysis when the ground truth image is not available. According to Zhang et al. [51],
HSI clustering algorithms [32] can be categorized into four groups. Centre-based
DOI: 10.1201/9781003283294-8 171

method is one of the most widely used clustering algorithms. In this type of clus-
tering, the data points are grouped based on their distance from the cluster centers.
K-means [32], is one of the most popular clustering algorithms based on Euclidean
distance. It is a hard clustering algorithm, which is sensitive to initial cluster centers
and membership values [28]. Fuzzy C means (FCM) [7], is a soft clustering algo-
rithm and produces better results than K -means, at the cost of increased iterations.
In both the algorithms, the main disadvantage is that the number of clusters should
be mentioned beforehand. In HSI, knowing the number of segments may always not
be possible.
Determining the number of clusters in HSI automatically is a challenging task.
Recently, researchers have started exploring methods to automatically detect cluster
numbers in various problems like image segmentation, data segmentation and others.
Very few works have been done on determining the number of clusters in HSI.
Clustering is considered to be a type of nondeterministic hard optimization prob-
lem or NP Hard problem [4]. Metaheuristics are found to be useful for solving NP
Hard problems in an efficient manner. Metaheuristics take reasonable time and pro-
duce near optimal solutions. Hence, many metaheuristic algorithms have been intro-
duced in the literature for solving clustering problems. Metaheuristic algorithms are
stochastic in nature and easy to implement [14]. They are mostly inspired by natural
phenomena like swarming of birds [42], ant’s colony, food finding behavior [12] and
others. Genetic Algorithm (GA) [20], Particle Swarm Optimization (PSO) [42], Ant
Colony Optimization [12] and Differential Evolution (DE) [44] are few well-known
metaheuristic algorithms.
In recent years, quantum computing has drawn the attention of a lot of researchers.
It was originally conceptualized by Sir Richard Feynman [16]. The various quan-
tum phenomena like superposition, entanglement, and interference can enhance the
computing capability of an algorithm exponentially [34]. Researchers have explored
these ideas to embed the basic principles of quantum computing with metaheuristic
algorithms [40]. Hence, a new category of metaheuristics was developed from these,
which are called quantum-inspired metaheuristics.
The main motivation of this work is to develop a fast and robust automatic cluster-
ing algorithm for HSI. Elephant Herd Optimization (EHO) [47] is a comparatively
new metaheuristic algorithm, based on the clan formation habit of elephants. The
simplicity of the algorithm and its good exploration abilities have inspired in devel-
oping the qubit and qutrit versions of the algorithm called the Qubit Elephant Herd
Optimization (QubEHO) and the Qutrit Elephant Herd Optimization (QutEHO), re-
spectively. The exploitation capability of a metaheuristic means that it is capable
of refining the result space. The EHO algorithm [47] lacks in this but the quan-
tum versions can easily achieve this. The parallel computing capability of a qubit or
qutrit enhances the exploitation property of the EHO algorithm [47]. Moreover, the
QubEHO and QutEHO are found to exhibit higher convergence speeds.
In this work, the Band Selection Convolutional Neural Network (BS-NET-
Conv) [8] is used in the pre-processing stage to reduce the number of bands in the
HSI. The Xie-Beni Index (XB-Index) [49] is used as the fitness function to detect the
optimal number of clusters. The Fuzzy C Means [7] algorithm is used to determine
the clusters.
Multilevel Quantum Elephant Herd Algorithm for Automatic Clustering of Hyperspectral Images173
The main contributions of this work are as follows:
• Two algorithms viz., the qubit and qutrit versions of Elephant Herd Optimization
are devised for optimal number of cluster detection in HSI.
• An algorithm for qubit based quantum rotation gate implementation for bringing
diversity in the population.
• An algorithm for qutrit based quantum rotation gate implementation for bringing
diversity in the population.
The chapter is organized as follows: Section 8.2 contains a brief literature survey of
the used methods. The important background concepts are discussed in Section 8.3.
Section 8.4 contains the details of the proposed methodology. The experimental
results and their analysis are provided in Section 8.5. A brief conclusion of the pro-
posed method has been drawn in Section 8.6.
8.2 LITERATURE SURVEY

HSI is a 3-dimensional data cube of spectral and spatial information. They are ca-
pable of providing detailed information of the captured area. In [53], research on
HSI is subdivided into two main categories, viz., supervised methods (classification
of HSI [10]) and unsupervised methods (clustering-based algorithms [48]). Dimen-
sionality reduction is also an important area in HSI pre-processing and lots of work
are done in this area. Principal Component Analysis [38], Local Discriminant Analy-
sis [30], Mutual Information [15] and Band Selection Convolutional Neural Network
(BS-NET-Conv) [8] are few widely used methods in this direction.
Clustering-based algorithms for HSI segmentation have caught the attention of
a lot of researchers lately. According to Wang et al. [47], existing clustering al-
gorithms can be classified into four groups, viz., Centroid method [32], Density
method [39], Biological methods [54], and Graph methods [52]. K means [32] and
Fuzzy C means [7] are examples of centroid methods. They are easy to implement
but are not always capable of providing robust results.
To determine the optimal number of clusters, CVI is used. A cluster validity index
(CVI) defines a relation between intracluster cohesion (within-group scatter) and
intercluster separation (between-group scatter) to estimate the quality of a clustering
solution [24]. Xie-Beni Index [49], Dunn index [13], Calinski - Harabasz Index [9],
Gamma index [6], I-Index [33] are few widely used indices. In [24], the authors have
provided an extensive study on CVI for automatic data clustering.
Finding the optimal number of clusters is a type of NP Hard problem. To solve
these problems, metaheuristic algorithms are found to be highly beneficial. Meta-
heuristic algorithms are developed based on some natural phenomena. They can
be widely classified into four categories [14], viz., Evolutionary Algorithms [20],
Physics-based Algorithms [26], Human-based Algorithms [37], and Swarm Algo-
rithms [42].
Swarm algorithms are a popular category of metaheuristic algorithms. They are
mostly inspired by the social behavior of animals or insects. The ability of a swarm to
reach an optimal solution collectively and using individual intelligence makes swarm
a powerful tool for optimization. Particle Swarm Optimization [42], Ant Colony Op-
timization [12], Cuckoo Search [50], Harris hawk Optimization [19], Border Collie
Optimization [14] and Elephant Herd Optimization [47] are few well-known swarm
intelligent algorithms. EHO [47], due to its easy implementation and good perfor-
mance, has drawn the attention of a lot of researchers. The algorithm has good explo-
ration capability but the exploitation of search space is less efficient. This also leads
to slower convergence. Hence, a lot of enhanced and hybrid versions of the EHO
algorithm have been researched [23] to overcome these disadvantages. In [23], three
different enhanced versions of EHO algorithm were proposed to overcome these de-
ficiencies. Li et al. [29] presented a detailed study on the different variants of the
EHO algorithm published so far, along with their different features.
Researchers working on metaheuristic algorithms, have been captivated by the
idea of designing algorithms with quantum advantage. These are also known as
quantum-inspired algorithms, as they are inspired by the principles of quantum com-
puting but the simulations are done on classical computers. Narayanan and Moore
were the first to conceptualize a quantum-inspired evolutionary algorithm for solv-
ing the Travelling Salesman problem [35]. In [18], a Quantum-Inspired Evolutionary
Algorithm was proposed with a better population diversity and concept of look-up
table for the application of rotation gates. Recently, a lot of researchers has devel-
oped quantum-inspired versions of metaheuristic algorithms like the Improved Bloch
Quantum Artificial Bee Colony algorithm, which was proposed in [22]. It involves
a complicated Bloch sphere representation, which makes it a little complex. Very
few works have been done on multilevel quantum systems. A qutrit-based Genetic
Algorithm was proposed in [46]. However, hardly any work has been done yet on
developing multilevel quantum-based EHO algorithm.
8.3 BACKGROUND CONCEPTS

A few important background concepts used for developing the proposed algorithm
are presented in this section.
8.3.1 ELEPHANT HERDING OPTIMIZATION

Elephant Herding Optimization [47] is a new nature-inspired metaheuristic algorithm
based on the herding behavior of elephants. The behavioral factors considered to
develop the mathematical model were:
• All elephants live in clans under the leadership of a matriarchal elephant.
• The number of elephants in a clan is fixed.
• Male elephants leave their clan and live alone after a certain time.
8.3.1.1 Clan Updation

For developing the EHO algorithm [47], initially few clans are considered. The ma-
triarch (Ebest) is the fittest elephant of the clan. Her position is updated using the
following equation:
Ebestcli = β × Ec,cli (8.1)
In Eqn. (8.1), cli is the ith clan, Ebestcli represents the updated position of the ma-
triarch in the clith clan, β is a factor that dictates the influence of the Ec,cli on new
Ebestnew,cli , j . The value of β ranges between [0, 1]. Ec,cli is the center of clan cli ,
calculated with the help of the following equation.
nci
1
Ec ,cli ,d = × ∑ Ecli , j,d (8.2)
ncli j=1
In Eqn. (8.2), d is the d th dimension and D is the total dimension. ncli is the number
of elephants in the clith clan. Eqn. (8.2) is used to calculate the center of a clan.
All the other elephants ( j) next positions are calculated using the following equa-
tion:
Enew,cli , j = Ecli , j + α × Ebestcli − Ecli , j × r (8.3)
In Eqn. (8.3), α is the influence that the matriarch has on the other elephants of the
clan. Its value ranges from [0, 1]. r is a random number.
8.3.1.2 Separation Operator

The male elephants on reaching puberty usually leave the clan and live on their own.
Usually, the most unfit individual is considered as the male elephant leaving the clan.
Its position is updated in the following manner:
Eworst,cli = Emin + (Emax − Emin + 1) × rand (8.4)
where, Emax and Emin are the upper and lower bounds of the elephant individuals.
Eworst,cli is the worst elephant individual in the ith clan. rand is a random number
generated between [0, 1].
8.3.1.3 Steps of EHO

The EHO algorithm [47] works in the following manner:
• Initialize an elephant population.

• Run the following steps for Ge number of generations.
- Find the fitness of the elephants and sort them.
- Implement clan updation using Eqns. (8.1),(8.2) and (8.3)
- Implement separation operator
8.3.2 BASIC CONCEPTS OF QUANTUM COMPUTING

A qubit is the smallest unit of quantum computation [34]. A physical system that
can be present in at least two orthogonal states can be used to represent a qubit. It
can be a spin particle in a magnetic field or a photon in polarized state [43]. The
orthogonal states are represented as |0⟩ and |1⟩. This notation is called the Dirac’s
Bra and Ket notation. Applying the laws of quantum mechanics, the orthogonal states
can be represented in their linear combination.
|ψ ⟩ = q0 |0⟩ + q1 |1⟩ (8.5)
In Eqn. (8.5), q0 is the probability of occurrence of |0⟩ and q1 is the probability

of occurrence of |1⟩. The probability constraints, q0 and q1 are complex numbers,
which satisfy the normalization condition as given below.
|q0 |2 + |q1 |2 = 1 (8.6)
This property of qubits is also called quantum superposition in which they are ca-
pable of existing in both basis states simultaneously. Hence, if an n qubit system is
considered, it will be able to exist in 2n states simultaneously. This can be explained
with the help of the following equation:
2n −1
|ψ ⟩ = ∑ qi |i⟩ (8.7)
i=0
Quantum computation like classical computation is not restricted to two states. Quan-
tum states can have D-dimensional states. This generalized state of quantum bits is
called qudit [3]. Hence, a qudit can exist in the following state.
Dn −1
|ψ ⟩ = ∑ qi |i⟩ (8.8)
i=0
A qutrit is the three-valued quantum state. It has three basis states, viz., |0⟩, |1⟩ and
|2⟩. The superposition of a qutrit state can be expressed as follows:
|ψ ⟩ = q0 |0⟩ + q1 |1⟩ + q2 |2⟩ (8.9)
The normalization of a qutrit state is expressed as follows:
|q0 |2 + |q1 |2 + |q2 |2 = 1 (8.10)
8.3.3 FUZZY C MEANS CLUSTERING ALGORITHM

Fuzzy C Means [7] is one of the most frequently used clustering algorithms. It is
a soft clustering method in which a pixel is given membership values based on the
different cluster centers [5]. The main aim of the FCM algorithm [7] is to find optimal
clusters, by minimizing the squared function given below [17].
k d k d
M = ∑ ∑ Uimj di2j = ∑ ∑ Uimj Pj −Vi
2
(8.11)
i=1 j=1 i=1 j=1
In Eqn. (8.11), Ui j represents the membership values of a pixel to k cluster centers.

The value of Ui j varies between 0 and 1. d represents the total number of pixels of the
image that is being clustered and m is the weighting exponent which ranges between
[1, ∞]. Vi represents the k cluster centers. di, j is the Euclidean distance between the
ith cluster center and the jth data point and Pj is the jth data point. Vi is calculated
using the following equation:
∑dj=1 Uimj Pj
Vi = (8.12)
∑dj=1 Uimj
The membership values are calculated as given below.

 !2/(m−1) −1
1 k Pj −Vi
Ui j = 2/(m−1) =  ∑  (8.13)
k di Pj −Vi
∑i=1 d ji i=1
The process is run for a number of iterations until a minimum value for Eqn. (8.11)
is obtained.
8.3.4 XIE-BENI INDEX

Xie-Beni Index is used for determining the optimum number of clusters [49]. It is a
quotient of mean quadratic error and minimal squared distance between the points in
the cluster. The XB-Index is represented as follows:
M2
XB = (8.14)
dmin
where, the compactness of a cluster is represented as follows:
1 k d m 2
Mm = ∑ ∑ ui j d (Pj , ki )
d i=1
(8.15)
j=1
and the minimum distance between two cluster centers is given by the following
equation:
dmin = min d 2 (ki , k j ) (8.16)
i, j
XB-Index [49] is a minimization function and the minimum value of Eqn. (8.14),
represents the optimal solution.
8.4 PROPOSED METHODOLOGY

The proposed method can be divided into two parts, viz., the hyperspectral image
pre-processing and the quantum versions of the Elephant Herd Optimization [47]
algorithm based determination of the optimal number of clusters.
8.4.1 HSI PREPROCESSING

The Band Selection Convolution Neural Network [8] is implemented in the pre-
processing stage, to reduce the number of bands in HSI. This algorithm chooses the
best informative bands, which are then used in the automatic clustering stage. The
network used, consists of a band attention module and a band re-weighing module.
This is followed by a reconstruction network. The attention module extracts infor-
mation from the HSI cube using 2-D convolution layers. To ensure that the learned
weights are non-negative, the Sigmoid function is used. Finally, a global pooling
layer and a fully connected layer are implemented. The best bands are selected based
on the average of weights of all the bands.
(a) (b)
Figure 8.1: Qubit and Qutrit representation in search space.
8.4.2 QUBIT AND QUTRIT BASED ELEPHANT HERD OPTIMIZATION

The proposed Qubit and Qutrit Elephant Herd Optimization are depicted with the
help of Algorithm 1. Initially, a quantum population consisting QP is taken consisting
of |QPα ⟩, |QPβ ⟩ for qubit or |QPα ⟩, |QPβ ⟩, |QPγ ⟩ for qutrit.
Result: Optimal Number o f Clusters(k)

Initialize: Number of Generation - Ge,
Size of Quantum Population - n,
Length of each particle -D,
No of clans - cl,
Qubit/Qutrit population initialization and corresponding
classical representation - QP and CR (using Algorithm 2 for
qubit or using Algorithm 3 for qutrit)
Applying Rotation Gate on QP using Algorithm 4 for qubit or using Algorithm 5 for qutrit
Allotting the elephants to cl different clans randomly
for t in 1,2,. . . ,Ge do
z=Random number of zeros; Randomly insert z zeros in CR; Calculate the z number of clusters using
FCM algorithm [7]; Calculate Fitness of Population using XB − Index - Fitt ; for i in 1,2,. . . ,cl do
Update QP value for Matriarch of the clan i ((QPα or QPβ ) for qubit or (QPα or QPβ or QPγ ) for
qutrit, for which CR is true) using Eqn. (8.1);
Update QP value for other elephants of the clan i ((QPα or QPβ ) for qubit or (QPα or QPβ or
QPγ ) for qutrit, for which CR is true) using Eqn. (8.3);
Update QP value for most unfit elephant of the clan i ((QPα or QPβ ) for qubit or (QPα or QPβ or
QPγ ) for qutrit, for which CR is true) using Eqn. (8.4);
end
if Fitt < Fitt−1 then
k=Optimal cluster number obtained in t th generation;
Update the quantum states to maintain superposition principle;
end
Algorithm 1: QubEHO/QutEHO
1 1
|QP⟩ = √ |0⟩ + √ |1⟩ (8.17)
2 2
1 1 1
|QP⟩ = √ |0⟩ + √ |1⟩ + √ |2⟩ (8.18)
3 3 3
Their corresponding basis state predictions (CR) are done using Algorithm 2 for
qubit or Algorithm 3 for qutrit. Algorithm 4 for qubit or Algorithm 5 for qutrit are
Result: Basis states -|0⟩,|1⟩

Initialize: QP using Eqn. (8.17)
for i in 1,2,. . . ,n do
for j in 1,2,. . . ,D do
r=random number between [0,1];
if r < α 2 then
CR = 0;
else
CR = 1;
end
end
end
Algorithm 2: Classical Representation of Quantum States
Result: Basis states -|0⟩,|1⟩,|2⟩

Initialize: QP using Eqn. (8.18)
for i in 1,2,. . . ,n do
for j in 1,2,. . . ,D do
if r < α 2 then
CR = 0;
else if r1 < β 2 + α 2 then
CR = 1;
else
CR = 2;
end
end
end
Algorithm 3: Classical Representation of Quantum States
proposed for rotating the quantum states without the help of look-up-tables. The
implementation of rotation gates brings diversity to the population. Each qubit or
Result: Rotated states -|QPα ⟩,|QPβ ⟩

for i in 1,2,. . . ,n do
for j in 1,2,. . . ,D do
if CRi, j == 0 then q
QPα i, j = QPα i, j ∗ r; QPβ i, j = 1 − QPα 2i, j ;
else if CRi, j == 1 then q
QPβ i, j = QPβ i, j ∗ r; QPα i, j =
2
1 − QPβ i, j ;
end
end
Algorithm 4: Rotation of Quantum States
qutrit represents the individual elephant. A fixed number of clans are considered and
the population is scattered randomly among the clans. Then, a random number of
zeros (z) are introduced in CR. The number of zeros represents the number of clusters
to be considered. The QP value for which CR is true, is taken as the cluster center.
z number of clusters are calculated using FCM algorithm [7]. The XB-Index [49] is
used as the fitness function to check the optimality of the number of clusters obtained.
Subsequently, the clan updation operation is done. The Matriarch of each clan
is identified. The QP value for which CR is true, is updated using Eqn. (8.1). Other
elephants of the clan are updated similarly using Eqn. (8.2). The separation operation
Result: Rotated states -|QPα ⟩,|QPβ ⟩,|QPγ ⟩

for i in 1,2,. . . ,n do
for j in 1,2,. . . ,D do
if CRi, j == 0 then q q
QPα i, j = QPα i, j ∗ r; QPβ i, j = 1 − QPα 2i, j ; QPγ i, j = 1 − QPα 2i, j + QPβ i, j ;
else if CRi, j == 1 then q q
QPβ i, j = QPβ i, j ∗ r; QPα i, j = 1 − QPβ i, j ; QPγ i, j =
2
1 − QPα 2i, j + QPβ i, j ;
else q q
QPγ i, j = QPγ i, j ∗ r; QPα i, j = 1 − QPγ 2i, j ; QPβ i, j = 1 − QPα 2i, j + QPγ i, j ;
end
end
end
Algorithm 5: Rotation of Quantum States
is executed on the least fit elephant using Eqn. (8.4). The whole process is run for
Ge number of generations. In Figure 8.2, the basic operations viz., formation of the
clan, separation of male elephant from the clan and how other elephants follow the
Matriarch can be visualized. The use of quantum bits helps to search the solution
space faster and more efficiently. This can be visualized in Figure 8.1. For a single
qubit, we can see that two different solutions points are considered. Simiarly, for a
single qutrit three different solutions points are taken. Hence, it enhances the speed
of the algorithm producing more robust results.
(a) Elephant Herd Optimization
Figure 8.2: Elephant Herd Optimization [47]-Clan formation, Clan members follow-
ing matriarch, Seperation from clan (Elephant picture from [1]).

The Salinas Dataset [2] is used for experimental purpose. A brief description of the
dataset is presented in Section 8.5.1. The proposed QubEHO and QutEHO algo-
rithms are compared with classical EHO [47] algorithm. Section 8.5.2 contains a
brief description of the fitness function used. The different parameters used to ana-
lyze the results are presented in Section 8.5.3.
8.5.1 SALINAS DATASET

The Salinas scene was collected by the AVIRIS sensor over the Salinas Valley Cal-
ifornia [2]. A total of 224 bands were captured and 20 were discarded due to water
absorption. The spatial resolution is 3.7m pixels consisting of 512 × 217 bands each.
The ground truth image of the Salinas dataset [2] consists of 16 classes, which in-
cludes vegetables, bare soils, and vineyards.
8.5.2 FITNESS FUNCTION

The Xie-Beni Index [49] is used as the fitness function for determining the optimal
number of clusters. Eqn. (8.14) is minimized for obtaining the optimal results.
8.5.3 ANALYSIS
The proposed QubEHO and QutEHO algorithms are compared with classical
EHO [47] algorithm. To obtain an impartial analysis, all the algorithms need to be
evaluated using the same parameters. Hence, all the algorithms were run 50 times
for 100 iterations each. The algorithms were executed in MATLAB R2019a, on Intel
(R) Core (TM) i7 8700 Processor in Windows 10 environment.
The mean, standard deviation, and best convergence time of all the three algo-
rithms are presented in Table 8.1. It is observed that the proposed QutEHO algo-
rithm arrives at optimal results in negligible time when compared to the other two
algorithms. The classical EHO [47] algorithm takes the highest time to converge.
In Table 8.2, the optimal cluster numbers (CL) obtained for EHO [47], QubEHO,
and QutEHO are reported. Their corresponding fitness values (FV) are also presented
in Table 8.2 where . The Salinas dataset [2], has 16 classes. The most optimal results
Table 8.1
Mean, Standard Deviation (Std) and the Best Reported Time for EHO [47],
QubEHO, and QutEHO
Parameters EHO [47] QubEHO QutEHO
Mean 0.2032 0.0725 0.0389
Std 0.1394 0.0124 0.0219
Time 181.9412 24.3839 3.3035
Table 8.2
Some Cluster Numbers and Best Fitness Values Obtained for EHO [47],
QubEHO and QutEHO.
Methods EHO [47] QubEHO QutEHO
Sr No CL FV CL FV CL FV
1. 4 0.0685 6 0.0733 10 0.0129
2. 4 0.0828 5 0.0693 9 0.0143
3. 4 0.0835 5 0.0787 9 0.0144
4. 5 0.0835 5 0.0660 7 0.0182
5. 4 0.0860 5 0.0614 7 0.0182
Table 8.3
Normalized F-Score [31] for EHO [47], QubEHO and QutEHO
Process EHO [47] QubEHO QutEHO
0.1998 0.0038 0.0016
are produced by the QutEHO algorithm, followed by the QubEHO algorithm. The
classical EHO [47] has comparatively lesser number of classes. In a real-life sce-
nario, where the ground truth cannot be obtained and the number of classes cannot
be estimated, the proposed algorithms, specially the qutrit version can be both bene-
ficial and time efficient.
To judge the quality of the segmented images, the normalized F score is used [31].
The F score is evaluated with the help of the following equation.
1 √ r e2i
F(SI) = r∑ √ (8.19)
1000(v × h) i=1 Ai
Here, SI is the final segmented image. The dimension of the image is v × h. The num-
ber of regions is designated by r. A and e stand for the area and the average color
error of the ith region, respectively. The results are normalized for easy representa-
tion. The results of all the three methods, viz., EHO [47], QubEHO, and QutEHO
are presented in Table 8.3. The proposed algorithm produces better results compared
to the other two algorithms. Few segmented images along with the ground truth im-
age and the pre-processed images are presented in Figure 8.3. In Figure 8.3, the
segmented images using EHO [47] has four clusters, QubEHO has five clusters and
QutEHO has nine clusters, respectively.
Another statistical test called the Kruskal-Wallis test [27] is applied to check the
null hypothesis with 1% significance level. The p-value obtained is less than 0.001
indicating that the results are highly significant. Hence, the null hypothesis that all
the results from all the three methods belong to the same distribution, stands rejected.
(a) Ground Truth Image (b) Pe-processed Image (c) EHO [47]
(d) QubEHO (e) QutEHO
Figure 8.3: (a) Ground Truth Image of Salinas Dataset [2], (b) Pre-processed Im-
age using BSCNN [8], (c)-(e) Clustered Images using EHO [47], QubEHO, and
QutEHO.
The results are presented in Table 8.4. The box-plot of the test is given in Figure 8.4.
The convergence curve of all the participating methods is presented in Figure 8.5.
From the convergence curve, it can be visually observed that the proposed algorithm
converges faster and with more optimal values compared to EHO [47] and QubEHO.
Table 8.4
Kruskal-Wallis Test [27]
Test p-value Significance
Kruskal-Wallis Test 2.8166e-25 Highly Significant
0.7
0.6
0.5
0.4
0.3
0.2
0.1
1 2 3
(a) Box-Plot of Kruskal-Wallis Test [27]
Figure 8.4: Box-Plot of Kruskal-Wallis Test [27] for EHO [47], QubEHO, and
QutEHO.
EHO
QubEHO
QutEHO
100
Fitness Values
10-1
10-2
0 20 40 60 80 100
Iterations
(a) Convergence Curve
Figure 8.5: Convergence curve for EHO [47], QubEHO, and QutEHO.
8.6 CONLUSION
In this chapter, qubit and qutrit-based Elephant Herd Optimization algorithms are
proposed for automatic clustering of hyperspectral images. The modified rotation
gate operation enhances the diversity of the population. The exploration and ex-
ploitation capabilities of the classical EHO algorithm is enhanced, with a faster
convergence capability. As automatic cluster detection is a tedious task in HSI pro-
cessing, these algorithms can be highly beneficial in real-life scenarios. The results
indicate that the proposed QubEHO, and QutEHO produce optimal results compared
to the classical version of EHO. The proposed algorithms produces better clustering
outcome when their comparative F scores are considered. The statistical tests also
establish the efficiency of the proposed algorithms. As a future direction, the qudit
version of the EHO algorithm can be developed.
REFERENCES
1. Elephant clipart. Free download transparent .PNG | Creazilla, Jul 2021. [Online; accessed
17. Jul. 2021].
2. Hyperspectral Remote Sensing Scenes - Grupo de Inteligencia Computacional (GIC), Jul
2021. [Online; accessed 17. Jul. 2021].
3. Qudits | Cirq | Google Quantum AI, Jul 2021. [Online; accessed 16. Jul. 2021].
4. Laith Abualigah, Amir H. Gandomi, Mohamed Abd Elaziz, Husam Al Hamad, Mahmoud
Omari, Mohammad Alshinwan, and Ahmad M. Khasawneh. Advances in meta-heuristic
optimization algorithms in big data text clustering. Electronics, 10(2), 2021.
5. P. Azimpour, R. Shad, M. Ghaemi, and H. Etemadfard. Hyperspectral image clus-
tering with albedo recovery fuzzy c-means. International Journal of Remote Sensing,
41(16):6117–6134, 2020.
6. Frank B. Baker and Lawrence J. Hubert. Measuring the power of hierarchical cluster
analysis. Journal of the American Statistical Association, 70(349):31–38, 1975.
7. J. C. Bezdek, R. Ehrlich, and W. Full. Fcm: The fuzzy c-means clustering algorithm.
Computers & Geosciences, 10(2):191 – 203, 1984.
8. Y. Cai, X. Liu, and Z. Cai. Bs-nets: An end-to-end framework for band selection of hyper-
spectral image. IEEE Transactions on Geoscience and Remote Sensing, 58(3):1969–1984,
2020.
9. T. Caliński and H. JA. A dendrite method for cluster analysis. Communications in Statistics
- Theory and Methods, 3:1–27, 01 1974.
10. Mulin Chen, Qi Wang, and Xuelong Li. Discriminant analysis with graph learning for
hyperspectral image classification. Remote Sensing, 10(6), 2018.
11. Samiran Das, Shubhobrata Bhattacharya, Aurobinda Routray, and Alok Kani Deb. Band
selection of hyperspectral image by sparse manifold clustering. IET Image Processing,
13(10):1625–1635, 2019.
12. M. Dorigo, V. Maniezzo, and A. Colorni. Ant system: optimization by a colony of cooper-
ating agents. IEEE Transactions on Systems, Man, and Cybernetics, Part B (Cybernetics),
26(1):29–41, 1996.
13. J. C. Dunn. A fuzzy relative of the isodata process and its use in detecting compact well-
separated clusters. Journal of Cybernetics, 3(3):32–57, 1973.
14. Tulika Dutta, Siddhartha Bhattacharyya, Sandip Dey, and Jan Platos. Border collie opti-
mization. IEEE Access, 8:109177–109197, 2020.
15. A. Elmaizi, H. Nhaila, E. Sarhrouni, A. Hammouch, and C. Nacir. A novel information

gain based approach for classification and dimensionality reduction of hyperspectral im-
ages. Procedia Computer Science, 148:126 – 134, 2019.
16. R.P. Feynman. Simulating physics with computers. International Journal of Theoretical
Physics, 21(6):467–488, 1982.
17. Saman Ghaffarian and Salar Ghaffarian. Automatic histogram-based fuzzy c-means clus-
tering for remote sensing imagery. ISPRS Journal of Photogrammetry and Remote Sensing,
97:46–57, 2014.
18. K-H Han and J-H Kim. Quantum-inspired evolutionary algorithm for a class of combi-
natorial optimization. IEEE Transactions on Evolutionary Computation, 6(6):580–593,
2002.
19. Ali Asghar Heidari, Seyedali Mirjalili, Hossam Faris, Ibrahim Aljarah, Majdi Mafarja, and
Huiling Chen. Harris hawks optimization: Algorithm and applications. Future Generation
Computer Systems, 97:849–872, 2019.
20. J.H. Holland. Adaptation in natural and artificial systems : an introductory analysis with
application to biology. Control and artificial intelligence. University of Michigan Press,
1975.
21. G. Hughes. On the mean accuracy of statistical pattern recognizers. IEEE Transactions on
Information Theory, 14(1):55–63, 1968.
22. F. Huo, X. Sun, and W. Ren. Multilevel image threshold segmentation using an im-
proved bloch quantum artificial bee colony algorithm. Multimedia Tools and Applications,
79:24472471, 2020.
23. Alaa A. K. Ismaeel, Islam A. Elshaarawy, Essam H. Houssein, Fatma Helmy Ismail, and
Aboul Ella Hassanien. Enhanced elephant herding optimization for global optimization.
IEEE Access, 7:34738–34752, 2019.
24. Adán José-García and Wilfrido Gómez-Flores. A survey of cluster validity indices for
automatic data clustering using differential evolution. In Proceedings of the Genetic and
Evolutionary Computation Conference, GECCO ’21, page 314322, New York, NY, USA,
2021. Association for Computing Machinery.
25. Chen Ke. Military object detection using multiple information extracted from hyperspec-
tral imagery. In 2017 International Conference on Progress in Informatics and Computing
(PIC), pages 124–128, 2017.
26. S. Kirkpatrick, C. D. Gelatt, and M. P. Vecchi. Optimization by simulated annealing.
Science, 220(4598):671–680, 1983.
27. W. H. Kruskal and W. A. Wallis. Use of ranks in one-criterion variance analysis. Journal
of the American Statistical Association, 47(260):583–621, 1952.
28. Tao Lei, Xiaohong Jia, Yanning Zhang, Shigang Liu, Hongying Meng, and Asoke K.
Nandi. Superpixel-based fast fuzzy c-means clustering for color image segmentation. IEEE
Transactions on Fuzzy Systems, 27(9):1753–1766, 2019.
29. Juan Li, Hong Lei, Amir H. Alavi, and Gai-Ge Wang. Elephant herding optimization:
Variants, hybrids, and applications. Mathematics, 8(9), 2020.
30. W. Liao, A. Pizurica, P. Scheunders, W. Philips, and Y. Pi. Semisupervised local dis-
criminant analysis for feature extraction in hyperspectral images. IEEE Transactions on
Geoscience and Remote Sensing, 51(1):184–198, 2013.
31. Jianqing Liu and Yee-Hong Yang. Multiresolution color image segmentation. IEEE Trans-
actions on Pattern Analysis and Machine Intelligence, 16(7):689–700, 1994.
32. J. Macqueen. Some methods for classification and analysis of multivariate observations. In
Proceedings of the Fifth Berkeley Symposium on Mathematical Statistics and Probability,
Volume 1: Statistics, pages 281–297, 1967.
33. U. Maulik and S. Bandyopadhyay. Performance evaluation of some clustering algorithms

and validity indices. IEEE Transactions on Pattern Analysis and Machine Intelligence,
24:1650–1654, 01 2002.
34. D. McMahon. Quantum Computing Explained. John Wiley & Sons, Inc., Hoboken,New
Jersey, 2008.
35. A. Narayanan and M. Moore. Quantum-inspired genetic algorithms. In Proceedings of
IEEE International Conference on Evolutionary Computation, pages 61–66, 1996.
36. Robert Pike, Guolan Lu, Dongsheng Wang, Zhuo Georgia Chen, and Baowei Fei. A min-
imum spanning forest-based method for noninvasive cancer detection with hyperspectral
imaging. IEEE Transactions on Biomedical Engineering, 63(3):653–663, 2016.
37. T. Ray and K.M. Liew. Society and civilization: An optimization algorithm based on the
simulation of social behavior. IEEE Transactions on Evolutionary Computation, 7(4):386–
396, 2003.
38. Craig Rodarmel and Jie Shan. Principal component analysis for hyperspectral image clas-
sification. Surv Land inf Syst, 62, 01 2002.
39. Alex Rodriguez and Alessandro Laio. Clustering by fast search and find of density peaks.
Science, 344(6191):1492–1496, 2014.
40. Maria Schuld, Ilya Sinayskiy, and Francesco Petruccione. An introduction to quantum
machine learning. Contemporary Physics, 56(2):172–185, 2015.
41. G. Shaw and Hsiao hua K. Burke. Spectral imaging for remote sensing. In Lincoln Labo-
ratory Journal, number 1, pages 3–28, 2003.
42. Y. Shi and R. Eberhart. A modified particle swarm optimizer. In 1998 IEEE Interna-
tional Conference on Evolutionary Computation Proceedings. IEEE World Congress on
Computational Intelligence (Cat. No.98TH8360), pages 69–73, May 1998.
43. M. Steffen, D. P. DiVincenzo, J. M. Chow, T. N. Theis, and M. B. Ketchen. Quantum
computing: An ibm perspective. IBM Journal of Research and Development, 55(5):13:1–
13:11, 2011.
44. Rainer Storn and Kenneth Price. Differential evolution – a simple and efficient
heuristic for global optimization over continuous spaces. J. of Global Optimization,
11(4):341–359, 1997.
45. Ming Yang Teng, Ruby Mehrubeoglu, Scott A. King, Kirk Cammarata, and James Simons.
Investigation of epifauna coverage on seagrass blades using spatial and spectral analysis of
hyperspectral images. In 2013 5th Workshop on Hyperspectral Image and Signal Process-
ing: Evolution in Remote Sensing (WHISPERS), pages 1–4, 2013.
46. V. Tkachuk. Quantum genetic algorithm based on qutrits and its application. Mathematical
Problems in Engineering, 2018(8614073), 2018.
47. Gai-Ge Wang, Suash Deb, and Leandro dos S. Coelho. Elephant herding optimization. In
2015 3rd International Symposium on Computational and Business Intelligence (ISCBI),
pages 1–5, 2015.
48. Huan Xie, Ang Zhao, Shengyu Huang, Jie Han, Sicong Liu, Xiong Xu, Xin Luo, Haiyan
Pan, Qian Du, and Xiaohua Tong. Unsupervised hyperspectral remote sensing image
clustering based on adaptive density. IEEE Geoscience and Remote Sensing Letters,
15(4):632–636, 2018.
49. X. L. Xie and G. Beni. A validity measure for fuzzy clustering. IEEE Transactions on
Pattern Analysis and Machine Intelligence, 13(8):841–847, Aug 1991.
50. Xin-She Yang and Suash Deb. Cuckoo search via lvy flights. In 2009 World Congress on
Nature Biologically Inspired Computing (NaBIC), pages 210–214, 2009.
51. Hongyan Zhang, Han Zhai, Liangpei Zhang, and Pingxiang Li. Spectralspatial sparse
subspace clustering for hyperspectral remote sensing images. IEEE Transactions on Geo-
science and Remote Sensing, 54(6):3672–3684, 2016.
52. Yang Zhao, Yuan Yuan, Feiping Nie, and Qi Wang. Spectral clustering based on iterative
optimization for large-scale and high-dimensional data. Neurocomputing, 318:227–235,
2018.
53. Yang Zhao, Yuan Yuan, and Qi Wang. Fast spectral clustering for unsupervised hyperspec-
tral image classification. Remote Sensing, 11(4), 2019.
54. Yanfei Zhong, Liangpei Zhang, and Wei Gong. Unsupervised remote sensing image clas-
sification using an artificial immune network. International Journal of Remote Sensing,
32(19):5461–5483, 2011.
9 Toward Quantum-Inspired
SSA for Solving
Multiobjective
Optimization Problems
9.1 INTRODUCTION
Many real-world optimization problems are multiobjective optimization problems. It

must find a uniformly distributed set of optimal solutions along the Pareto-optimal
front in the search space. Multiobjective optimization problem (MOP) is an empir-
ical area of research. It has been applied in many sciences and engineering fields,
including mathematical optimization and the controller placement problem, where
ideal choice requires to be made in the existence of trade-offs between two or more
conflicting objectives. In general, a multiobjective optimization problem can be a
minimization or maximization problem f (x) subject to xε X, where f (x) is a real-
valued scalar function often known as objective or cost function and X is the feasible
set or subset of the search space. Without the loss of assumptions, it can be mathe-
matically presented as below:


 minimize f (x) = ( f1 (x) , f1 (x) , . . . fn (x) )

where x ε X

 n≥2

set X includes all the feasible solutions
where x = (x1 , x1 , . . . xd ) is a d-dimensional vector in the search space Rd , f (x) =

( f1 (x) , f1 (x) , . . . fn (x) ) is an n -dimensional vector in the search space Rn . After all,
the objectives in MOP are commonly counter to one and all. The Pareto dominion
association is often applied to differentiate distinct solutions. According to the Pareto
′
dominion association, for any two given solutions x and x , it can be said that x
′
dominates another solution x ,
′
if ∀iε {1, 2, 3, . . . n}, fi (x) ≤ fi (x ) and
′
∃iε {1, 2, 3, . . . n}, fi (x) < fi (x )
In case solution x dominates any other solutions, x, x could be said as a non-

dominated solution. The set of all other than dominated solutions in the search space
DOI: 10.1201/9781003283294-9 189

is called estimated Pareto-optimal (PE) [1]. The set of vectors in the search space
that correspond to the PE is mentioned as the actual Pareto-optimal (PA) [2][3]. The
essence of solving a multiobjective optimization problem is identifying a set of well-
distributed optimal solutions along the Pareto-optimal front in the search space.
Today, a metaheuristic algorithm is a popular approach to resolve compound opti-
mization problems in the field of optimization in science and engineering disciplines.
Optimization is an intelligent method to explore the optimal solution among all ob-
tainable ones of a specific problem [4][5]. The conventional approach for the mul-
tiobjective optimization problem is to scalarize the vector of multiobjective into a
single objective by averaging the value returned by objective functions with a weight
vector. Converting MOP into a single objective optimization problem allows a single-
objective algorithm to be used more straightforwardly. Still, the obtained optimal
solution is primarily dependent on the weight vector considered during the scalariza-
tion process. Moreover, a decision-maker has to know the problem in advance and
provide a weight against each objective.
Furthermore, the decision-maker would be more interested in learning alternate
solutions if obtainable. Algorithms are associated with such a predefined assump-
tion, making them ineffective in solving today’s compounded optimization problems.
Solving a multiobjective optimization problem is one of such kind where a set of uni-
formly distributed solutions are required along the Pareto-optimal front in the search
space.
Swarm-based optimization algorithm is a metaheuristic algorithm that works with
a group of solutions and attempts to achieve the optimal solution in each iteration.
It is often inspired by biological groups’ sociality and has been widely applied for
many real-world optimization problems to overcome the traditional optimization ap-
proach’s restrictions. Such an intelligent system brings the advantages of being work-
able and straightforward to deal with different optimization problems. Additionally,
it has the natural tendency to work with problem represented as the population of
the optimal solution. Moreover, it takes place with the benefits of inherent move-
ment, exploration, and exploitation, that lessen the possibility of entrapment into
local optima. A number of Pareto-optimal sets can be obtained using a swarm-based
optimization method for the multiobjective optimization problem.
The Salp swarm algorithm (SSA) is a new swarm-based metaheuristic method
that imitates Salps flocking behavior in the ocean by making a chain. The SSA is
similar to other evolutionary algorithms in many features, and it works proficiently
for numerous real-world optimization problems. The flocking behavior of SSA can
avoid entrapment of each solution into local optima up to some point due to its salp
chain behavior. However, there are optimization problems where SSA cannot obtain
a solution and easily trap into a local or deceptive optimum. The multiobjective op-
timization problem is one of such kind. The difficulty for SSA lies mainly due to a
good search strategy for the multiobjective optimization problem. The original de-
sign of SSA is to save only one solution and update the positions based on the food
source to obtain the best solution. However, there is no single best solution for the
MOPs. It is required to obtain a set of well-distributed optimal solutions along the
Toward Quantum-Inspired SSA for Solving Multiobjective Optimization Problems 191
Pareto-optimal front in the search space. Thus, there is a need to modify the SSA
algorithms’ original design to perform the overall searching process well with a bal-
ance of exploration and exploitation propensity to achieve the expected results.
The quantum-inspired algorithm is a new branch of study in the area of evolution-
ary computation. It is characterized by the particular principles of quantum physics
such as uncertainty, superposition, interference, etc. The approach to merge and de-
sign the quantum-inspired algorithms for classical computers represents the solutions
into quantum representation. The principles of quantum computing offer better di-
versity during the optimization process. The quantum search strategy intelligently
guides the individuals toward the global optima by significantly improving conver-
gence speed and solution efficiency. The variety can be derived from the representa-
tion model of the population. A probabilistic model of a linear superposition of states
presents better characteristics of generating diversity in the population. Maintaining
a good assortment in the population increases the searchability of the algorithm and
resolves search stagnation. Sun et al. presented Quantum-behaved Particle Swarm
Optimization (QPSO) to improve the performance of PSO, including good conver-
gence and global search ability [6]. The approach guaranteed subjectively to discover
a reasonable optimal solutions in the search space. The experimental results indicate
that the QPSO works better than PSO and is a promising approach. Hence, in this
study, we propose to employ quantum’s inspiration to standard SSA for the same
reason as QPSO for multiobjective optimization problems and compare it to MSSA
and NSGA-II.
This article introduces a new presentation for the Multiobjective Quantum in-
spired Salp Swarm Algorithm (MQSSA) to ameliorate the overall performance of
SSA. The approach is a hybrid of novel paradigms: SSA and Quantum Computing.
Besides many other essential properties, this model can find a suitable solution faster
using fewer individuals. This approach reduces the required number of estimation
dramatically, which is a predominant effecting factor for solving the optimization
problems. The Delta potential-well model (DPWM) representation of SSA in this
paper enhances the convergence speed faster than traditional SSA. It maintains the
population’s diversity, preventing the population from stagnating in deceptive optima
and increasing the algorithm’s searchability. According to DPWM, if individual Salp
in the SSA algorithm has quantum behavior, the algorithm is bound to work differ-
ently due to Heisenberg’s uncertainty principle of quantum physics. To the best of
our knowledge, a Quantum-inspired approach to improving the performance of SSA
for the multiobjective optimization problems is introduced the first time. In this new
approach, a simplified representation with DPWM integration to improve original
SSA performance is introduced for MOP, making this algorithm more comfortable
to understand and implement. The proposed algorithm’s performance is evaluated
on the multiobjective domain’s complex benchmark problems in this work. A com-
parative study is performed to assess the performance with well-regarded algorithms
MSSA and NSGA-II.
The rest of this paper is structured as follows: Standard Salp Swarm Algorithm
(SSA) is presented in the next section, followed by the proposed algorithm with Delta
potential-well model (DPWM). Experimental procedure followed by the empirical

study and analysis, then a section discussing the results followed with conclusion.
9.2 SALP SWARM ALGORITHM

Salp Swarm Algorithm (SSA) is recently introduced swarm intelligence optimiza-
tion algorithm. It is introduced by Mirjalili et al. in December 2017 [7]. After that,
many types of research work for the enhancement and implementation of SSA are
employed in engineering and technology, feature selection, and the multi-controller
placement problem. A simpler version of SSA is introduced with a random search
radius, in which the efficiency of the algorithm was improved [8]. Another particle-
based SSA proposes balancing the exploration and exploitation propensity, which
helps to boost the convergence speed and estimation accuracy [9]. Similarly, to ame-
liorate the searchability of the algorithm, the SSA is hybridized with the gravitational
search algorithm [10]. Although the Salp Swarm Algorithm has been used broadly in
many areas of engineering fields such as power system optimization and multilevel
color image segmentation [11] [12], the recent literature shows lukewarm interest in
the optimization problems where multiple scenarios are required to considered, that
is, a multiobjective SSA (MS-MOSS) studied for sizing the photovoltaic system in
which the number of PV modules and storage battery is pivotal and influences the
soundness and cost of the system [13].
Figure 9.1: Salp chain in SSA.
Figure 9.1 shows a salp chain that is created by group of salps in deep ocean.
According to the research, this chain follows the leader and follower pattern, where
leaders’ responsibility is to direct the chain toward food source and followers align
with the leader in turn. For more details, refer to [7]. The mathematical representa-
tion and implementation of this chain categorize the Salps population into the front
salps as leader, which position being updated based target and the rear salps as fol-
lower, which updates its position to align with adjacent individuals. This complete
procedure is divided into three parts: Initialization, Define leaders, and Salps position
updates.
9.2.1 INITIALIZATION
F in equation (9.1) represents an objective function of minimization problem:
F = f (x1 , x2 , x3 , x4 , x5 , ..., xD ) (9.1)

D is the dimension in equation (9.1) and xi is in constant boundaries such as

lbi ≤ xi ≤ ubi (i = 1, 2, 3, ..D), where lbi and ubi are the lower and upper bounds of ith
variable. Equation (9.2) calculates the initial position of each Salp for the population
size N as the number of search agents, in which rand (N, D) is a N ∗ D matrix with
each element as a random number between 0 and 1:
XN∗D = rand (N, D) . ∗ (ub − lb) + lb (9.2)
9.2.2 LEADERS’ SPECIFICATION

On initial position, each individual fitness is calculated using equation (9.1) and N
fitness values obtained, respectively. Then the fitness values are sorted in ascending
order and the minimum value from the vector, i.e., first position is regarding as the
leader position, which is nearest to the global optimum. The rows of matrix X are re-
arranged based on sorted fitness values consequently, to appropriately distinguished
between the leaders and followers salps. Mathematically, one from all N salps, the
leader salp is constituted by row vector x1 whose dimension is D. Therefore, the
jth part of the leader salp represented as x1j where j = 1, 2, 3, . . . D. Similarly, the ith
follower in dimension jth represented as xij where i = 1, 2, 3, . . . N.
9.2.3 UPDATING POSITION

The leader’s and followers’ positions are updated as equations (9.3) and (9.4), re-
spectively, the coefficient c1 of the leader’s equation is calculated using the math-
ematical equation (9.5). The position vector x1j in the equation (9.3) represents the
position of leaders in jth dimension value and first individual with the best fitness
value; similarly, Fj represents the position of the food source (F) in jth dimension.
The algorithm parameters c2 and c3 are evenly distributed random numbers between
0 and 1. The upper bound and lower bound of the search space are represented as
ub j and lb j respectively in jth dimension.
Fj + c1 ((ub j − lb j ) c2 + lb j ) c3 ≥ 0
x1j = (9.3)
Fj − c1 ((ub j − lb j ) c2 + lb j ) c3 ≤ 0
1 i
xij = x j + xi−1
j (9.4)
2
4t 2
c1 = 2e−( T ) (9.5)
This is important to note that updating leaders’ position by determining the updat-
ing direction c3 by equation (9.3), only related to individual toward global optimum
and it has nothing to do with historical position of salps. However, equation (9.4) up-
dates the followers position with a mechanism of adjacent salps, where jth variable of
the ith individual is xij and the jth variable of the ith position of the nearby individual
in front of the ith individual is xi−1

j . Further in equation (9.5) c1 is the most important
parameter of the algorithm, which regulates the steps of position updates, which is
defined with the current iteration t, maximum iteration T, and a natural constant e or
Euler’s number. From equation (9.5), it can be clearly observed that the value of c1
is larger for the smaller number of iterations. Hence, in the initial iteration, the value
of c1 is for the exploration and at the later iterations for the exploitation of the search
space.
9.2.4 RE-EVALUATION AND DECISION-MAKING

On the updated position, each individual fitness is calculated again using equation
(9.1) and N fitness values obtained correspondingly. The leader position is revised
again based on the updated fitness value to move the salp chain toward global op-
timum. The steps of iterations repeated for the maximum number of iterations to
complete the optimization process.
9.3 PROPOSED MULTIOBJECTIVE QUANTUM-INSPIRED SALP SWARM

ALGORITHM
Standard Salp Swarm Algorithm (SSA) has the superior ability to simplify the updat-
ing positions of salps by straightforward position update equations. However, SSA
cannot produce expected results for the multiobjective optimization problems due to
the same reason, as mentioned in Section 9.1. In this section, a new algorithm-based
on SSA and Quantum Computing (QC) is presented. In QC, when considering the
dynamics of particles and to avoid explosion with guaranteed convergence, particles
must be inbound state during the run, that is, moving in a potential attraction field
towards a center point [14]. There is potential field model such as Delta potential-
well and Quantum oscillator to ascertain the particles movements in a bound state.
In this paper, we employ the Delta potential-well model for the convergence of each
particle having quantum behavior because particles’ convergence is much faster in
the quantum oscillator, leading to prematurity [6].
In this proposed algorithm, the algorithm’s representation is changed based on
the principle of DPWM and integrated into the standard SSA. The inspiration of the
proposed MQSSA is based on a similar approach to QPSO; the aim is to improve
the overall performance of SSA for multiobjective optimization problems. Further,
this modification aims to adapt SSA and enhance the exploration ability and increase
the algorithm’s overall performance by an appropriate balance between exploration
and exploitation tendency. This embedding approach of SSA and quantum comput-
ing is a promising strategy to improve the performance of SSA. One side, a DPWM
representation known for increasing the speed of convergence and diversity in the
population and on the other side integration with SSA to design a better search strat-
egy increases the population’s global searchability.
9.3.1 DELTA POTENTIAL-WELL MODEL FOR SSA

The Delta potential-well model describes using Dirac delta function as a theorized
function in quantum physics. It approximately corresponds to the potential, which is
zero throughout omitting a single point where it takes an infinite value. Each Salp
(X i ) has a quantum state, and it can be formulated using the wavefunction Ψ (x, t).
From the analysis of the movements of salps toward a center point Ad , we assume
that each Salp moves in a Delta potential well of search space for which the center
point is Ad , that is, described as in equation (9.9).
For the simplicity of the presentation, considering in one-dimensional space with
center point A, the Delta potential well can be represented as follows:
Z((x)) = −γδ (x − A) = −γδ (y) (9.6)

where γ is a constant positive value and γδ (y) denotes the Dirac delta function for y
= (x A).
9.3.2 SALP POSITION MEASUREMENT

Multiobjective quantum-inspired SSA algorithm employs the Delta potential-well
model representation from the quantum mechanics and considers salps movements
towards a center point Ad as per equation (9.9), which is updated based on the lead
position (F) stated in Section II as the best location. The potential field model and
quantum oscillator in quantum physics ascertain the particles movements in a bound
state due to the characteristics that each particle oscillate in a potential attraction
field. The salps in the delta potential-well model must move in the bound state ac-
cording to this strategy. To evaluate the fitness of each Salp, the position of the Salp
needs to be calculated. However, only the probability of position for each Salp (X i
can be learned from the probability density function |Ψ (x, t) |2 , which depends on
the potential field the particle lies into. The probability of position indicates that a
salp appears at position x relative to a point Ad . Hence, it is required to measure
salp position using the collapsing technique, that is, transformation from a quantum
state to the classical form. The Monte Carlo method can reproduce this process of
measurement [15][16].
In this paper, the following iterative equations were used to measure each salp
according to the Monte Carlo method. As stated, salps moved round and swayed
toward Ad with its kinetic energy declining to zero. With this characteristic in the
modified algorithm, the salps are inbound states and avoid explosion with guaranteed
convergence.
k Ad + Bl × |BestMeanl − X jk |log( urd ) c3 > 0

X j+1 = d (9.7)
Ad − Bl × |BestMeanl − X jk |log( urd ) c3 < 0
d
where d is the current dimension and l is the current iteration.

We assumed here the first iteration l = 0 with maximum iteration size L. The
rd and ud are random numbers in the range [0, 1] in d − dimension . Bl describes
equation (9.8), a contraction-expansion coefficient that is adjusted dynamically dur-

ing the optimization process. It means gradually decreasing or increasing iteration
wise according to individual salp’s convergence speed and algorithm performance.
Ad described as in equation (9.9), a center point for salps to move toward locally.
Also known as the learning inclination point for salps to oscillate around. BestMeanl
describes mathematically as equation (9.10), mean of the best position. X jk is the kth
salp in j − dimension and X j+1
k is the new position of salp.

0.59 ∗ (L − l)
β=
l + 0.59
where the current iteration is l, and L is the maximum number of iterations
Bl = (β + (1 − β )) (9.8)

r1d ∗ X jk + r2d ∗ Leader j
Ad = (9.9)
r1d + r1d
where r1d and r2d are random numbers in the range [0, 1]. Leader j is the leader
position and represented as the best location.
d
1
BestMeanl =
N ∑ lead j (l) (9.10)
j=1
where N is the maximum number of populations.
9.3.3 THE NEW ALGORITHM BEHAVIOR

This hybrid approach with quantum computing to enhance the standard SSA per-
formance boosts the overall performance and maintains individual diversity for the
multiobjective optimization problems. In which it produces a set of well-distributed
Pareto-optimal solutions. Preserving diversity is extremely important for such op-
timization problems as it is often entrapment into non-distributed solutions. The
searching tendency of evolutionary algorithms is determined based on the population
diversity. It means lower exploration ability in the populations identical elements.
In MQSSA, equation (9.7) updates the salps position and produces new solutions.
Pseudo-code for the MQSSA is presented in Algorithm 1 of Figure 9.2. After the
initialization phase of the population and first approximation calculation in steps 2
and 3, the algorithm takes up into the optimization process’ main loop. This loop ex-
ecutes for a maximum number of MAX generations and is formed alongside to per-
form various tasks. First, the control parameters for the standard SSA and MQSSA
calculated according to equation (9.5) and (9.8), respectively, followed by calculating
the objective function values for each Salp and identify other than dominated ones at
step 6. Step 7 handles the repository to update for the obtained non-dominated solu-
tions. The other than dominated solutions to be added in repository if it is not full.
At this stage, the non-dominated solutions are compared with available solutions in
the repository to keep only other than dominated solutions in the repository. After
that, it also deletes the solutions from the crowded region to maintain the number of
solutions. For this, it first ranked the solutions and then selected using the roulette
wheel technique.
Step 10 executes only once at first iteration and generates a salp chain where a
division of population happens into leaders and followers. After that, this algorithm
is used in step 8, followed by evaluating solutions for the coverage and updating
the search boundary. The process repeats for MAX generation, and in each gener-
ation, it produces Pareto optimal solutions, which is a set of uniformly distributed
solutions in the search space along the estimated Pareto optimal front. At step 8, the
interference process and proposed quantum-based equations for SSA are executed to
generate new solutions. This process consists of updating the individual position us-
ing equation (9.7), calculating the contraction-expansion coefficient, evaluating the
converging points and best mean along with the fitness of salps. At step 7, the other
than dominated solutions to be put into an archive if the archive has an empty space,
i.e., its not full. At this stage, the other than dominated solutions are compared with
available solutions in the repository to keep only other than dominated obtained so-
lutions in the repository. After that, at step 11 CoverageSelection() function used to
delete the solutions from the crowded region to maintain the number of solutions.
For this, it first ranked all the solutions and then selected using an intelligent roulette
wheel technique. At step 14, the algorithm returns the obtained Pareto-optimal set as
the best solution along with the first approximation.
9.4 EXPERIMENTAL PROCEDURE

9.4.1 COMPUTING ENVIRONMENT
The source code of MQSSA is developed in MATLAB R2017b, and the experimental
computer environment for the simulation is used as follows:
1. Intel Core i7-3520M CPU @ 2.90 GHz

2. 16 GB of RAM
3. macOS Catalina v.10.15.7 operating system
9.4.2 PERFORMANCE ASSESSMENT METRICS

To evaluate the performance of the MQSSA algorithm, two metrics are used in this
paper: Pareto Sets Closeness (PSC) and Inverted Generational Distance metric (IGD)
[17]. PSC assesses the similarity between the approximated PF and the obtained
Pareto-optimal set. The IGD indicator assesses diversity and convergence both for the
solutions obtained in the search space. The more considerable value of PSC means
a well-distributed obtained solution in the search space. The lower value of IGD
means that the diversity and convergence of the obtained solutions in search space
are adequate.
The IGD can be defined as an average Euclidean distance between the approxi-
mated Pareto-optimal set and obtained Pareto-optimal set in the search space. Let U ∗
Figure 9.2: The Algorithm MQSSA.
denotes a set of uniformly distributed exact Pareto-optimal location and H represents

a set of obtained Pareto-optimal solutions of the final population. The IGD can be
computed as the average distance from U ∗ to H in the decision space using the below
equation:
∑ min d(u, H)
IGD (H, U ∗ ) = uεU ∗ |U∗|
where u represents an exact point in U ∗ , min d (u, H) is the minimum Euclidean

distance between u and the points in H.
Similarly, the PSC value can be calculated using the below equation, which indi-
cates the closeness between the approximated PSs and obtains PSs.
CR
PSC =
IGD
CR shows the overlapping ratio between the approximated Pareto-optimal set and
the obtained Pareto-optimal set, and IGD is the inverted generational distance in the
search space. A higher PSC value means the obtained solutions are well distributed
in the search space and is adequate.
Table 9.1
Benchmark Problems Used in This Study
9.4.3 MULTIOBJECTIVE BENCHMARK PROBLEMS

The MQSSA algorithm’s performance is evaluated on the seven challenging well-
known multiobjective optimization benchmark problems whose Pareto-optimal front
is convex-shaped and concave-shaped. These multiobjective optimization benchmark
functions are presented in Table 9.1 with the name of the function, dimension, ranges
of variables, and PF characteristics. The benchmark functions are divided into func-
tions according to the features of their shape of Pareto-optimal front, i.e. convex, con-
cave, and convex disconnected. Benchmark functions SCH1 and ZDT1 are convex-
shaped, whose surface is curved like the exterior of a circle. Functions ZDT2, ZDT4,
FON, and KUR are concave-shaped, whose surface is curved inward like the circle’s
interior. Function ZDT3 is a convex disconnected problem, where the Pareto-optimal
front is in the shape of convex but with isolated regions. These Pareto fronts are reg-
ular in many real-world optimization problems and are very difficult to be set on
by the optimization approaches. All these benchmark problems have two objectives,
with having no constraints.
Table 9.2
Statistical Results of Pareto Sets Proximity (PSP)
9.4.4 EVALUATING METHOD AND ALGORITHMS PARAMETERS

The MQSSA requires five parameters: Max-Iteration, Current-Iteration, lower
bound, upper bound, dimension, and objective function reference. All the necessary
parameters for the algorithm are being managed using the other functions such as
the number of salps, best solution, etc. This paper performed a comparative study of
MQSSA with MSSA [7] and NSGA-II [18].
For a fair performance comparison, the following parameters are used in this ex-
periment:
1. Population size 60
2. Maximum number of generations 1000 for all the test functions
3. The same number of function evaluation is considered for all the test problems.
4. For NSGA-II, Crossover = 0.7, Mutation = 0.4 as percentage, and mu=0.02 as
mutation rate.
The average results of 30 independent runs and average mean, the standard devia-
tion of metrics PSP and IGD are summarized in Tables 9.2 and 9.3. The average norm
indicates how the MQSSA performs on average, and the standard deviation shows
how stable it is during all the runs. The experimental results demonstrate that the
overall performance of MQSSA is competitive as compared with other approaches.
9.5 EXPERIMENTS AND DISCUSSION

The MQSSA is executed and evaluated for the multiobjective optimization bench-
mark problems, and quantitative results are calculated using the PSP and IGDX of
each trial. They are recorded in the form of average and standard deviation (SD),
presented in Tables 9.2 and 9.3 as results for the analysis. Also, the obtained solution
as true Pareto optimal set is shown in Figures 9.3–9.9. The maximum size of repos-
itory is set to 100 for MQSSA. In this paper, the two highly regarded multiobjective
Table 9.3
Statistical Results of Inverted Generational Distance (IGD)
optimization algorithms in this domain are selected for the results endorsement and
for the comparative study: MSSA and NSGA-II.
Tables 9.2 and 9.3 indicate that the MQSSA algorithm remarkably performs better
than MSSA and NSGA-II on most of the ZDT functions. Further, when examined,
the obtained PF in Figures 9.4–9.6 indicates that MQSSA shows a superior conver-
gence than MSSA and NSGA-II for the ZDT benchmark functions. The obtained
solutions are well distributed uniformly for MQSSA algorithm, which means the al-
gorithm coverage is high. A space in midway of the obtained Pareto optimal set by
MSSA and NSGA-II in some of the ZDT problems indicates that how the coverage
of the algorithm is negatively impacted. The SCH1 and ZDT1 functions are convex-
shaped as indicated by the Pareto optimal front. Hence, the coverage and convergence
propensity of the algorithms can be benchmarked. The obtained Pareto-optimal set
in Figures 9.3 and 9.4 indicate that MQSSA performs anew better than MSSA and
NSGA-II algorithms. The coverage is deficient for the NSGA-II on ZDT2 function.
It looks that the convergence and coverage of MQSSA are better than MSSA.
Further, it can be observed that the MSSA algorithm find a gap in the obtained Pareto
optimal set, but the obtained solutions are well distributed in rest of the true Pareto-
optimal set. Additionally, the problems ZDT2 and ZDT4 shown to be Pareto optimal
front of type concave-shaped, which is invariably difficult for the algorithm designed
based on aggregation approach. But, the obtained results indicates that MQSSA can
efficiently approximate these functions’ right front with exceedingly good coverage
and convergence. When it is compared between MQSSA and NSG-II in Figures
9.4–9.6, results indicate that the coverage and convergence of MQSSA are better
and keen.
Tables 9.2 and 9.3 and Figure 9.6 analysis of ZDT3 function has Pareto optimal
set with isolated regions. Such kind of Pareto-optimal set is usual in the real-world
optimization problems. It is very difficult for the algorithms to obtain the Pareto op-
timal set of such problems. There is a highly possibility that the algorithm is failed
to obtain the Pareto-optimal set in all the separated areas and pin down in one of
the region. The comparison outcome on ZDT3 and other previously discussed ZDT
problems indicate that the achievement of NSGA-II is comparatively low, the cov-
erage is good, but the convergence is inferior. The obtained Pareto optimal set of
several isolated regions is far away from the true Pareto-optimal front. In the view
of obtained Pareto-optimal front by MQSSA and MSSA, it is evident that MQSSA
performs comparatively better than MSSA with regard to coverage and convergence.
These outcomes exemplify that MQSSA can successfully discover all the isolated
regions of the Pareto optimal front with well-distributed solutions in all the areas.
Figure 9.3: SCH1 Pareto front obtained by MQSSA, NSGA-II, and MSSA.
Figure 9.4: ZDT1 Pareto front obtained by MQSSA, NSGA-II, and MSSA.
The experimental results and above discussion prove that MQSSA can estimate
Pareto optimal front of type concave and convex shaped with a reasonable coverage
and convergence on four ZDT series benchmark problems: ZDT1, ZDT2, ZDT3, and
ZDT4.
Figure 9.5: ZDT2 obtained Pareto front by MQSSA, NSGA-II, and MSSA.
Figure 9.6: ZDT3 Pareto-front obtained by MQSSA, NSGA-II, and MSSA.
Figure 9.7: ZDT4 Pareto-front obtained by MQSSA, NSGA-II, and MSSA.

Figure 9.8: FON obtained Pareto-front by MQSSA, NSGA-II, and MSSA.
Figure 9.9: KUR obtained Pareto front by MQSSA, NSGA-II, and MSSA.
Further study on FON and KUR, which is equipped with nonconvex-shaped, can
benchmark the convergence and coverage algorithm. The analysis outcome of Fig-
ures 9.8–9.9 shows that all three algorithms, MQSSA, MSSA, and NSGA-II, are
unable to converge for KURs Pareto optimal front. Most of the obtained solutions
are far from the Pareto-optimal front. However, the results from Tables 9.2–9.3 show
that MQSSA performs slightly better. Inspecting the fronts of MQSSA, NSGA-II,
and MSSA for FON shows that MQSSA and NSGA-II provide better convergence.
However, the results from Tables 9.2 and 9.3 show NSGA-II performs slightly better.
These results exemplify that MQSSA successfully guide salp chain with regard
to separate regions of true Pareto optimal front and is competitive as compared with
other approaches.
9.6 CONCLUSION
This chapter presents a novel Multiobjective Quantum-inspired Salp Swarm Al-
gorithm (MQSSA) with Delta potential-well model presentation, which is a better
alternative than the binary presentation for the multiobjective optimization problems.
The proposed approach is evaluated on several multiobjective optimization bench-
mark problems having convex-shaped and concave-shaped Pareto-optimal front. It
shows a favorable outcome, with better performance than other well-regarded al-
gorithms in the multiobjective domain. Further investigations would be required to
evaluate the robustness of MASSA on solving different kinds of optimization prob-
lems.
Besides, the experimental study showed that the proposed algorithm, MQSSA:
1. Has the ability to achieve the Pareto-optimal front for the large dimensional mul-
tiobjective optimization problems.
2. Excellent speed as compared to the traditional SSA.
3. An appropriate balance between exploitation and exploration propensities.
4. Appropriate convergence rate and good coverage.
Future study would include applying the proposed algorithm to solve more bench-
mark problems and real-world optimization problems from different domains. It
would also be interesting to study the relationship and the differences between this
algorithm and other optimization approaches, such as several improved PSO and
other swarm intelligence techniques.
REFERENCES
1. H. Li, K. Deb, and Q. Zhang, “Variable-length Pareto optimization via decomposition-
based evolutionary multiobjective algorithm,” IEEE Trans. Evol. Comput., vol. 23, no.
6, pp. 987–999, Dec. 2019.
2. K. Li, R. Chen, G. Fu, and X. Yao, “Two-archive evolutionary algorithm for constrained
multiobjective optimization,” IEEE Trans. Evol. Comput., vol. 23, no. 2, pp. 303–315,
Apr. 2019.
3. Y. Sun, B. Xue, M. Zhang, and G.G. Yen, “A new two-stage evolutionary algorithm for
many-objective optimization,” IEEE Trans. Evol. Comput., vol. 23, no. 5, pp. 748–761,
Oct. 2019.
4. X. Yang, “Review of meta-heuristics and generalized evolutionary walk algorithm”, Int.
J Bio-Inspir. Com., vol. 3, no. 2, pp. 77–84, 2011.
5. Z. Michalewicz, Genetic algorithms + data structures = evolution programs (2nd, ex-
tended ed.). New York, NY: Springer-Verlag New York, Inc., 1994.
6. J. Sun, B. Feng, and W. Xu, “Particle swarm optimization with particles having quantum
behavior,” in Proc. Congr. Evol. Comput., vol. 1, pp. 325–331, 2004.
7. S. Mirjalili, A.H. Gandomi, S.Z. Mirjalili, S. Saremi, H. Faris, S.M. Mirjalili, “Salp
swarm algorithm: A bio-inspired optimizer for engineering design problems,” Adv. Eng.
Softw., vol. 114, pp. 163–191, 2017.
8. B. Xiao, R. Wang, Y. Xu, J. Wang, W. Song, and Y. Deng, “Simplified Salp
Swarm Algorithm,” 2019 IEEE International Conference on Artificial Intelligence
and Computer Applications (ICAICA), Dalian, China, 2019, pp. 226–230, doi:
10.1109/ICAICA.2019.8873515.
9. B. Xiao, R. Wang, Y., Xu, et al. “Salp Swarm Algorithm based on Particle-best”, In:
IEEE 3rd Information Technology, Networking, Electronic and Automation Control
Conference (ITNEC). Chengdu, pp. 1383–1387, 2019.
10. S. Li, Y. Yu, D. Sugiyama, D., et al. “A Hybrid Salp Swarm Algorithm With Gravita-
tional Search Mechanism,” In: 5th IEEE International Conference on Cloud Computing
and Intelligence Systems (CCIS), Nanjing, pp. 257–261, 2018.
11. S. Ekinci and B. Hekimoglu, “Parameter optimization of power system stabilizer via
Salp Swarm algorithm,” In: 5th International Conference on Electrical and Electronic
Engineering (ICEEE), Istanbul, pp. 143–147, 2018.
12. Z. Xing and H. Jia. “Multilevel Color Image Segmentation Based on GLCM and Im-
proved Salp Swarm Algorithm,” IEEE Access, vol. 7, pp. 37672–37690, 2019.
13. H.M. Ridha, C. Gomes, H. Hizam, and S. Mirjalili. “Multiple scenarios multiobjective
salp swarm optimization for sizing of standalone photovoltaic system,” Renewable En-
ergy, Elsevier, vol. 153(C), pp. 1330–1345, 2020.
14. M. Clerc and J. Kennedy, “The Particle Swarm: Explosion, stability and convergence in
a multi-dimensional complex space”, IEEE Trans. Evol. Comput., 6: 58–73, 2002.
15. J. Liu, W. Xu, and J. Sun, “Quantum behaved particle swarm optimization with mu-
tation operator”, In: Proc. of IEEE International Conference on Tools with Artificial
Intelligence, pp. 240–244, 2005.
16. J. Sun et al, “A global search strategy of quantum-behaved particle swarm optimization”,
IEEE Conference on Cybernetics and Intelligent Systems, pp 111–116, 2004.
17. G. Li, L. Yan, and B. Qu, “Multiobjective particle swarm optimization based on Gaus-
sian sampling,” IEEE Access, vol. 8, pp. 209717–209737, 2020, doi: 10.1109/AC-
CESS.2020.3038497.
18. K. Deb, S. Agrawal, A. Pratap, and T. Meyarivan. A Fast Elitist Non-dominated
Sorting Genetic Algorithm for Multiobjective Optimization: NSGA-II. In: Schoe-
nauer M. et al. (eds) Parallel Problem Solving from Nature PPSN VI. PPSN 2000.
Lecture Notes in Computer Science, vol. 1917. Springer, Berlin, Heidelberg, 2000.
https://doi.org/10.1007/3-540-45356-383.
10 Quantum-Inspired
Multi-Objective NSGA-II
Algorithm for Automatic
Clustering of Gray Scale
Images
10.1 INTRODUCTION
Clustering [1][2][3] is a process of partitioning a heterogeneous dataset into some
groups of homogeneous data points or elements. It can be considered as a challeng-
ing task to generate appropriate number of cluster from a given dataset due to lack
of proper knowledge of the dataset. By addressing the problem of automatic clus-
tering, several research works have been done so far, some of them are presented in
[4][5][6][7]. While considering the purpose of clustering, sometimes the clustering
algorithms provide good results with one type of datasets but not be able to provide
good results with other types of dataset, which yields more challenges in the field of
automatic clustering [12]. Though, there exist many automatic clustering algorithms
[8][9][10][11], but those algorithms have focused only on optimizing a single ob-
jective, whereas in real-world scenario, many problems may have more than one
objective which need to be taken for consideration [13]. In this aspect, the require-
ment of multi-objective optimization has been realized by which multiple objectives
which are not only similar but even possibly conflicting can be tackled properly.
Nowadays, multi-objective optimization algorithms are becoming popular for
their capability of searching a highly complex search space. In the last decade,
researchers have developed many nature-inspired multi-objective optimization al-
gorithms, which include non-dominated sorting GA (NSGA-II) [14][15], Pareto
envelope-based selection algorithm (PESAII) [16], Strength Pareto Evolutionary Al-
gorithm (SPEA) [17], and its improved version, SPEA2 [18]. The overview and ap-
plicability of the multi-objective algorithms for clustering are presented by Maulik
et al. in [19]. Zhou et al. addressed the basic principles, advancements, and applica-
tions of multi-objective algorithms to solve several real-world optimization problems
in [20].
In recent years, the concepts of quantum computing [21] are being incor-
porated with the evolutionary algorithms for effectively exploring the search
space for multi-objective optimization problems. Quantum-inspired evolutionary
algorithms have been developed for performing quasi quantum operations on
DOI: 10.1201/9781003283294-10 207

classical computers by implementing the concept of quantum mechanical phenom-

ena. Few research works on quantum-inspired evolutionary algorithms are available
in [17][23][24][25][26][27][28]. Though, the quantum-inspired evolutionary algo-
rithms can efficiently explore and exploit search space for a global optimal solution,
still there remains a challenge to handle the problem of multi-objective optimization.
It can be handled by the help of quantum-inspired algorithms to improve the prox-
imity of the non-dominated Pareto optimal front and by employing advantages of
quantum-inspired algorithms to preserve the diversity.
In this paper, NSGA-II [15] has been taken as the base algorithm for implement-
ing the proposed work as NSGA-II [15] uses a strong elitist mechanism for maintain-
ing the diversity of the search space efficiently. It also uses non-dominated sorting
and crowding distance assignment for identifying the non-dominated Pareto opti-
mal front easily. With the help of quantum-inspired mechanism incorporated with
NSGA-II [15], it is possible to increase the efficiency of the existing elitism method
as, for a single solution, there exists multiple observations of qubit individuals, which
allow a local search in the neighborhood of the non-dominated solutions. Further-
more, the possibility of losing high quality individuals can be reduced by selecting
the best qubit individuals in every generation. The convergence and preservation of
diversity being the main issues under scrutiny, the proposed algorithm is expected
to improve the performance of its counterpart when used for automatic clustering.
In this paper, an effort has been made to identify the optimal number of clusters
from a gray scale image on the run by implementing a quantum-inspired algorithm,
viz., Quantum-Inspired Multi-objective NSGA-II (QIMONSGA-II), which basically
performs quasi quantum computation and optimizes two different objectives, viz.,
CS-Measure (CSM) [7] and DB index [30] simultaneously. All the experiments have
been performed over six Berkeley [31] gray scale images of different sizes.
The rest of the paper is organized as follows: The fundamental concepts of quan-
tum computing are presented in Section 10.2. The computational processes of both
the objectives functions are presented in Section 10.3. Section 10.4 presents a brief
introduction regarding the multi-objective optimization along with the basic steps of
NSGA-II algorithm for automatic clustering of gray scale images. Section 10.5 pro-
vides an elaborate description of the proposed work followed by Section 10.6 with
the discussion on experimental results and analysis. Finally, the paper is concluded
with a conclusion and future guidance in Section 10.7.
10.2 QUANTUM COMPUTING FUNDAMENTAL

In recent years, quantum computing, a new computational paradigm has been in-
vented by Deutsch and Feynman [21][32] in 1982, which utilizes the fundamental
principles, viz., quantum bits or qubits, superposition, and entanglement, of quan-
tum mechanics for performing any calculation. Unlike classical computer, a quantum
computer uses quantum bits qubit as a basic computational unit. A single qubit may
have any one of the three states, viz., |1⟩, |0⟩ or any superposition state of these two
Quantum-Inspired Multi-Objective NSGA-II for Automatic Clustering 209
states, at the same time [21]. The state of a qubit can be represented as
|Ψ⟩ = α |0⟩ + β |1⟩ (10.1)
where α and β are complex numbers. The probability amplitudes and specify α2 β2
the probabilities of the state to be in |0⟩ or |1⟩, respectively. The superposition state
|Ψ⟩ can be realized by the following:

|0⟩ , if |α |2 > |β |2
|Ψ⟩ = (10.2)
|1⟩ , Otherwise
where α 2 and β 2 should always satisfy the following equation:
|α |2 + |β |2 = 1 (10.3)
10.3 COMPUTING THE OBJECTIVES

Two different cluster validity indices, viz., CS-Measure (CSM) [7] and DB index
[30], have been used as the optimization functions, which need to be optimize simul-
taneously.
10.3.1 CS-MEASURE (CSM) INDEX

In 2004, Chou et al. proposed a cluster validity index , viz., CS-Measure (CSM) [7]
index, which can efficiently handle the clusters of different densities and/or sizes by
computing the ratio of sum of within cluster scatter and between cluster separation. It
initially identifies the cluster centroid Ci from a data set by averaging the data points.
The cluster centroid Ci can be created by the following equation.
1
Ni Di∑
Ci = Di (10.4)
∈DSi
where the total number of data points (Di ) is represented by Ni in the ith cluster of the
data set DS . Now, the mathematical representation for the CSM index can be defined
by the following equation in which the data set DS contains NC number of clusters.
" #
1
NC
1

NC ∑ |Ni | ∑ max Di f f (Di , Dmx )
i=1 Di ∈ DSi Dmx ∈DSi
CSM = (10.5)
1
NC
NC ∑ min Di f f (Ci ,C j )
i=1 j∈NC , j̸=i
" #
NC
1

∑ |Ni | ∑ max Di f f (Di , Dmx )
i=1 Di ∈ DSi Dmx ∈DSi
∴ CSM = (10.6)
NC
∑ min Di f f (Ci ,C j )
i=1 j∈NC , j̸=i
where the difference between any two data points Di and D j is defined as
Di f f (Di , D j ). The optimal result is achieved for a minimum value of CSM index
[7].
10.3.2 DAVIES–BOULDIN (DB) INDEX

In the year 1979, David L. Davies and Donald W. Bouldin proposed a cluster validity
index, viz., Davies–Bouldin (DB) index [30]. It computes a similarity measure Ri j
between the clusters Ci and C j , which is basically defined based on a measure of
dispersion of a cluster Ci . Mathematically, the index Ri j is defined by the following
equation.
Si + S j
Ri j = (10.7)
Di j
where the dispersion measure of the ith cluster is represented by Si as follows:
" #1
1 Ni
2
2
∑ Xl − Zi
(i)
Si = (10.8)
Ni l=1
where the ith cluster center of a cluster Ci is represented by Zi and the total number
(i)
of objects is represented by Ni with all Xl belonging to Ci . The distance between
clusters Ci and C j is measured in terms of the cluster dissimilarity measure Di j , which
can be defined as follows:
Di j Zi − Z j (10.9)
Finally, the Davies Bouldin (DB) index [30] is defined by the following equation.
1 Nc
DB = ∑ Ri
Nc i=1
(10.10)
where Ri = max (Ri j ) , i = 1, 2, ..., Nc . While computing Ri j , the following

j=1,2,...,Nc ,i̸= j
constrains should be satisfied.
1. Ri j ≥ 0
2. Ri j = R ji
3. if Si = S j = 0 then Ri j = 0
4. if S j = Sk and Di j < Dik then Ri j > Rik
5. if S j > Sk and Di j = Dik then Ri j > Rik
The optimal result is achieved by minimizing the value of DB index [30].
10.4 MULTI-OBJECTIVE OPTIMIZATION

Multi-objective optimization is a process of identifying a set of solutions, called
the Pareto optimal set, by optimizing more than one objective functions simulta-
neously [33]. In many real-life problems, there exists no single optimal solution
for a given problem but may have a set of possible solutions of equivalent quality
[13]. Such problems are addressed by multi-objective optimization algorithms by si-
multaneously optimizing the multiple and possibly competing objective functions.
Unlike single-objective optimization problem, the goodness of a solution in a multi-
objective optimization problem is determined by the dominance of the solutions.
The multi-objective optimization can be formally stated as [33] follows: In gen-

eral, for a multi-objective optimization problem, M number of objectives are need
to be optimized simultaneously by the decision variable vector x̄∗ = [x1∗ , x2∗ , ..., xn∗ ]T
while satisfying m inequality and n equality constraints as follows:
gi (x̄) ≥ 0, i = 1, 2, ..., m (10.11)
and
hi (x̄) ≥ 0, i = 1, 2, ..., n (10.12)
and the optimization vector is
f (x̄) = [ f1 (x̄), f2 (x̄), ..., fM (x̄)]T (10.13)
While considering a maximization problem, a solution x̄i is said to dominate x̄i if the
following conditions are satisfied.
∀k ∈ 1, 2, ..., M, fk (x̄i ) ≥ fk (x¯j ) (10.14)
and
∃k ∈ 1, 2, ..., M, s.t. fk (x̄i ) > fk (x¯j ) (10.15)
The non-dominated set of solutions are those solutions, which are not dominated by
any member of the set of solutions. This non-dominated set of solutions is called the
Pareto optimal front.
10.4.1 NSGA-II
In the year 2002, a fast elitist Non-dominated Sorting Genetic Algorithm (NSGA-II)
[15] was proposed by Deb, et al. It provides a mechanism for better sorting and in-
corporates elitism and diversity preservation mechanism to improve the performance
of the NSGA algorithm [14]. It shows good performance in solving critical prob-
lems. The working principle of the NSGA-II algorithm for automatic clustering of
gray scale images is discussed as follows:
10.4.2 POPULATION INITIALIZATION AND CHROMOSOME REPRESENTA-

TION
Initially, a population P with N number of chromosomes has been produced by the
normalized values (between 0 and 1) of randomly selected pixel intensity values of
an input image. The length L of each chromosome has been taken as the square root
of the highest intensity value of the input image.
10.4.3 CREATING CLUSTER CENTROIDS

A predefined threshold value Th = 0.5 has been used to identify the active cluster
centroids from each chromosome. In a particular generation t, for any chromosome,
if the value of Th is greater than the normalized intensity value of Pi,t j , where i =
{1, 2, ..., N} and j = {1, 2, ..., L}, then that value is selected for the active cluster
center of that chromosome.
10.4.4 GENETIC OPERATION

In this chapter, in order to implement classical NSGA-II for automatic clustering of
gray scale images, three genetic operators, viz., tournament selection, conventional
crossover and mutation, have been used. The elaborate description of the different
genetic operations is available in [34]. The elitism operation has also been carried
out to identify the next-generation population. The last generation provides different
solutions, which are considered as the near-Pareto-optimal solutions.
10.4.5 FAST NON-DOMINATED SORTING

The new fitness values from both the objective functions have been evaluated after
performing the conventional crossover and mutation operation. At the end of these
operations, total 2N number of solutions has been generated. The first non-dominated
sorting is then done on these 2N number of solutions to create the near Pareto-optimal
front. A detailed explanation of the first non-dominated sorting procedure is available
in [34].
10.4.6 CROWDING DISTANCE

Once the non-dominated sort has been completed, the population is sorted according
to each objective function in an ascending order based on the computation of the
crowding distance. The computational details of the crowding distance calculation is
available in [34]. Thereafter, the first N number of solutions are considered from the
front to prepare the next-generation population Pt+1 .
10.4.7 BASIC STEPS OF CLASSICAL NSGA-II ALGORITHM FOR AUTOMATIC

CLUSTERING OF GRAY SCALE IMAGES
The experimental steps of classical NSGA-II for automatic clustering of gray scale
images are presented as follows:
1. A population Pt is created from the input gray scale image as described in Sec-
tion 10.4.2.
2. The active cluster centroids of all the chromosomes are identified belonging to
Pt as discussed in Section 10.4.3.
3. Both the fitness values FV 1 by Equation (10.6) and FV 2 by Equation (10.10)
are evaluated simultaneously for all the chromosomes as explained in Section
10.3.
4. The tournament selection is performed and the population is updated by main-
taining elitism as discussed in Section 10.4.4.
5. The conventional crossover operation is performed based on a predefined
crossover probability to produce new offspring NPt .
6. The conventional mutation is performed based on a predefined mutation proba-
S
bility over some randomly selected chromosomes belonging to Pt NPt .
S
7. After that, the fast non-dominated sorting is performed on Pt NPt to produce
near Pareto-optimal front as discussed in Section 10.4.5.
8. The crowding distance of all the elements from near Pareto-optimal front is
conducted to identify the first N number of chromosomes to produce the next-
generation population Pt+1 . Both the fitness values, FV 1, and FV 2, along with
their number of cluster centroids, corresponding to those solutions, are memo-
rized as explained in Section 10.4.6.
9. Steps 4 to 8 are repeated until the stopping criteria is met.
10. Finally, the obtained output is reported.
10.5 PROPOSED TECHNIQUE

The proposed quantum-inspired multi-objective NSGA-II (QIMONSGA-II) algo-
rithm for automatic clustering of gray scale images is elaborated below along with
the flowchart, presented in Figure 10.4.
10.5.1 QUANTUM STATE POPULATION INITIALIZATION

In the proposed algorithm, the quantum state population Q is created to encode the
original population P. Hence, Q consists of N number of strings and each string
stores L number of quantum state information. The quantum state population is cre-
ated by the realization of probabilistic representation of qubits [17]. A single qubit
is defined with a pair of number, (α , β ), as

α
(10.16)
β
where |α |2 + |β |2 = 1. In a particular t th generation, the quantum state population

(Qt ) is defined as
Qt = {qt1 , qt2 , ..., qtN } (10.17)
where N is the size of the population and qti , i = {1, 2, ..., N} represents each individ-
ual chromosome. Therefore, each chromosome with length L belonging to Qt can be
defined as t
α αi2
t ... αiL
t
qti j = i1 (10.18)
βi1t βi2t ... βiL
t
where i = {1, 2, ..., N}. Another way of representing Equation (10.18) is as follows:

cos θi1t cos θi2t ... cos θiL
t
qti j = (10.19)
sin θi1t sin θi2t ... sin θiL
t
where i = {1, 2, ..., N} and L is the length of each string.

Initially, the quantum state population Qt has been generated by random values
of θi,t j , where t = 0. Each chromosome belonging to original population Pt has been
encoded by some binary strings BtiL , where i = {1, 2, ..., N} and L is the length of
each string. The binary strings BtiL are then generated after observing the values of
Qt by the following equation.
( 2 2
t
Bi, j = 1, if β t
i, j > α t
i, j (10.20)
0, Otherwise.
where i = {1, 2, ..., N} and j = {1, 2, ..., L}.
10.5.2 CREATING CLUSTER CENTROIDS IN QUANTUM-INSPIRED FRAME-

WORK
The active cluster centroids are identified for each chromosome belonging to Pt by
observing the values of BtiL , s.t. {∀BtiL = 1}, i = {1, 2, ..., N} and L is the length of
each chromosome.
10.5.3 GENETIC OPERATORS IN QUANTUM-INSPIRED FRAMEWORK

Unlike NSGA-II, the proposed QIMONSGA-II uses quantum behaved selection,
followed by quantum behaved crossover and quantum behaved mutation. The next
generation is created from the non-dominated solutions among the parent and child
populations by these elitism operations. The elaborate description of the different
conventional genetic processes is available in [34]. Different solutions are identified
from the near-Pareto-optimal strings of the last generation.
10.5.3.1 Quantum-Behaved Selection

The concept of quantum rotation gate operation has been used for selecting the most
promising chromosome within a population. The quantum rotation gate operation has
been applied on each individual belonging to Qt in order to produce a new quantum
state population NQt for the same generation t. The mathematical description of
quantum rotation gate is given by

cos(∆θ ) − sin(∆θ )
R(δ θ ) = (10.21)
sin(∆θ ) cos(∆θ )
where ∆θ is a very small rotation angle which has been taken randomly between
[−0.5, 0.5] for updating the value of each qubit in Qt to produce NQt . It is depicted
by

α cos θ cos(∆θ ) − sin(∆θ ) cos θ
R(δ θ ) = R(δ θ ) =
β sin θ sin(∆θ ) cos(∆θ ) sin θ
′ (10.22)
cos(θ + δ θ ) α
= = ′
sin(θ + δ θ ) β
Figure 10.1 depicts the effect of quantum rotation gate operation, which is responsi-
ble for creating new quantum states.
Figure 10.1: Quantum rotation gate operation.
Now, each individual from NQt produces new binary string NBtiL , where i =
{1, 2, ..., N} and L is the length of each chromosome. Then, each string belonging
to NBt is used to evaluate the new fitness from both the objective functions of each
chromosome of Pt . Thereafter, the best solutions are identified between Qt and NQt
to update Qt along with Bt . An extensive explanation of the working principle of
quantum rotation gate operation is presented in [28][35].
10.5.3.2 Quantum-Behaved Crossover

The quantum-behaved crossover is performed on Qt to create NQt . The crossover
between any two individuals (say i and j, where (i, j) ∈ N and i ̸= j) from Qt has been
carried out depending upon a predefined crossover probability. Figure 10.2 depicts
as to how any two individuals are responsible for creating new individuals.
Figure 10.2: Quantum-behaved crossover.
10.5.3.3 Quantum-Behaved Mutation

The quantum-behaved mutation is implemented by utilizing the concept of Pauli-X
gate. The Pauli-X gate operation has been performed over some qubits depending
upon a predefined mutation probability, which reverses the probability amplitude
values of a single qubit in Qt . Mathematically Pauli-X gate (PX) can be defined as

0 1
PX = (10.23)
1 0
The Pauli-X gate operation on a single qubit, responsible for reversing the probability
amplitude values of that qubit, is demonstrated as follows:

α 0 1 α β
PX = = (10.24)
β 1 0 β α
Figure 10.3 depicts the effect of quantum behave mutation.

Figure 10.3: Quantum-behaved mutation.
10.5.4 FAST NON-DOMINATED SORTING IN QUANTUM-INSPIRED FRAME-

WORK
The new fitness values from both the objective functions are evaluated after perform-
ing the quantum-behaved crossover and mutation operations. At the end of these
operations, total 2N numbers of solutions are generated. Like classical NSGA-II
algorithm, the first non-dominated sorting is done on these 2N numbers of solu-
tions to create the near Pareto-optimal front. A detailed explanation of the fast non-
dominated sorting procedure is available in [34].
10.5.5 CROWDING DISTANCE COMPUTATION IN QUANTUM-INSPIRED

FRAMEWORK
Once the non-dominated sort has been completed, the crowding distance compu-
tation is required to sort the population according to each objective function in an
ascending order. The computational details of the crowding distance calculation are
similar to the classical NSGA-II algorithm, which is elaborately presented in [34].
Thereafter, the first N number of solutions have been considered from the front to
prepare the next-generation population Pt+1 and the corresponding qubits have been
chosen for Qt+1 .
10.5.6 QIMONSGA-II ALGORITHM FOR AUTOMATIC CLUSTERING OF

GRAY SCALE IMAGES
This section presents the proposed Quantum-Inspired Multi-Objective NSGA-II
(QIMONSGA-II) algorithm for automatic clustering of gray scale images.
Input Parameters
Maximum Generation:= Max G
Population Size := N
Crossover Probability := C p
Mutation Probability := µ p
Output Parameters
Optimum Cluster Number:= ONC
Optimum Fitness Value1 := FV 1
Optimum Fitness Value2 := FV 2
1. t ← 0
2. Create original population Pt from the input gray scale image as described in
Section 10.4.2
3. Create quantum state population Qt to encode Pt as discussed in Section 10.5.1.
4. Identify active cluster centroids of all the chromosomes belonging to Pt by the
guidance of Qt as described in Section 10.5.2.
5. Evaluate both the fitness values, FV 1 by Equation (10.6) and FV 2 by Equation
(10.10) simultaneously of all the chromosomes belonging to Pt as discussed in
Section 10.3.
6. Perform quantum-behaved rotation gate operation on Qt to create new NQt as
elaborated in Section 10.5.3.1.
7. Again identify active cluster centroids of all the chromosomes belonging to Pt by
the guidance of NQt .
8. Evaluate both the fitness values, FV 1 and FV 2, simultaneously of all the chromo-
somes belonging to Pt .
S
9. Identify N number of best solutions from Qt NQt by performing fast non-
dominated sorting followed by crowding distance calculation and update Qt by
those solutions along with their fitness values and number of cluster centroids.
10. Depending upon a predefined crossover probability C p , perform quantum-
behaved crossover operation on Qt to produce offspring NQt as elaborated in
Section 10.5.3.2.
11. Depending upon a predefined mutation probability µ p , perform quantum-behaved
mutation operation on some strings of Qt as discussed in Section 10.5.3.3.
12. Perform fast non-dominated sorting followed by crowding distance calculation to
generate the near Pareto-optimal front and thereafter consider the first N number
of solutions from the front to prepare the next-generation population Pt+1 with
its corresponding quantum state population Qt+1 . Memorize the corresponding
fitness values along with their number of cluster centroids as described in Sections
10.5.4 and 10.5.5 repeatedly.
13. t ← t + 1
14. a. If t < Max G then
b. Repeat Steps from 6 to 14
15. Finally, report the obtained output.
Figure 10.4: Flowchart of QIMONSGA-II algorithm for automatic clustering of gray

scale images.

This section presents an elaborate discussion of the experimental results of the pro-
posed quantum-inspired multi-objective NSGA-II algorithm for automatic clustering
of gray scale images along with its classical counter parts. All the experiments have
been conducted on a personal computer with Intel(R) Core (TM) i5-8250U proces-
sor, 8.00GB RAM and 1.60GHz CPU speed. The executable programs have been
developed in Windows 10 operating system and in Python language.
10.6.1 USED DATASET

In this paper, all the experiments have been conducted over six Berkeley [31] gray
scale images of different sizes. Among these six test images, the first three images,
viz., (a)#86000, (b)#92059 and (c)#89072 are of dimensions 321 × 481 and the last
three images, viz., (d)#86016, (e)#87046 and (f)#94079 are of dimensions 481 ×
321. All the test images are shown in Figure 10.5.
10.6.2 PARAMETER SETTINGS

The settings of the input parameters for both the algorithms, viz., QIMONSGA-II
and NSGA-II for automatic clustering of gray scale images are presented in Table
10.1.
10.6.3 PERFORMANCE EVALUATION

In this paper, the well-known Minkowski score [36] has been evaluated for under-
standing the results of clustering. Mathematically, the Minkowski score (MS ) for a
computed solution (Comp) is defined as follows:
M True − MComp
MS = (10.25)
M True
where MComp represents a binary matrix s.t. Mi,Comp

j = 1, iff i and j belong to the
same cluster of a computed solution (Comp) and M True represents the corresponding
matrix for the true solutions True. The better clustering results are achieved for the
smaller values of this score.
Additionally, the Silhouette index (SIL) [37] has been used to measure the per-
formance of clustering by the proposed algorithm. In the year 1987, Rousseeuw pro-
posed the Silhouette index (SIL) [37]. It compares the pairwise difference of between
and within cluster distances and is defined by
1 K
SIL = ∑ S(Ci )
K i=1
(10.26)
Figure 10.5: Original test images [31]: (a)#86000, (b)#92059, (c)#89072, (d)#86016,
(e)#87046, (f)#94079.
where the Silhouette width for the given cluster Ci is represented by S(Ci ). The value
of S(Ci ) is computed as
1 b(x) − a(x)
S(Ci ) = ∑ max(a(x), b(x))
Ni x∈C
(10.27)
i
Table 10.1
Input Parameters for QIMONSGA-II and NSGA-II [36]
Parameters QIMONSGA-II NSGA-II
Population Size : 50 50
Maximum Generation : 50 100
Crossover Probability : 0.8 0.8
1 1
Mutation Probability : ChromosomeLength ChromosomeLength
Small Rotation Angle : [-0.1 to 0.1] -
Table 10.2
Results of Mean Fitness Values from CSM [7] and DB [30]
Data Sets CVI QIMONSGA-II NSGA-II
#86000 CSM 0.29673 0.33841
DB 0.274e7 0.355e7
#92059 CSM 0.27115 0.36954
DB 0.335e7 0.248e7
#89072 CSM 5.07643 6.32973
DB 6.92591 7.19382
#86016 CSM 0.13794 0.26844
DB 1.35473 1.90035
#87046 CSM 0.30133 0.31472
DB 0.352e7 0.287e6
#94079 CSM 0.51692 0.39961
DB 8.19665 8.94837
where Ni is the number of patterns belonging to Ci , a(x) represents the within cluster
mean distance by averaging the distance between x and the rest of the patterns from
the same cluster, whereas, b(x) represents the smallest mean distance of x to the
patterns from another cluster. The SIL [37] index value generally lies between -1 to
1. The optimal result is achieved for the maximum value of SIL index.
10.6.4 EXPERIMENTAL RESULTS

Both the algorithms, viz., QIMONSGA-II and NSGA-II, have been executed for
30 runs. Table 10.2 presents the mean fitness values of both the objective func-
tions, viz., CSM [7] and DB [30] corresponding to the best solutions which have
been obtained from the final stage of non-dominated Pareto optimal front. The non-
dominated Pareto optimal front obtained by QIMONSGA-II and NSGA-II are pre-
sented in Figures 10.6 and 10.7, respectively.
The result of standard deviation (σ ), standard error (ε ), and optimal computa-
tional time (τ in second) by CSM [7] and DB [30] are presented in Table 10.3. After
Figure 10.6: Non-dominating Pareto optimal front of test images [31] (a)#86000,
(b)#92059, (c)#89072, (d)#86016, (e)#87046, (f)#94079 by QIMONSGA-II.
Figure 10.7: Non-dominating Pareto optimal front of test images [31] (a)#86000,
(b)#92059, (c)#89072, (d)#86016, (e)#87046, (f)#94079 by NSGA-II.
analyzing the results of Table 10.3, it is established that the proposed algorithm per-
forms better than its classical counterpart as most of the results are in favor of the
proposed algorithm.
While performing the unpaired t-test [38] between QIMONSGA-II and NSGA-II
[15], it is seen that eight results are “extremely significant,” two are “very significant”
and one is “significant” but one result is identified as “not significant” result. The
result of unpaired t-test [38] is presented in Table 10.4.
Table 10.3
Results of Standard Deviation (σ ), Standard Error (ε ) and Optimal computa-
tional Time (τ in second) by CSM [7] and DB [30]
Data Sets CVI QIMONSGA-II NSGA-II
σ ε τ σ ε τ
#86000 CSM 0.02682 0.00489 63 0.03783 0.00504 97
DB 0.01775 0.00324 85 0.02764 0.00573 118
#92059 CSM 0.00784 0.00143 54 0.05941 0.01084 72
DB 0.03965 0.00729 77 0.08451 0.01543 65
#89072 CSM 0.07834 0.01431 61 0.01919 0.00350 89
DB 0.23968 0.04375 67 0.38936 0.07108 94
#86016 CSM 0.16543 0.03020 45 0.25828 0.04715 94
DB 0.08164 0.01492 59 0.09739 0.01778 78
#87046 CSM 0.49472 0.09032 56 0.20398 0.03726 56
DB 0.03657 0.00667 83 0.05826 0.01064 126
#94079 CSM 0.21943 0.04006 38 0.22559 0.04186 61
DB 0.05828 0.01064 52 0.04694 0.00857 87
Table 10.4
Results of Unpaired t-test [38] between QIMONSGA-II and NSGA-II for CSM
[7] and DB [30]
Data Sets CVI QIMONSGA-II vs. NSGA-II
P −Value Significance Level
#86000 CSM <0.0001 Extremely Significant
DB <0.0001 Extremely Significant
DB 0.0022 Very Significant
#86016 CSM 0.0233 Significant
#87046 CSM 0.8915 Not Significant
#94079 CSM 0.0457 Significant
Table 10.5 presents the obtained values of MS [36] and SIL [37], which have
been used to compare the performance of the proposed algorithm with its classical
counterpart. While considering the score value of MS [36] and SIL, it is found that
all the test images excluding #87046 have scored better MS [36] score values and SIL
Table 10.5
Results of Performance Evaluation by MS [36] and SIL [37]
Data Sets Performance Metrics QIMONSGA-II NSGA-II
#86000 MS 0.50127 0.58982
SIL 0.65294 0.52369
#92059 MS 0.39421 0.43024
SIL 0.58392 0.44993
#89072 MS 0.36375 0.36333
SIL 0.70025 0.65047
#86016 MS 0.53285 0.59346
SIL 0.68327 0.63284
#87046 MS 0.47293 0.49211
SIL 0.45982 0.49328
#94079 MS 0.32691 0.43973
SIL 0.68485 0.47226
[37] values. This proves the supremacy of the proposed algorithm over its classical
counterpart.
Finally, the clustered images obtained from the proposed algorithm and its clas-
sical counterparts are presented in Figures 10.8 and 10.9, respectively. The corre-
sponding threshold values obtained by QIMONSGA-II, used for creating the clus-
tered images, are presented in Table 10.6.
10.7 DISCUSSIONS AND CONCLUSION

In this paper, a Quantum-Inspired Multi-Objective NSGA-II (QIMONSGA-II) algo-
rithm has been presented for automatic clustering of gray scale images. In this work,
the proposed algorithm has been compared with its classical counterpart. All the ex-
periments have been performed over six Berkeley gray scale images. The superiority
of the proposed algorithm over its classical counterpart has been justified by some
metrics, viz., computational time, mean fitness values obtained from both the cluster
validity indices, viz., CSM [7] and DB [30], standard deviation, and standard error.
Moreover, a statistical superiority tests, viz., unpaired t-test has been performed to
prove the efficacy of the proposed algorithm. Finally, the superiority of the proposed
algorithm has been proved with the help of Minkowski score [36] and Silhouette in-
dex (SIL) [37]. Though the proposed algorithm has been applied only on gray scale
Berkeley images, there still remains a scope of research for using them on true color
images. The authors are currently engaged in this direction.
Figure 10.8: Clustered images corresponding to test images [31]: (a)#86000,

(b)#92059, (c)#89072, (d)#86016, (e)#87046, (f)#94079, by QIMONSGA-II.
Figure 10.9: Clustered images corresponding to test images [31]: (a)#86000,

(b)#92059, (c)#89072, (d)#86016, (e)#87046, (f)#94079, by NSGA-II.
Table 10.6
Results of Number of Clusters and Threshold Values
Data Sets Number of Clusters Threshold Values
#86000 3 [55, 102, 180]
#92059 4 [46, 89, 119, 223]
#89072 3 [56, 120, 193]
#86016 4 [82, 140, 176, 209]
#87046 4 [75, 112, 127, 239]
#94079 5 [57, 103, 141, 168, 217]
REFERENCES
1. A. K. Jain and R. C. Dubes. Algorithms for Clustering Data. Prentice-Hall, Inc., USA,
1988.
2. A. K. Jain, M. N. Murty, and P. J. Flynn. Data clustering: A review. ACM Computing
Surveys, 31(3):264–323, 1999.
3. Ashwini Gulhane, Prashant Paikrao, and D. Chaudhari. A review of image data clustering
techniques. International Journal of Soft Computing and Engineering (IJSCE), 2(1):212–
215, 2011.
4. J. C. Platt, M. Czerwinski, and B.A. Field. Phototoc: Automatic clustering for browsing
personal photographs. In Fourth International Conference on Information, Communica-
tions and Signal Processing, 2003 and the Fourth Pacific Rim Conference on Multimedia.
Proceedings of the 2003 Joint, volume 1, pages 6–10, 2003.
5. J.H. Chen, Y.C. Chang, and W.L. Hung. A robust automatic clustering algorithm for prob-
ability density functions with application to categorizing color images. Communications
in Statistics – Simulation and Computation, 47(7):2152–2168, 2018.
6. T. Geraud, P. Strub, and J. Darbon. Color image segmentation based on automatic mor-
phological clustering. In Proceedings 2001 International Conference on Image Processing
(Cat. No.01CH37205), volume 3, pages 70–73, 2001.
7. T. Lei, P. Liu, X. Jia, X. Zhang, H. Meng, and A.K. Nandi. Automatic fuzzy clustering
framework for image segmentation. IEEE Transactions on Fuzzy Systems, 28(9):2078–
2092, 2020.
8. S. Bandyopadhyay and U. Maulik. Genetic clustering for automatic evolution of clusters
and application to image classification. Pattern Recognition, 35(6):1197–1208, 2002.
9. S. Das, A. Abraham, and A. Konar. Automatic clustering using an improved differential
evolution algorithm. IEEE Transactions on Systems, Man, and Cybernetics – Part A: Sys-
tems and Humans, 38(1):218–237, 2008.
10. A.E. Ezugwu. Nature-inspired metaheuristic techniques for automatic clustering: A survey
and performance study. SN Applied Sciences, 2, 2020.
11. A. Jose-Garca and W. Gomez-Flores. Automatic clustering using nature-inspired meta-
heuristics: A survey. Applied Soft Computing, 41:192–213, 2016.
12. S. Saha and S. Bandyopadhyay. A generalized automatic clustering algorithm in a multi-
objective framework. Applied Soft Computing, 13:89–108, 2013.
13. K. Suresh, D. Kundu, S. Ghosh, S. Das, and A. Abraham. Data clustering using multiob-
jective differential evolution algorithms. Fundamenta Informaticae, 97:381–403, 2009.
14. N. Srinivas and K. Deb. Muiltiobjective optimization using nondominated sorting in ge-
netic algorithms. Evolutionary Computation, 2(3):221–248, 1994.
15. K. Deb, A. Pratap, S. Agarwal, and T. Meyarivan. A fast and elitist multiobjective genetic
algorithm: Nsgaii. IEEE Transactions on Evolutionary Computation, 6:182–197, 2002.
16. D.W. Corne, N.R. Jerram, J.D. Knowles, and M.J. Oates. PESA-II: Region-based selection
in evolutionary multiobjective optimization. In Proceedings of the 3rd Annual Conference
on Genetic and Evolutionary Computation, GECCO01, pages 283–290, San Francisco,
CA, USA, 2001. Morgan Kaufmann Publishers Inc.
17. E. Zitzler and L. Thiele. Multiobjective evolutionary algorithms: A comparative case study
and the strength pareto approach. IEEE Transactions on Evolutionary Computation, vol. 3,
no. 4, 257–271, 1999.
18. M. Kim, T. Hiroyasu, M. Miki, and S. Watanabe. SPEA2+: Improving the performance
of the strength pareto evolutionary algorithm 2. In Xin Yao, Edmund K. Burke, Jose A.
Lozano, Jim Smith, Juan Julian Merelo-Guervos, John A. Bullinaria, Jonathan E. Rowe,
Peter Tino, Ata Kaban, and Hans-Paul Schwefel, editors, Parallel Problem Solving from
Nature – PPSN VIII, pages 742–751, Springer, Berlin, Heidelberg, 2004.
19. U. Maulik, S. Bandyopadhyay, and A. Mukhopadhyay. Multiobjective genetic algorithms
for clustering: Applications in data mining and bioinformatics. Springer Science & Busi-
ness Media, 2011.
20. Aimin Zhou, Bo-Yang Qu, Hui Li, Shi-Zheng Zhao, Ponnuthurai Nagaratnam Suganthan,
and Qingfu Zhang. Multiobjective evolutionary algorithms: A survey of the state of the art.
Swarm and Evolutionary Computation, 1(1):32–49, 2011.
21. T. Hey. Quantum computing: An introduction. Computing & Control Engineering Journal,
10:105–112, June 1999.
22. K.H. Han and J.H. Kim. Quantum-inspired evolutionary algorithm for a class of combina-
torial optimization. IEEE Transactions on Evolutionary Computation, 6(6):580–593, 2002.
23. T. Gandhi, Nitin, and T. Alam. Quantum genetic algorithm with rotation angle refinement
for dependent task scheduling on distributed systems. In 2017 Tenth International Confer-
ence on Contemporary Computing (IC3), pages 1–5. IEEE, Aug 2017.
24. H.P. Chiang, Y.H. Chou, C.H. Chiu, S.Y. Kuo, and Y.M. Huang. A quantum-inspired
tabu search algorithm for solving combinatorial optimization problems. Soft Computing,
18:1771–1781, 2013.
25. M. Ross and H. Oscar. A review of quantum-inspired metaheuristics: Going from classical
computers to real quantum computers. IEEE Access, 8:814–838, 2019.
26. C. Wojciech and K. Joanna. Quantum-inspired evolutionary approach for the quadratic
assignment problem. Entropy, 20(10):781, Oct 2018.
27. S. Dey, S. Bhattacharyya, and U. Maulik. Quantum inspired automatic clustering for multi-
level image thresholding. In 2014 International Conference on Computational Intelligence
and Communication Networks, pages 247–251, 2014.
28. A. Dey, S. Dey, S. Bhattacharyya, J. Platos, and V. Snasel. Novel quantum inspired ap-
proaches for automatic clustering of gray level images using particle swarm optimization,
spider monkey optimization and ageist spider monkey optimization algorithms. Applied
Soft Computing, 88(106040), 2020.
29. C.-H. Chou, M.-C. Su, and E. Lai. A new cluster validity measure and its application to
image compression. Pattern Analysis and Applications, 7(2):205–220, Jul 2004.
30. D.L. Davies and D.W. Bouldin. A cluster separation measure. IEEE Transactions on Pat-
tern Analysis and Machine Intelligence, 1:224–227, February 1979.
31. Berkley images. Accessed on 15/12/2020.
32. R. Blatt, H. Haiffner, C.F. Roos, C. Becher, and F. Schmidt-Kaler. Course 5 – quantum
information processing in ion traps. In Daniel EstOEeve, Jean-Michel Raimond, and Jean
Dalibard, editors, Quantum Entanglement and Information Processing, volume 79 of Les
Houches, pages 223–260. Elsevier, 2004.
33. S. Bandyopadhyay, S. Saha, U. Maulik, and K. Deb. A simulated annealing-based multi-
objective optimization algorithm: AMOSA. IEEE Transactions on Evolutionary Computa-
tion, 12(3):269–283, 2008.
34. K. Deb. Multiobjective Optimization Using Evolutionary Algorithms. Wiley, New York,
2001.
35. A. Dey, S. Dey, S. Bhattacharyya, J. Platos, and V. Snasel. Quantum inspired meta-heuristic
approaches for automatic clustering of color images. International Journal of Intelligent
Systems, 2021.
36. A. Mukhopadhyay, S. Bandyopadhyay, and U. Maulik. Clustering using multi-objective
genetic algorithm and its application to image segmentation. In 2006 IEEE International
Conference on Systems, Man and Cybernetics, vol. 3, pp. 2678–2683, 2006.
37. P.J. Rousseeuw. Silhouettes: A graphical aid to the interpretation and validation of cluster
analysis. Journal of Computational and Applied Mathematics, 20:53–65, 1987.
38. B. Flury. A First Course in Multivariate Statistics. Springer Texts in Statistics.
11 Conclusion
A metaheuristic is a heuristic (partial search) algorithm that is more or less an ef-
ficient optimization algorithm to real-world problems. Hybrid metaheuristics refer
to a proper and judicious combination of several other metaheuristics and machine
learning algorithms. The hybrid metaheuristics have been found to be more robust
and failsafe owing to the complementary character of the individual metaheuristics
in the resultant combination. This is primarily due to the fact that the vision of hy-
bridization is to combine different metaheuristics such that each of the combination
supplements the other in order to achieve the desired performance.
Quantum computer, as the name suggests, principally works on several quantum
physical features. These could be used as an immense alternative to today’s appo-
site computers since they possess faster processing capability (even exponentially)
than classical computers. A number of researchers coupled the underlying princi-
ples of quantum computing into various metaheuristic structures to introduce dif-
ferent quantum-inspired algorithmic approaches [1]–[5]. The evolution of the quan-
tum computing paradigm has led to the evolution of time efficient and robust hybrid
metaheuristics by means of conjoining the principles of quantum mechanics with
the conventional metaheuristics, thereby enhancing the real-time performance of the
hybrid metaheuristics.
This volume is a novel effort to bring together the recent advances and trends
in designing new and novel quantum-inspired metaheuristics to solve real-life prob-
lems in various branches of science and engineering. This volume introduces the
principles of quantum mechanics to evolve hybrid metaheuristics-based optimiza-
tion techniques useful for real-world engineering and scientific problems. Starting
from the introductory chapter, which presents an outline of the basic theory and con-
cepts pertaining to quantum-inspired metaheuristics, the chapter also throws light on
several types of quantum-inspired metaheuristics in details. It also comes up with
a bird’s eye view on the different bi-level/multi-level quantum system-based opti-
mization techniques. In addition to that, several entanglement-induced optimization
techniques and W-state encoding of optimization methods have also been discussed.
The applications related to the theme of the topic have been provided that would also
certainly be bring up to date the readers.
With the development of machine learning theory and the accumulation of prac-
tical experience of using various algorithms, it became clear that there is no ideal
classification method that would be better than all others for all sizes of the training
sample, for any percentage of noise in data, for any complexity of the boundaries of
dividing objects into classes, etc. Therefore, at present, ensemble classification meth-
ods that combine many different classifiers trained on different data samples. One of
the most accurate and fast parallelization methods available today is bagging, which
turns out to be useful in the case of heterogeneous classifiers and instability, when
small changes in the initial sample lead to significant changes in the classification.
DOI: 10.1201/9781003283294-11 231

Of late, several quantum-inspired methods for collective decision-making based on

metaheuristic quantum algorithms have been proposed. These algorithms are found
to increase the speed of combining decisions of basic classifiers.
As per definition, optimization is the technique of finding an alternative with the
most cost-effective or highest achievable performance under the given constraints, by
maximizing desired factors and minimizing undesired ones. Since many years, lot of
efforts have been incorporated to resolve the optimization problems, specifically the
NP-hard and multi-objective problems. It becomes slight easy after the invention of
quantum computer, which is very costly and sensitive. IBM makes it available freely
for the common people by introducing IBM Q Experience online platform [6]. The
IPB Q circuit composer and QISkit can be used to resolve optimization problems
using IBM Q.
Minimization of power loss in a DN is one of the challenging areas of research for
the distribution utilities. In recent times, power losses are reduced by implementing
DGs into distribution network. However, majority of research has been done on this
important optimization problem with CP load model. Majority of consumers at load
center uses VDLMs such as CZ, CC, IL, RL, and CL, whereas CP load model is inde-
pendent of voltage. If the optimal placement and capacity of DG with CP load model
are used on practical distribution system, it induces high power losses and poor volt-
age regulation in the system. In this study, an investigation has been performed to
reduce the losses in the distribution system with DG for different VDLMs. Optimal
location and capacity of DG is a difficult nondifferentiable, non-linear, complex com-
binatorial optimization problem. A Multipartite Adaptive Quantum-inspired Evolu-
tionary Algorithm is proposed for optimal location and sizing of DG. MAQiEA uses
probabilistic approach with Q-bits, and is an updated version of AQiEA that has in-
troduced two Q-bits per solution vector and entanglement inspired adaptive crossover
operator. MAQiEA has introduced a Multipartite Adaptive Crossover operator as a
variation operator for better convergence.
A Quantum-Inspired Manta Ray Foraging Optimization algorithm can be evolved
for automatic clustering of color images. A novel quantum rotation gate and Pauli-X
gate strategy is used to achieve the exploration and exploitation. Several tests con-
ducted among the competitive algorithms establish the effectiveness of the proposed
algorithm.
A Quantum Genetic Algorithm (QGA) can be evolved to perform an automatic
feature Cancer Modelling and Simulation selection in order to select only those fea-
tures that has a strong influence on a Support Vector Machine-based (SVM) classi-
fier. The QGA performs a search over the space formed by the feature set in order to
find an optimal combination of features and at the same time, keeping or decreasing
the loss rate in the training stage. The test image database is balanced in terms of the
positive and negative stenosis cases. After the feature selection process ends, a subset
of the features is able to keep the classification rate in terms of the accuracy metric
and the Jaccard index, compared with the original set with all the features. In addi-
tion, the reduction of features has effect on the time required to perform an exhaus-
tive feature extraction of new angiograms. The proposed method can be applied in
Conclusion 233
clinical practice to assists cardiologists in the evaluation and finding of possible

stenosis cases in X-ray coronary angiograms.
A Hybrid Quantum-Convolutional Neural Network method for atherosclerosis
detection in XCA images can be envisaged. It includes a Quantum Convolutional
Layer used as a preprocessing to improve the atherosclerosis detection performance
of typical CNNs. Numerical experiments, based on two distinct CNN architectures
(DenseNet-based and VGG-based architectures) and trying on two different train-
ing algorithms: Stochastic Gradient Descent and Stochastic Gradient Descent with
Momentum, have demonstrated that using a Quantum Convolutional Layer on a lim-
ited XCA dataset performs efficiently for atherosclerosis detection. The introduced
hybrid methodology improved the five evaluation metrics concerning the classical
CNN architectures trained with the normalized-raw XCA images.
Both qubit- and qutrit-based Elephant Herd Optimization algorithms are efficient
for automatic clustering of hyperspectral images. A modified rotation gate opera-
tion enhances the diversity of the population. The exploration and exploitation ca-
pabilities of the classical EHO algorithm are enhanced, with a faster convergence
capability. As automatic cluster detection is a tedious task in HSI processing, these
algorithms can be highly beneficial in real-life scenarios.
A novel Multiobjective Quantum-inspired Salp Swarm Algorithm (MQSSA) with
Delta potential-well model presentation, a better alternative than the binary presen-
tation for the multiobjective optimization problems, can be envisaged. The proposed
approach is evaluated on several multiobjective optimization benchmark problems
having convex-shaped and concave-shaped Pareto-optimal front. It shows a favor-
able outcome, with better performance than other well-regarded algorithms in the
multiobjective domain.
Finally, a Quantum-Inspired Multi-Objective NSGA-II (QIMONSGA-II) algo-
rithm has been presented for automatic clustering of gray scale images. The proposed
algorithm has been compared with its classical counterpart. All the experiments have
been performed over six Berkeley gray scale images. The superiority of the proposed
algorithm over its classical counterpart has been justified by some metrics, viz., com-
putational time, mean fitness values obtained from two cluster validity indices, viz.,
CSM [7] and DB [8], standard deviation, and standard error. This volume is expected
to come to the benefit of senior researchers, practitioners, and aspiring researchers,
who intend to work on quantum metaheuristic foundations and applications.
REFERENCES
1. Wang, L. & Niu, Q. & Fei, M. R. (2008). A novel quantum ant colony optimization algo-
rithm and its application to fault diagnosis. Transactions of the Institute of Measurement
and Control, 30(3–4), 313–329.
2. Dey, S. & Bhattacharyya, S. & Maulik, U. (2013). Quantum-inspired metaheuristic algo-
rithms for multi-level thresholding for true colour images. In Proceeding of 2013 Annual
IEEE India Conference (INDICON).
3. Dey, S. & Bhattacharyya, S. & Maulik, U. (2017). Efficient quantum-inspired metaheuris-
tics for multi-level true colour image thresholding. Applied Soft Computing, 56, 472–513.
4. Dey, S. & Bhattacharyya, S. & Maulik, U. (2016). New quantum-inspired metaheuristic

techniques for multi-level colour image thresholding. Applied Soft Computing, 46, 677–
702.
5. https://qiskit.org/
6. Mishra, N. & Bisarya, A. et. al. (2020). Breast Cancer Detection Using Quantum Convo-
lutional Neural Networks: A Demonstration on a Quantum Computer. medRxiv.
7. Chou, C.-H. & Su, M.-C. & Lai, E. (2004). A new cluster validity measure and its applica-
tion to image compression. Pattern Analysis and Applications, 7(2), 205–220.
8. Davies, D. L. & Bouldin, D. W. (1979). A cluster separation measure. IEEE Transactions
on Pattern Analysis and Machine Intelligence, 1, 224–227.
A Automatic
Selection
Feature
for Coronary Stenosis

Detection in X-Ray
Angiograms Using
Quantum Genetic
Algorithm
A.1 MATLAB CODE TO EXTRACT VESSEL SEGMENTS
function segments = findsegments(m, endpoints)
%Find Segments Function
% Find the segments in a logical matrix using the
% information present in the neighborhood of a
% center element.
%------------------------------------------------
%Artifact: findsegments.m
%Version: 1.0
%Date: 15/03/2020 12:20:00
%Author: Miguel Angel Gil Rios
%Email: [email protected]
%------------------------------------------------
%Usage:
% segments = findsegments(m, endpoints)
%
%Inputs:
% m: The logical matrix
% endpoints: A Nx2 matrix containing the N
% endpoints positions which this
% function will try to follow a
% connection path between them.
%------------------------------------------------
DOI: 10.1201/9781003283294-A 235

%Outputs:
% segments: A Nx4 matrix containing the
% segment path located by the row
% and column index. Third column
% represents the corresponding
% pixel slope. Fourth column
% represents the resultant angle
% in order to search structure
% borders.
%------------------------------------------------
segments = {};
k = 0;
% Check if at least two end-points exists in the

% endpoints matrix:
if size(endpoints, 1) < 2 || size(endpoints, 2) ˜= 2
return;
end
bp = zeros(0);
%Find segments while there exists endpoints:
while size(endpoints, 1) > 1
%Take the first endpoint:

start_point = endpoints(1, :);
%Find the segment:

[current_segment, m, end_point, bp] = ...
findsegment(m, start_point(1), start_point(2), bp);
%Check if a segment was formed:

if end_point ˜= start_point
%Store the segment:

k = k + 1;
segments{k}.points = current_segment;
%Check if end_point exists in endpoints matrix:

pos = frindex(m, end_point);
if pos > 0
%Remove end-point:
endpoints(pos, :) = [];
else
%Replace end point:
endpoints(1, :) = end_point;
end
else
%Remove end-point:
endpoints(1, :) = [];
end
end
%After segments were computed, each pixel angle

%must be calculated.
%For each segment
for k = 1 : size(segments, 2)
v1 = segments{k}.points(1, :);
v2 = segments{k}.points(2, :);
[slope, angle] = computeSlope(v1, v2);
segments{k}.points(1, 3) = slope;
segments{k}.points(1, 4)= angle;
for kp = 2 : size(segments{k}.points, 1) - 1
v1 = segments{k}.points(kp - 1, :);
v2 = segments{k}.points(kp, :);
v3 = segments{k}.points(kp + 1, :);
[slope, angle] = computeSlope(v1, v2, v3);
segments{k}.points(kp, 3) = slope;
segments{k}.points(kp, 4) = angle;
end
i2 = size(segments{k}.points, 1);
i1 = i2 - 1;
v1 = segments{k}.points(i1, :);
v2 = segments{k}.points(i2, :);
[slope, angle] = computeSlope(v1, v2);
segments{k}.points(i2, 3) = slope;
segments{k}.points(i2, 4) = angle;
end
end %End Function
function [slope, angle] = computeSlope(v1, v2, v3)

%
%˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜
% S1 = 0, S2 = 0
% m=0 m=0 angle=0 Degrees
%[v1]----[v2]----[v3]
%˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜
% S1 = 0, S2 = 1
% m=0
% [v3] [v1]----[v2]
% / / angle=45 Degrees
% / m=1 / m=1
% [v1]----[v2] [v3]
% m=0
%˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜
% S1 = 0, S2 = Inf
% m=0
% [v3] [v1]----[v2]
% | m=Inf | m=Inf angle=90 Degrees
% | |
%[v1]----[v2] [v3]
% m=0
%˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜
% S1 = 0, S2 = -1
% m=0
% [v3] [v1]----[v2]
% \ m=-1 \ m=-1 angle=135 Degrees
% \ \
%[v1]----[v2] [v3]
% m=0
%˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜
% S1 = Inf, S2 = Inf
% [v1]
% | m=Inf
% | angle=0 Degrees
% [v2]
% | m=Inf
% |
% [v3]
%˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜
% S1 = Inf, S2 = 0
% [v1] [v1]
% | m=Inf |m=Inf
% | | angle=90 Degrees
% [v2]----[v3] [v3]----[v2]
% m=0 m=0
%˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜
% S1 = Inf, S2 = 1
% [v1] [v3] [v1]
% m=Inf | / | m=Inf
% | / m=1 | angle=45 Degrees
% [v2] [v2]
% /
% / m=1
% [v3]
%˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜
% S1 = Inf, S2 = -1
% [v3][v1] [v1]
% m=-1 \ |m=Inf | m=Inf
% \| | angle=135 Degrees
% [v2] [v2]
% \
% \ m=-1
% [v3]
%˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜
% [v1]
% \ m=-1
% \
% [v2]
% \ m=-1
% \
% [v3]
%˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜
% [v1]
% / m=1
% /
% [v2]
% / m=1
% /
% [v3]
%
s1 = getSlope(v1, v2);
if ˜exist(’v3’)
slope = s1;
switch slope
case 0
angle = 90;
case 1
angle = 135;
case -1
angle = 45;
case Inf
angle = 0;
otherwise
angle = 99;
end
else
s2 = getSlope(v2, v3);
if s1 == 0 && s2 == 0
slope = 0;
angle = 90;
elseif (s1 == 0 && s2 == 1) || (s2 == 0 && s1 == 1)
slope = 0;
angle = 90;
elseif (s1 == 0 && s2 == Inf) || (s2 == 0 && s1 == Inf)
slope = 1;
angle = 135;
elseif (s1 == 0 && s2 == -1) || (s2 == 0 && s1 == -1)
slope = 0;
angle = 90;
elseif s1 == Inf && s2 == Inf
slope = Inf;
angle = 0;
elseif (s1 == Inf && s2 == 0) || (s2 == Inf && s1 == 0)
slope = 1;
angle = 45;
elseif (s1 == Inf && s2 == 1) || (s2 == Inf && s1 == 1)
slope = Inf;
angle = 0;
elseif (s1 == -1 && s2 == Inf) || (s2 == -1 && s1 == Inf)
slope = Inf;
angle = 0;
elseif s1 == -1 && s2 == -1
slope = -1;
angle = 45;
elseif s1 == 1 && s2 == 1
slope = 1;
angle = 135;
else
slope = 99;
angle = 99;
end
end
end
function slope = getSlope(v1, v2)

slope = (v2(1) - v1(1)) / (v2(2) - v1(2));
end
A.2 MATLAB CODE TO FIND PIXEL POSITIONS

function [points, m_result, end_point, bp] = findsegment(m, row,
col, bp)
%Find Segment Function
% Search and detect all possible pixels that are part of a
$segment.
%---------------------------------------------------------------
%Artifact: findsegment.m
%Version: 1.0
%Date: 15/03/2020 12:35:00
%---------------------------------------------------------------
%Usage:
% [points, m_result, end_point] = findsegment(m, row,
% col, bp)
%
%Inputs:
% m: A Logical matrix.
% row: The row index where the start point pixel is located.
% col: The column index where the start point pixel is located.
% bp: An optional array of Nx2 indicating the current
% found branch points locations (row, col).
%
%Outputs:
% points:A Nx2 matrix containing the positions (row,
$ column) with pixels that are part of the segment.
% m _result:A copy of the m input matrix with zeros in the
% positions that were identified as part of the segment.
% end_point:A 1x2 vector containing the position (row,
% col) of the segment end reached by the function.
% Segment end can be reached under next conditions:
% 1. No more ways to explore are available
% (all neighborgs are 0).
% 2. We fall in a branch point from which a non unique
% path is possible to follow.
% 3. We reach some of the matrix boundaries and
% there is no more remaining positions to explore.
% However, you can assume additional considerations:
% 1. If row and col values are similar to those stored
% in end_point there is probably to exists
% only an isolated pixel but not a segment.
% 2. You can check if end_point values are similar to
% another end-point that you are keeping in your
% computing code and you could consider to remove it.
%
%---------------------------------------------------------------
points = zeros(0);
m_result = m;
end_point(1, 1:2) = -1;
row_current = row;
col_current = col;
k = 0;
flag = true;
if ˜exist(’bp’)
bp = zeros(0);
end
%We reach the end of the segment if (flag == false) under next
%conditions:
% 1. No more ways to explore are available(all neighborgs
% are 0).
% 2. We fall in a branch point from which a non unique path
% is possible to follow.
% 3. We reach some of the matrix boundaries and there is no
% more remaining positions to explore.
first_time = true;
while flag == true
k = k + 1;
points(k, 1) = row_current;
points(k, 2) = col_current;
%The last point (end-point) explored is (row_current,

% col_current):
end_point(1, 1) = row_current;
end_point(1, 2) = col_current;
%Extract a 3x3 window centered at (row_current,

% col_current):
sw = eswm(m_result, row_current, col_current, 3, 3, true);
sw(2, 2) = 0;
sum_pos = sum(sw(:));
if sum_pos == 0 %First flag case
if row ˜= row_current || col ˜= col_current
if frindex(points, [row_current col_current]) == 0
points(k + 1, 1) = row_current;
points(k + 1, 2) = col_current;
end
m_result(row_current, col_current) = 0;
end
flag = false;
else
%Find the positions from 1 to 8 where is posible to
% follow a path:
vpos = findpixdirs(sw);
pos = find(vpos == 1);
%If only one direction is possible

if sum_pos == 1
m_result(row_current, col_current) = 0;
%Check if current position falls in a branch point:

pos_bp = frindex(bp, [row_current, col_current]);
if pos_bp > 0 && ˜first_time
flag = false;
else
% Find the next position to explore:
switch pos(1, 1)
case 1
col_current = col_current + 1;
case 2
row_current = row_current - 1;
case 3
case 4
col_current = col_current - 1;
case 5
case 6
row_current = row_current + 1;
case 7
case 8
end
%Check if new position is out of matrix boundaries:

%Third flag case
flag = row_current > 0 & row_current <= size(m_result,
1) & ... col_current > 0 & col_current <= size(m_result,
2);
end
%Second flag case: we fall in a branch point

else
%Add the position of branch point to the bp array:
if frindex(bp, [row_current col_current]) < 1
bp(size(bp, 1) + 1, :) = [row_current col_current];
end
flag = false;
end
end
first_time = false;
end
end
A.3 MATLAB CODE TO EXTRACT A WINDOW FROM A MATRIX

function subwindow = eswm(m, ri, ci, h, w, c)
% ESWM Extract a subwindow from a matrix.
%
%
% subwindow = eswm(m, ri, ci, h, w, c)
% Extracts a subwindow of size h*w.
% Params:
% m: The matrix fom which subwindow will be extracted.
% ri: The row index corresponding to the interest point.
% ci: The column index corresponding to the interest point.
% h: The number of points that will be taken for the
% subwindow vertically.
% w: The number of points that will be taken for the
% subwindow horizontally.
% c: A logical value indicating if the interest point will
% be located at the center of the sub-window (true) or it
% will be located at the upper left corner of the
% sub-window (false).
% If the intetest point will be located at the center
% of the sub-window, the h and w values will be scaled
% to the nearest upper odd value in order to take:
% floor(h/2) points above of the interest point
% floor(h/2) points bellow of the interest point
% floor(w/2) points at left of the interest point

% floor(w/2) points at right of the interest point
%
if ˜exist(’c’,’var’)
c = false;
end
if c == true
subwindow = eswm_center (m, ri, ci, h, w);
else
subwindow = eswm_top_left(m, ri, ci, h, w);
end
end
function subwindow = eswm_top_left(m, ri, ci, h, w)

if (ri + h) > size(m, 1) || (ci + w) > size(m, 2)
subwindow = zeros(h, w);
if (ri + h) > size(m, 1)
h = size(m, 1) - ri;
end
if (ci + w) > size(m, 2)
w = size(m, 2) - ci;
end
for i = 1 : h + 1
for j = 1 : w + 1
subwindow(i, j) = m(ri + (i - 1), ci + (j - 1));
end
end
else
subwindow = m((ri : ri + (h - 1)), (ci : ci + (w - 1)));
end
end
function subwindow = eswm_center(m, ri, ci, h, w)

if mod(h, 2) == 0
h = h + 1;
end
if mod(w, 2) == 0
w = w + 1;
end
subwindow = zeros(h, w);
r_ini = ri - floor(h / 2);

for i = 1 : h
c_ini = ci - floor(w / 2);
for j = 1 : w
if r_ini > 0 && c_ini > 0 && r_ini <= size(m, 1)
&& c_ini <= size(m, 2)
subwindow(i, j) = m(r_ini, c_ini);
end
c_ini = c_ini + 1;
end
r_ini = r_ini + 1;
end
end
A.4 MATLAB CODE TO FIND ROW VECTOR

function row_index = frindex(m, vector)
%Find Row Index
% Find the index where the values of a row vector are the
% same inside a matrix.
%-----------------------------------------------------------
%Artifact: frindex.m
%Version: 1.0
%Date: 17/03/2020 20:05:00
%-----------------------------------------------------------
%Usage:
% row_index = frindex(m, vector)
%
%Inputs:
% m: A numerical matrix.
% vector: The row vector with the values to be found
% inside the matrix.
%
%Outputs:
% row_index: An Integer representing the row number
% where the vector values were coincident
% inside the matrix. If no coincident row
% vector was found, 0 is returned. If there
% exists more than one coincidence, the first
% occurence index is returned.
%-----------------------------------------------------------
row_index = 0;
for k = 1 : size(m, 1)
if isequal(m(k, :), vector)
row_index = k;
k = size(m, 1) + 1;
end
end
end
Index
W state, 8, 9 crossover, 212, 215
t-test, 107 crowding distance, 208
F score, 182 Cuckoo search algorithm, 42
69 bus system, 76
Delta Potential-well Model, 192, 195
Quantum Ant Colony Multi-Objective Differential Evolution, 1, 2, 6
Routing, 13 Distributed Generator, 58
Activation function, 152 Economic Load Dispatch, 13

Ant Colony Optimization, 1, 2, 5 Elephant Herd Optimization, 172
Ant colony optimization (ACO), 41 elitism, 214
Artificial neural network (ANN), 43 ensemble, 21, 22, 27, 28, 32, 33
Artificial Neural Networks, 2 Entanglement, 43, 147
Atherosclerosis, 141, 142 Entanglement-QTS, 8
automatic clustering, 95–97, 100, 207 Evolution Strategies, 6
Evolutionary Algorithm, 2
bagging, 21, 22, 32 Evolutionary Computation, 6
Bat algorithm, 42 Evolutionary Computing, 2
Bee colony optimization (BCO), 41
bi-level optimization, 9 feature extraction, 120, 130
Binary Cross-Entropy, 157 feature selection, 124
Bloch sphere, 43 Frangi method, 123, 124
Branch Current to Bus Voltage, 66 Friedman test, 107
Bus Injection to Branch Current, 66 Fully connected layer, 151
Fuzzy logic controller (FLC), 42
CIRCUIT COMPOSER, 45 Fuzzy system, 42
Circuit composer, 46, 53 Fuzzy Systems, 2
Clan Updation, 174
classifier, 21, 22, 24, 27–30, 32, 33 Genetic Algorithm, 2, 6
cluster validity index, 209 Genetic algorithm (GA), 40
Clustering, 171 Genetic quantum algorithm (GQA), 44
clustering, 95, 96 Greedy Randomized Adaptive Search
color image thresholding, 10 Procedure, 5
Combinatorial Optimization, 5 Grover’s adaptive search (GAS), 53
commercial load, 60 Grover’s search algorithm, 2
constant current, 60 Guided Local Search, 4
constant impedance, 60
Constructive metaheuristics, 4 Hybrid metaheuristics, 7
Convolutional layer, 150 Hybrid Quantum-Convolutional Neural
Convolutional Neural Network, 142, 143, Network, 145
149 IBM Q EXPERIENCE, 45
coronary stenosis, 117, 120, 125 IBM quantum computer, 53
249
250 INDEX
IBMs quantum computing, 43, 44 Non-nature inspired metaheuristics, 2

image thresholding, 9 Nurse scheduling problem (NSP), 45
industrial load, 60
Interference, 43 Optimization, 1, 9
Iterated Local Search, 2, 4
Pareto optimal, 210
Jupyter notebook, 48 Pareto-optimal Set, 201
Particle Swarm Optimization, 1, 2
Kruskal-Wallis test, 182 Particle swarm optimization (PSO), 41
Path Relinking, 6
learning, 21, 25, 27, 31–33 Pauli-X, 96, 98, 101, 102, 108
Local search, 3 Pauli-X gate, 216
Local search metaheuristics, 3 PBM index, 99
Pooling layer, 150
manta ray, 100, 101
Population-based metaheuristics, 6
mean, 181
Population-based search, 2
Measurement, 149
power losses, 59
Memory-based methods, 2
Memory-less methods, 2 QGA, 126, 128–131, 134, 136
Metaheuristic, 1, 2 QISKIT, 47
metaheuristic, 21, 30–33, 95–97 QISKit, 53
Metaheuristic Algorithm, 190 Qiskit Aqua, 51
Metaheuristic method, 60 Qiskit Terra, 51
Metaheuristics, 2 Quadratic unconstrained binary opti-
morphologic features, 122, 131 mization (QUBO), 45
Multi-agent Metaheuristic Architecture, quantum, 172
7 Quantum adiabatic algorithm (QAA), 45
Multi-agent Metaheuristic Hybridization, Quantum annealing algorithm, 45
7 Quantum Binary Particle Swarm Opti-
multi-level thresholding, 10 mization, 9
multi-objective, 207, 210 quantum bits, 3
Multi-objective optimization, 1 Quantum Circuits, 144, 149
Multi-objective optimization problems Quantum Computing, 2, 194
(MOOPs), 38 Quantum computing, 43
Multi-start Local Search, 4 quantum computing, 3
Multiobjective Optimization, 189, 199, Quantum Convolutional Layer, 142, 144,
201 154, 159
Multipartite Adaptive Quantum-inspired Quantum entanglement, 7
Evolutionary algorithm, 58 Quantum gates, 147
mutation, 212, 216, 217 quantum gates, 97
Quantum Genetic Algorithm, 126, 136
Nature-inspired metaheuristics, 2
Quantum Inspired Ant Colony Optimiza-
Network Reconfiguration, 58
neural networks, 22, 27 tion, 10
Neuro-fuzzy systems, 43 Quantum Inspired Metaheuristics, 3
non-dominated, 208, 214 Quantum logic gates, 44
INDEX 251
Quantum Machine Learning, 142 Transfer Learning, 143

Quantum processing unit (QPU), 45 Travelling salesman problem (TSP), 52
quantum rotation gate, 214
Quantum system, 147 Uncertainty, 38
Quantum tabu search (QTS), 52
quantum-inspired, 21, 31–33, 96, 172, Variational quantum eigensolver (VQE),
208 52
Quantum-inspired Algorithm, 191 Vehicle routing problem (VRP), 52
Quantum-inspired evolutionary algorithms, voltage dependent load model, 60
7 Voltage Stability Index, 63
Quantum-Inspired Simulated Annealing,
Wireless Sensor Network, 13
10
Quantum-inspired tabu search, 8 X-ray Coronary Angiography, 141, 143
Qubit, 43, 145 Xie-Beni Index, 172
qubit, 98, 208, 213, 216
Qubit Elephant Herd Optimization, 172
qubits, 3, 51
Qutrit Elephant Herd Optimization, 172
residential load, 60
rotation, 96, 98, 102, 108
rotation gates, 179
Scatter Search, 6
sensitivity analysis, 96, 106
Shor’s factorization algorithm, 2
Simulated Annealing, 1, 4
Simulated annealing (SA), 41
Single objective optimization, 1
Single objective optimization problems
(SOOPs), 38
Single point based search, 2
Single point-based search, 2
standard deviation, 181
Stenosis, 141
Stochastic Gradient Descent, 142, 157
Stochasticity, 38
Superposition, 43, 145
Support Vector Machine, 125
SVM, 125, 137
Swarm Optimization, 2
Tabu Search, 1, 2, 4
Tabu search, 52
texture features, 121, 130

(Quantum Machine Intelligence) Siddhartha Bhattacharyya, Mario Köppen, Elizabeth Behrman, Ivan Cruz-Aceves - Hybrid Quantum Metaheuristics - Theory and Applications-CRC Press (2022)

Uploaded by

Copyright:

Available Formats

(Quantum Machine Intelligence) Siddhartha Bhattacharyya, Mario Köppen, Elizabeth Behrman, Ivan Cruz-Aceves - Hybrid Quantum Metaheuristics - Theory and Applications-CRC Press (2022)

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

(Quantum Machine Intelligence) Siddhartha Bhattacharyya, Mario Köppen, Elizabeth Behrman, Ivan Cruz-Aceves - Hybrid Quantum Metaheuristics - Theory and Applications-CRC Press (2022)

Uploaded by

Copyright:

Available Formats

Hybrid Quantum

Quantum Machine Intelligence Series

Siddhartha Bhattacharyya, Rajnagar Mahavidyalaya, Birbhum, India

Elizabeth C. Behrman, Wichita State University, USA

Hybrid Quantum Metaheuristics

Quantum Machine Intelligence

For more information about this series, please visit: https://www.rout-

First edition published [2022]

and by CRC Press

CRC Press is an imprint of Taylor & Francis Group, LLC

ISBN: 9780367751562 (hbk)

Typeset in Nimbus font CMR10

Siddhartha would like to dedicate this volume to late Prof. John

Mario would like to dedicate this book to the Schrödinger’s cat.

Chapter 1 An Introductory Illustration to Quantum-Inspired Metaheuristics.. 1

Chapter 2 A Quantum-Inspired Approach to Collective Combine Basic

Chapter 3 Function Optimization Using IBM Q............................................37

Chapter 4 Multipartite Adaptive Quantum-Inspired Evolutionary Algorithm

Chapter 5 Quantum-Inspired Manta Ray Foraging Optimization Algorithm

5.3 Fundamentals of Quantum Computing.................................97

Chapter 6 Automatic Feature Selection for Coronary Stenosis Detection in

Chapter 7 Quantum Preprocessing for DCNN in Atherosclerosis

7.3.2.4 Activation Functions.............................. 152

Chapter 8 Multilevel Quantum Elephant Herd Algorithm for Automatic

Chapter 9 Toward Quantum-Inspired SSA for Solving Multiobjective

9.3.1 Delta Potential-Well Model for SSA ..................... 195

Chapter 10 Quantum-Inspired Multi-Objective NSGA-II Algorithm for

10.6.1 Used Dataset .......................................................... 220

Chapter 11 Conclusion................................................................................... 231

Appendix A Automatic Feature Selection for Coronary Stenosis Detection in

Dr. Elizabeth Behrman earned her Bachelor’s degree

Mario Köppen studied Physics at the Humboldt-

Ivan Cruz-Aceves received a PhD in Electrical Engi-

quantum-inspired metaheuristics. A brief summary of entanglement-induced opti-

number of clusters in color images by using a quantum-inspired framework incorpo-

Dora Elisa Alvarado-Carrillo Debashis De

D K Chaturvedi Sergey Gorbachev

Soumyajit Goswami Sanjai Pathak

Martha Alicia Hernandez-Gonzalez Jan Platos

1.2 QUANTUM-INSPIRED METAHEURISTICS

1.2.1 LOCAL SEARCH METAHEURISTICS

1.2.2 CONSTRUCTIVE METAHEURISTICS

There exists few popular constructive metaheuristics in the literature, out of

1.2.3 POPULATION-BASED METAHEURISTICS

1.2.4 HYBRID METAHEURISTICS

1.3 ENTANGLEMENT-INDUCED OPTIMIZATION

1.4 W-STATE ENCODING OF OPTIMIZATION ALGORITHMS

The entanglement feature is an encouraging topic in the ﬁeld of research now

1.5 QUANTUM SYSTEM-BASED OPTIMIZATION

1.5.1 BI-LEVEL QUANTUM SYSTEM-BASED OPTIMIZATION

et al. [61] proposed an elite quantum-inspired PSO (EQPSO) algorithm, where an

1.5.2 MULTI-LEVEL QUANTUM SYSTEM-BASED OPTIMIZATION

1.6 APPLICATIONS OF QUANTUM-INSPIRED METAHEURISTICS

quantum-inspired genetic algorithm-based FCM [82][83]. Liu et al. [84] presented

QEA. Konar et al. [96] proposed a Hybrid Quantum-Inspired Genetic Algorithm

91. A. K. Al-Othman, F. S. Al-Fares, and K. M. EL-Nagger. Power system security constrained

2.2 BAGGING METHOD

1. Classiﬁers based on the similarity of objects.

2.3 CLASSIFIERS BASED ON SIMILARITY OF OBJECTS