Table of Contents

1. Cover
2. Table of Contents
3. Series Page
4. Title Page
5. Copyright Page
6. Foreword
7. Preface
8. 1 Metaheuristic Algorithms and Their Applications in
Different Fields: A Comprehensive Review
1. 1.1 Introduction
2. 1.2 Types of Metaheuristic Algorithms
3. 1.3 Application of Metaheuristic Algorithms
4. 1.4 Future Direction
5. 1.5 Conclusion
6. References
9. 2 A Comprehensive Review of Metaheuristics for
Hyperparameter Optimization in Machine Learning
1. 2.1 Introduction
2. 2.2 Fundamentals of Hyperparameter Optimization
3. 2.3 Overview of Metaheuristic Optimization Techniques
4. 2.4 Population-Based Metaheuristic Techniques
5. 2.5 Single Solution-Based Metaheuristic Techniques
 6. 2.6 Hybrid Metaheuristic Techniques
7. 2.7 Metaheuristics in Bayesian Optimization
8. 2.8 Metaheuristics in Neural Architecture Search
9. 2.9 Comparison of Metaheuristic Techniques for
Hyperparameter Optimization
10. 2.10 Applications of Metaheuristics in Machine Learning
11. 2.11 Future Directions and Open Challenges
12. 2.12 Conclusion
13. References
10. 3 A Survey of Computer-Aided Diagnosis Systems for Breast
Cancer Detection
1. 3.1 Introduction
2. 3.2 Procedure for Research Survey
3. 3.3 Imaging Modalities and Their Datasets
4. 3.4 Research Survey
5. 3.5 Conclusion
6. 3.6 Acknowledgment
7. References
11. 4 Enhancing Feature Selection Through Metaheuristic Hybrid
Cuckoo Search and Harris Hawks Optimization for Cancer
Classification
1. 4.1 Introduction
2. 4.2 Related Work
3. 4.3 Proposed Methodology
 4. 4.4 Experimental Setup
5. 4.5 Results and Discussion
6. 4.6 Conclusion
7. References
12. 5 Anomaly Identification in Surveillance Video Using
Regressive Bidirectional LSTM with Hyperparameter
Optimization
1. 5.1 Introduction
2. 5.2 Literature Survey
3. 5.3 Proposed Methodology
4. 5.4 Result and Discussion
5. 5.5 Conclusion
6. References
13. 6 Ensemble Machine Learning-Based Botnet Attack Detection
for IoT Applications
1. 6.1 Introduction
2. 6.2 Literature Survey
3. 6.3 Proposed System
4. 6.4 Results and Discussion
5. 6.5 Conclusion
6. References
14. 7 Machine Learning-Based Intrusion Detection System with
Tuned Spider Monkey Optimization for Wireless Sensor
Networks
 1. 7.1 Introduction
2. 7.2 Literature Review
3. 7.3 Proposed Methodology
4. 7.4 Result and Discussion
5. 7.5 Conclusion
6. References
15. 8 Security Enhancement in IoMT‑Assisted Smart Healthcare
System Using the Machine Learning Approach
1. 8.1 Introduction
2. 8.2 Literature Review
3. 8.3 Proposed Methodology
4. 8.4 Conclusion
5. References
16. 9 Building Sustainable Communication: A Game-Theoretic
Approach in 5G and 6G Cellular Networks
1. 9.1 Introduction
2. 9.2 Related Works
3. 9.3 Methodology
4. 9.4 Result
5. 9.5 Conclusion
6. References
17. 10 Autonomous Vehicle Optimization: Striking a Balance
Between Cost-Effectiveness and Sustainability
1. 10.1 Introduction
 2. 10.2 Methods
3. 10.3 Results
4. 10.4 Conclusions
5. References
18. 11 Adapting Underground Parking for the Future:
Sustainability and Shared Autonomous Vehicles
1. 11.1 Introduction
2. 11.2 Related Works
3. 11.3 Methodology
4. 11.4 Analysis
5. 11.5 Conclusion
6. References
19. 12 Big Data Analytics for a Sustainable Competitive Edge: An
Impact Assessment
1. 12.1 Introduction
2. 12.2 Related Works
3. 12.3 Hypothesis and Research Model
4. 12.4 Results
5. 12.5 Conclusion
6. References
20. 13 Sustainability and Technological Innovation in
Organizations: The Mediating Role of Green Practices
1. 13.1 Introduction
2. 13.2 Related Work
 3. 13.3 Methodology
4. 13.4 Discussion
5. 13.5 Conclusions
6. References
21. 14 Optimal Cell Planning in Two Tier Heterogeneous
Network through Meta-Heuristic Algorithms
1. 14.1 Introduction
2. 14.2 System Model and Formulation of the Problem
3. 14.3 Result and Discussion
4. 14.4 Conclusion
5. References
22. 15 Soil Aggregate Stability Prediction Using a Hybrid Machine
Learning Algorithm
1. 15.1 Introduction
2. 15.2 Related Works
3. 15.3 Proposed Methodology
4. 15.4 Result and Discussion
5. 15.5 Conclusion
6. References
23. Index
24. Also of Interest
25. End User License Agreement

List of Tables
1. Chapter 1
1. Table 1.1 Strengths and weaknesses of metaheuristic
algorithms.
2. Chapter 2
1. Table 2.1 Breakdown of popular metaheuristics and their
I&D components [75]....
2. Table 2.2 Performance comparison of four different
metaheuristics based on a...
3. Table 2.3 Performance comparison of eight population-
based metaheuristics fo...
3. Chapter 3
1. Table 3.1 Summary of the medical jargon used.
2. Table 3.2 Advantages and disadvantages.
4. Chapter 4
1. Table 4.1 Information regarding the six cancer microarray
data.
2. Table 4.2 Parameter settings of the proposed algorithm.
3. Table 4.3 Accuracies of the proposed algorithm with the
mRMR, mRMR+CSA, and ...
4. Table 4.4 Accuracies of the proposed algorithm with the
mRMR, mRMR+CSA, and ...
5. Table 4.5 Accuracies of the proposed algorithm with the
mRMR, mRMR+CSA, and ...
 6. Table 4.6 Comparison of the different published methods
with the proposed me...
5. Chapter 8
1. Table 8.1 Comparison of the accuracy.
2. Table 8.2 Comparison of the precision.
3. Table 8.3 Comparison of the sensitivity.
4. Table 8.4 Comparison of the specificity.
5. Table 8.5 Comparison of the security.
6. Chapter 10
1. Table 10.1 Example of demand and supply data from
expert interviews with the...
2. Table 10.2 Provides an analysis of the logistic network
situation.
7. Chapter 11
1. Table 11.1 Index of the driver, status, and response model
system.
2. Table 11.2 The DSR indexes’ weights and value
attributions.
3. Table 11.3 Rankings of function replacement for each UPS
type.
8. Chapter 12
1. Table 12.1 Results of validity and reliability tests.
2. Table 12.2 HTMT values.
3. Table 12.3 Examine the legitimacy of differences.
 4. Table 12.4 The model’s fit outcomes.
9. Chapter 13
1. Table 13.1 Description of the companies.
2. Table 13.2 Presentation of illustrative information.
3. Table 13.3 Impact on the various aspects.
4. Table 13.4 Credibility, dependability, and relevance.
5. Table 13.5 Inferential statistics.
6. Table 13.6 Evaluation of interactions.
10. Chapter 14
1. Table 14.1 Parameter values.

List of Illustrations

1. Chapter 1
1. Figure 1.1 Flowchart of the genetic algorithm.
2. Figure 1.2 Flowchart of simulated annealing.
3. Figure 1.3 Flowchart of the particle swarm optimization.
4. Figure 1.4 Flowchart of the ant colony optimization.
2. Chapter 2
1. Figure 2.1 Tabu search for optimizing the tour cost for a
city plotted vs. ite...
2. Figure 2.2 A Gaussian process approximation of an
objective function being ite...
3. Figure 2.3 Convergence comparison of four metaheuristics
based on the first 10...
 4. Figure 2.4 Best score convergence profiles vs. iterations for
eight renowned a...
5. Figure 2.5 Accuracy of metaheuristics for different ML
models [83].
3. Chapter 3
1. Figure 3.1 Pictorial representation of the imaging
modalities.
2. Figure 3.2 CNN architecture as illustrated by Mohamed et
al. in [29].
4. Chapter 4
1. Figure 4.1 The proposed research methodology.
2. Figure 4.2 Hybrid flowchart of the HHO and CSA.
3. Figure 4.3 Error comparison with the SVM classifier.
4. Figure 4.4 The variance observed in the proposed
algorithm (mRMR+CSAHHO) compa...
5. Figure 4.5 Error comparison with the KNN classifier.
6. Figure 4.6 The variance observed in the proposed
algorithm (mRMR+CSAHHO) compa...
7. Figure 4.7 Error comparison with the NB classifier.
8. Figure 4.8 The variance observed in the proposed
algorithm (mRMR+CSAHHO) compa...
5. Chapter 5
1. Figure 5.1 Schematic architecture of our proposed system.
 2. Figure 5.2 Normal and abnormal clips from the
ShanghaiTech dataset.
3. Figure 5.3 Accuracy comparison between the suggested
and current techniques.
4. Figure 5.4 Precision comparison between the suggested
and current techniques....
5. Figure 5.5 Recall comparison between the suggested and
current techniques.
6. Figure 5.6 Error rate comparison between the suggested
and current techniques....
6. Chapter 6
1. Figure 6.1 The proposed methodology.
2. Figure 6.2 The dataset’s distribution.
3. Figure 6.3 Architecture of the ANN.
4. Figure 6.4 Results of accuracy.
5. Figure 6.5 Results of precision.
6. Figure 6.6 Results of recall.
7. Figure 6.7 Results of the F-measure.
7. Chapter 7
1. Figure 7.1 Flowchart of the proposed SVM-TSMO model.
2. Figure 7.2 The support vector machine.
3. Figure 7.3 Accuracy of the existing and proposed methods.
4. Figure 7.4 Precision of the existing and proposed methods.
5. Figure 7.5 Recall % of the existing and proposed methods.
 6. Figure 7.6 F1-measure of the existing and proposed
methods.
8. Chapter 8
1. Figure 8.1 The IoMT-smart healthcare system.
2. Figure 8.2 A systematic diagram of security enhancement
in the IoMT using mach...
3. Figure 8.3 Diagrammatic representation of the proposed
method.
4. Figure 8.4 The linear SVM model.
5. Figure 8.5 The MLPSO algorithm flowchart.
9. Chapter 9
1. Figure 9.1 IDO based on a game model.
2. Figure 9.2 Spectrum use ratio.
3. Figure 9.3 Offload ratio.
4. Figure 9.4 Throughput analysis.
5. Figure 9.5 Response delay analysis.
6. Figure 9.6 Energy consumption analysis.
10. Chapter 10
1. Figure 10.1 Framework for logistic clusters that limits
supply chain managemen...
2. Figure 10.2 Illustrates the assumptions of the logistic
network model.
3. Figure 10.3 Distribution model simulations with simulated
annealing.
11. Chapter 11
1. Figure 11.1 Weight matrix of attributes.
2. Figure 11.2 Ranking of the factors.
3. Figure 11.3 UPS-type characteristics.
4. Figure 11.4 Renewal time outcomes.
5. Figure 11.5 Analyzing renewal timing and UPS properties.
6. Figure 11.6 Distribution of renewal times.
12. Chapter 12
1. Figure 12.1 Suggested research design.
2. Figure 12.2 Reliability and validity of the CA.
3. Figure 12.3 Reliability and validity of the CR.
4. Figure 12.4 Reliability and validity of the AVE.
13. Chapter 13
1. Figure 13.1 Method of measuring model.
2. Figure 13.2 Model of structure.
14. Chapter 14
1. Figure 14.1 System model.
2. Figure 14.2 Flowchart of the proposed model.
3. Figure 14.3 Optimal user association to BSs with data suit-
1.
4. Figure 14.4 Optimal user association to BSs with data suit-
2.
5. Figure 14.5 Network utility maximization graph.
15. Chapter 15
1. Figure 15.1 Block diagram of soil aggregation.
2. Figure 15.2 C5.0’s algorithm flow.
3. Figure 15.3 Comparative analysis of the RMSE.
4. Figure 15.4 Comparative analysis of the R2.
5. Figure 15.5 Comparative analysis of the nRMSE.
6. Figure 15.6 Comparative analysis of the MAE.
Scrivener Publishing
100 Cummings Center, Suite 541J
Beverly, MA 01915-6106

Artificial Intelligence and Soft Computing for Industrial

Transformation

Series Editor: Dr S. Balamurugan ([email protected])

The book series is aimed to provide comprehensive handbooks

and reference books for the benefit of scientists, research
scholars, students and industry professional working towards
next generation industrial transformation.

Publishers at Scrivener
Martin Scrivener ([email protected])
Phillip Carmical ([email protected])
Metaheuristics for Machine Learning

Algorithms and Applications

Edited by

Kanak Kalita
Vel Tech University, Avadi, India

Narayanan Ganesh
Vellore Institute of Technology, Chennai, India

and

S. Balamurugan
Intelligent Research Consultancy Services, Coimbatore,
Tamilnadu, India
This edition first published 2024 by John Wiley & Sons, Inc., 111 River Street,
Hoboken, NJ 07030, USA and Scrivener Publishing LLC, 100 Cummings Center, Suite
541J, Beverly, MA 01915, USA
© 2024 Scrivener Publishing LLC
For more information about Scrivener publications please visit
www.scrivenerpublishing.com.

All rights reserved. No part of this publication may be reproduced, stored in a

retrieval system, or transmitted, in any form or by any means, electronic,
mechanical, photocopying, recording, or otherwise, except as permitted by law.
Advice on how to obtain permission to reuse material from this title is available at
http://www.wiley.com/go/permissions.

Wiley Global Headquarters

111 River Street, Hoboken, NJ 07030, USA

For details of our global editorial offices, customer services, and more information
about Wiley products visit us at www.wiley.com.

Limit of Liability/Disclaimer of Warranty

While the publisher and authors have used their best efforts in preparing this work,
they make no representations or warranties with respect to the accuracy or
completeness of the contents of this work and specifically disclaim all warranties,
including without limitation any implied warranties of merchantability or fitness for
a particular purpose. No warranty may be created or extended by sales
representatives, written sales materials, or promotional statements for this work.
The fact that an organization, website, or product is referred to in this work as a
citation and/or potential source of further information does not mean that the
publisher and authors endorse the information or services the organization, website,
or product may provide or recommendations it may make. This work is sold with the
understanding that the publisher is not engaged in rendering professional services.
The advice and strategies contained herein may not be suitable for your situation.
You should consult with a specialist where appropriate. Neither the publisher nor
authors shall be liable for any loss of profit or any other commercial damages,
including but not limited to special, incidental, consequential, or other damages.
Further, readers should be aware that websites listed in this work may have changed
or disappeared between when this work was written and when it is read.

Library of Congress Cataloging-in-Publication Data

ISBN 978-1-394-23392-2

Cover image: Pixabay.Com

Cover design by Russell Richardson
Foreword
In the dynamic landscape of today’s technological revolution,
machine learning and its applications span multiple domains,
offering both opportunities and challenges. As we navigate this
terrain, the significance of data has shifted; it has transformed
from merely passive entities to active drivers influencing
decisions, sculpting perceptions, and determining collective
trajectories. This book serves a pivotal reference that sheds
light upon complex computational arenas and provides clarity
to those navigating this domain.

This book is more than an aggregation of knowledge. It

epitomizes the expertise and adaptability of current
computational researchers and accentuates the potential of
metaheuristics. For those unfamiliar with the term, envision
metaheuristics as high-level strategists that steer a multitude of
heuristic methodologies toward their zenith. They offer the
requisite tools to address complex challenges where
conventional algorithms might be inadequate.

Throughout the book, you will find a wide range of applications

and potential uses of metaheuristics that span across domains
from machine learning to the cutting-edge fields of
sustainability, communication, and networking. It is fascinating
to note that the algorithms aren’t just theoretical entities; they
resonate with pressing real-world challenges. For instance,
consider the pivotal role of metaheuristics in life-saving
applications like breast cancer detection, or in ensuring
security through anomaly identification in surveillance systems
and botnet attack detection.

Moreover, as we delve deeper, we witness the subtle yet

profound synergies between metaheuristics and contemporary
technological innovations. The chapters dedicated to the
advancements in 5G and 6G communication, and the future of
autonomous vehicles, are prime examples. These sections
underline the intricate balance and interdependence of the
challenges we face today and the innovative solutions
metaheuristics can offer.

For researchers who dedicate their lives to exploration,

practitioners at the frontline of technological innovations, and
students who look with hopeful eyes toward the future, this
book will be a pivotal tool. Let it guide you, as it did for me,
through the mesmerizing world of algorithms and their real-
world applications.

Diego Alberto Oliva

Universidad de Guadalajara, Mexico
Preface
While compiling this book, we were guided by a singular vision:
to sculpt a resource that seamlessly melds the theoretical
intricacies of metaheuristics with their myriad practical
applications. Our aspiration was to produce a reference that not
only delves deeply into the subject, but is also accessible to
readers across spectra, offering a holistic understanding that is
both profound and practical.

With every chapter, we strived to weave a narrative, oscillating

between the vast expanse of the topic and the intricate
minutiae that define it. The book commences with a
foundational introduction, leading readers through the
labyrinthine world of metaheuristics. Going forward, the
narrative transitions, diving deeper into their multifaceted
applications—spanning from the dynamic domain of machine
learning to the ever-evolving spheres of technology,
sustainability, and the intricate web of communication
networks.

Metaheuristics present a promising solution to many

formidable optimization conundrums. Yet, their true allure
comes not just from their theoretical promise but their practical
prowess. This book attempts to unveil this allure, transforming
nebulous algorithms into tangible entities with real-world
resonances—whether in the life-saving realm of healthcare or
the cutting-edge world of vehicular communications.

We extend our endless gratitude to the brilliant authors,

reviewers, and countless others whose relentless dedication,
insight, and expertise are evident in these pages. The editorial
journey has been one of profound learning and growth for all
involved. With each chapter, we have gleaned new
perspectives, and we hope this book becomes a wellspring of
knowledge, inspiration, and introspection for both scholars and
professionals.

In closing, we offer our sincere thanks to the Scrivener and

Wiley publishing teams for their help with this book. We
entreat you to immerse your intellect and curiosity in the
mesmerizing world of metaheuristics and their applications.
Here’s to an enlightening reading journey ahead!

Kanak Kalita

Narayanan Ganesh

S. Balamurugan
1
Metaheuristic Algorithms and
Their Applications in Different
Fields: A Comprehensive Review
Abrar Yaqoob1*, Navneet Kumar Verma2 and Rabia Musheer
Aziz1

1
School of Advanced Science and Language, VIT Bhopal
University, Kothrikalan, Sehore, India

2
State Planning Institute (New Division), Planning Department
Lucknow, Utter Pradesh, India

Abstract

A potent method for resolving challenging optimization issues

is provided by metaheuristic algorithms, which are heuristic
optimization approaches. They provide an effective technique
to explore huge solution spaces and identify close to ideal or
optimal solutions. They are iterative and often inspired by
natural or social processes. This study provides comprehensive
information on metaheuristic algorithms and the many areas in
which they are used. Heuristic optimization algorithms are
well-known for their success in handling challenging
optimization issues. They are a potent tool for problem-solving.
Twenty well-known metaheuristic algorithms, such as the tabu
search, particle swarm optimization, ant colony optimization,
genetic algorithms, simulated annealing, and harmony search,
are included in the article. The article extensively explores the
applications of these algorithms in diverse domains such as
engineering, finance, logistics, and computer science. It
underscores particular instances where metaheuristic
algorithms have found utility, such as optimizing structural
design, controlling dynamic systems, enhancing manufacturing
processes, managing supply chains, and addressing problems in
artificial intelligence, data mining, and software engineering.
The paper provides a thorough insight into the versatile
deployment of metaheuristic algorithms across different
sectors, highlighting their capacity to tackle complex
optimization problems across a wide range of real-world
scenarios.

Keywords: Optimization, metaheuristics, machine learning,

swarm intelligence

1.1 Introduction
Metaheuristics represent a category of optimization methods
widely employed to tackle intricate challenges in diverse
domains such as engineering, economics, computer science,
and operations research. These adaptable techniques are
designed to locate favorable solutions by exploring an extensive
array of possibilities and avoiding stagnation in suboptimal
outcomes [1]. The roots and advancement of metaheuristics can
be traced back to the early 1950s when George Dantzig
introduced the simplex approach for linear programming [2].
This innovative technique marked a pivotal point in
optimization and paved the way for the emergence of
subsequent optimization algorithms. Nonetheless, the simplex
method’s applicability is confined to linear programming issues
and does not extend to nonlinear problems. In the latter part of
the 1950s, John Holland devised the genetic algorithm, drawing
inspiration from concepts of natural selection and evolution [3].
The genetic algorithm assembles a set of potential solutions and
iteratively enhances this set through genetic operations like
mutation, crossover, and selection [4]. The genetic algorithm
was a major milestone in the development of metaheuristics
and opened up new possibilities for resolving difficult
optimization issues. During the 1980s and 1990s, the field of
metaheuristics experienced significant expansion and the
emergence of numerous novel algorithms. These techniques,
which include simulated annealing (SA), tabu search (TS), ant
colony optimization (ACO), particle swarm optimization (PSO),
and differential evolution (DE), were created expressly to deal
with a variety of optimization issues. They drew inspiration
from concepts like simulated annealing, tabu search, swarm
intelligence, and evolutionary algorithms [5].

The term “meta-” in metaheuristic algorithms indicates a higher

level of operation beyond simple heuristics, leading to
enhanced performance. These algorithms balance local search
and global exploration by using randomness to provide a range
of solutions. Despite the fact that metaheuristics are frequently
employed, there is not a single definition of heuristics and
metaheuristics in academic literature, and some academics
even use the terms synonymously. However, it is currently
fashionable to classify as metaheuristics all algorithms of a
stochastic nature that utilize randomness and comprehensive
exploration across the entire system. Metaheuristic algorithms
are ideally suited for global optimization and nonlinear
modeling because randomization is a useful method for
switching from local to global search. As a result, almost all
metaheuristic algorithms can be used to solve issues involving
nonlinear optimization at the global level [6]. In recent years,
the study of metaheuristics has developed over time and new
algorithms are being developed that combine different concepts
and techniques from various fields such as machine learning,
deep learning, and data science. The development and
evolution of metaheuristics have made significant contributions
to solving complex optimization problems and have led to the
development of powerful tools for decision-making in various
domains [7]. In order to find solutions in a huge search area,
metaheuristic algorithms are founded on the idea of mimicking
the behaviors of natural or artificial systems. These algorithms
are particularly valuable for tackling problems that are
challenging or impossible to solve using traditional
optimization methods. Typically, metaheuristic algorithms
involve iterations and a series of steps that modify a potential
solution until an acceptable one is discovered. Unlike other
optimization techniques that may become stuck in local optimal
solutions, metaheuristic algorithms are designed to explore the
entire search space. They also exhibit resilience to noise or
uncertainty in the optimization problem. The adaptability and
plasticity of metaheuristic algorithms are two of their main
features. They can be modified to take into account certain
limitations or goals of the current task and are applicable to a
wide variety of optimization situations. However, for complex
problems with extensive search spaces, these algorithms may
converge slowly toward an optimal solution, and there is no
guarantee that they will find the global optimum. Metaheuristic
algorithms find extensive application in various fields including
engineering, finance, logistics, and computer science. They have
been successfully employed in solving diverse problems such as
optimizing design, control, and manufacturing processes,
portfolio selection, and risk management strategies [8].

1.2 Types of Metaheuristic Algorithms

We shall outline some of the most popular metaheuristic

methods in this section.

1.2.1 Genetic Algorithms

Genetic algorithms (GAs) fit to a cluster of metaheuristic

optimization techniques that draw inspiration from natural
selection and genetics [9–11]. In order to find the optimal
solution for a particular issue, the core idea underlying the GA
is to mimic the evolutionary process. The genetic algorithm has
the capability to address challenges spanning various fields
such as biology, engineering, and finance [12–14]. In the
methodology of the GA, a potential solution is denoted as a
chromosome, or a collection of genes. Each gene within the
context of the problem signifies an individual variable, and its
value corresponds to the potential range of values that the
variable can take [15, 16]. Subsequently, these chromosomes
undergo genetic operations like mutation and crossover. This
process can give rise to a fresh population of potential
solutions, resulting in a novel set of potential outcomes [17–19].

The following are the major steps in the GA:

Initialization: The algorithm initializes a set of potential
responses first. A chromosome is used to symbolize each
solution, which is a string of genes randomly generated
based on the problem domain [20].

Evaluation: The suitability of each chromosome is assessed

based on the objective function of the problem. The quality
of the solution is evaluated by the fitness function, and the
objective is to optimize the fitness function by either
maximizing or minimizing it, depending on the particular
problem [21].

Selection: Chromosomes that possess higher fitness values

are chosen to form a fresh population of potential solutions.
Various techniques, such as roulette wheel selection,
tournament selection, and rank-based selection, are
employed for the selection process [22].

Crossover: The selected chromosomes are combined

through crossover to generate new offspring chromosomes.
The crossover operation exchanges the genetic information
from the parent chromosomes and is utilized to generate
novel solutions [23].
Mutation: The offspring chromosomes are subjected to
mutation, which introduces random changes to the genetic
information. Mutation aids in preserving diversity within
the population and preventing the occurrence of local
optima [24].

Replacement: As the child chromosomes multiply, a new

population of potential solutions is formed and replaces the
less fit members of the prior population.

Termination: The technique proceeds to iterate through the

selection, crossover, mutation, and replacement phases until
a specific termination condition is satisfied. Reaching a
predetermined maximum for iterations is one scenario for
termination, attaining a desired fitness value, or exceeding a
predetermined computational time limit.
Figure 1.1 Flowchart of the genetic algorithm.

The GA has several advantages, such as being capable of solving

complex issues, locating the global optimum, and being
applicable to various domains. However, the GA also has some
limitations, such as the need for a suitable fitness function, the
possibility of premature convergence, and the high
computational cost for complex problems. Figure 1.1 shows the
flowchart of the genetic algorithm.

1.2.2 Simulated Annealing

Simulated annealing is a probabilistic method for optimizing

complex multidimensional problems by seeking the global best
solution. It draws inspiration from the metallurgical technique
of annealing, which includes heating and gradually cooling a
metal to enhance its strength and resilience [25]. Similarly,
simulated annealing commences at an elevated temperature,
enabling the algorithm to extensively investigate a vast array of
possible solutions, and then slowly decreases the temperature
to narrow down the search to the most promising areas. SA
works by maintaining a current solution and repeatedly
making small changes to it in search of a better solution. At
each iteration, the algorithm calculates a cost function that
measures how good the current solution is. The cost function
can be any function that assigns a score to a potential solution,
such as a distance metric or a likelihood function. Subsequently,
the algorithm determines whether to embrace or disregard a
new solution by utilizing a probability distribution that relies
on the existing temperature and the disparity between the costs
of the current and new solutions [26]. High temperatures
increase SA’s propensity to embrace novel solutions even if they
have a higher cost than the current solution. This is because the
algorithm is still exploring the space of potential solutions and
needs to be open to new possibilities. As the temperature
decreases, SA becomes increasingly discriminating and admits
novel solutions solely if they surpass the existing solution. By
employing this approach, SA prevents itself from becoming
trapped in local peaks and eventually achieves convergence
toward the global peak [27]. SA offers a notable benefit by
effectively addressing non-convex optimization problems,
characterized by numerous local optima. By permitting the
acceptance of solutions with greater costs, SA can navigate
diverse areas within the solution space and prevent
entrapment in local optima. Moreover, SA boasts ease of
implementation and independence from cost function
gradients, rendering it suitable for scenarios where the cost
function lacks differentiability. However, SA does have some
limitations. It can be slow to converge, especially for large or
complex problems, and may require many iterations to find the
global optimum. SA’s effectiveness is also influenced by the
decision of cooling schedule, which determines how quickly the
temperature decreases. If the cooling schedule is too slow, the
algorithm may take too long to converge, while if it is too fast,
the algorithm may converge too quickly to a suboptimal
solution [28, 29].
Figure 1.2 Flowchart of simulated annealing.

To put it briefly, simulated annealing is a highly effective

optimization method that has the capability to address intricate
problems, multidimensional problems with multiple local
optima. It works by exploring the solution space and gradually
narrowing down the search for the most promising regions.
While it has some limitations, SA is a helpful tool for a variety
of optimization issues in the real world. Figure 1.2 shows the
flowchart of simulated annealing.
1.2.3 Particle Swarm Optimization

Particle swarm optimization is a technique for optimization

that employs a population-based strategy to address a wide
range of optimization problems. First introduced by Kennedy
and Eberhart in 1995, this concept takes inspiration from the
coordinated movements observed in the flocking of birds and
the schooling of fish [30–32]. This algorithm emulates the social
dynamics exhibited by these creatures, where each member
learns from its own encounters and the experiences of its
nearby peers, with the aim of discovering the best possible
solution [33]. The PSO method begins by generating a
population of particles, each of which acts as a potential
solution to the optimization issue at hand. These particles,
which have both a location and a velocity vector, are randomly
distributed throughout the search space [34]. The location
vector represents the particle’s current solution, whereas the
velocity vector represents the particle’s moving direction and
magnitude inside the search space. Through iterative steps,
each particle’s location and velocity vectors undergo constant
modification and adjustment in the PSO algorithm, guided by its
own best solution encountered thus far and the solutions of its
neighboring particles [35]. Collaborative learning continues
until a predetermined stopping condition is met, such as when
the desired outcome is attained or the maximum number of
iterations has been reached. Compared to other optimization
algorithms, the PSO algorithm boasts various advantages,
including simplicity, rapid convergence, and robustness [36, 37].
PSO has found applications in diverse problem domains,
spanning function optimization, neural network training, image
processing, and feature selection. Nevertheless, the algorithm
does come with certain limitations. These include the risk of
premature convergence, where the algorithm may converge to
suboptimal solutions prematurely, and challenges in effectively
handling problems with high-dimensional spaces [38].

Figure 1.3 Flowchart of the particle swarm optimization.

In general, the particle swarm optimization algorithm is a

robust and effective optimization method capable of addressing
numerous practical optimization problems. Its simplicity and
intuitive approach make it an appealing choice compared to
more intricate optimization methods. Figure 1.3 shows the
flowchart of the particle swarm optimization.

1.2.4 Ant Colony Optimization

Ant colony optimization is a nature-inspired method that

addresses difficult optimization problems by mimicking the
behavior of ant colonies. The program takes its cues from the
behavior of ant colonies, specifically the way ants communicate
to discover the shortest path toward food sources. The
fundamental idea behind the ACO is to simulate the foraging
behavior of ants to solve optimization problems effectively. A
simulated group of ants is put on a graph representing the
problem space in the ACO. These ants navigate the graph by
selecting the next node to visit based on the pheromone trails
left behind by other ants. The strength of the pheromone trail
represents the quality of the solution that passed through that
edge. As more ants traverse the same edge, the pheromone trail
becomes stronger. This is similar to how ants communicate with
each other in real life by leaving pheromone trails to signal the
location of food sources [39, 40]. The ACO algorithm has several
key parameters, such as the amount of pheromone each ant
leaves, the rate at which pheromones evaporate, and the
balance between exploiting the best solution and exploring new
solutions. The optimal values of the parameters in the
algorithm are determined through a process of
experimentation and refinement to obtain the best possible
results for a specific problem [41].

The ACO has showcased impressive achievements in resolving

diverse optimization challenges, including but not limited to the
traveling salesman problem, vehicle routing, and job
scheduling. One notable advantage of the algorithm is its ability
to swiftly discover favorable solutions, even when confronted
with extensive search spaces. Furthermore, because the ACO
belongs to the category of metaheuristic algorithms, it can be
applied to a variety of situations without requiring a deep
understanding of the underlying structure of those problems
[42]. Figure 1.4 shows the flowchart of the ant colony
optimization.
Figure 1.4 Flowchart of the ant colony optimization.

1.2.5 Tabu Search

The tabu search is a metaheuristic technique utilized for

optimization problems, initially proposed by Fred Glover in
1986. It has gained significant popularity across diverse
domains, including operations research, engineering, and
computer science. The core concept behind the tabu search
involves systematically traversing the search space by
transitioning between different solutions in order to identify
the optimal solution. However, unlike other local search
algorithms, the tabu search incorporates a memory structure
that records previous moves executed during the search. These
data are then used to steer the search to potential places within
the search space [43]. The tabu list, a memory structure that
plays an important part in the algorithm, is at the heart of the
tabu search. This list serves to store and remember previous
moves made during the search process, ensuring that the
algorithm avoids revisiting solutions that have already been
explored. By utilizing the tabu list, the tabu search effectively
restricts the search to new and unexplored regions of the
solution space, promoting efficient exploration and preventing
repetitive or redundant searches. This list is used to enforce a
set of constraints, known as the tabu tenure, which determines
how long a move is considered tabu. By imposing this
constraint, the algorithm is compelled to investigate diverse
regions within the search space and evade being trapped in
local optima. This ensures that the algorithm remains dynamic
and continually explores new possibilities, preventing it from
being overly fixated on suboptimal solutions [43]. The tabu
search is a versatile optimization algorithm applicable to both
continuous and discrete optimization problems. When
addressing continuous optimization problems, the algorithm
typically uses a neighborhood function to generate new
solutions by perturbing the present solution. In the event of
discrete optimization problems, the neighborhood function is
typically defined in terms of specific moves that can be made to
the solution, such as swapping two elements in a permutation.
The effectiveness of the tabu search is based on a number of
variables, such as the choice of neighborhood function, the tabu
tenure, and the stopping criterion. The algorithm can be
enhanced by using various strategies, such as diversification
and intensification, which balance the search space’s
exploitation and exploration [44].

The tabu search, a metaheuristic approach introduced by Fred

Glover in 1986, is utilized for optimizing problems. Its
application spans a wide array of domains, including
operations research, engineering, and computer science,
establishing it as a widely recognized technique. The
fundamental principle of the tabu search involves a systematic
traversal of the solution space, shifting between different
solutions to ascertain the optimal one. Distinguishing itself from
conventional local search algorithms, the tabu search
incorporates a memory structure that logs prior moves
executed during the search process. This stored information
guides the search toward unexplored areas within the solution
space [45]. Central to the tabu search is the tabu list, a pivotal
memory structure. This list retains and recalls previous moves
executed during the search, ensuring that revisiting already
explored solutions is avoided. The tabu list effectively restricts
the exploration to untrodden regions, preventing redundant
searches and fostering efficient investigation. Governing the
tabu list is the concept of the tabu tenure, setting the duration
for which a move remains prohibited. This constraint compels
the algorithm to explore diverse solution space regions, eluding
entrapment in local optima. This dynamic approach
continuously explores novel avenues, counteracting fixation on
suboptimal solutions [46]. The tabu search is a versatile
optimization algorithm suitable for both continuous and
discrete optimization challenges. For continuous optimization, a
neighborhood function is commonly used to generate new
solutions by perturbing the current one. In the context of
discrete optimization, the neighborhood function is typically
defined by specific permissible moves, such as element swaps
in a permutation. The efficacy of the tabu search hinges on
factors like neighborhood function choice, tabu tenure, and the
stopping criteria. The algorithm can be augmented through
strategies like diversification and intensification, ensuring a
balance between exploiting and exploring the search space.
In general, the tabu search is a robust and adaptable
optimization technique that has demonstrated its effectiveness
in addressing diverse problem sets. It can be employed
independently or integrated into more intricate optimization
algorithms. Its popularity stems from its versatility and
straightforwardness, making it a favored option for tackling
real-life challenges in various domains.

1.2.6 Differential Evolution

The DE is an optimization algorithm based on populations,

originally created by Storn and Price in 1997 [47]. It fit to the
category of evolutionary algorithms that iteratively grow a
population of potential solutions to find the optimal solution.
The algorithm adheres to the fundamental steps of mutation,
crossover, and selection, which are key elements commonly
shared among numerous evolutionary algorithms [48].

In the process of the differential evolution, a population of

potential solutions undergoes iterative evolution through the
implementation of the following sequential steps:

1. Initialization: A population of N possible solutions is

produced at random.
2. Mutation: Involves randomly selecting three candidate
solutions and modifying them to create a trial vector.
3. Crossover: It is a technique used in optimization algorithms
to create a new candidate solution by combining the trial
vector with the target vector.
4. Selection: If the new candidate solution has a higher fitness,
it will take the place of the target vector.

The success of the differential evolution depends on the

selection of the optimization technique’s adjustable settings,
such as mutation rate, crossover rate, and population size [49,
50]. Several variants of the DE have been proposed, including
the SHADE (success history-based adaptive differential
evolution) and JADE (adaptive differential evolution)
algorithms, which incorporate adaptive control parameters to
improve the algorithm’s performance.

1.2.7 Harmony Search

The harmony search (HS) is an optimization technique

motivated by the musical improvization process. Geem [125]
had put forward the idea in 2001 and has now been used to
solve several optimization issues. The techniques mimic the
process of improvization by a group of musicians, where they
adjust their pitches (or notes) to create harmony. In the HS, the
choice variables of an optimization problem in high school are
comparable to musical notes, and the value of the goal function
reflects harmony [51].

Starting individuals of decision variable vectors (i.e., the notes)

are used in the HS approaches and iteratively search for better
solutions by generating new solutions through the following
steps:

1. Harmony memory: A set of the best candidate solutions (i.e.,

the harmonies) is maintained.
2. Harmony creation: A potential solution is created by chance
selecting values from the harmony memory.
3. Pitch adjustment: The values in the new candidate solution
are adjusted with a probability based on a pitch adjustment
rate.
4. Acceptance: The new candidate solution is accepted if it
improves the objective function value.

The control variables, such as the harmony memory size, pitch

adjustment rate, and number of iterations, affect how well the
HS performs. The method follows the core phases of mutation,
crossover, and selection, and the approach has been utilized to
tackle diverse optimization challenges, such as managing water
resources, designing structures, and operating power systems
[52, 53].

1.2.8 Artificial Bee Colony

The artificial bee colony (ABC) is a population-based

optimization method that draws inspiration from honey bees’
feeding habits. Since its introduction by Karaboga in 2005, the
technique has been used to solve a number of optimization
issues [54]. The ABC mimics the foraging process of bees, where
they search for food sources by visiting the flowers in the
vicinity of the hive [55].

The artificial bee colony technique starts with an arbitrarily

generated population of candidate solutions (i.e., food sources)
and iteratively searches for better solutions by simulating the
foraging process of bees through the following steps:

1. Phase of employed bees: The employed bees develop new

candidate solutions by modifying the values of current
solutions.
2. Phase of the onlooker bees: The onlooker bees choose the
candidate solutions with the highest fitness values and send
this information to the employed bees.
3. Phase of scout bees: The scout bees search for new candidate
solutions by randomly generating new solutions.

The success of the ABC depends on the control parameters, such

as the population size, the number of iterations, and the
probability of abandoning a food source. The algorithm has
been applied to a wide range of optimization problems,
including image processing, wireless sensor networks, and
fuzzy control systems [56].

1.2.9 Firefly Algorithm

A metaheuristic optimization technique called the firefly

algorithm (FA) is based on how fireflies behave. Dr. Xin-She
Yang first presented the FA in 2008. The social behavior of
fireflies, which is characterized by their flashing light to attract
mates or prey, served as an inspiration for the algorithm [57].
The firefly technique’s primary role is to increase light intensity,
and in order to maximize light intensity, fireflies travel toward
the brighter fireflies. Starting with a random population of
fireflies, using the goal function, the program calculates the
light intensity of each firefly [58]. The movement of fireflies is
governed by their attractiveness, which is determined by the
brightness of their light, and the distance between them. The
fireflies move toward the brighter fireflies and update their
positions until the maximum light intensity is achieved [59].
Numerous optimization issues, including those involving
machine learning, image processing, and function optimization,
have been effectively solved using the firefly algorithm.

1.2.10 Gray Wolf Optimizer

The gray wolf optimizer (GWO) is a metaheuristic optimization

system based on wolves’ social structure and hunting
techniques. The algorithm was developed in 2014 by Seyed Ali
Mirjalili, Seyed Mohammad Mirjalili, and Andrew Lewis [60].
The optimization issue is viewed in the gray wolf optimizer as a
wolf pack’s prey–predator dynamic. The wolf population used
by the method is randomly divided into four groups: alpha,
beta, delta, and omega. While the omega wolf is the weakest
and has the lowest fitness value, the alpha wolf assumes the
role of the leader within the wolf pack and has the greatest
fitness value [61–63]. The movement of the wolves is governed
by three different types of hunting techniques: hunting for prey,
following the alpha wolf, and surrounding the prey. The wolves
adjust their locations based on these techniques until the
optimal solution is accomplished. The gray wolf optimizer’s
application has been expanded to address a variety of
optimization issues, such as function optimization, engineering
design, and feature selection [64, 65].
1.2.11 Imperialist Competitive Algorithm

The imperialist competitive algorithm (ICA), a metaheuristic

optimization method designed to address numerous
optimization issues, was inspired by the political and economic
rivalry of imperialist nations. Amir Hossein Gandomi and Ali
Alavi introduced the method in 2010 [66]. In the imperialist
competitive algorithm, the optimization problem is considered
as a competition between empires. The algorithm starts with a
random population of empires, which are categorized into two
groups: imperialist and colonies. The imperialist empires have
higher fitness values, and they expand their territories by
annexing colonies. The movement of empires is governed by
two different types of actions: assimilation and revolution. In
the assimilation process, the imperialist empires try to improve
the fitness of their colonies, while in the revolution process, the
colonies rebel against their imperialist empires and become
independent. Various optimization problems, including
function optimization, image segmentation, and parameter
estimation, have been effectively solved through the successful
application of the imperialist competitive algorithm [67, 68].

1.2.12 Bat Algorithm

Xin-She Yang created the bat algorithm (BA) in 2010, which is a
metaheuristic optimization technique used to tackle a variety of
optimization challenges [69]. The bat algorithm is inspired by
bats’ echolocation activity, which uses ultrasonic noises to
navigate and locate prey in the dark. It replicates the activity of
bats in their hunt for prey in order to discover the best solution
to a given optimization issue [70].

To solve an optimization problem using the bat algorithm, a

population of bats is created in the search space with random
placements and velocities. The bats move randomly, emitting
frequencies proportional to their fitness values. Bats with better
fitness emit higher-frequency sounds that attract other bats
toward their position in the search space [71]. In adding to the
frequency-based attraction mechanism, the bat algorithm
includes a random walk component that allows the bats to
explore uncharted regions of the search space. During each
iteration, the algorithm updates the velocity and position of
each bat using information on the best solution found thus far,
as well as the loudness and frequency of its emitted signal. The
algorithm iterates until it reaches a present stopping point, such
as a maximum number of iterations or a goal fitness value [72].

1.2.13 Cuckoo Search

The cuckoo search, a metaheuristic optimization method
inspired by the cuckoo bird reproductive behavior, was
introduced in 2009 by Xin-She Yang and Suash Deb. Cuckoo
birds use an unusual approach in which they deposit their eggs
in the nests of other bird species, leaving the host birds to care
for and rear their children [73]. The cuckoo search algorithm
imitates the foraging behavior of cuckoos as they search for
food, employing the strategy of brood parasitism. The objective
is to discover the optimal solution for an optimization problem.
The technique commences by initializing a population of
cuckoos with random positions and velocities. Subsequently,
each cuckoo deposits an egg in a nest that is chosen randomly,
with the likelihood of selecting a specific nest being
proportionate to its fitness level [74–76]. The cuckoo search
algorithm incorporates a random walk element, allowing the
cuckoos to explore unexplored regions within the search space.
During each iteration, the algorithm adjusts the position and
speed of each cuckoo based on the best solution found thus far
and the potential for uncovering a superior solution by
depositing an egg in a new nest. To enhance the solution quality
by exploring the vicinity of the current answer, the algorithm
also integrates a local search component. The process persists
until a predetermined stopping condition is fulfilled, such as
reaching the maximum iteration count or achieving the desired
fitness level [77, 78].

1.2.14 Flower Pollination Algorithm

The flower pollination algorithm (FPA), created by Xin-She Yang

in 2012, is a metaheuristic optimization algorithm that draws
pollinators like butterflies and bees with their aroma [79]. The
flower pollination algorithm emulates flower pollination
behaviors to discern the optimal solution for a given problem.
In 2012, Xin-She Yang introduced this algorithm, a
metaheuristic optimization technique inspired by the way
flowers attract pollinators like butterflies and bees through
their fragrance. This involves initializing a group of flowers
with random positions and fragrances. Each flower releases a
fragrance that draws pollinators, with the likelihood of
attraction determined by its fitness level. A random walk
element is also integrated, enabling pollinators to explore novel
sections of the search space [80]. Throughout each iteration, the
flower pollination algorithm adjusts the position and fragrance
of each flower based on the current best solution and the
probability of enticing new pollinators to potentially uncover
an improved solution. Furthermore, the algorithm incorporates
a local search component that investigates the neighboring
vicinity to enhance solution quality. The algorithm continues
until a predefined stopping condition is met, which could be
achieving a specific fitness target or reaching the maximum
iteration count [81].

1.2.15 Krill Herd Algorithm

In 2012, Amir H. Gandomi and Amir H. Alavi introduced the

krill herd algorithm (KHA), a metaheuristic optimization
technique inspired by the coordinated motions and interplays
of krill within their oceanic habitat. This algorithm endeavors
to replicate the herding conduct of krill as they forage for
sustenance and navigate survival challenges, all with the aim of
identifying the finest solution for a specified problem [82].

The krill herd algorithm emulates the synchronized movements

and interactions of krill in their pursuit of nourishment and
companionship, striving to pinpoint the optimal solution for a
given task. Commencing with the creation of a krill population
featuring random positions and velocities, each individual krill
engages in random movements across the search space. The
course and speed of these movements are shaped by forces of
attraction, repulsion from other krill, and environmental
influences [83].
Depending on the best solution found thus far, the impact of
other krill and the environment and the location and velocity of
each krill are adjusted by the algorithm at each iteration. The
technique also has a random walk component that enables krill
to explore new regions of the search space [84]. The algorithm
continues to run until a predefined stopping condition is met,
such as a predetermined number of iterations or a
predetermined goal fitness value [85].

1.2.16 Whale Optimization Algorithm

In 2016, Seyedali Mirjalili and Andrew Lewis introduced the

whale optimization algorithm (WOA), a metaheuristic
optimization technique. This concept drew inspiration from the
hunting behaviors of humpback whales, characterized by a
blend of independent and cooperative movements and
vocalizations [86]. The core concept of the whale optimization
algorithm is to replicate the foraging conduct of humpback
whales as they seek sustenance, aiming to uncover the optimal
solution for a particular problem. The algorithm commences by
establishing a group of whales, each assigned random positions
and velocities. Subsequently, every whale undertakes random
movements across the search space, with the course and speed
influenced by the forces of attraction, repulsion exerted by
other whales, and environmental conditions [87]. Throughout
each iteration, the whale optimization algorithm modifies the
positions and velocities of individual whales based on the
current best solution and the effects of other whales and the
environment. Moreover, the algorithm integrates a random
walk aspect to facilitate exploration of uncharted regions
within the search space. Termination of the algorithm occurs
upon meeting predefined cessation criteria, such as attaining
the maximum iteration count or reaching a desired fitness level
[88–90].

1.2.17 Glowworm Swarm Optimization

The glowworm swarm optimization (GSO) is an optimization

technique inspired by nature and was introduced by
Kalyanmoy Deb and Samir Dasgupta in 2006. It draws
inspiration from the bioluminescent communication exhibited
by fireflies and glowworms to emulate the coordinated
behavior of swarms [91]. The glowworm swarm optimization
serves as a metaheuristic optimization algorithm that replicates
the bioluminescent actions of glowworms as they navigate their
surroundings in pursuit of sustenance and companionship, all
with the aim of identifying the optimal solution for a given
problem. Commencing with the creation of a group of
glowworms, each assigned random positions and luminosities,
every glowworm emits light that attracts other individuals, with
the degree of attraction being proportional to both the
luminosity and the distance of neighboring glowworms [92].
Throughout each iteration, the glowworm swarm optimization
algorithm updates the position and luminosity of each
glowworm based on the likelihood of drawing new glowworms
and the best solution attained up to that point. A random walk
component is also integrated into the algorithm, allowing
glowworms to venture into unexplored territories of the search
space. The algorithm’s execution continues until reaching a
specific termination criterion, such as achieving a target fitness
level or reaching the maximum permissible number of
iterations [93].

1.2.18 Cat Swarm Optimization

The cat swarm optimization (CSO) method, which is a

metaheuristic optimization algorithm based on the cooperative
hunting behavior of a colony of cats, was first proposed by Ying
Tan and Yuhui Shi in 2006. The method is modeled after the
collaboration and communication that occur among cats while
they are hunting [94]. The fundamental concept behind the cat
swarm optimization involves emulating the collective hunting
actions of a group of cats to ascertain the finest solution for a
given problem. The approach commences by placing and
endowing a population of cats with random positions and
velocities. Subsequently, each cat undertakes randomized
movements within the search space, wherein the course and
speed of movement are influenced by the forces of attraction
and repulsion exerted by fellow cats, alongside environmental
variables [95].

In each iteration of the cat swarm optimization algorithm,

adjustments are made to the positions and velocities of the cats
based on the prevailing best solution, impact of other cats, and
environmental conditions. Furthermore, a random walk
element is infused into the algorithm, granting cats the
capability to explore previously uncharted territories within
the search space. This optimization process continues until a
predetermined halting criterion, such as reaching a desired
fitness threshold or attaining the maximum allowable number
of iterations, is met. This algorithm was introduced by Ying Tan
and Yuhui Shi in 2006, with inspiration drawn from the
collaborative hunting demeanor of cats [96].

1.2.19 Grasshopper Optimization Algorithm

The grasshopper optimization algorithm (GOA), created in 2014

by Seyedali Mirjalili and Andrew Lewis, is a nature-inspired
optimization algorithm that imitates the swarming behavior of
grasshoppers. The GOA bases its optimization on the collective
behavior of grasshoppers, which involves interpersonal
communication and cooperation [97].

Throughout each iteration of the grasshopper optimization

algorithm, adjustments are made to the position and velocity of
every grasshopper, guided by the prevailing best solution and
the influences stemming from fellow grasshoppers and
environmental factors. Moreover, a random walk attribute is
introduced, empowering grasshoppers to explore and uncover
previously unexplored sectors within the search space [98, 99].
This algorithm persists until a predetermined cessation
criterion is satisfied, such as reaching a maximum iteration
count or achieving a designated fitness target.

1.2.20 Moth–Flame Optimization

The moth–flame optimization (MFO) algorithm draws

inspiration from the natural behavior of moths, which exhibit
an inherent attraction to flames and utilize celestial cues for
navigation. This technique emulates the search pattern of
moths, with the goal of uncovering the most optimal solution
for a specific problem [100].

In the initial stages of the optimization procedure, the

algorithm commences by generating a cluster of moths, each
assigned random positions and luminosities. Subsequent to this,
every moth traverses the search space via randomized
movements, adjusting its velocity and trajectory based on two
key factors: its inclination toward the brightest moth and the
impact of environmental conditions [101]. As the moth–flame
optimization algorithm progresses through each iteration, the
position and luminosity of each moth are updated. These
adjustments hinge upon the current best solution achieved and
the attraction toward the brightest moth. Additionally, the
algorithm incorporates a random walk feature to allow moths
to explore new areas of the search space. The algorithm stops
when a predefined stopping criterion is achieved, such as
reaching a target fitness value or a maximum number of
iterations [102].

Apart from this algorithm, several other algorithms like the sine
cosine algorithm [103, 104], dragonfly algorithm [105, 106], and
symbiotic organism search [107, 108] have widely been used in
literature.

1.3 Application of Metaheuristic

Algorithms
Hu et al. presented the crisscross optimisation arithmetic
optimization algorithm (CSOAOA), an improved hybrid
arithmetic optimization algorithm that seeks to solve the
constraints of the original arithmetic optimization algorithm
(AOA). The CSOAOA combines numerous strategies to enhance
the speed at which convergence occurs, search efficiency, and
calculation accuracy of the AOA, including the point set
strategy, optimal neighborhood learning approach, and
crisscross strategy. In terms of precision, convergence rate, and
solution quality, the testing findings revealed that the CSOAOA
outperformed other optimization algorithms, proving its
potential as a viable approach for handling complicated
engineering optimization issues [109].

Ghafil and Jármai proposed the dynamic differential annealed

optimization (DDAO) in 2020, a unique metaheuristic
optimization technique that uses random search and simulated
annealing. The algorithm’s efficacy was evaluated on 51
benchmark functions, and it beat many well-known
optimization techniques, confirming its higher performance.
The DDAO successfully converged to the global minimum of two
real-world engineering optimization problems and produced a
better feasible solution for the spring design problem than
several other algorithms [110].
In 2022, Zhao et al. introduced a novel swarm intelligence
optimization algorithm known as the dandelion optimizer (DO).
This algorithm imitates the flight patterns of dandelion seeds
and comprises three distinct stages: rising, descending, and
landing. Each stage is characterized by its own set of rules and
attributes. During the falling and landing stages, the algorithm
simulates the seed’s journey. Brownian motion and a Levy
random walk were employed, and the efficiency of the
differential evolution (DO) was evaluated using the CEC 2017
benchmark functions, and its performance was compared
against nine other nature-inspired metaheuristic algorithms.
The DO exhibited superior optimization accuracy, stability,
convergence, and scalability, outperforming the competing
algorithms. The study further demonstrated the robustness and
exceptional performance of the DO by applying it to solve four
real-world optimization problems [111].

The smart flower optimization algorithm (SFOA) is a

metaheuristic optimization algorithm created by Sattar et al.
(2021) that is inspired by the heliotropic motions of immature
sunflowers as well as growth mechanisms based on sun-
tracking and the biological clock. The success of the method is
proven by benchmark testing utilizing CEC 2015 and
comparisons with other optimization algorithms, which
indicated improved optimization efficiency and performance.
In addition, the SFOA was applied to real-world engineering
design issues, yielding encouraging results and demonstrating
its utility in addressing problems with unknown search spaces
[112].

Azizi et al. introduced a novel metaheuristic algorithm named

the fire hawk optimizer (FHO) in 2023. The FHO draws
inspiration from the foraging behavior of fire hawks and
exhibits remarkable efficiency in optimizing problems. Through
comprehensive benchmarking against 233 mathematical test
functions and two Competition on Evolutionary Computation
(CEC) problems, the FHO outperformed other metaheuristic
algorithms. Furthermore, the algorithm’s efficacy was
demonstrated in solving realsize structural frames, yielding
promising outcomes and emphasizing its potential for
addressing real-world problems. This research not only
contributes to the advancement of metaheuristic algorithms but
also underscores the capabilities of nature-inspired approaches
[113].

In 2020, Ghasemian et al. introduced a new metaheuristic

algorithm inspired by human behavior to address urbanization
and improve the quality of life. The algorithm incorporates
combined searching and population management strategies,
and evaluation results demonstrate its superiority over other
high-performing algorithms in a majority of search spaces
[114].

Abollah Ansar et al. (2020) developed a hybrid metaheuristic

algorithm for predicting corporate bankruptcy by combining
the magnetic optimization algorithm (MOA) with the particle
swarm optimization. The resulting algorithm, the MOA-PSO,
was utilized to train an artificial neural network (ANN) for
bankruptcy prediction and was found to outperform four
existing algorithms, achieving an accuracy rate of 99.7%.
Numerous optimization strategies have their benefits, and
hybrid algorithms have proven to be useful in solving
complicated problems more quickly and accurately [115].

Using a combination of artificial neural networks and

optimization algorithms based on metaheuristics, Farahani et
al. conducted a research study in 2022 with the goal of
anticipating the Tehran Price Index (TEPIX). To determine the
main market determinants influencing the stock market, the
inquiry used a variety of algorithms, including GA, HS, PSO,
MFO, and WOA. The outcomes showed that, in comparison to
the traditional ANN model, the combined algorithms exhibited
improved prediction skills. It is noteworthy that the study
identified the whale optimization algorithm as the most precise
technique for forecasting TEPIX [116].
In their 2022 research, Hadi Gholizadeh et al. introduced an
environmentally friendly approach to the reverse logistics
process for polystyrene disposable items. This method
addresses uncertainties in demand and recovery costs by
employing a mixed-integer nonlinear programming (MINLP)
model. To tackle the problem, the researchers employed
heuristic methods such as simulated annealing, cross-entropy,
and genetic algorithms to generate initial solutions.
Subsequently, they utilized the response surface method, robust
optimization, and Taguchi method to refine the algorithm’s
performance. By comparing the heuristics’ effectiveness with
robust optimization using the best–worst technique and
conducting a case study, the study confirmed the model’s
viability and highlighted the efficacy of optimization techniques
in enhancing algorithmic performance. This work offers an
ecologically sound approach to reverse logistics while
showcasing the positive impact of optimization methods on
algorithmic efficiency [117].

A hybrid metaheuristic method combining the discrete particle

swarm optimization (DPSO) and the Harris hawks optimization
(HHO) with three improvement strategies was first presented
by Yan et al. in 2023. The purpose was to address the location
routing problem (LRP) in the distribution of relief supplies in
the immediate aftermath of disasters. They validated their
proposed multiobjective model using a case study on COVID-19
in Wuhan, showing that the algorithm outperformed previous
metaheuristic algorithms in terms of accuracy and the capacity
to avoid local optima. This study offers a viable method for
planning disaster response that will optimize the distribution of
relief supplies [118].

Yang et al. (2022) talked about the importance of information

systems (ISs) in logistics and distribution management and
suggested using the metaheuristic techniques (MHTs) to
efficiently handle information across the supply chain. The
study emphasized the value of cooperation, coordination,
information-sharing, and communication among supply chain
actors and stated that metaheuristics can be helpful in solving
challenging logistical issues. The paper highlights the role of
advanced decision support systems and the development of
information and communication technology in adapting to the
challenges of integration [119].

Adel Pourghader Chobar et al. (2022) focused on reverse

logistics planning in response, recovery, and reconstruction
phases during earthquakes, where a large amount of waste is
generated. The problem is multiobjective and uncertain, which
is solved using the non dominated sorting algorithm (NSGA-II)
metaheuristic algorithm. The study indicates that increasing the
capacity reduces the number of distribution centers needed but
may lead to increased transportation costs. The case study is
based on Tehran city [120].

Simran Gibson et al. published a review of the literature on

identifying spam emails with machine learning algorithms and
bio-inspired optimization strategies in 2020. On seven different
email datasets, the researchers used numerous machine
learning models, including Naive bayes, support vector
machine, random forest, decision tree, and multilayer
perceptron, along with feature extraction and preprocessing
processes. The performance of the classifiers was improved by
using bio-inspired techniques such as the particle swarm
optimization and the genetic algorithm. The analysis found that
the multinomial naive Bayes model, optimized with the genetic
algorithm, performed better across several criteria.
Furthermore, the scientists conducted a comparison with
different machine learning and bio-inspired models to establish
the best technique [121].

Oyelade et al. (2022) introduced a novel optimization algorithm

named the Ebola optimization search algorithm (EOSA) that is
inspired by the Ebola virus disease propagation mechanism.
The technique was evaluated using two sets of benchmark
functions, including both conventional and limited IEEE-CEC
benchmark functions, by fusing the SEIR-HVQD mathematical
model with the propagation model. The outcomes
demonstrated that the EOSA beat popular metaheuristic
algorithms including the artificial bee colony algorithm, genetic
algorithm, and particle swarm optimization algorithm. The
method was also used to resolve the issue of hyperparameter
optimization in the categorization of digital mammography
images, with a level of 96.0% accuracy [122].

The difficulty of resource scheduling for client tasks in the

cloud environment, which has grown in popularity over the
past 10 years, was the topic of the study by Bindu et al. (2020).
To cut down on energy use, costs, and time, they created the
optimal sequential ant colony optimisation algorithm (OSACO),
an optimized scheduling method that takes into account the
quality of service (QoS) requirements of customers. The OSACO
algorithm was compared to other algorithms, and simulation
results demonstrated its efficacy in resource scheduling
optimization [123].
Table 1.1 Strengths and weaknesses of metaheuristic algorithms.

Strengths Weaknesses

Metaheuristic algorithms are Metaheuristic

flexible and can be applied to a algorithms may
wide range of optimization converge slowly to an
problems. They can also be optimal solution,
easily adapted to specific particularly for
problem constraints or complex problems with
objectives. a large search space.

Metaheuristic algorithms are There is no guarantee

designed to explore the entire that a metaheuristic
search space, unlike other algorithm will find the
optimization techniques that global optimum.
may get stuck in local optima.
Strengths Weaknesses

Unlike some optimization Metaheuristic

methods, metaheuristic algorithms often
algorithms do not require the require careful
computation of derivatives or selection of algorithm
gradients of the objective parameters, which can
function. be a time-consuming
and difficult process.

Metaheuristic algorithms are Metaheuristic

usually robust to noise or algorithms often
uncertainty in the optimization involve complex
problem. mathematical
operations, making it
difficult to understand
how they arrive at a
solution.
 Strengths Weaknesses

Many metaheuristic algorithms In some cases,

can be parallelized, allowing metaheuristic
the use of high-performance algorithms may overfit
computing resources to solve the training data,
complex problems. resulting in a
suboptimal solution.

Gharehchopogh (2022) discusses the incorporation of quantum

computing (QC) concepts into metaheuristic algorithms and the
ensuing development of quantum-inspired metaheuristic
algorithms. The paper highlights the benefits of using QC-
inspired metaheuristic algorithms in solving numerical and
combinational optimization problems, achieving superior
performance compared to conventional metaheuristic
algorithms. The article presents a comprehensive review of the
various applications of QC in metaheuristics, their classification
in optimization problems, and their use in science and
engineering. The study provides valuable insights into the
potential of QC-inspired metaheuristic algorithms in solving
optimization problems [124].
From the above discussion on different metaheuristic
algorithms, we find that there are several strengths and
weaknesses that are important to consider when choosing an
optimization method. Some of the strengths and weaknesses of
metaheuristic algorithms are shown in Table 1.1.

1.4 Future Direction

In the coming years, metaheuristic algorithms are anticipated

to evolve along several trajectories. One promising path
involves the creation of hybrid metaheuristic algorithms, which
amalgamate the strengths of multiple techniques to enhance
performance and resilience. Another trajectory involves
adapting metaheuristics to grapple with optimization
challenges posed by vast amounts of data, addressing the
computational hurdles presented by extensive datasets.
Moreover, the parallel and distributed implementation of
metaheuristic algorithms can harness computational resources
and expedite the optimization process. Subsequent research
should also emphasize dynamic and multiobjective
optimization, where metaheuristics can adjust to shifting
problem conditions and manage numerous conflicting
objectives simultaneously. Additionally, the call for explainable
artificial intelligence (AI) necessitates metaheuristic algorithms
that not only optimize but also shed light on decision-making
processes. Real-time optimization demands metaheuristics that
swiftly converge to nearly optimal solutions within strict time
constraints.

Lastly, delving into innovative applications in domains like

renewable energy, healthcare, smart cities, and cybersecurity
will broaden the horizons of metaheuristic algorithms. By
advancing in these directions, metaheuristics can persist in
providing potent tools for resolving intricate optimization
problems across diverse domains.

1.5 Conclusion

In summary, this article presents a comprehensive overview of

metaheuristic algorithms and their broad spectrum of
applications across diverse domains. The paper outlines well-
known metaheuristic algorithms and underscores their efficacy
in tackling intricate optimization issues. It extensively surveys
the utilization of metaheuristic algorithms in engineering,
finance, logistics, computer science, and related fields. The
provided instances of employing these algorithms for tasks like
devising optimal structures, governing dynamic systems,
refining manufacturing procedures, orchestrating supply
chains, and resolving challenges in artificial intelligence, data
mining, and software engineering serve as concrete
demonstrations of their capability to address complex problems
across a variety of contexts. In essence, the paper underscores
the significance and adaptability of metaheuristic algorithms in
surmounting optimization hurdles in different spheres and
offers a comprehensive comprehension of their practical
application.

References

1. Kvasov, D.E. and Mukhametzhanov, M.S., Metaheuristic vs.

deterministic global optimization algorithms: The univariate
case. Appl. Math. Comput., 318, 245–259, 2018, doi:
10.1016/j.amc.2017.05.014.
2. Sinuany-Stern, Z., Foundations of operations research: From
linear programming to data envelopment analysis. Eur. J.
Oper. Res., 306, 3, 1069–1080, 2023, doi:
10.1016/j.ejor.2022.10.046.
3. Ghneamat, R., Genetic algorithms and application to adaptive
automata for game theory, Al-Balqa University, As-Salt,
Jordan, 2005.
4. Holland, J.H., Genetic algorithms. Sci. Am., 267, 1, 66–72, 1992,
doi: 10.1038/ scientificamerican0792-66.
5. Hammouche, K. et al., A comparative study of various meta-
heuristic techniques applied to the multilevel thresholding
problem. Eng. Appl. Artif. Intell., 23, 5, 676–688, 2010, doi:
10.1016/j.engappai.2009.09.011.
6. Gandomi, A.H. et al., Metaheuristic algorithms in modeling
and optimization, in: Metaheuristic Appl. Struct. Infrastruct.,
pp. 1–24, 2013, doi:10.1016/ b978-0-12-398364-0.00001-2.
7. Bandaru, S. and Deb, K., Metaheuristic techniques, in: Decis.
Sci., pp. 693–750, 2016, doi:10.1201/9781315183176-12.
8. Yang, X.-S., Engineering optimization: An introduction with
metaheuristic applications. John Wiley & Sons, Hoboken,
2010, https://doi.org/10.1002/9780470640425.
9. Katoch, S. et al., A review on genetic algorithm: Past, present,
and future. Multimed. Tools Appl., 80, 5, 8091–8126, 2021, doi:
10.1007/ s11042-020-10139-6.
10. Slowik, A. and Kwasnicka, H., Evolutionary algorithms and
their applications to engineering problems. Neural Comput.
Appl., 32, 16, 12363–12379, 2020, doi: 10.1007/s00521-020-
04832-8.
11. Ghadai, R.K. and Kalita, K., Accurate estimation of DLC thin
film hardness using genetic programming. Int. J. Mater. Res.,
111, 6, 453–462, 2020, doi: 10.3139/146.111911.
12. Ghadai, R.K. et al., Genetically optimized diamond-like
carbon thin film coatings. Mater. Manuf. Processes, 34, 13,
 1476–1487, 2019, doi: 10.1080/10426914.2019.1594273.
13. Kalita, K. et al., Robust genetically optimized skew laminates.
Proc. Inst. Mech. Eng. C, 233, 1, 146–159, 2019, doi:
10.1177/0954406218756943.
14. Kalita, K. et al., Genetic programming assisted multi-scale
optimization for multi-objective dynamic performance of
laminated composites: The advantage of a more elementary-
level analysis. Neural Comput. Appl., 32, 12, 7969–7993, 2020,
doi: 10.1007/s00521-019-04280-z.
15. Kalita, K. et al., Data-driven genetic programming-based
symbolic regression metamodels for EDM process, in: Data-
Driven Optimization of Manufacturing Processes, IGI Global,
Scopus. USA, 2020, doi: 10.4018/978-1-7998-7206-1.ch009.
16. Kalita, K. et al., Sensitivity analysis of GFRP composite
drilling parameters and genetic algorithm-based
optimization. Int. J. Appl. Metaheuristic Comput., 13, 1, 1–17,
2022.
17. Kumar, V. et al., A hybrid genetic programming-grey wolf
optimizer approach for process optimization of biodiesel
production. Processes, 9, 3, 442, 2021, doi: 10.3390/pr9030442.
18. Pal, S. et al., Optimization of frequency separation of
laminated shells carrying transversely distributed mass
using genetic algorithm. J. Vib. Control, 2023, doi:
10.1177/10775463231190277.
19. Pal, S. et al., Genetic algorithm-based fundamental
frequency optimization of laminated composite shells
carrying distributed mass. J. Inst. Eng. (India) Ser. C, 103, 3,
389–401, 2022, doi: 10.1007/s40032-021-00801-9.
20. Todorovski, M. and Rajicic, D., An initialization procedure in
solving optimal power flow by genetic algorithm. IEEE Trans.
Power Syst., 21, 2, 480–487, 2006, doi:
10.1109/TPWRS.2006.873120.
21. Ji, S. et al., Optimal tolerance allocation based on fuzzy
comprehensive evaluation and genetic algorithm. Int. J. Adv.
Manuf. Technol., 16, 7, 461–468, 2000, doi:
10.1007/s001700070053.
22. Kumar, M., Husian, M., Upreti, N., Gupta, D., Genetic
algorithm: Review and application. Int. J. Inf. Technol. Knowl.
Manag., 2, 451–454, 2010.
23. Umbarkar, and Sheth, Crossover operators in genetic
algorithms: A review. ICTACT J. Soft Comput., 06, 1, 1083–
1092, 2015, doi: 10.21917/ijsc.2015.0150.
24. Mirjalili, S. et al., Genetic algorithm: Theory, literature
review, and application in image reconstruction, in: Studies
in Computational Intelligence. Nature-Inspired Optimizers, pp.
69–85, 2020, doi:10.1007/978-3-030-12127-3_5.
25. Suman, B. and Kumar, P., A survey of simulated annealing as
a tool for single and multiobjective optimization. J. Oper. Res.
 Soc., 57, 10, 1143–1160, 2006, doi:
10.1057/palgrave.jors.2602068.
26. Amine, K., Multiobjective simulated annealing: Principles
and algorithm variants. Adv. Oper. Res., 2019, 1–13, 2019, doi:
10.1155/2019/8134674.
27. Eglese, R.W., Simulated annealing: A tool for operational
research. Eur. J. Oper. Res., 46, 3, 271–281, 1990, doi:
10.1016/0377-2217(90)90001-R.
28. Baykasoğlu, A. and Gindy, N.N.Z., A simulated annealing
algorithm for dynamic layout problem. Comput. Oper. Res.,
28, 14, 1403–1426, 2001, doi: 10.1016/S0305-0548(00)00049-6.
29. Ming, C. and Raghav, N., Simulated annealing for mixture
distribution analysis and its applications to reliability testing,
in: Simul. Annealing, 2008, doi:10.5772/5834.
30. Banks, A. et al., A review of particle swarm optimization.
Part I: Background and development. Nat. Comput., 6, 4, 467–
484, 2007, doi: 10.1007/ s11047-007-9049-5.
31. Kalita, K. et al., Optimizing drilling induced delamination in
GFRP composites using genetic algorithm & particle swarm
optimization. Adv. Compos. Lett., 27, 1, 1–9, 2018.
32. Shankar, R. et al., Hybridized particle swarm - Gravitational
search algorithm for process optimization. Processes, 10, 3,
616, 2022, doi: 10.3390/ pr10030616.
33. Jain, N.K. et al., A review of particle swarm optimization. J.
Inst. Eng. (India) Ser. B, 99, 4, 407–411, 2018, doi:
10.1007/s40031-018-0323-y.
34. Ganesh, N. et al., Multi-objective high-fidelity optimization
using NSGAIII and MO-RPSOLC. Comput. Model. Eng. Sci., 129,
2, 443–464, 2021, doi: 10.32604/cmes.2021.014960.
35. Kalita, K. et al., Optimizing process parameters for laser
beam micro-marking using genetic algorithm and particle
swarm optimization. Mater. Manuf. Processes, 32, 10, 1101–
1108, 2017, doi: 10.1080/10426914.2017.1303156.
36. Wang, D. et al., Particle swarm optimization algorithm: An
overview. Soft Comput., 22, 2, 387–408, 2018, doi:
10.1007/s00500-016-2474-6.
37. Narayanan, G. et al., PSO-tuned support vector machine
metamodels for assessment of turbulent flows in pipe bends.
Eng. Comput., 37, 3, 981–1001, 2019, doi: 10.1108/EC-05-2019-
0244.
38. Poli, R. et al., Particle swarm optimization. Swarm Intell., 1, 1,
33–57, 2007, doi: 10.1007/s11721-007-0002-0.
39. Dorigo, M. and Blum, C., Ant colony optimization theory: A
survey. Theor. Comput. Sci., 344, 2-3, 243–278, 2005, doi:
10.1016/j. tcs.2005.05.020.
40. Blum, C., Ant colony optimization: Introduction and recent
trends. Phys. Life Rev., 2, 4, 353–373, 2005, doi:
 10.1016/j.plrev.2005.10.001.
41. Maniezzo, V. and Carbonaro, A., Ant colony optimization: An
overview, in: Operations Research/Computer Science
Interfaces Series. pp. 469–492, 2002, doi:10.1007/978-1-4615-
1507-4_21.
42. Dorigo, M. and Stutzle, T., Ant Colony Optimization. MIT
Press, Cambridge, UK, 2004.
43. Glover, F., Tabu search and adaptive memory programming
—Advances, applications and challenges, in: Operations
Research/Computer Science Interfaces Series. pp. 1–75, 1997,
doi:10.1007/978-1-4615-4102-8_1.
44. Cangalovic, M. M., Kovacevic-Vujcic, V. V., Ivanovic, L.,
Drazic, M., Asic, M. D., Tabu search: A brief survey and some
real-life applications. Yugosl. J. Oper. Res., 6, 1, 5–18, 1996.
45. Jaeggi, D.M. et al., The development of a multi-objective Tabu
Search algorithm for continuous optimisation problems. Eur.
J. Oper. Res., 185, 3, 1192–1212, 2008, doi:
10.1016/j.ejor.2006.06.048.
46. Glover, F., Tabu search for nonlinear and parametric
optimization (with links to genetic algorithms). Discrete Appl.
Math., 49, 1-3, 231–255, 1994, doi: 10.1016/0166-
218X(94)90211-9.
47. Fleetwood, K., An introduction to differential evolution, in:
Proc. Mathematics and Statistics of Complex Systems
 (MASCOS) One Day Symposium, Brisbane, Australia, pp. 785–
791, 2004.
48. Bilal, M. et al., Differential evolution: A review of more than
two decades of research. Eng. Appl. Artif. Intell., 90, 103479,
2020, doi: 10.1016/j. engappai.2020.103479.
49. Ganesh, N. et al., Efficient feature selection using weighted
superposition attraction optimization algorithm. Appl. Sci.
MDPI, 13, 5, 3223, 2023, doi: 10.3390/app13053223.
50. Panigrahi, B.K., Suganthan, P.N., Das, S., Dash, S.S., Swarm,
evolutionary, and memetic computing. In Proceedings of the
Third International Conference SEMCCO, pp. 16–18, Springer,
Chennai, India, December 2010.
51. Alia, O.M. and Mandava, R., The variants of the harmony
search algorithm: An overview. Artif. Intell. Rev., 36, 1, 49–68,
2011, doi: 10.1007/ s10462-010-9201-y.
52. Guo, Z. et al., Adaptive harmony search with best-based
search strategy. Soft Comput., 22, 4, 1335–1349, 2018, doi:
10.1007/s00500-016-2424-3.
53. Yadav, P. et al., An intelligent tuned harmony search
algorithm for optimization. Inf. Sci., 196, 47–72, 2012, doi:
10.1016/j.ins.2011.12.035.
54. Kaur, G. and Oberoi, A., Novel approach for brain tumor
detection based on naïve bayes classification. In Data
 Management, Analytics and Innovation, pp. 451–462, Springer,
Singapore, 2020.
55. Karaboga, D. and Basturk, B., A powerful and efficient
algorithm for numerical function optimization: Artificial bee
colony (ABC) algorithm. J. Glob. Optim., 39, 3, 459–471, 2007,
doi: 10.1007/s10898-007-9149-x.
56. Karaboga, D. and Ozturk, C., A novel clustering approach:
Artificial bee colony (ABC) algorithm. Appl. Soft Comput., 11,
1, 652–657, 2011, doi: 10.1016/j.asoc.2009.12.025.
57. Yang, X.-S. and He, X.-S., Why the firefly algorithm works?.
Nature-Inspired Algorithms and Applied Optimization, pp.
245–259, Springer, 2018.
58. Yang, X.-S., Nature-inspired optimization algorithms:
Challenges and open problems. J. Comput. Sci., 46, 101104,
2020, doi: 10.1016/j.jocs.2020.101104.
59. Osaba, E. and Yang, X.S. (Eds.), Applied optimization and
swarm intelligence, in: Springer Tracts in Nature-Inspired
Computing, Springer, Singapore, 2021.
60. Mirjalili, S. et al., Grey wolf optimizer. Adv. Eng. Software,
Barking, London, England, 69, 46–61, 2014, doi:
10.1016/j.advengsoft.2013.12.007.
61. Faris, H. et al., Grey wolf optimizer: A review of recent
variants and applications. Neural Comput. Appl., 30, 2, 413–
435, 2018, doi: 10.1007/ s00521-017-3272-5.
62. Rajendran, S. et al., A conceptual comparison of six nature-
inspired metaheuristic algorithms in process optimization.
Processes, 10, 2, 197, 2022, doi: 10.3390/pr10020197.
63. Kalita, K. et al., A hybrid TOPSIS-PR-GWO approach for
multi-objective process parameter optimization. Process
Integr. Optim. Sustain., 6, 4, 1011–1026, 2022, doi:
10.1007/s41660-022-00256-0.
64. Kalita, K. et al., Grey wolf optimizer-based design of
ventilated brake disc. J. Braz. Soc. Mech. Sci. Eng., 43, 8, 405,
2021, doi: 10.1007/s40430-021-03125-y.
65. Tu, Q. et al., Hierarchy strengthened grey wolf optimizer for
numerical optimization and feature selection. IEEE Access
Pract. Innov. Open Solut., 7, 78012–78028, 2019, doi:
10.1109/ACCESS.2019.2921793.
66. Hosseini, S. and Al Khaled, A., A survey on the imperialist
competitive algorithm metaheuristic: Implementation in
engineering domain and directions for future research. Appl.
Soft Comput., 24, 1078–1094, 2014, doi: 10.1016/j.
asoc.2014.08.024.
67. Lucas, C. et al., Application of an imperialist competitive
algorithm to the design of a linear induction motor. Energy
Convers. Manage., 51, 7, 1407–1411, 2010, doi:
10.1016/j.enconman.2010.01.014.
68. Ebtehaj, I. and Bonakdari, H., Comparison of genetic
algorithm and imperialist competitive algorithms in
predicting bed load transport in clean pipe. Water Sci.
Technol., 70, 10, 1695–1701, 2014, doi: 10.2166/wst.2014.434.
69. Yang, X.-S., A new metaheuristic bat-inspired algorithm, in:
Studies in Computational Intelligence. Nature Inspired
Cooperative Strategies for Optimization (NICSO 2010), pp. 65–
74, 2010, doi:10.1007/978-3-642-12538-6_6.
70. Yang, X.-S. and Hossein Gandomi, A., Bat algorithm: A novel
approach for global engineering optimization. Eng. Comput.,
29, 5, 464–483, 2012, doi: 10.1108/02644401211235834.
71. Mirjalili, S. et al., Binary bat algorithm. Neural Comput. Appl.,
25, 3-4, 663–681, 2014, doi: 10.1007/s00521-013-1525-5.
72. Fister, I. et al., Bat algorithm: Recent advances. Presented at
the 2014 IEEE 15th International Symposium on
Computational Intelligence and Informatics (CINTI), Budapest,
vol. 2014, IEEE, 2014, Nov., doi: 10.1109/ CINTI.2014.7028669.
73. Yaqoob, A. et al., A review on nature-inspired algorithms for
cancer disease prediction and classification. Mathematics, 11,
5, 1081, 2023, doi: 10.3390/ math11051081.
74. Yang, X.-S. and Deb, S., Cuckoo search: Recent advances and
applications. Neural Comput. Appl., 24, 1, 169–174, 2014, doi:
10.1007/s00521-013-1367-1. 75. Gandomi, A.H. et al., Cuckoo
search algorithm: A metaheuristic approach to solve
 structural optimization problems. Eng. Comput., 29, 1, 17–35,
2013, doi: 10.1007/s00366-011-0241-y.
76. Kalita, K. et al., Optimizing frequencies of skew composite
laminates with metaheuristic algorithms. Eng. Comput., 36, 2,
741–761, 2020, doi: 10.1007/ s00366-019-00728-x.
77. Kalita, K. et al., A comparative study on the metaheuristic-
based optimization of skew composite laminates. Eng.
Comput. Springer, 38, 4, 3549–3566, 2022, doi: 10.1007/s00366-
021-01401-y.
78. Valian, E. et al., Improved cuckoo search for reliability
optimization problems. Comput. Ind. Eng., 64, 1, 459–468,
2013, doi: 10.1016/j.cie.2012.07.011.
79. Yang, X.-S., Flower pollination algorithm for global
optimization, in: Lecture Notes in Computer Science.
Unconventional Computation and Natural Computation, pp.
240–249, 2012, doi: 10.1007/978-3-642-32894-7_27.
80. Abdel-Basset, M. and Shawky, L.A., Flower pollination
algorithm: A comprehensive review. Artif. Intell. Rev., 52, 4,
2533–2557, 2019, doi: 10.1007/ s10462-018-9624-4.
81. Yang, X.S. (Ed.), Nature-Inspired Algorithms and Applied
Optimization, vol. 744, Springer, Berlin/Heidelberg, Germany,
2017.
82. Hossein, A. and Hossein, A., A new bio-inspired optimization
algorithm. Commun. Nonlinear Sci. Numer. Simul., 17, 12,
 4831–4845, 2012, doi: 10.1016/j.cnsns.2012.05.010.
83. La, A., Al-betar, M.A., Awadallah, M.A., Tajudin, A.,
Mohammad, L., A comprehensive review: Krill Herd
algorithm (KH) and its applications. Appl. Soft Comput. J., 49,
437–446, 2016, doi: 10.1016/j.asoc.2016.08.041.
84. Abualigah, L.M.Q., Feature selection and enhanced krill herd
algorithm for text document clustering, Springer,
Berlin/Heidelberg, Germany, 2019.
85. Wang, G.-G. et al., Hybrid krill herd algorithm with
differential evolution for global numerical optimization.
Neural Comput. Appl., 25, 2, 297–308, 2014, doi:
10.1007/s00521-013-1485-9.
86. Mirjalili, S. and Lewis, A., The whale optimization algorithm.
Adv. Eng. Software Barking, London, England, 95, 51–67, 2016,
doi: 10.1016/j. advengsoft.2016.01.008.
87. Mohammed, H.M. et al., A systematic and meta-analysis
survey of whale optimization algorithm. Comput. Intell.
Neurosci., 2019, 8718571, 2019, doi: 10.1155/2019/8718571.
88. Rana, N. et al., Whale optimization algorithm: A systematic
review of contemporary applications, modifications and
developments. Neural Comput. Appl., 32, 20, 16245–16277,
2020, doi: 10.1007/s00521-020-04849-z.
89. Kalita, K. et al., A comparative study on multi-objective
pareto optimization of WEDM process using nature-inspired
 metaheuristic algorithms. Int. J. Interact. Des. Manuf., 17, 2,
499–516, 2023, doi: 10.1007/s12008-022-01007-8.
90. Shanmugasundar, G. et al., Optimization of variable stiffness
joint in robot manipulator using a novel NSWOA-Marcos
approach. Processes, 10, 6, 1074, 2022, doi:
10.3390/pr10061074.
91. Ab Wahab, M.N. et al., A comprehensive review of swarm
optimization algorithms. PLoS One, 10, 5, e0122827, 2015, doi:
10.1371/journal.pone.0122827.
92. Puttamadappa, and Parameshachari, Demand side
management of small scale loads in a smart grid using glow-
worm swarm optimization technique. Microprocess.
Microsyst., 71, 102886, 2019, doi: 10.1016/j.
micpro.2019.102886.
93. Kalaiselvi, T. et al., A review on glowworm swarm.
Optimization, 3, 2, 49–56, 2017.
94. Tsai, P.-W. and Istanda, V., Review on cat swarm
optimization algorithms. Presented at the 2013 3rd
International Conference on Consumer Electronics,
Communications and Networks (CECNet), 2013, Nov., vol.
2013, Xianning, China, doi: 10.1109/CECNet.2013.6703394.
95. Ahmed, A.M. et al., Cat swarm optimization algorithm – A
survey and performance evaluation. Comput. Intell.
Neurosci., 2020, 4854895, 2020, doi: 10.1155/2020/4854895.
96. Darwish, A., Bio-inspired computing: Algorithms review,
deep analysis, and the scope of applications. Future
Computing Inform. J., 3, 2, 231–246, 2018, doi:
10.1016/j.fcij.2018.06.001.
97. Saremi, S. et al., Grasshopper optimisation algorithm:
Theory and application, in: Adv. Eng. Softw., vol. 105, pp. 30–
47, Barking, London, England, 2017, doi:
10.1016/j.advengsoft.2017.01.004.
98. Pijarski, P. and Kacejko, P., A new metaheuristic optimization
method: The algorithm of the innovative gunner (AIG). Eng.
Optim., 51, 12, 2049–2068, 2019, doi:
10.1080/0305215X.2019.1565282.
99. Joshi, M. et al., A conceptual comparison of dragonfly
algorithm variants for CEC-2021 global optimization
problems. Arab. J. Sci. Eng., 48, 2, 1563–1593, 2023, doi:
10.1007/s13369-022-06880-9.
100. Shehab, M. et al., Moth–flame optimization algorithm:
Variants and applications. Neural Comput. Appl., 32, 14, 9859–
9884, 2020, doi: 10.1007/ s00521-019-04570-6.
101. Mirjalili, S., Moth-flame optimization algorithm: A novel
nature-inspired heuristic paradigm. Knowl. Based Syst., 89,
228–249, 2015, doi: 10.1016/j. knosys.2015.07.006.
102. Khurma, R.A. et al., An enhanced evolutionary software
defect prediction method using island moth flame
 optimization. Mathematics, 9, 15, 1722, 2021, doi:
10.3390/math9151722.
103. Narayanan, R.C. et al., A novel many-objective sine cosine
algorithm (MaOSCA) for engineering. Appl. Math., 11, 10,
2301, 2023.
104. Priyadarshini, J. et al., Analyzing physics-inspired
metaheuristic algorithms in feature selection with K-nearest-
neighbor. Appl. Sci., 13, 2, 906, 2023, doi:
10.3390/app13020906.
105. Joshi, M. et al., Comparison of NSGA-II, MOALO and MODA
for multi-objective optimization of micro-machining
processes. Mater. (Basel), 14, 17, 5109, 2021, doi:
10.3390/ma14175109.
106. Kalita, K. et al., A novel MOALO-MODA ensemble approach
for multi-objective optimization of machining parameters for
metal matrix composites. Multiscale Multidiscip. Model. Exp.
Des., 6, 1, 179–197, 2023, doi: 10.1007/s41939-022-00138-5.
107. Ganesh, N. et al., Efficient feature selection using weighted
superposition attraction optimization algorithm. Appl. Sci.,
13, 5, 3223, 2023, doi: 10.3390/ app13053223.
108. Ganesh, N. et al., A novel decomposition-based multi-
objective symbiotic organism search optimization algorithm.
Mathematics, 11, 8, 1898, 2023, doi: 10.3390/math11081898.
109. Hu, G. et al., An enhanced hybrid arithmetic optimization
algorithm for engineering applications. Comput. Methods
Appl. Mech. Eng., 394, 114901), 2022, doi:
10.1016/j.cma.2022.114901.
110. Ghafil, H.N. and Jármai, K., Dynamic differential annealed
optimization: New metaheuristic optimization algorithm for
engineering applications. Appl. Soft Comput., 93, 106392,
2020, doi: 10.1016/j.asoc.2020.106392.
111. Zhao, S. et al., Dandelion optimizer: A nature-inspired
metaheuristic algorithm for engineering applications. Eng.
Appl. Artif. Intell., 114, 105075, 2022, doi:
10.1016/j.engappai.2022.105075.
112. Sattar, D. and Salim, R., A smart metaheuristic algorithm for
solving engineering problems. Eng. Comput., 37, 3, 2389–2417,
2021, doi: 10.1007/ s00366-020-00951-x.
113. Azizi, M. et al., Fire hawk optimizer: A novel metaheuristic
algorithm. Artif. Intell. Rev., 56, 1, 287–363, 2023, doi:
10.1007/s10462-022-10173-w.
114. Ansari, A. et al., A hybrid metaheuristic method in training
artificial neural network for bankruptcy prediction. IEEE
Access, 8, 176640–176650, 2020, doi:
10.1109/ACCESS.2020.3026529.
115. Chobar, A.P. et al., Multi-objective hub-spoke network
design of perishable tourism products using combination
 machine learning and meta-heuristic algorithms. Environ.
Dev. Sustain., 1–28, 2022, doi: 10.1007/ s10668-022-02350-2.
116. Ghasemian, H. et al., Human urbanization algorithm: A
novel metaheuristic approach. Math. Comput. Simul., 178, 1–
15, 2020, doi: 10.1016/j. matcom.2020.05.023.
117. Gholizadeh, H. et al., Fuzzy data-driven scenario-based
robust data envelopment analysis for prediction and
optimisation of an electrical discharge machine’s parameters.
Expert Syst. Appl., 193, 116419, 2022, doi: 10.1016/j.
eswa.2021.116419.
118. Gibson, S. et al., Detecting spam email with machine
learning optimized with bio-inspired metaheuristic
algorithms. IEEE Access, 8, 187914–187932, 2020, doi:
10.1109/ACCESS.2020.3030751.
119. Peng, B., Digital leadership: State governance in the era of
digital technology. Cult. Sci., 5, 4, 210–225, 2022, doi:
10.1177/2096608321989835.
120. Shahvaroughi Farahani, M. and Razavi Hajiagha, S.H.,
Forecasting stock price using integrated artificial neural
network and metaheuristic algorithms compared to time
series models. Soft Comput., 25, 13, 8483–8513, 2021, doi:
10.1007/s00500-021-05775-5.
121. Yan, T. et al., A hybrid metaheuristic algorithm for the
multi-objective location-routing problem in the early post-
 disaster stage. J. Ind. Manage. Optim., 19, 6, 4663–4691, 2023,
doi: 10.3934/jimo.2022145.
122. Bindu, G.B. et al., Optimized resource scheduling using the
meta heuristic algorithm in cloud computing. IAENG Int. J.
Comput. Sci., 47, 3, 360–366, 2020.
123. Gharehchopogh, F.S., Quantum-inspired metaheuristic
algorithms: Comprehensive survey and classification. Artif.
Intell. Rev., 56, 6, 5479–5543, 2023, doi: 10.1007/s10462-022-
10280-8.
124. Oyelade, O.N. et al., Ebola optimization search algorithm: A
new nature-inspired metaheuristic optimization algorithm.
IEEE Access, 10, 16150–16177, 2022, doi:
10.1109/ACCESS.2022.3147821.
125. Geem, Z.W. (Ed.), Music-Inspired Harmony Search
Algorithm: Theory and Applications, vol. 191, Springer,
Berlin/Heidelberg, Germany, 2009.

Note

1. *Corresponding author: [email protected]

2
A Comprehensive Review of
Metaheuristics for
Hyperparameter Optimization in
Machine Learning
Ramachandran Narayanan1* and Narayanan Ganesh2

1
Department of Computer Science and Engineering, Sona College
of Technology, Salem, India

2
School of Computer Science & Engineering, Vellore Institute of
Technology, Chennai, India

Abstract

Hyperparameter optimization is a critical step in the

development and fine-tuning of machine learning (ML) models.
Metaheuristic optimization techniques have gained significant
popularity for addressing this challenge due to their ability to
search the hyperparameter space efficiently. In this review, we
present a detailed analysis of various metaheuristic techniques
for hyperparameter optimization in ML, encompassing
population-based, single solution-based, and hybrid
approaches. We explore the application of metaheuristics in
Bayesian optimization and neural architecture search, two
prominent areas within the field. Moreover, we provide a
comparative analysis of these techniques based on established
criteria and evaluate their performance in diverse ML
applications. Finally, we discuss future directions and open
challenges with special emphasis on the opportunities for
improvement in metaheuristics. Other crucial issues like
adaptability to new ML paradigms, computational complexity,
and scalability issues are also discussed critically. This review
aims to provide researchers and practitioners with a
comprehensive understanding of the state-of-the-art
metaheuristic optimization techniques for hyperparameter
tuning, thereby facilitating informed decisions and
advancements in the field.

Keywords: Metaheuristics, hyperparameter tuning, machine

learning, population-based techniques, hybrid techniques

2.1 Introduction

Hyperparameter optimization (HPO) is an integral aspect of

machine learning (ML) advancements—the idea of which
revolves around finding the optimum values of parameters that
are externally set by the user or the model and are not learned
by the machine learning model during the training phase [1].
These are called hyperparameters, and they hold great power
in determining the performance of the ML model that has been
designed to solve a complex problem at hand. The optimization
is not an easy job itself and faces critical challenges due to the
large data space that needs to be sorted and simplified and the
inherent nonlinear constraints. Metaheuristic algorithms that
can provide effective routes for fast and decisive HPO serve as
an enabling beacon of light for this task. This paper provides a
consolidated and detailed review of these algorithms and
techniques and provides a complete insight into the approaches
that govern the systematic routes to reach the desired
optimization while expanding on the scope and potential of the
growing applications of metaheuristics.

2.1.1 Background and Motivation

Hyperparameter optimization (HPO) is rapidly becoming a

trending global force in the machine learning (ML)
development community, and harnessing its power using
metaheuristic algorithms is becoming a new priority at the
industrial level. This prominent change in R&D methods for
obtaining more efficient solutions in much lesser time by
minimizing iterations conducted for complex optimization
problems demands effective metaheuristic integration for
higher yields. The detail and scale of data and data structures
follow an ever-increasing profile, and to support the growing
race of enormous bit streams in every system, intelligent
models need to be developed to select optimized parameters to
save on computation time. Hutter et al. [2] have highlighted this
potential area for growth, and the paper describes various
metaheuristic algorithms and approaches for enabling
automated machine learning. Developed on similar grounds,
this comprehensive review aims to provide in-depth insight
into the state-of-the-art metaheuristic optimization algorithms
for HPO in ML to facilitate R&D efforts and accelerate the
growth of this newly strengthening phenomenon.

2.1.2 Scope of the Review

This review encompasses a discussion of a wide range of

metaheuristic optimization algorithms that can be used directly
or as building blocks to enable HPO. These include the broad
categories of the genetic algorithms (GAs), particle swarm
optimization (PSO), differential evolution (DE), animal colony
and group behavior-based models, tabu search (TS), harmony
search (HS), simulated annealing (SA), etc. The paper not only
mentions the types of the algorithms but comments on their
pros and cons, conducts a comparative analysis with other
types, discusses their efficacy and efficiency in relation to real-
life scenario applications, and gauges their performance using
various parameters when integrated for HPO roles. This is
brought together by sifting through a vast compilation of
research journals to form the literature basis, such as this
comparative analysis for approaches like the grid search,
manual search, and random search cited in Bergstra and
Bengio [3]. This review also offers guidelines for choosing the
best-suited metaheuristic algorithm depending on the demand
of the application and offers insight into general standard
practices for dealing with complex HPO problems.

2.1.3 Organization of the Paper

This paper has been intelligently divided into 12 sections, each

highlighting a core concept or application set in the field of HPO
or associated topics. Section 2.1 is the overall introduction to the
paper, its background and motivation, the scope of the review,
and the organizational structure of the contents. Section 2.2
discusses the fundamentals of HPO, while section 2.3 dives into
an overview of various metaheuristic optimization techniques.
Sections 2.4, 2.5, and 2.6 explore the subcategories of the
techniques further, taking on population-based, single solution-
based, and hybrid metaheuristic techniques. Sections 2.7 and
2.8 are focused on applying metaheuristic applications in
Bayesian optimization and neural architecture search. Section
2.9 compares different techniques that are employed at the
broader industrial level using factors such as applications and
performance characteristics. Section 2.10 discusses the
applications that are brought about by the integration of
metaheuristics into machine learning models, and section 2.11
discusses future prospects and possible fields that hold great
potential in embracing metaheuristics and HPO while
addressing challenges and threats posed to the expansion of
ML. Section 2.12 rounds off the review with a consolidated
conclusion to comment on the paper and to bridge any gaps in
ideas for future research and development.

2.2 Fundamentals of Hyperparameter

Optimization

HPO can be considered to be the most fundamental step in

designing an efficient methodology for forming a smart and
complete ML algorithm. Neglecting to optimize
hyperparameters tends to increase the overall runs needed in
the learning phase, causes persistent inaccuracies and errors to
remain stagnant in the code, plagues the speed of the program,
and consumes resources that hold the potential to handle more
data and efficiently process other commands. HPO is a sensitive
and tedious task, with the challenges mainly originating from
the nature of the datasets, including the vast high-dimensional
search space and the non-linear constraints in the system.
Section 2.2 provides an overview of hyperparameters, their
importance, insight into the performance metrics used for HPO,
and the challenges associated with HPO.

2.2.1 Introduction to Hyperparameters

Hyperparameters differ from ordinary input variables or

parameters in the way they are addressed to the ML model
system. While a regular parameter set is inherent to the data
available to the ML model and is learned during the execution
of the training phase, hyperparameters are selected by the user
or by the governing algorithm. Hyperparameters form the basis
for HPO and can control the overall behavior of the algorithm
model, its convergence rate and performance, and the ability
for generalization. Popular examples of hyperparameters that
are frequently employed in applications include the rate of
machine learning, the batch size, the number of hidden layers
and sub-layers, and the number of neurons in the sub-layers.

2.2.2 Importance of Hyperparameter Optimization

HPO is a critical need in today’s world of smart and automated
process optimization and technological process or system
design. Traditionally, HPO was conducted by humans—with all
of the cumbersome calculations being done by hand [4]. As
demonstrated by the authors of Bergstra et al. [5], modern CPU
clusters and GPU modules can be programmed using various
algorithm-based approaches, i.e., using random search and
greedy sequential methods to automate HPO, yielding
remarkable results even for complicated systems such as neural
networks and deep belief networks (DBNs) [6, 7]. Optimizing
hyperparameters can directly reflect a great increase in
enhancing the performance, capability, and utility of such ML
models [8]. Additionally, it can reduce the model training time,
boost convergence accuracy, and help impart better
generalization ability. These improvements can lower the rate
and magnitude of the inherent error that is output by the model
and can also increase performance metric statistics such as F1
scores [9].

2.2.3 Performance Metrics for Hyperparameter

Optimization

Performance metrics are indicator statistics that are employed

to evaluate the performance, effectiveness, and efficiency of
various hyperparameter settings in the HPO process. A few
metrics that are frequently used in correlation to HPO include
accuracy, precision, F1 score, recall, and area under the
receiver operating characteristic curve (AUC-ROC). These
metrics also form the basis to assess the ability of a model to
generalize unseen data. While conventional user-dependent
HPO approaches had their performance slumps, modern
research efforts have yielded fruitful results in boosting the
overall effectiveness, accuracy, and convergence rate of the
process [10]. Thornton et al. [11] have employed a fully
automated Bayesian optimization approach for boosting the
performance of HPO in simultaneously selecting the learning
algorithm and associated hyperparameters. This was enabled
by taking into account various feature selection techniques (a
combination of three search and eight evaluator methods) and
WEKA-implemented classification approaches with two
ensemble methods, 10 meta-methods, 27 base classifiers, and
individual hyperparameter settings for each classifier. Snoek et
al. [12] have researched gaining higher tunability performance
for HPO by varying the nature of a Gaussian process, the type of
kernel, and the treatment of hyperparameters. This was
enabled by integrating Bayesian optimization techniques.

2.2.4 Challenges in Hyperparameter Optimization

HPO is a newly developing methodology and faces new and
unique challenges at present, including the increasingly
accumulating dimensions of the search space data, the huge
periods required for intensive computations, and the cost of the
needed hardware to perform well. Domhan et al. [13] have
worked to resolve the issue of poor HPO algorithm performance
and delayed run termination on large and complex datasets
with the help of learning curve extrapolation via probabilistic
models. Additionally, the challenge to tune hyperparameter sets
is an emerging issue that has proven to be slightly technical to
be dealt with. Probst et al. [14] have researched the latter and
have evaluated the problem of tuning from a statistical
standpoint. In addition, the paper acts as a guide for
quantifying the tunability of hyperparameters for HPO. The
existence of several local optima also adds to the difficulty of
finding an optimal solution, and this is a direct consequence of
the non-convexity of the objective function.

2.3 Overview of Metaheuristic

Optimization Techniques

Metaheuristic algorithm-based optimization techniques have

gained popularity in recent years in the ML development
industry—being an effective approach for solving evolving
complex problems, with the list including HPO. Section 2.3 dives
deeper into the world of metaheuristics to form a basis of the
fundamental concepts needed to grasp its definition,
characteristics, and classification.

2.3.1 Definition and Characteristics of Metaheuristics

Metaheuristics are problem-solving techniques or procedures

that provide a flexible approach to solving optimization
problems. Their superiority over the regular class of heuristic
algorithms is apparent by the prefix added up front in the word
itself as explained by Gandomi et al. [15] who define the term
“meta” as “beyond or higher level.” These algorithms are robust
in nature, are easy to implement into existing pieces of code,
and can be applied to a wide range of optimization problems,
including HPO. They are algorithms that explore the given
search space in incremental steps and guide the search based
on a heuristic criterion to find a good solution—the
convergence quality of which is often defined by the user.

They are characterized by their ability to escape local optima

during the search phase and by their adaptability in handling
complex and dynamic optimization problems. Gendreau et al.
[16] have extensively researched the field of metaheuristics to
compile their review paper and have keenly highlighted
important concepts and properties of the algorithms. These
include the ideas of intensification and diversification.
Intensification is the act of concentrating the search on areas
that are likely to contain good solutions, while diversification
involves exploring different areas of the search space to
prevent getting trapped in suboptimal solutions. Researchers
working on metaheuristics use these two strategies together to
find better-quality solutions more efficiently.

2.3.2 Classification of Metaheuristic Techniques

Metaheuristic techniques are generally inspired by forces of

nature, animal social behavior, and other observable
phenomena that might be natural, artificial, or digital and
might exist at the macro or micro scale. For simplicity,
metaheuristics can be categorized into three broad types, i.e.,
population-based, single solution-based, and hybrid
metaheuristics.

Population-based metaheuristics iteratively improve a

population of solutions for optimization. These are built upon
the principles of natural selection and the survival of the fittest.
Popular examples of population-based metaheuristics include
the genetic algorithms, differential evolution, and particle
swarm optimization.
Single solution-based metaheuristics focus on improving a
single solution by iteratively exploring its neighborhood as
opposed to a population solution set. These techniques are
particularly useful when convergence to a sole solution is
required and the state is proving to be difficult to obtain under
regular conditions. Examples of single solution-based
metaheuristics include simulated annealing, tabu search,
harmony search, and bat algorithm (BA).

Hybrid metaheuristics are the pinnacle of modern

developments in the world of metaheuristics. These techniques
are based on the combination of two or more metaheuristics to
harvest synergized benefits. This results in improved
inclusivity, performance, and accuracy when compared to the
characteristics of a single-parent term involved. For example,
the ant colony optimization (ACO) can be combined with a local
search-based algorithm to improve the quality of the solutions
obtained.

Presenting another classification perspective, Abdel-Basset et al.

[17] employ two major classification methods for segregating
metaheuristics. The first divides them into metaphor-based and
non-metaphor-based metaheuristics, and the second method
divides them into the lines of improved, adaptive, and
hybridized metaheuristics.
The range and versatility of metaheuristics cannot be defined,
since new forms are continuously being developed and rolled
out around the clock. It is important to note that no
metaheuristic technique can be considered universal, and the
real challenge in today’s world is finding the right algorithm
and the right proportions to get things done.

2.4 Population-Based Metaheuristic

Techniques

Section 2.4 of this paper focuses on population-based

metaheuristic techniques that are the foremost picks for HPO
applications, including the go-to options and new innovative
emerging methods. While each of them is unique in its own
sense, these algorithms explore a set of candidate solutions
called a population and gradually improve the solutions in each
run by iteratively applying selection, crossover, mutation, and
other operators. A few important population-based
metaheuristics are discussed below.

2.4.1 Genetic Algorithms

The genetic algorithms are robust metaheuristics based on the

principles of natural selection, evolution, and genetics [18–20].
These algorithms use a population of candidate solutions that
evolve and refine over time to form generations, and this is
achieved through selection, crossover, and mutation operations
[21, 22]. To demonstrate the versatility of the GA under the most
unique applications, Di Francescomarino et al. [23] have
developed a predictive process monitoring framework
integrated with the GA to forecast the performance of ongoing
businesses, basing the study on traces of past successes [24, 25].

2.4.2 Particle Swarm Optimization

The particle swarm optimization simulates the behavior of

social animals such as birds and fish that move and huddle in
swarms or groups [26] in order to find the best-fitting solution
for the given scenario [27, 28]. The PSO maintains a population
of particles in focus and observes their movements and mutual
interactions as they move through the search space to yield
results [29–32]. The PSO has been shown to be effective in HPO
for various machine learning algorithms as revealed by
Lorenzo et al. [33] in the research paper who successfully
obtained a high classification accuracy for the case of CIFAR-10
machine vision datasets, an integral part of a deep neural
network (DNN) study.

2.4.3 Differential Evolution

The differential evolution is an optimization technique that
uses a population of candidate solutions and differential
operators to explore the search space. Just like the rest, the DE
updates the population of solutions iteratively to improve the
accuracy of the results until the satisfactory convergence
criterion is reached. The DE has proven its utility under the
toughest scenarios, such as in the case mentioned by Zulfiqar et
al. [34] who have employed an adaptive differential evolution
(ADE) algorithm with a support vector machine (SVM) to
estimate the demand capacity of an electrical distribution grid
system. Combined with a multivariate empirical modal
decomposition (MEMD) algorithm, the MEMD-ADE-SVM
forecasting model delivered a high accuracy of 93.145% and
offered great stability and a faster convergence rate than
conventional approaches.

2.4.4 Ant Colony Optimization

The ant colony optimization is a population-based

metaheuristic technique inspired by the foraging behavior of
ants that swarm the search space and choose a hyperparameter
set. Each generation of these digital ants deposits pheromones
to guide existing and subsequent ants toward high-potential
holding areas, guiding the search effectively and speeding up
the optimization process. Lohvithee et al. [35] integrated the
ACO for optimizing X-ray computed tomography (XCT)
reconstruction and found the algorithm to have over 10 times
faster convergence than regular cross-validation methods.

2.4.5 Biogeography-Based Optimization

The biogeography-based optimization (BBO) is based on

mathematical equations that explain the geographical
distribution of organisms, and the same analogy is applied to
dataset populations for hunting optimal solutions. In this
technique, the search space is represented as a set of islands,
and the immigration and emigration rates between the islands
are used to update the candidate solutions. Simon [36] have
validated the accuracy of the BBO on a set of 14 standard
benchmarks and have demonstrated its performance in
comparison with seven popular biology-based optimization
algorithms such as the GA and PSO.

2.4.6 Cuckoo Search

The cuckoo search (CS) is an emerging metaheuristic technique

inspired by the natural brood parasitic behavior of cuckoo
birds [37]. The technique is based on generating a population of
candidate solutions through a combination of random walk and
Lévy flight methodology for exploring the search space [38, 39].
Yang and Deb [40] have successfully validated the newly
developed technique against test function checks and found it
superior to existing genetic and PSO algorithms for multimodal
objective functions.

2.4.7 Gray Wolf Optimizer

The gray wolf optimizer (GWO) is a metaheuristic technique

inspired by the hunting behavior of gray wolves. Replicating the
territorial behavior of wolves, the technique divides the search
space into territories, and a social hierarchy of four types of
gray wolves (i.e., alpha, beta, delta, and omega) is used to
update the population of candidate solutions. This is done in
accordance with the hunting behavior of the animals who hunt
in packs [41, 42]. According to Mirjalili et al. [43], the
metaheuristic holds the capability of simulating three main
steps of hunting: searching for prey, encircling the prey, and
attacking the prey. The authors have also benchmarked on 29
renowned test functions and found the GWO to match
mainstream metaheuristics such as the PSO and DE in terms of
result characteristics [44, 45].

2.4.8 Whale Optimization Algorithm

The whale optimization algorithm (WOA) is an emerging
metaheuristic technique that is based on employing three
operators for emulating the hunting behavior of humpback
whales including the search for prey, encircling the prey, and
bubble-net foraging. The methodology involves generating a
population of candidate solutions that is updated using three
types of whales (i.e., leader, challenger, and follower) to explore
the search space for best-fitting points [46]. Ashraf et al. [47]
have overcome the intense challenge of conducting HPO for a
deep deterministic policy gradient (DDPG) algorithm used to
regulate autonomous driving control, obtaining maximized
total rewards and maintaining a stable driving policy.

2.4.9 Recent Developments in Population-Based

Metaheuristics

Population-based metaheuristics seem to have been saturated

for single-objective optimization (SOO) owing to the consistent
research efforts being made in the field in recent decades.
Modern research developments have focused on the use of
hybrid techniques, multi-objective optimization (MOO), and
parallel computing to improve the performance of these
techniques further [48]. As established earlier, hybrid
techniques involve combining two or more metaheuristics to
boost HPO performance and convergence rate. MOO differs
from SOO in terms of setting goals and involves optimizing
multiple objectives simultaneously. The quality and success of
such developments can be seen in an example of a swarm
intelligence-based artificial bee colony (ABC) metaheuristic
algorithm. Karaboga and Basturk [49] have researched to
demonstrate the use of ABC for optimizing multi-variable
functions to yield promising results that outperformed the GA
and PSO. Parallel computing involves using multiple processors
to speed up the optimization process [50].

2.5 Single Solution-Based Metaheuristic

Techniques

Section 2.5 discusses the strategy of single solution-based

metaheuristics that aim to optimize a unitary solution instead
of refining a whole population of candidates. These techniques
are often chosen by coders when the solution space under
evaluation is continuous, and there is no need for multiple
solutions to exist in order to refine the quality of the resulting
solution.

2.5.1 Simulated Annealing

Simulated annealing is a technique inspired by the physical

process of annealing in metallurgy where a material (most
often a metal) is heated and slowly allowed to cool to reach a
low-energy state. Similarly, SA minimizes an objective function
by gradually decreasing toward a global minimum in refining
steps. As explained by the authors of [16], it accepts worse
solutions in hill-climbing moves with a certain probability to
escape local optima and decreases this probability over time—
this being the inherent minimization property of the SA
technique. SA is indeed a powerful tool to use, and this
statement can be backed by the work of Kirkpatrick et al. [51]
who solved the traveling salesman problem, a challenge that
belongs to a set of nondeterministic polynomial time complete
problem family, using SA.

2.5.2 Tabu Search

The tabu search is a local search-based technique that

overcomes the inherent weakness of local search methods
where the algorithm gets trapped between local optima. The TS
does this by using what is called a tabu list, storing a directory
of recently explored solutions in order to avoid revisiting them
in the following iterations. It explores the neighborhood of the
current solution and selects the best one that is not on the tabu
list. It can escape local optima by allowing moves that worsen
the objective function. Glover [52] have presented an in-depth
review of the TS and exhibit the hill-climbing behavior of the TS
in Figure 2.1 where multiple local optima can be seen and at
times the cost did worsen in search for the next minima.

Figure 2.1 Tabu search for optimizing the tour cost for a city plotted vs. iterations
[52].

2.5.3 Harmony Search

The harmony search is an emerging metaheuristic technique

that is based on the music industry, simulating the
improvisation process conducted by musicians to improve
soundtracks. It generates a set of random solutions called
harmony and improves them by iteratively selecting the best
elements from other harmonies. Like the TS and other
developed techniques, the HS can escape local optima by hill-
climbing. Geem et al. [53] have reflected on the choice of the HS
for HPO and solved the traveling salesman problem with great
results using the HS.

2.5.4 Bat Algorithm

The bat algorithm (BA) is a newly developed metaheuristic

generated by Yang [54]. It is inspired by the echolocation
technique employed by bats to locate prey in the dark and
works by generating a set of solutions that it then moves toward
the best solution found so far. Keeping the hunt on, it introduces
random walks to explore new regions of the solution space. The
inventors have found it superior to the GA and PSO and intend
on developing hybridized variants based on the BA.

2.5.5 Recent Developments in Single Solution-Based

Metaheuristics

Single solution-based metaheuristics have been developed for

improved local search performance in recent times, integrating
the use of machine learning to escape local minima entrapment
as proposed by Jaddi and Abdullah [55]. Additionally, the ever-
continuing hybridization research efforts have borne fruit in
improving single solution-based techniques in terms of
accuracy and efficiency. These have led to various
combinations, and the room for creativity seems to be endless.
An example would be a hybridized synthesis of SA with the TS
for improved exploration capability.

2.6 Hybrid Metaheuristic Techniques

Hybrid metaheuristic techniques have been briefly discussed

earlier in this paper several times, all due to their strong
presence and keen influence in today’s metaheuristic
approaches. Building a stronger foundation on the concept,
these techniques can be formally introduced as an approach
that combines two or more metaheuristic algorithms to
synergize the strengths of each algorithm and overcome their
weaknesses. It is important to note that while it is desired that
the pair would be able to sufficiently bridge the gaps inherent
to the individual components, it is not necessary to reach a
perfect combination, and this requires trial and error between
the choice of algorithms and in the degree of fractional
composition. This section provides an overview of three
commonly used hybrid metaheuristic techniques, i.e., GA-PSO,
GA-SA, and PSO-TS hybrids.

2.6.1 Genetic Algorithm and Particle Swarm

Optimization Hybrid
One of the most widely used hybrid metaheuristic techniques is
the genetic algorithm and particle swarm optimization (GA-
PSO) hybrid, combining the most fundamental algorithms to
produce a robust and powerful output. Jeong et al. [56] aimed to
improve the diversity of the GA and the convergence rate of the
PSO with this combination. The GA-PSO hybrid technique was
applied to two test functions, and after proper tuning, it was
found that the search performance was raised by a satisfactory
level. It is important to note that the GA-PSO hybrid has raised
its status to hold popular applications in solving optimization
problems for engineering design, scheduling, and image
processing.

2.6.2 Genetic Algorithm and Simulated Annealing

Hybrid

Another popular hybrid technique is the genetic algorithm and

simulated annealing (GA-SA) hybrid. The GA-SA hybrid has
been successfully applied to various optimization problems,
including the traveling salesman problem, machine learning,
and wireless sensor networks. Sanagooy Aghdam et al. [57]
solved the challenge of optimized asset tracking that was to be
done using radio frequency identification (RFID) systems. Using
the pros of the GA, a multi-objective function was developed for
maximizing the network coverage of the antennas, and using
the pros of SA, the hyperparameters of the total cost, rate of tag
reader collision, interference, and total energy consumption of
the RFID antennas were minimized. The GA-SA hybrid
metaheuristic allowed for savings (on average) in the total cost,
making use of three types of readers with one, two, and four
antenna ports in the most efficient way.

2.6.3 Tabu Search and Particle Swarm Optimization

Hybrid

The tabu search and particle swarm optimization (TS-PSO)

hybrid is another effective hybrid technique that combines the
global search capabilities of the PSO with the local search
capabilities of the TS. The TS-PSO hybrid has found great
applicability to various optimization problems, including
feature selection, image segmentation, and clustering. Zhang
and Wu [58] conducted experiments on four test functions
including Easom, Rosenbrock, Hump, and Rastrigin to validate
the TS-PSO combination for increasing the probability of
finding the optimal minima in the search space. It was
discovered that the TS-PSO hybrid had higher performance
than conventional standalone integrations of the GA, TS, and
PSO for the same task.

2.6.4 Recent Developments in Hybrid Metaheuristics

The concept of hybridization itself is the ultimate direction of
growth and development that has been selected for all
preceding metaheuristics discussed above [59]. However, this
did not rule out progress in the field of hybrid metaheuristics,
and they have rapidly improved both vertically in detail and
complexity and laterally in inclusivity. De et al. [60] have
worked tirelessly to compile recent advances in hybrid
metaheuristics, with the most notable research being conducted
on developing hybrid algorithms for complex data clustering to
raise computational intelligence.

2.7 Metaheuristics in Bayesian

Optimization

Metaheuristics are gaining an increasing role in Bayesian

optimization applications for solving complex optimization
problems, including HPO. Section 2.7 focuses to deliver insight
into the fundamentals of Bayesian optimization and related
concepts such as Gaussian process regression and acquisition
functions. Lastly, recent developments in metaheuristic-based
Bayesian optimization are discussed to provide a complete
overview.

2.7.1 Background of Bayesian Optimization

Bayesian optimization is a search-guiding approach based on
statistical methods for global optimization. It is the top choice
for evaluating the objective function when it is complicated,
non-differentiable, discontinuous, or time-consuming. The
approach involves constructing a probabilistic model of the
objective function and using it to make decisions about where
to sample next, and this is done until the solution has been
adequately refined. Bayesian optimization has proven to be
highly resourceful for optimizing design choices. Shahriari et al.
[61] have researched the technique at an industrial level and
have analyzed the benefits that Bayesian optimization can
produce in big data applications. They found it to provide a
greater automation potential that has boosted both product
quality and human productivity.

2.7.2 Gaussian Process Regression

Gaussian process regression is a statistical modeling technique

that is often used in the Bayesian optimization to model the
objective function. It involves constructing a Gaussian process
(GP), which is a probabilistic model that can be used to make
predictions about the function values at newly established
input points. The model’s predictions are not as accurate and
require gauging for the level of confidence—marked by
estimates. Being a versatile tool, GPs can be used to model both
linear and nonlinear relationships between the input
parameters and the corresponding output variables. The shape
and behavior of the GP are governed by the estimated
hyperparameters that were used to train the model [62, 63].

2.7.3 Acquisition Functions

Acquisition functions are used in Bayesian optimization as a

tool to guide the decision process for locating regions for the
next sampling. These functions incorporate a balance between
exploration (the idea of sampling in regions where the function
has high uncertainty) and exploitation (the idea of sampling in
regions where the function is likely to be optimal and have local
maxima or minima). This can be seen in Figure 2.2 where the
acquisition function is high in regions of high exploration and
exploitation.
Figure 2.2 A Gaussian process approximation of an objective function being
iteratively refined over iterations. The green shaded region shows the acquisition
function [64].

Common acquisition functions used in the Bayesian

optimization technique include expected improvement (EI),
probability of improvement (PI), Thompson sampling (TS),
knowledge gradient (KG), and upper confidence bound (UCB).

2.7.4 Recent Developments in Metaheuristic-Based

Bayesian Optimization
Recent developments in Bayesian optimization include the use
of metaheuristics to optimize the acquisition function
parameters and the development of efficient parallel
algorithms. Many attempts have been made to enhance the
effectiveness and broaden the scope of Bayesian optimization.
It is important to note that GPs involved in Bayesian
optimization tend to scale cubically with the number of
observations, and this does create challenges in handling the
objectives when several evaluations are needed per objective.
Snoek et al. [65] have solved this parallelizing issue using
linearly scaling deep neural networks as a competitive
alternative to GP. Additionally, Brochu et al. [64] have
researched to propose two developed extensions of Bayesian
optimization: active user modeling with preferences and
hierarchical reinforcement learning for improving the
performance, inclusivity, and degree of control.

2.8 Metaheuristics in Neural Architecture

Search

Neural architecture search (NAS) is a rapidly growing field that

deals with finding the optimal architecture for neural networks.
At present, the majority of neural architectures are created
through manual development by human experts, which is
prone to errors and can be a time-consuming process.
Consequently, there is a growing demand for automated
methods of neural architecture search, and metaheuristic
algorithms have gained global attention due to their ability to
efficiently explore a large search space and find optimal
solutions. Section 2.8 aims to provide an introduction to NAS,
highlight key applications of metaheuristics in NAS, and the
recent developments in this emerging field.

2.8.1 Introduction to Neural Architecture Search

Neural architecture search is a technique used to automate

machine learning and the search tasks for hunting optimal
architectures of a neural network. NAS involves sifting through
a large space of candidate architectures to search for the
optimal element in the array that would give out the best
performance for the given task at hand. Owing to the
complexity of modern systems and the large bandwidth
requirements, the search space can be very large, making the
process computationally expensive and very exhaustive in
terms of time and resources. Developments in metaheuristic
algorithms have allowed efficient searches in the space to locate
optima in the field of possible architectures.
2.8.2 Applications of Metaheuristics in Neural
Architecture Search

Metaheuristic algorithms have been successfully applied in

many NAS studies to find optimal neural network architectures.
Some of the popular metaheuristic algorithms used in NAS
include the GA, PSO, SA, and TS. Convolutional neural networks
(CNNs), deep neural networks, and recurrent neural networks
are examples of popular neural architectures that have been
found to be highly compatible with metaheuristic integration
for automating the search process for the ideal design. The
authors of [66] have applied the GA on CIFAR10 to demonstrate
the ability to search for network architectures via a small
dataset experiment. The proposed solution allowed the GA to
automatically learn the structure of deep convolutional neural
networks [67]. The authors of [68] have found NAS methods to
have outperformed manually designed architectures in areas of
image classification, object detection, or semantic segmentation.

2.8.3 Recent Developments in Metaheuristic-Based

Neural Architecture Search

Recent research has focused on improving the efficiency and

accuracy of NAS with metaheuristic algorithms. One approach
is to use a neural architecture search space that is more
compact, allowing for a more efficient search. Additionally,
there have been considerable research efforts on hybrid
metaheuristic algorithms in recent times—aimed at combining
the strengths of multiple metaheuristics to achieve better
performance in NAS. Atop of the usual approaches, the authors
of [69] have developed the GA-based NeuroEvolution of
Augmenting Topologies (NEAT) that boasts increased efficiency
over the current best fixed-topology methods for reinforcement
learning tasks [70]. The major increase in performance
efficiency can be attributed to employing a principled method
of crossover of different topologies using speciation for
protecting structural innovation and growing from minimal
structures incrementally. The authors of [71] have introduced a
novel algorithm called DARTS (Differentiable ARchiTecture
Search) that is based on bilevel optimization, designed to take
on the scalability challenge encountered by conventional
architecture search methods. Instead of applying evolution or
reinforcement learning over a discrete and non-differentiable
search space, DARTS allows for continuous relaxation of the
architecture representation to enable the use of gradient
descent. The algorithm has proven itself on several leading
benchmarks, yielding competitive results on CIFAR-10 and
excelling in performance on PTB dataset (it is the name of a
dataset, PTB stands for Physikalisch-Technische Bundesanstalt
(PTB)). The algorithm has increased the overall efficiency of the
architecture search process and has reduced the exploration
time to a few GPU days only. Keeping with the boosted interest
of researchers in the field of NAS in recent times, the continued
exploration and advancement of metaheuristic algorithms in
NAS hold great promise for the future.

2.9 Comparison of Metaheuristic

Techniques for Hyperparameter
Optimization

Countless metaheuristic techniques have been widely used for

HPO in the defined context, but there has been no definition
made so far to provide an insight into how each ranks in
comparison and which algorithm is to be selected for a given
problem. Therefore, there is a prominent need to evaluate and
compare these techniques to determine the most effective ones,
providing both a generalist view and a specific view under
different cases. This section hence focuses on the comparison of
the abovementioned metaheuristic techniques for HPO.

2.9.1 Criteria for Comparison

To perform a fair and objective comparison of metaheuristic

techniques for HPO, it is essential to establish agreed-upon
criteria for comparison that follows the defined objectives and
scope of the desired application in which the metaheuristic will
be implemented. These criteria include the efficiency, accuracy,
and robustness of the techniques [72]. Efficiency refers to the
smart use of computational resources required by the
technique to obtain the optimal hyperparameters, wasting
minimal time while doing so [73]. Accuracy is a measure of the
trueness of the output of the metaheuristic technique in
locating and identifying the optimized global value. The
robustness of the algorithm is the ability to perform well under
different conditions and for different types of models without
requiring large changes to the code and without requiring
excessive retraining altogether [74]. Choosing a unique
approach, Blum and Roli [75] have researched the
intensification and diversification (I&D) characteristics of
commonly used metaheuristics and chose to use these as a
comparative criterion for functional behaviors [76, 77].

2.9.2 Comparative Analysis of Metaheuristic

Techniques

Based on existing literature and empirical relation-based

research studies, it is difficult to conclude which metaheuristic
technique is the best for HPO. In general, the GA and PSO have
shown good performance on a wide range of hyperparameter
optimization problems, while SA and DE have shown promising
results on specific problem domains. However, there always
exists a need for in-depth analyses to support an empirical idea.
Several studies have evaluated different metaheuristic
algorithms and the way they process the search space to look
for optima, and the results vary case by case, depending on the
optimization problem and the specific hyperparameters being
optimized. Table 2.1 summarizes the research findings of Blum
and Roli [75], who have broken down the I&D components of
the most frequently used metaheuristic techniques for a
comparison of sub-methods and steps used in each [78].
Table 2.1 Breakdown of popular metaheuristics and their I&D components [75].

Metaheuristics I&D component

SA acceptance criterion + cooling schedule

TS neighbor choice (tabu lists)

aspiration criterion

EC recombination

mutation

selection

ACO pheromone update

probabilistics construction

ILS black box local search

kick-move
 Metaheuristics I&D component

acceptance criterion

VNS black box local search

neighborhood choice

shaking phase

acceptance criterion

GRASP black box local search

restricted candidate list

GLS penalty function

While the study concluded with the identification of these

subtasks and a guideline to run experiments to evaluate the
differences originating due to the variations in these
mechanisms, several other studies have followed through to
further the vision.
Table 2.2 Performance comparison of four different metaheuristics based on
average, median and standard deviation [80].

Loss on Best value Worst

Algorithm
LeNet-5 (loss) value (loss

ABC ƒAVG 0.039988012 0.036257669 0.04684992

ƒMEDIAN 0.039912244

ƒSTD 0.002684507

ALO ƒAVG 0.038987691 0.03411367 0.04337282

ƒMEDIAN 0.039091125

ƒSTD 0.002267059

BA ƒAVG 0.041151504 0.03392095 0.04544504

ƒMEDIAN 0.041771095

ƒSTD 0.00265998

PSO ƒAVG 0.040463663 0.04651067 0.04651067

 Loss on Best value Worst
Algorithm
LeNet-5 (loss) value (loss

ƒMEDIAN 0.039871245

ƒSTD 0.002462824

Traditionally, comparative analysis of metaheuristic techniques

has been conducted through various benchmark problems,
including optimization of the hyperparameters of convolutional
neural networks, decision trees (DTs), and support vector
machines [79]. Gaspar et al. [80] have studied the HPO for CNNs
using four different metaheuristics, i.e., the particle swarm
optimization, ant lion optimization (ALO), artificial bee colony,
and bat algorithm. Each metaheuristic used the same CNN
architecture and the same MNIST database—running 30 times
to accumulate the needed data for comparison. It was found
that the ALO stood out the best for the case of CNN optimization
based on statistical parameters of mean, median, and standard
deviation.

Another comparison can be attributed to the work of Ezugwu et

al. [81] who have devoted their efforts to providing a
differentiating insight on eight prominent algorithms, i.e., the
particle swarm optimization, spotted hyena optimizer (SHO),
gray wolf optimizer, ant colony optimization, whale
optimization algorithm, gravitational search algorithm (GSA),
moth–flame optimization (MFO), and bat algorithm. After
running 23 benchmark tests and analyzing the exploration and
exploitation of these algorithms’ convergence curves, it was
concluded that the SHO and GWO stood out from the rest in
terms of the quality and accuracy of the global optima value.

2.9.3 Performance Evaluation of Metaheuristic

Techniques

The performance evaluation of metaheuristics, specifically for

the case of HPO, is usually conducted using benchmark datasets
and ML models. This involves a comparison based on a variety
of established and recognized metrics, including the accuracy of
the resulting model’s convergence speed, success rate, and
quality of the final solution. The experiments’ outcomes are
subsequently evaluated to identify the merits and demerits of
each technique. In certain cases, trade-offs between strong
areas are required to reach a compromised overall
performance needed for the given application, since no
metaheuristic can generally perform well in all directions
needed.
Continuing discussion on the previously stated study by Gaspar
et al. [80], Table 2.2 shows the performance comparison of four
algorithms based on the average mean, median, and standard
deviation, tabulated across losses. The results show that the
ALO performs the best among the rest in terms of the
abovementioned metrics.

Another comparison metric for performance is the rate and

ease of convergence, keeping with the least loss incurred.
Figure 2.3 shows the convergence comparison of four
metaheuristics based on the first 10 iterations that shows that
the ALO and other algorithms stabilized in under two iterations
while the ABC was found to be the worst for the given
application, stabilizing after five iterations. It can also be seen
that for the first 10 iterations, the ALO had the least overall
losses incurred during the runs.

Similarly, the work of Ezugwu et al. [81] includes an evaluation

for a convergence-based comparison of the eight renowned
algorithms studied. Figure 2.4 shows the SHO and GWO
converging in the least number of iterations—going below 10-20
in just over 200 iterations. The GSA can be seen to flatline after
approximately 200 iterations after little initial decay, while the
rest move out of the window for 1,000 iterations without
reaching satisfactory convergence. The WOA proves to be the
next best thing to the GWO in the benchmark test.

Figure 2.3 Convergence comparison of four metaheuristics based on the first 10

iterations [80].
Figure 2.4 Best score convergence profiles vs. iterations for eight renowned
algorithms [81].

Sonmez [82] compared the performance of eight population-

based metaheuristic algorithms (GA, ACO, PSO, ABC, GSA, firefly
algorithm (FA), GWO, Jaya) for the application of optimal truss
design evaluation in three-dimensional (3D) space. It was found
that for double-layer grid (DLG) designs with over 53 variables,
the PSO, GSA, and FA converged poorly and were highly
sensitive to increment compared to the rest. Overall, the GWO,
JAYA, and ABC had better results for runs having over 20
dimensional variables, as shown in Table 2.3 that shows the
performance comparison of eight population-based
metaheuristics for the DLG truss problem below.
It was also discovered that the nature and count of control
parameters can be another influencing factor in gauging
performance and would require additional studies.

At times, choosing the best-performing algorithm becomes

highly necessary. Such is the case of critical applications that
are highly sensitive. This includes fraudulent bank transaction
detection, taken up by Tayebi and Kafhali [83]. The paper
involves the performance comparison of industry-leading
metaheuristic algorithms including the GA, DE, ABC, GWO, PSO,
teaching learning-based optimization (TLBO), and GS, with a
range of ML models such as decision tree, mlp classifier (MLP),
k-nearest neighbors (KNN), logistic regression (LR), random
forest (RF), support vector machine classifier, and AdaBoost
(AD). The optimizers were compared based on accuracy,
precision, recall, F1 score, and the area under the ROC curve
(AUC) metrics.
Table 2.3 Performance comparison of eight population-based metaheuristics for the
DLG truss problem [82].

GA ACO PSO ABC GSA

ϕbest 32.154 33.721 73.255 29.994 169.431

(tons)

Viol 1.3E - 0 0 0 0
08

37.34 46.38 95.30 34.57 183.06

σ 3.06 10.63 10.33 2.75 10.73

(TONS)

# SA 54,110 53,010 53,000 53,045 53,000

Time 0.268 0.267 0.247 0.239 0.203

(h)

The most important characteristic of global optima is accuracy.

According to Figure 2.5, different metaheuristic techniques
have varying performances for different ML models. LR
produced the same results for most metaheuristic techniques
except the PSO, while the DT provided higher accuracy
compared to using the grid search (GS). RF performed better
when using the ABC, DE, and TLBO. The SVM had better
accuracy using the TLBO, while AD showed better accuracy
using the GS. Meanwhile, MLP had better accuracy using the GA
and GS, and KNN demonstrated better accuracy using the GA.

Figure 2.5 Accuracy of metaheuristics for different ML models [83].

Therefore, it can be concluded that there are high variations in

the overall performance of metaheuristics, affected strongly by
several parameters, the nature of the application, and the type
of ML model being used. While it may be possible to collectively
group tasks and applications of similar nature and see what
algorithm generally stands out the best for each, it is still a
better practice to run personalized trial-and-error routines for
the task at hand.

2.10 Applications of Metaheuristics in

Machine Learning

Metaheuristics play an important role in automating processes

in the digital industry, including ML. Machine learning is a
rapidly growing field that aims to create intelligent algorithms
capable of learning from datasets and making decisions or
predictions based on newly acquired information. Section 2.10
focuses to discuss various applications of metaheuristics in the
broad categories of ML, including supervised, unsupervised,
reinforcement, and deep learning for integrating HPO in
intelligent systems.

2.10.1 Supervised Learning

Supervised learning is an ML technique that is frequently

employed to train a model in applications where the program is
to discern between easily predictable outcomes. These usually
have a consistency built-in such as a recurring pattern or
progressive sequence. The technique is enabled by providing
the system with inputs and their corresponding outputs (the
concept of labeled data) in order to establish functional
relations between the two. This trained function can then be
applied to make predictions on unseen data. Metaheuristics
have been applied in the supervised learning domains mainly
for feature selection (FS) [84, 85], training ANNs and SNNs, HPO,
and for model selection (MS). Javanshir et al. [86] have used a
metaheuristic-based supervised learning technique for training
spiking neural networks (SNNs), overcoming the challenge of
discontinuities and the non-differentiable inherent nature of
the spiking neuron (SN). Metaheuristics such as the HS, CS, DE,
PSO, GA, ABC, and gene expression (GE) were implemented to
run the trials, with the CS standing out from the rest to exhibit
the best performance across four classification benchmarks.
Turning toward HPO, Xian et al. [87] have improved a sine
cosine algorithm (SCA) for finding global optima by integrating
a teacher supervision learning (TSL) model. The hybridized
TSL-SCA has better population control and distributivity, higher
convergence speed, and balanced exploration and exploitation.
The TSL-SCA proved to be more competitive than the DE, PSO,
MFO, CS, and WOA algorithm types.

2.10.2 Unsupervised Learning

Unsupervised learning differs from the supervised mode in

terms of data labels. It needs to be smart enough to identify
patterns or structures in the data while training the model
using unlabeled data. Metaheuristic algorithms have been used
for clustering, dimensionality reduction, and anomaly detection
in unsupervised learning. Clustering, being the most important
application of all, involves grouping similar data points
together for means of easier handling and efficiency for pattern
finding. Metaheuristic algorithms possess the ability to optimize
this process smoothly, and several research efforts have been
dedicated in this direction. Jahwar and Abdulazeez [88] have
discussed the use of SI approaches, including the PSO, ACO, and
ABC for K-means clustering, with the PSO performing the best.
Similarly, Kaur and Kumar [89] have devised a new water wave
optimization (WWO)-based metaheuristic to solve the clustering
problem. The WWO has been known to be effective in tackling
constrained and unconstrained optimization issues but has a
shortcoming that causes premature convergence. The authors
have improved the algorithm with a modified search
mechanism and decay operator and tested the new WWO in 13
benchmarks, gaining higher accuracy and F-score rates for
clustering problems than usual metaheuristics.

2.10.3 Reinforcement Learning

Reinforcement learning involves training a model to adopt

positive behaviors that are encouraged and reject negative ones
that were discouraged by feedback received from the user. The
model then makes decisions based on this feedback rewarding
and correcting received. Reinforcement models require high-
dimensional storage spaces and are computationally intensive,
since they require data to be generally sorted into two classes
as they operate in an accept/reject manner. Metaheuristic
algorithms have been used for policy search, global
optimization in HPO problems, and deep reinforcement
learning. Seyyedabbasi [90] have combined a sand cat swarm
optimization (SCSO) algorithm with integrated reinforcement
learning to create an RL-SCSO metaheuristic hybrid for efficient
exploration and exploitation for global optima in the search
space. The study was validated on 20 benchmark tests, and the
results were found appreciable. Seyyedabbasi et al. then rose
the bar further in [91], developing three new hybrid
metaheuristics, i.e., RLI-GWO, RLEx-GWO, and RLWOA. By
implementing a control mechanism, the reward and penalty
values for each action were obtained, and the algorithms were
simulated over 30 benchmark functions to generate statistical
data on their performance. It was found that the RL-based
hybrids performed better than conventional algorithms in
terms of exploration and exploitation abilities, with RLWOA
outdoing the other two hybrids.

2.10.4 Deep Learning

Deep learning involves training deep neural networks to learn
complex patterns in data. Metaheuristic algorithms have been
used for various tasks in deep learning, including HPO,
architecture search, and training DNNs. Tian and Fong [92]
provide a comprehensive review of deep learning integration
for DNNs, discussing the overall process, pros and cons, the
importance and need of metaheuristic techniques, and the use
of the GA and PSO for HPO in ANNs. In a similar effort, Rere et
al. [93] present the use of SA to boost the performance of a CNN.
Providing success with an implementation of an MNIST dataset,
the proposed method did have to make an increased
computation time trade-off. Another important discovery in the
field is made by Kingma and Ba [94], who have created Adam,
an algorithm that can handle first-order gradient-based
optimization of stochastic objective functions. The research
successfully demonstrated the versatility of the Adam algorithm
that holds the ability to scale to large-scale high-dimensional
machine learning problems, compares well against existing ML
models to handle CNNs and DNNs, and can even efficiently
solve practical deep learning problems.

2.11 Future Directions and Open

Challenges
The field of metaheuristics is an emerging one, exploring new
directions, taking new paths, and developing deep in a variety
of unique ways. Entering into the world of ML opens up a lot
more possibilities for the future, and just as more advanced
techniques are developed and new applications are discovered,
the need for improvement grows exponentially. This allows for
R&D into metaheuristics to hold great potential. However, the
pros come with cons. Branched-out opportunities in new
ventures also pose an increased number of open challenges that
must be conquered to fully master metaheuristics and their
true abilities. These include challenges of complex model
training, high-dimensional data handling, and ANN and DNN
optimization.

2.11.1 Opportunities for Improvement in

Metaheuristics

Metaheuristic algorithms have remarkably proven themselves

to be useful tools for optimization problems in ML, but there is
always room for improvement. The development of new
metaheuristic techniques that can address the challenges of
high-dimensional data and complex optimization landscapes is
the foremost approach. Refining the art of creating better-
hybridized algorithms stands second in the list – finding the
right algorithms to combine and in the right proportions. Ting
et al. [95] have discussed the past, present, and future of hybrid
metaheuristics, extending hope for wider research horizons
and countless possibilities of research and development to
harness infinite opportunities. Lastly, improving existing
algorithms in terms of inclusivity and capabilities remains a
forever green task that is to remain a continual growth process
till the end of time. Areas of growing metaheuristic applications
include advanced quantum computing applications,
nanotechnology, NFT and cryptocurrency transactions, space
exploration simulations, and others.

2.11.2 Adapting Metaheuristics to New Machine

Learning Paradigms

As ML continues to evolve, new paradigms are emerging that

require the development of new optimization techniques to
keep up with industrial benchmarks and with the intense
competition in the global markets. Philosophies such as lean
entrepreneurship, and other resource-managing methodologies
require utmost optimization throughout all processes involved
in the producer-to-consumer chain. Pairing metaheuristics with
techniques such as continual learning, cognitive computing,
quantum machine learning, adversarial machine learning,
explainable AI, federated learning, human-in-the-loop learning,
meta-learning and more might be the key to major
breakthroughs in the waiting. Developing metaheuristics that
can work with these new ML paradigms will definitely be an
important area of research in the future.

2.11.3 Addressing the Computational Complexity and

Scalability

With the increasing size and complexity of ML models,

scalability and computational efficiency become significant
challenges to counter in the current digital era of smart
optimization and control. Keeping with the trends, the future is
likely to add more challenges in terms of increased bandwidth
and dimensional requirements for processing and storage.
Although time taking, this needs to be addressed and
substantial efforts should be concentrated on the issue.
Additionally, there is a need for more research on parallel and
distributed metaheuristic algorithms to speed up HPO. There is
also a need to develop metaheuristics that can synergize with
newer computing architectures, such as GPUs and FPGAs to
exploit hardware efficiently and speed up optimization.

2.12 Conclusion

It can be agreed that metaheuristic algorithms and HPO are

inseparable ideas in today’s digital world. With the rapid
developments being made in the industrial and artificial
technology sectors around the clock, it is becoming increasingly
important for companies to outdo the competition and to keep
on improving product quality and diversity while keeping
factors such as costs and process time low. This is where HPO
comes into play for finding global optima for process
parameters to boost desired metrics. Throughout this paper, it
was seen that manual optimization is indeed a laborious task
and requires an eternity of corrections, iterations, and error
finding. Metaheuristics are indeed a powerful tool that can be
used as a real-life hack to execute HPO in a smart and efficient
manner. To build a sound foundation, this review paper
covered important concepts needed to comprehend the
complete idea of HPO, including terms in the field of
metaheuristics, ML, AI, etc., and also examined unique
optimization techniques employed by various researchers. The
paper has also identified several areas of future research for
improving the performance of metaheuristics in ML and also
discusses challenges that need to be conquered to keep the push
going on. Overall, it can be said with high confidence that the
field holds significant potential for improving the accuracy and
efficiency of ML algorithms – through both innovation of newer
techniques and by hybridizing combinations of existing
variants, and can help to address the challenges associated with
HPO, feature selection, and model selection.

References

1. Gupta, K.K. et al., Machine learning-based predictive

modelling of biodiesel production-A comparative
perspective. Energies, 14, 4, 1122, 2021, doi:
10.3390/en14041122.
2. Hutter, F., Kotthoff, L., Vanschoren, J., Automated Machine
Learning, Springer: New York, NY, USA, 2019.
3. Bergstra, J. and Bengio, Y., Random search for hyper-
parameter optimization. J. Mach. Learn. Res., 13, 281–305,
2012.
4. Feurer, M. and Hutter, F., Hyperparameter optimization, in:
Automated Machine Learning: Methods, Systems, Challenges,
pp. 3–33, 2019, doi:10.1007/978-3-030-05318-5_1.
5. Bergstra, J.S., Bardenet, R., Bengio, Y., Kégl, B., Algorithms for
hyper-parameter optimization. In Advances in Neural
Information Processing Systems 24, J. Shawe-Taylor, R.S.
Zemel, P.L. Bartlett, F. Pereira, K.Q. Weinberger (Eds.), pp.
2546–2554, Curran Associates, Inc., Granada, Spain, 2011.
6. Eggensperger, K., others, Towards an empirical foundation for
assessing bayesian optimization of hyperparameters, in: NIPS
 Workshop on Bayesian Optimization in Theory and Practice,
2013.
7. Hutter, F. et al., Sequential model-based optimization for
general algorithm configuration, in: Learning and Intelligent,
Optimization: 5th International Conference, LION 5, Rome,
Italy, January 17-21, 2011, 2011, Selected Papers 5.
8. Falkner, S. et al., BOHB: Robust and efficient hyperparameter
optimization at scale, in: International Conference on Machine
Learning, 2018.
9. Bergstra, J. et al., Implementations of algorithms for hyper-
parameter optimization, in: NIPS Workshop on Bayesian
Optimization, 2011.
10. Chen, T. and Guestrin, C., Xgboost: A scalable tree boosting
system, in: Proc. 22nd ACM SIGKDD International Conference
on Knowledge Discovery and Data Mining, pp. 785–794, 2016,
doi: 10.1145/2939672.2939785.
11. Thornton, C. et al., Auto-WEKA: Combined selection and
hyperparameter optimization of classification algorithms, in:
Proc. 19th ACM SIGKDD International Conference on
Knowledge Discovery and Data Mining, pp. 847–855, 2013, doi:
10.1145/2487575.2487629.
12. Snoek, J., Larochelle, H., Adams, R.P., Practical bayesian
optimization of machine learning algorithms. in: Proceedings
of the 25th International Conference on Neural Information
 Processing Systems, vol. 2, pp. 2951–2959, Granada, Spain, 20
November 2020.
13. Domhan, T. et al., Speeding up automatic hyperparameter
optimization of deep neural networks by extrapolation of
learning curves, in: Twenty-Fourth International Joint
Conference on Artificial Intelligence, 2015.
14. Probst, P. et al., Tunability: Importance of hyperparameters
of machine learning algorithms. J. Mach. Learn. Res., 20,
1934–1965, 2019.
15. Gandomi, A.H. et al., Metaheuristic algorithms in modeling
and optimization, in: Metaheuristic Appl. Struct. Infrastruct.,
vol. 1, 2013, doi:10.1016/B978-0-12-398364-0.00001-2.
16. Gendreau, M. and Potvin, J.-Y., Handbook of metaheuristics,
in: International Series in Operations Research &
Management Science, 2nd ed., M. Gendreau, J.-Y. Potvin
(Eds.), p. 648, Springer US, Boston, MA, USA, 2010.
17. Abdel-Basset, M. et al., Metaheuristic algorithms: A
comprehensive review, in: Comp. Intell. Multimedia Big Data
Cloud Eng. Appl., pp. 185–231, 2018, doi:10.1016/B978-0-12-
813314-9.00010-4.
18. Holland, J.H., Adaptation in natural and artificial systems: An
introductory analysis with applications to biology, control, and
artificial intelligence, MIT Press, Cambridge, MA, USA, 1975.
19. Kalita, K. et al., Genetic programming-assisted multi-scale
optimization for multi-objective dynamic performance of
laminated composites: The advantage of more elementary-
level analyses. Neural Comput. Appl., 32, 12, 7969–7993, 2020,
doi: 10.1007/s00521-019-04280-z.
20. Kalita, K. and Chakraborty, S., An efficient approach for
metaheuristic-based optimization of composite laminates
using genetic programming. Int. J. Interact. Des. Manuf., 17, 2,
899–916, 2023, doi: 10.1007/ s12008-022-01175-7.
21. Kalita, K. et al., Optimizing drilling induced delamination in
GFRP composites using genetic algorithm& particle swarm
optimization. Adv. Compos. Lett., 27, 1, 2018, doi:
10.1177/096369351802700101.
22. Kalita, K. et al., Optimizing process parameters for laser
beam micro-marking using genetic algorithm and particle
swarm optimization. Mater. Manuf. Processes, 32, 10, 1101–
1108, 2017, doi: 10.1080/10426914.2017.1303156.
23. Di Francescomarino, C.D. et al., Genetic algorithms for
hyperparameter optimization in predictive business process
monitoring. Inf. Syst., 74, 67–83, 2018, doi:
10.1016/j.is.2018.01.003.
24. Kalita, K. et al., Parametric optimization of non-traditional
machining processes using multi-criteria decision making
techniques: Literature review and future directions.
 Multiscale Multidiscip. Model. Exp. Des., 6, 1, 1–40, 2023, doi:
10.1007/s41939-022-00128-7.
25. Kalita, K. et al., Robust genetically optimized skew laminates.
Proc. Inst. Mech. Eng. C, 233, 1, 146–159, 2019, doi:
10.1177/0954406218756943.
26. Kennedy, J. and Eberhart, R., Particle swarm optimization,
in: Proc. ICNN’ 95-International Conference on Neural
Networks, pp. 1942–1948, 1995, doi:
10.1109/ICNN.1995.488968.
27. Tasgetiren, M.F. et al., Particle swarm optimization
algorithm for single machine total weighted tardiness
problem, in: Proc. 2004 Congress on Evolutionary
Computation, 2004, (IEEE, Cat. No. 04TH8753).
28. Narayanan, G. et al., PSO-tuned support vector machine
metamodels for assessment of turbulent flows in pipe bends.
Eng. Comput., 37, 3, 981–1001, 2019, doi: 10.1108/EC-05-2019-
0244.
29. Shankar, R. et al., Hybridized particle swarm-Gravitational
search algorithm for process optimization. Processes, 10, 3,
616, 2022, doi: 10.3390/ pr10030616.
30. Price, K.V., Storn, R.M., Lampinen, J.A., Differential evolution:
A practical approach to global optimization. Springer Science
& Business Media, Berlin/ Heidelberg, Germany, 2006.
31. Kalita, K. et al., A response surface modelling approach for
multi-objective optimization of composite plates. Steel
Compos. Struct. Int. J., 32, 455–466, 2019.
32. Kalita, K. et al., A comparative study on the metaheuristic-
based optimization of skew composite laminates. Eng.
Comput., 38, 4, 3549–3566, 2022, doi: 10.1007/s00366-021-
01401-y.
33. Lorenzo, P.R. et al., Particle swarm optimization for hyper-
parameter selection in deep neural networks, in: Proc.
Genetic and Evolutionary Computation Conference, pp. 481–
488, 2017, doi: 10.1145/3071178.3071208.
34. Zulfiqar, M. et al., Hyperparameter optimization of support
vector machine using adaptive differential evolution for
electricity load forecasting. Energy Rep., 8, 13333–13352, 2022,
doi: 10.1016/j.egyr.2022.09.188.
35. Lohvithee, M. et al., Ant colony-based hyperparameter
optimisation in total variation reconstruction in X-ray
computed tomography. Sensors (Basel), 21, 2, 591, 2021, doi:
10.3390/s21020591.
36. Simon, D., Biogeography-based optimization. IEEE Trans.
Evol. Comput., 12, 6, 702–713, 2008, doi:
10.1109/TEVC.2008.919004.
37. Kalita, K. et al., Optimizing frequencies of skew composite
laminates with metaheuristic algorithms. Eng. Comput., 36, 2,
 741–761, 2020, doi: 10.1007/ s00366-019-00728-x.
38. Kalita, K. et al., Metamodel based multi-objective design
optimization of laminated composite plates. Struct. Eng.
Mech. Int. J., 67, 301–310, 2018.
39. Kalita, K. et al., Memetic cuckoo-search-based optimization
in machining galvanized iron. Mater. (Basel), 13, 14, 3047,
2020, doi: 10.3390/ma13143047.
40. Yang, X.-S. and Deb, S., Cuckoo search via Lévy flights, in:
World Congress on Nature & Biologically Inspired Computing
(NaBIC), vol. 2009, 2009, doi: 10.1109/NABIC.2009.5393690.
41. Kalita, K. et al., A hybrid TOPSIS-PR-GWO approach for
multi-objective process parameter optimization. Process
Integr. Optim. Sustain., 6, 4, 1011–1026, 2022, doi:
10.1007/s41660-022-00256-0.
42. Kalita, K. et al., A comprehensive review on high-fidelity and
metamodel-based optimization of composite laminates. Arch.
Comput. Methods Eng., 29, 5, 3305–3340, 2022, doi:
10.1007/s11831-021-09699-z.
43. Mirjalili, S. et al., Grey wolf optimizer. Adv. Eng. Software, 69,
46–61, 2014, doi: 10.1016/j.advengsoft.2013.12.007.
44. Kalita, K. et al., Grey wolf optimizer-based design of
ventilated brake disc. J. Braz. Soc. Mech. Sci. Eng., 43, 8, 1–15,
2021, doi: 10.1007/s40430-021-03125-y.
45. Kumar, V. et al., A hybrid genetic programming-gray wolf
optimizer approach for process optimization of biodiesel
production. Processes, 9, 3, 442, 2021, doi: 10.3390/pr9030442.
46. Shanmugasundar, G. et al., Optimization of variable stiffness
joint in robot manipulator using a novel NSWOA-Marcos
approach. Processes, 10, 6, 1074, 2022, doi:
10.3390/pr10061074.
47. Ashraf, N.M. et al., Optimizing hyperparameters of deep
reinforcement learning for autonomous driving based on
whale optimization algorithm. PloS One, 16, 6, e0252754,
2021, doi: 10.1371/journal.pone.0252754.
48. Kalita, K. et al., A novel MOALO-MODA ensemble approach
for multi-objective optimization of machining parameters for
metal matrix composites. Multiscale Multidiscip. Model. Exp.
Des., 6, 1, 179–197, 2023, doi: 10.1007/s41939-022-00138-5.
49. Karaboga, D. and Basturk, B., A powerful and efficient
algorithm for numerical function optimization: Artificial bee
colony (ABC) algorithm. J. Glob. Optim., 39, 3, 459–471, 2007,
doi: 10.1007/s10898-007-9149-x.
50. Kaveh, A. and Talatahari, S., A novel heuristic optimization
method: Charged system search. Acta Mech., 213, 3-4, 267–
289, 2010, doi: 10.1007/ s00707-009-0270-4.
51. Kirkpatrick, S. et al., Optimization by simulated annealing.
Science, 220, 4598, 671–680, 1983, doi:
 10.1126/science.220.4598.671.
52. Glover, F., Tabu search-part I. ORSA J. Comput., 1, 3, 190–206,
1989, doi: 10.1287/ijoc.1.3.190.
53. Geem, Z.W. et al., A new heuristic optimization algorithm:
Harmony search. Simulation, 76, 60–68, 2001, doi:
10.1177/003754970107600201.
54. Yang, X.-S., A new metaheuristic bat-inspired algorithm, in:
Nature Inspired Cooperative Strategies for Optimization, pp.
65–74, 2010, doi:10.1007/978-3-642-12538-6_6.
55. Jaddi, N.S. and Abdullah, S., Global search in single-solution-
based meta-heuristics. Data Technol. Appl., 54, 3, 275–296,
2020, doi: 10.1108/ DTA-07-2019-0115.
56. Jeong, S. et al., Development and investigation of efficient
GA/PSO-hybrid algorithm applicable to real-world design
optimization. IEEE Comput. Intell. Mag., 4, 3, 36–44, 2009, doi:
10.1109/MCI.2009.933099.
57. Sanagooy Aghdam, A. et al., A hybrid GA-SA multiobjective
optimization and simulation for RFID network planning
problem. J. Appl. Res. Ind. Eng., 8, 1–25, 2021.
58. Zhang, Y. and Wu, L., A hybrid TS-PSO optimization
algorithm. J. Converg. Inf. Technol., 6, 5, 169–174, 2011, doi:
10.4156/jcit.vol6.issue5.18.
59. Liu, H. et al., Chaotic dynamic characteristics in swarm
intelligence. Appl. Soft Comput., 7, 3, 1019–1026, 2007, doi:
 10.1016/j.asoc.2006.10.006.
60. De, S., Dey, S., Bhattacharyya, S. Recent advances in hybrid
metaheuristics for data clustering, John Wiley & Sons,
Hoboken, NJ, USA, 2020.
61. Shahriari, B. et al., Taking the human out of the loop: A
review of Bayesian optimization. Proc. IEEE, 104, 1, 148–175,
2015, doi: 10.1109/ JPROC.2015.2494218.
62. Bhattacharya, S. et al., A comparative analysis on prediction
performance of regression models during machining of
composite materials. Mater. (Basel), 14, 21, 6689, 2021, doi:
10.3390/ma14216689.
63. Jain, P. et al., Random forest regression-based machine
learning model for accurate estimation of fluid flow in
curved pipes. Processes, 9, 11, 2095, 2021, doi:
10.3390/pr9112095.
64. Brochu, E. et al., A tutorial on Bayesian optimization of
expensive cost functions, with application to active user
modeling and hierarchical reinforcement learning, arXiv
Preprint ArXiv:1012.2599, 2010.
65. Snoek, J. et al., Scalable bayesian optimization using deep
neural networks, in: International Conference on Machine
Learning, 2015.
66. Xie, L. and Yuille, A., Genetic CNN, in: Proc. IEEE
International Conference on Computer Vision, pp. 1388–1397,
 2017, doi: 10.1109/ICCV.2017.154.
67. Krizhevsky, A. et al., Imagenet classification with deep
convolutional neural networks. Commun. ACM, 60, 6, 84–90,
2017, doi: 10.1145/3065386.
68. Elsken, T. et al., Neural architecture search: A survey. J.
Mach. Learn. Res., 20, 1997–2017, 2019, doi: 10.1007/978-3-
030-05318-5_3.
69. Stanley, K.O. and Miikkulainen, R., Evolving neural networks
through augmenting topologies. Evol. Comput., 10, 2, 99–127,
2002, doi: 10.1162/106365602320169811.
70. Hansen, P. et al., Variable neighbourhood search: Methods
and applications. Ann. Oper. Res., 175, 1, 367–407, 2010, doi:
10.1007/s10479-009-0657-6.
71. Liu, H. et al., Darts: Differentiable architecture search, arXiv
Preprint ArXiv:1806.09055, 2018.
72. Ying, C. et al., NAS-Bench-101: Towards reproducible neural
architecture search, in: International Conference on Machine
Learning, 2019.
73. Narayanan, R.C. et al., A novel many-objective sine-cosine
algorithm (MaOSCA) for engineering applications.
Mathematics, 11, 10, 2301, 2023, doi: 10.3390/math11102301.
74. Silver, D. et al., Lanctot and others, Mastering the game of Go
with deep neural networks and tree search. Nature, 529, 484–
489, 2016, doi: 10.1038/ nature16961.
75. Blum, C. and Roli, A., Metaheuristics in combinatorial
optimization: Overview and conceptual comparison. ACM
Comput. Surv., 35, 3, 268–308, 2003, doi:
10.1145/937503.937505.
76. Ganesh, N. et al., A novel decomposition-based multi-
objective symbiotic organism search optimization algorithm.
Mathematics, 11, 8, 1898, 2023, doi: 10.3390/math11081898.
77. Joshi, M. et al., A conceptual comparison of dragonfly
algorithm variants for CEC-2021 global optimization
problems. Arab. J. Sci. Eng., 48, 2, 1563–1593, 2023, doi:
10.1007/s13369-022-06880-9.
78. Shaik, K. et al., Big data analytics framework using squirrel
search optimized gradient boosted decision tree for heart
disease diagnosis. Appl. Sci., 13, 9, 5236, 2023, doi:
10.3390/app13095236.
79. Real, E. et al., Regularized evolution for image classifier
architecture search. AAAI, 33, 1, 4780–4789, 2019, doi:
10.1609/aaai.v33i01.33014780.
80. Gaspar, A. et al., Hyperparameter optimization in a
convolutional neural network using metaheuristic
algorithms, in: Metaheuristics in Machine Learning: Theory
and Applications, pp. 37–59, Springer, Boston, USA, 2021, doi:
10.1007/978-3-030-70542-8_2.
81. Ezugwu, A.E. et al., A conceptual comparison of several
metaheuristic algorithms on continuous optimisation
problems. Neural Comput. Appl., 32, 10, 6207–6251, 2020, doi:
10.1007/s00521-019-04132-w.
82. Sonmez, M., Performance comparison of metaheuristic
algorithms for the optimal design of space trusses. Arab. J.
Sci. Eng., 43, 10, 5265–5281, 2018, doi: 10.1007/s13369-018-
3080-y.
83. Tayebi, M. and El Kafhali, S., Performance analysis of
metaheuristics based hyperparameters optimization for
fraud transactions detection. Evol. Intell., 1–22, 2022, doi:
10.1007/s12065-022-00764-5.
84. Priyadarshini, J. et al., Analyzing physics-inspired
metaheuristic algorithms in feature selection with K-nearest-
neighbor. Appl. Sci., 13, 2, 906, 2023, doi:
10.3390/app13020906.
85. Ganesh, N. et al., Efficient feature selection using weighted
superposition attraction optimization algorithm. Appl. Sci.,
13, 5, 3223, 2023, doi: 10.3390/ app13053223.
86. Javanshir, A. et al., Training spiking neural networks with
metaheuristic algorithms. Appl. Sci., 13, 8, 4809, 2023, doi:
10.3390/app13084809.
87. Xian, H. et al., A modified sine cosine algorithm with teacher
supervision learning for global optimization. IEEE Access, 9,
 17744–17766, 2021, doi: 10.1109/ACCESS.2021.3054053.
88. Jahwar, A.F. and Abdulazeez, A.M., Meta-heuristic algorithms
for K-means clustering: A review. PalArchs J. Archaeol. Egypt
Egyptol., 17, 12002–12020, 2020.
89. Kaur, A. and Kumar, Y., A new metaheuristic algorithm based
on water wave optimization for data clustering. Evol. Intell.,
15, 1, 759–783, 2022, doi: 10.1007/s12065-020-00562-x.
90. Seyyedabbasi, A., A reinforcement learning-based
metaheuristic algorithm for solving global optimization
problems. Adv. Eng. Software, 178, 103411, 2023, doi:
10.1016/j.advengsoft.2023.103411.
91. Seyyedabbasi, A. et al., Hybrid algorithms based on
combining reinforcement learning and metaheuristic
methods to solve global optimization problems. Knowl. Based
Syst., 223, 107044, 2021, doi: 10.1016/j.knosys.2021.107044.
92. Tian, Z. and Fong, S., Survey of meta-heuristic algorithms for
deep learning training, in: Optim. Algor.-Methods Appl., 2016,
doi:10.5772/63785.
93. Rere, L.M.R. et al., Simulated annealing algorithm for deep
learning. Proc. Comput. Sci., 72, 137–144, 2015, doi:
10.1016/j.procs.2015.12.114.
94. Kinga, D. and Adam, J.B., A method for stochastic
optimization, in: Proceedings of the International Conference
 on Learning Representations–ICLR, vol. 5, p. 6, San Diego, CA,
USA, 7–9 May 2015.
95. Ting, T.O. et al., Hybrid metaheuristic algorithms: Past,
present, and future, in: Recent Advances in Swarm
Intelligence and Evolutionary Computation, pp. 71–83, 2015,
doi:10.1007/978-3-319-13826-8_4.

Note

1. *Corresponding author: [email protected]

3
A Survey of Computer-Aided
Diagnosis Systems for Breast
Cancer Detection
Charu Anant Rajput*, Leninisha Shanmugam and Parkavi K.

School of Computer Science and Engineering, Vellore Institute of

Technology, Chennai, India

Abstract

Computer-aided diagnosis (CAD) has been the most critical and

vital approach concerning the medical domain in recent times.
Also, with the trend of telemedicine in place, the reliability and
efficiency of CAD systems have become the need of the hour.
Dedicated research concerning CAD systems for the detection of
breast cancer is being promoted globally. However, given the
vast range of imaging modalities and the availability of diverse
datasets concerning the aforementioned domain, choosing the
most efficient and optimum implementation methodology
becomes a challenge. The proposed survey paper, therefore,
aims to provide a comprehensive overview of all of the existing
imaging modalities and their corresponding datasets. Apart
from this, some notable works concerning each modality
representing both the machine learning and deep learning
domains would also be presented in order to give a head start
for the upcoming novel research. The proposed work would
shed light on all the possible categories of classifications that
can be performed concerning the domain of breast cancer. We
finally conclude by pointing out new possible research gaps,
thereby opening up avenues for future upcoming research.

Keywords: Benign, malignant, in situ, invasive ductal

carcinoma, CAD

3.1 Introduction

Various indications and abnormalities come up during image

screening and may prove to be early indicators of breast cancer.
On further deep-dwelling, they are further classified into
benign or malignant wherein malignant represents the
potential danger of development into breast cancer shortly.

According to a recently conducted statistics study, one woman is

diagnosed with breast cancer every 4 minutes according to the
Indian context. Breast cancer is annually registering a figure of
approximately 1,78,000 as of 2020 and has therefore even
overtaken cervical cancer to become the leading cause of
cancer among Indian women. The point to be noted is that, in
cases of early diagnosis, the survival rates can increase up to
98%. However, the irony lies in the fact that approximately
50%–80% of breast cancer cases in India are detected in the
advanced stages.

This makes it necessary to put some light on the potential

indicators of breast cancer. The occurrence of calcifications and
masses are considered to be red flags and thereby involves
clinical interpretation for facilitating proper diagnosis and
treatment. Chaudhury et al. [1] in his work has termed
calcifications as calcium deposits found in the breast. Probable
reasons are listed as aging, historical infections, and wounds
and are not necessarily linked to calcium intake. Calcifications
have two main categories, namely, macrocalcifications and
microcalcifications, wherein the former is not risky while
microcalcifications raise an early suspicion of breast cancer
and thereby need multiple follow-up tests. These days, there are
automated systems in place for detecting instances of
calcifications in the input ultrasound scans. Chaudhury et al. [1]
have discussed a novel segmentation approach that is
postulated in terms of four phases. The first phase involves
preprocessing wherein the input image is converted into a
grayscale image. The image segmentation is done by the
application of K-means, and the subsequent feature extraction
is performed by making use of the gray-level co-occurrence
matrix (GLCM). Finally, for performing the classification
concerning normal cells and calcified cells, the model makes
use of Hidden Markov Model (HMM) classification algorithm
wherein resemblance has precedence over contextual
knowledge for classification purposes. The aforementioned
proposed work made use of ultrasound scans. Loizidou et al. [2]
have further provided a scope of extension to the
aforementioned work wherein the proposed work not only
detects microcalcifications but also further classifies whether
they are benign or suspicious by making use of temporally
sequential mammograms. The proposed work emphasizes
temporal subtraction as the main accuracy-building factor. The
methodology initiates with the application of preprocessing
techniques like pixel normalization and border removal
subsequently followed by the image registration of the
temporally prior scan. Temporal subtraction is followed
thereafter and then comes feature extraction wherein different
categories like shape features, intensity-based features, first
order statistic features (FOS), and GLCM features are extracted.
Finally, for classification, different classifiers like linear
discriminant analysis (LDA), kNN, naive Bayes, support vector
machine (SVM), decision trees, and ensemble of decision trees
are used.

Another significant indicator is the presence of masses or lumps

in the breast scans. Here, too, the presence of CAD systems has
provided support and reassurance to the manual practice being
carried out presently. Jeong et al. [3] made use of three-
dimensional ultrasound images to detect the masses. The
proposed methodology made use of the 3D Hough transform for
identifying the spherical hyperplanes in the input images. The
recognized spheres are then converted into a circumscribing
parallelepiped cube as mass lesion candidates, and then the
final classification is done based on the geometrical
overlapping. Hamed et al. [4] made use of the You Only Look
Once (YOLO)-based CAD system that works on full-field and
cropped mammograms. This is followed by the application of
feature extractors like ResNet, VGG, and Inception. The system
successfully detects mass location with an accuracy of 98% and
also performs classification between benign and malignant
with an accuracy of 95%.

The aforementioned discussion puts light on some early

symptoms like calcifications and masses; however, another
judgment category that plays a significant role in determining
the role of treatment is deciding whether the cancer is in the
invasive ductal carcinoma (IDC) phase or undergoing the in situ
stage. In medical terms, in situ means that the cancerous cells
are limited inside the milk ducts; however, IDC means that the
cancerous cells have started spreading to the surrounding
tissues.

Automated works concerning the aforementioned criteria

majorly make use of the histopathological whole slide imaging
(WSI) dataset. For instance, Celik et al. [5] have implemented
the detection of IDC by making use of standard deep learning
(DL) architectures like ResNet-50 and DenseNet-161. The
training was performed over the last layers of the model, and
the performance produced was almost the same wherein
DenseNet produced an F-score of 92.38% and a balanced
accuracy of 91.75%, whereas the ResNet model obtained an F-
score of 94.11% and an accuracy of 90.96%; thus, it can be
concluded that both models have performed considerably with
the same efficiency. Bejnordi et al. [6] on the other hand have
proposed a three-step approach for detecting the presence of
Ductal Carcinoma In Situ (DCIS) in H&E-stained
histopathological WSIs. The proposed algorithm performed
multi-scale superpixel classification in an attempt to identify
the epithelial regions. This is followed by graph-based
clustering of the epithelium-labeled superpixels and the
subsequent delineation of the ROIs. Finally, the segmented
regions are classified as benign or DCIS. Lastly comes the CAD
systems predicting the classification between benign and
malignant, and according to the survey carried out, the
majority of the research works are focused upon this
classification modality.

Table 3.1 represents the definitions of all of the medical jargon

described in the aforementioned discussion for the convenience
of the readers. Hereafter, the structure of the paper is framed in
the following order. Section 3.2 describes the methodology and
ideology adopted for successfully surveying all of the research
works concerning the medical domain of breast cancer
detection. Section 3.3 gives an account of different imaging
modalities considering both the biological and computational
perspectives. Additionally, this section also describes all of the
publicly available datasets corresponding to each of the
imaging modalities mentioned. Additionally, a mention of all of
the private datasets used in the papers included in the survey
has also been represented. Section 3.4 gives an account of the
research works systematically bifurcated based on the type of
technique adopted, i.e., machine learning (ML) or deep
learning. They have been further arranged based on the
imaging modality employed for the work. Finally, we conclude
in Section 3.5 by highlighting the research gaps identified across
the survey, thereby opening up new avenues for future
research.
Table 3.1 Summary of the medical jargon used.

Medical
Definition Source
term

Calcification Minute solid bits of calcium Chaudhury

found in the breast region et al. [1]

Masses Excessive growth of tissues Wook

leads to the formation of Jeong et al.
round structures that are [3]
further classified based on
the region of the formation.

Ductal Represents the condition Bejnordi et

carcinoma wherein the cancerous al. [6]
in situ outgrowth is limited to the
milk duct region
 Medical
Definition Source
term

Invasive Represents the medical Celik et al.

ductal condition wherein the [5]
carcinoma cancerous region starts
spreading in the
surrounding regions and is
not limited to the duct
region.

Benign Cancer that is non-spreading Deblee et

in nature al. [7]

Malignant Cancer tends to spread and Deblee et

affect other organs al. [7]

3.2 Procedure for Research Survey

We surveyed research papers of recent years covering a span

from 2019 to 2022. A few papers from 2016 were also included
in the survey, keeping in mind the novelty and complexity of
the proposed approach. These papers were evaluated keeping
in mind the following parameters: 1) Imaging modality used, 2)
Complexity and novelty of the proposed approach, 3) Nature of
the dataset employed, 4) Nature of technique used, i.e., whether
ML or DL. The general search criteria for this survey made use
of keywords like “automated breast cancer detection,” “CAD
systems using ML and (name of the imaging modality),” “CNN-
based CAD systems for breast cancer detection,” and “ROI
segmentation-based papers for BC detection.” The searches
were carried out in acclaimed research databases and websites
like 1) PubMed, 2) Science Direct, 3) IEEE Xplore Digital Library,
4) Google Scholar, and 5) MDPI (Multidisciplinary Digital
Publishing Institute). A few survey papers were also reviewed to
analyze and understand the procedure for drafting a survey
paper concerning the subject of breast cancer. The references
and details concerning the datasets were obtained from Kaggle.

3.3 Imaging Modalities and Their Datasets

3.3.1 Histopathological WSI

This imaging modality involves taking samples of soft tissues

from the suspicious regions with the help of surgical experts
using fine-needle aspiration (FNA). According to Mahmood et al.
[8], these extracted tissues are then pigmented and examined
under a microscope for further diagnosis, classification, and
assertion. These are then converted into digital records, as
microscopic slides are perishable, which is indeed useful for the
application of automated algorithms. The digitized copies not
only help in portability and record-storing but also give an edge
over manual examination using a microscope in terms of
enhancing magnification flexibility and also help view different
regions without much eye-straining. The disadvantage that is
faced is during the processing phase. The WSI file is typically
memory-intensive and cannot be fed into the
processing/classification model. Typically, this requires
cropping out the ROI and augmenting the image that can lead to
loss of diagnostically critical information. But on medical
grounds, this is the most preferred and reliable imaging
modality used for confirming the diagnosis of breast cancer.

Some of the popular public datasets concerning the

aforementioned imaging modality are as follows:

1. IDC Breast Histopathological Image Dataset: This dataset

contains approximately 277,524 RGB image patches of
dimensions 50 × 50 pixels, and the primary aim is to identify
the presence of invasive ductal carcinoma.
2. BACH Grand Challenge (ICIAR 2018 conference): The acronym
BACH stands for breast cancer histology images that
essentially consist of 400+ labeled microscopy images, and
 the task is to perform classification corresponding to four
classes: normal, benign, in situ carcinoma, and invasive
carcinoma.
3. BreakHis (Breast Cancer Histopathological Database): This
dataset contains approximately 9,109 microscopic images
wherein there were 2,480 benign and 5,429 malignant
samples. These samples are available in the RGB mode with
the dimensions 700 × 460 pixels.
4. CAMELYON17: This dataset is a part of the challenge for
detecting and differentiating between the metastatic and
normal regions, and the annotations are available on the
lesion level and patient level. The dataset contains 1,000 slide
images covering different sections of the lymph node.
5. PCam dataset (Patch Camelyon): This dataset contains
approximately 327,680 images wherein the ROI patch region
of dimension 96 × 96 pixels has been extracted and labeled
based on the presence of metastatic tissue.

3.3.2 Digital Mammography

This imaging modality involves capturing the insides of the

breast by exposing the area under examination to low doses of
X-rays. It is similar to CT wherein the images are taken in slices
and then assembled for getting the whole image. CAD systems
examine abnormalities by assessing the overall density and
morphological features of the abnormal areas. As opposed to
histopathological WSI, here, there is no need for manual effort
for the conversion of digitization; however, as far as CAD
systems are concerned, the mammograms require a lot of
image preprocessing and annotations before feeding into the
classification model.

Some of the popular public datasets concerning the

aforementioned imaging modality are as follows:

1. DDSM-400 and CBIS-DDSM: These datasets are often used

together to increase the volume of the input data. The dataset
contains ROI-extracted images of dimensions 299 × 299 pixels
with a sample size of 55,890 out of which 14% of the samples
are positive for the presence of malignancy while the
remaining 86% are negative. Preprocessing techniques like
cropping, random flips, and rotations accompanied by
resizing have been performed.
2. MIAS Mammography (Mammographic Image Analysis
Society): The dataset contains approximately 327 images with
detailed annotations describing: 1) the character of the
background tissue, 2) the class of abnormality present, 3) the
severity of abnormality, and 4) coordinate location of the
abnormality. The dimensions of the samples are 1,024 pixels
× 1,024 pixels.
3.3.3 Ultrasound

This imaging modality is normally used for capturing and

examining early signs of cancer like masses and calcifications.
It is complementary to both mammograms and MRI and
therefore can be used for deep diving into the abnormalities
detected by the aforementioned scans. For instance,
ultrasonography is recommended for whether a detected breast
mass is a fluid-filled cyst or a solid tumor. However, the
classification concerning benign and malignant is not very clear
using this imaging modality. As far as the CAD systems are
concerned, effective and precise feature selection tends to be an
issue because of inaccurate contour identification and
segmentation of the ROI. Additionally, the majority of the works
concerning the use of ultrasonography have represented the
need for manual annotations for training purposes, and
therefore, this imaging modality is not fully supportive of
automation.

Some of the popular public datasets concerning the

aforementioned imaging modality are as follows:

1. OASBUD (Open Access Series of Breast Ultrasonic Data): The

dataset contains US scans of approximately 100 patients
wherein 52 of them are malignant and 48 are benign.
 Additionally, two perpendicular scans in the transverse and
longitudinal directions were performed for each lesion.
2. UDIAT: The dataset contains 163 US images wherein 110 of
them belong to the benign category and 53 of them are
malignant. The dataset was acquired from UDIAT Diagnostic
Center in Spain by making use of a Siemens ACUSON scanner.
3. BUSI (Breast Ultrasound Images Dataset): Contains
approximately 780 ultrasound scan images of dimensions
500 × 500 pixels and are collected from approximately 600
patients. The images are categorized into three classes,
namely, normal, benign, and malignant. This dataset was
developed by Dhabyani et al. [9].

3.3.4 Magnetic Resonance Imaging

This imaging modality is primarily performed on high-risk

women who cannot undergo procedures like mammography
and ultrasonography because of the nature of their tumors. This
technique is based on the absorption of radio frequencies by
the cellular nuclei in the presence of magnetic fields. However,
concerning their implementation using CAD systems, the main
issue faced is the segmentation and separation of the breast
region from the non-required body parts. Additionally, because
of the use of different frequencies while imaging, complex
feature fusion techniques have to be used for extracting useful
diagnostic conclusions from the scans using classification
models.

The majority of the works concerning this domain consist of a

privately collected dataset from hospitals mainly because of the
need for manual annotation by a radiologist. However, the point
to be noted is that majority of these works make use of
dynamically contrast-enhanced MRI (DCE-MRI). One such
benchmark report describing the practices and preprocessing
methods utilized while collecting this dataset has been put up
by Witowski et al. [10].

3.3.5 Infrared Breast Thermal Images

This imaging modality is noninvasive and nonionizing and

makes use of temperature distribution across the breast region
for detecting abnormality. According to Venkatachalam et al.
[11], thermography is based upon the fact that precancerous
regions exhibit higher temperatures as compared to
surrounding regions because of a high metabolic rate due to
angiogenesis. This activity is captured using an infrared
thermal camera. However, the main issues posed by this
imaging modality as rightly identified and as worked upon by
Venkatachalam in [11] are intensity inhomogeneity, ROI
overlapping, poor contrast, and low signal-to-noise ratio (SNR).
A reference to a refined privately collected dataset as utilized
by Venkatachalam et al. [11] is as follows:

The Infrared Breast Thermal Images (IBTI) dataset used for this
research work was collected from “Harshmitra Super Speciality
Cancer Centre and Research Institute,” Trichy, Tamil Nadu,
India, and consisted of images captured using the infrared
camera DITI CX320 having a resolution of 320 × 240. It
contained data collected from approximately 50 patients.
Table 3.2 Advantages and disadvantages.

Imaging
Advantages Disadvantages
modality

Histopathological Useful for Large resolutions

WSI minute make it difficult
observation and for processing
classification purposes.
concerning in
situ and
invasive ductal
carcinoma

Digital Noninvasive The multitude of

mammography and efficient for image
the detection of preprocessing
tumors and required
masses
Imaging
Advantages Disadvantages
modality

Ultrasound Noninvasive Radiologist

and painless expertise and
method of image mandatory
screening manual
annotation
required

Magnetic Is the most Use of complex

resonance detailed imaging feature fusion
imaging (MRI) modality and is techniques for the
preferred for purpose of
women at the extracting
high-risk stage. diagnostically
critical
information
 Imaging
Advantages Disadvantages
modality

Infrared breast noninvasive, Intensity

thermal images nonionizing inhomogeneity,
(IBTI) imaging ROI overlapping,
modality and is poor contrast, low
therefore signal- to-noise
potentially ratio (SNR).
harmless.

Another public dataset being used is DMR-IR that has been

published by Fluminense Federal University, Brazil. This dataset
too contained an account collected from 50 patients where the
images were captured by making use of a FLIR infrared camera
having a resolution of 640 × 480 pixels.

A summary of different imaging modalities along with an

account of their advantages and disadvantages is represented
in Table 3.2. Figure 3.1 represents a pictorial representation of
all of the aforementioned imaging modalities with the
specification of benign and malignant.
Figure 3.1 Pictorial representation of the imaging modalities.

3.4 Research Survey

This section gives a brief description of all of the notable
research works concerning each imaging modality as
mentioned in the aforementioned section. For the ease of
identification of interest, these works have been systematically
bifurcated depending on whether the proposed methodology
was based on machine learning or deep learning.

3.4.1 Histopathological WSI

Given the complexity involved in the interpretation of the input

image, most of the research works are based on deep-learning
CNN models. However, an account of one of the noteworthy
works involving the use of machine learning techniques is as
follows.

3.4.1.1 Machine Learning-Based Histopathological WSI

Roy et al. [12] have made use of ML classifiers and ensemble

learning for drawing out the classification between IDC-positive
and IDC-negative samples. The classification has been
performed by the extraction of textural features like SIFT, SURF,
and ORB. Haralick features are the statistical features that are
taken into account for performing the classification. Pearson’s
coefficient has been used for weeding out the redundancy of
features. The features thus extracted are subsequently fed into
six classifiers, namely, random forest, extra trees, XGBoost,
AdaBoost, CatBoost, and multilayer perceptron to present a
comparative study. The proposed work concluded with the fact
that CatBoost was able to deliver the highest accurate
performance registering a figure of 92.55%. The proposed work
made use of BreakHis dataset.

3.4.1.2 Deep Learning-Based Histopathological WSI

Sanyal et al. [13] have based the proposed work on the BACH
(Breast Cancer Histology Images) dataset that is posed as a
challenge organized as a part of the ICIAR 2018 conference. The
proposed methodology involves fine-tuning of principal and
popular CNN architectures, namely, VGG 19, InceptionV3,
Inception-ResNetv2, and ResNet-101 to extract classification
critical and discriminative features. As a part of ensemble
learning, XGBoost has been incorporated into the softmax layer
for improving the classification performance. The work
proposes a model performing fourclass classification at the
image and patch level. The classification categories are in situ,
invasive, normal, and benign. The inference of image level is
drawn from the fact that the individual patches together make
an image and the classification at the image level is essentially
made out by using the patches as participants during the voting
procedure.
Das et al. [14] have proposed a methodology for automating the
preprocessing stage wherein manual annotations are required.
It makes use of a deep multiple instance learning-based CNN
framework. Its working can be closely mapped to the popular
bags of words approach that is adopted and widely used for
sentiment analysis in the domain of natural language
processing (NLP). Here, the input test images are represented as
a bag of patches that are obtained during the training phase.
The input test image is classified as benign if and only if all of
the patches are benign and classified as malignant even if
anyone of the patch comes out to be malignant. Additionally,
the versatility of the proposed model has been represented by
testing it on diverse and popular datasets like BreakHis, IUHPL,
and UCSB, registering accuracy figures of 93.06%, 96.63%, and
95.83%, respectively.

Han et al. [15] have come up with a noteworthy methodology

targeting the multi-classification problem involving the
identification of subordinate classes of malignant tissues that
are named ductal carcinoma, lobular carcinoma, adenosis,
fibroadenoma, phyllodes tumor, tubular adenoma, mucinous
carcinoma, and papillary carcinoma. The novelty of the
proposed work is the introduction of a new deep learning
model called the structured deep learning model that achieves
an accuracy of 93.2%. The novelty of class structure-based deep
convolutional neural network (CSDCNN) is that it systematically
bypasses the stage involving feature engineering for the
training process, and instead of that, it makes use of space
distance constraints for classification.

3.4.2 Digital Mammogram

3.4.2.1 Machine Learning-Based Digital Mammogram

Mohiyuddin et al. [16] have made use of the popular object

detection algorithm YOLO for scanning digital mammograms
and classifying them as malignant or benign. The proposed
model aims to decrease the false positive rate (FPR) and false
negative rate (FNR) while increasing Matthew’s correlation
coefficient (MCC). Initially, the preprocessing stage involves the
removal of rough white borders and pectoral muscles. This is
followed by the usage of contrast limited adaptive histogram
equalization (CLAHE) for image enhancement. CLAHE is
primarily used for viewing the hidden features wherein the
image is fragmented into tiles and the adjacent tiles are then
blended using bilinear interpolation for removing the false
boundaries. The prepared data are then fed as input to the
proposed model of YOLOv5 that utilizes all four versions of
YOLOv5. This is followed by presenting a comparative study
with YOLOv3 and faster RCNN. The accuracy produced as a
result is a whopping 96.5%. The dataset used for the
aforementioned work was CBIS-DDSM.

Darweesh et al. [17] have proposed a two-level hierarchical

classification wherein the first classification is distinguishing
between normal and abnormal tissues, while the second level
dives deep into the abnormality for further classification as
benign or malignant. Feature extraction in the first phase uses
GLCM, while the second level makes use of local binary
patterns (LBP). The point to be noted here is that random forest
is employed as the classifier in both stages of classification. The
first stage classification gives an accuracy of 97%, whereas the
second stage classification gives an accuracy of 75%. The model
together produces an overall accuracy of 85% and an MCC
coefficient value of 0.76. MIAS was employed as the dataset for
the aforementioned work.

3.4.2.2 Deep Learning-Based Digital Mammogram

Tsochatzidis et al. [18] have proposed a methodology to

integrate mammographic segmentation information along with
the input image to improve the diagnosis. This involves
modification in each of the CNN layers such that the
information of the input image along with the segmentation
map is considered.
Either the segmentation maps can be established from the
ground truth or an automatic segmentation system is put in
place using U-Net, which by its features of upsampling,
downsampling, and skip layers come up with the segmented
lesion. Furthermore, a new loss function is introduced, which
adds an extra term to the standard cross-entropy, aiming to
steer the attention of the network toward the mass region,
penalizing strong feature activations based on their location.
The main CNN model thereafter employed is ResNet. DDSM-400
and CBIS-DDSM were used for implementing the proposed
model.

Mohapatra et al. [19] have come up with a survey paper for

analyzing and comparing the performance of various popular
CNN architectures like AlexNet, VGG16, and ResNet. All these
models were trained on the CBIS-DDSM dataset, and the ratio of
90:10 is used as the validation strategy. According to the
observations, AlexNet emerged as the best-performing model,
putting up an accuracy of 65% when fine-tuned and pretrained
with weights. VGG16 and ResNet-50 registered accuracy figures
of 65% and 61%, respectively, under the same circumstances.

3.4.3 Ultrasound

3.4.3.1 Machine Learning-Based Ultrasound

Cheng et al. [20] have presented an effective survey about the
techniques being used by CAD systems based on ultrasound
images. The survey explores methods being applied for
automated breast cancer detection under four main phases: 1)
image preprocessing, 2) image segmentation, 3) feature
extraction and selection, and 4) classification. The image
preprocessing stage aims at targeting and solving image
modality issues like low contrast and speckle interference. The
techniques covered under the image segmentation phase are
aimed at identifying and isolating the ROI. Under the section
concerning feature extraction and selection, the author
presents a detailed description of the most effective features
that are detrimental to the accurate performance of classifiers.
It is under the classifiers section that the application of popular
machine learning techniques like linear discriminant analysis,
logistic regression, Bayesian neural networks, and decision
trees is discussed at length. The research works providing an
application of the aforementioned techniques have also been
pondered upon as a part of this survey.

Sadad et al. [21] have proposed a methodology that employs the

use of ML-based classifiers like a decision tree and KNN. It
makes use of Hilbert transform for the conversion of original
ultrasound images into B-mode. These B-mode images are in
turn used for ease of tumor identification and detection given
the low contrast of the ultrasound images. Segmentation of the
lesions is done using marker-controlled watershed
transformation, and a hybrid set of features including shape
and texture parameters is used for feature extraction. In
addition to the public dataset OASBUD, the work also makes use
of a private dataset collected from Baheya Hospital, Egypt.

3.4.3.2 Deep Learning-Based Ultrasound

Byra [22] proposed novel changes in transfer learning

techniques that are applied to CAD systems in the medical
domain. Traditionally, transfer learning involves fine-tuning
wherein the weights of pretrained networks are modified to
address the given medical problem. However, fine-tuning
becomes difficult in case the number of trainable parameters
becomes large and therefore the work proposes a novel
transfer learning technique based on deep representation
scaling (DRS) that is inserted in between the CNN blocks for
better flow of information. With the application of DRS, we
need to just change the DRS parameters and significantly
reduce the number of trainable parameters. To showcase the
accuracy of the proposed model, the authors have implemented
it to be used on a system of breast mass classification
implemented upon a dataset of ultrasound images. Indeed, the
method seems to be pretty effective to be implemented on other
medical domains as well seeing it registering the accuracy
figures of approximately 0.915.

Li et al. [23] have proposed an entirely novel CNN architecture

for the detection of malignant lesions in ultrasound images. It is
called BUSnet. This architecture has been specifically developed
for effectively solving and terminating accuracy-depleting
factors like speckle noise and acoustic shadow. ROI
segmentation makes use of unsupervised region proposal and
bounding box regression algorithms. The unsupervised region
proposal method makes use of the Canny Edge Detector, while
ResNet-50 serves as the backbone for the bounding box
regression technique. Aggregation is used as the postprocessing
technique that effectively integrates all of the regressive
bounding boxes into one and thereby eliminates redundant
information.

3.4.4 MRI-Based

3.4.4.1 Machine Learning-Based MRI Analysis

Dewangan et al. [24] have proposed an ML-based approach by

making use of MRI images. The novelty of the work lies in its
preprocessing stage wherein an effective combination of
Wiener and median filters is used. Together, this combination is
named Wienmed filter. The combined application of these two
filters ensures the replacement of the noisy neighborhood
pixels, thereby improving the quality of the MRI breast image
portion. Additionally, it also leads to the enhancement of some
of the important image attributes that play a vital role during
the classification process. After the preprocessing stage, the
images are fed to the backpropagation recurrent Wienmed
model that is used for performing the classification and
identification of breast tumors in the image. GLCM is used for
feature extraction. Additionally, the proposed work makes use
of a hybrid algorithm called hybrid krill herd African buffalo
optimization for improving classification accuracy. The
accuracy thus produced is 99.6% with a 0.12% lower error rate
that is quite appreciable given the complexity of the dataset
under examination. This work can further be explored for
understanding the formulation and implementation of hybrid
algorithms.

3.4.4.2 Deep Learning-Based MRI Analysis

Hu et al. in [25] have made a private dataset by acquiring

examination samples over 8 years spanning from 2005 to 2013
at a medical institute. The MRI images have been collected in
two sequences, DCE-MRI and T2-weighted. The extraction of
features from these two sequences of MRI scan can be
incorporated in the following three ways:

Image fusion – fusing the DCE and T2w images for producing
the RGB composite
Feature fusion – merging the CNN features extracted from
both the DCE and T2w images and then serving it to the SVM
for classification
Classifier fusion – Aggregation of the malignancy probability
output from both the DCE and T2w via soft voting

The proposed work makes use of the feature fusion technique.

ROI location has been done using fuzzy C-means with the
provision of seed indication done manually. The identification
of 4D characteristics concerning temporal and volumetric has
been done uniquely for each frequency sequence. For DCE,
maximum intensity projection (MIP) was performed over the
images obtained from the second postcontrast subtraction.
However, for the T2w sequence images, the largest lesion area
was selected and cropped out. These were subsequently fed into
a deep-learning CNN model for getting the results.

Ha and Vahedi [26] have proposed an automated tumor

detection system based on hybrid CNN architecture and
implementation of an improved deer hunting optimization
algorithm (DHOA). The first phase involves preprocessing that
involves normalization of the intensities followed by noise
removal. The filter used here is based on the fuzzy-neural
network as proposed by Ma et al. [27], and the primary aim
here is the removal of acoustic noise. The novelty of the
proposed work lies in the implementation of an improved
version of the deer hunting optimization algorithm that is used
in CNN architecture replacing the traditional back-propagation
approach to reduce network error. The proposed work also
makes use of Haralick texture and LBP for feature extraction.

Jiao et al. [28] have proposed an automated breast mass

detection approach consisting of two main parts: breast region
segmentation and breast mass detection. Here, for the
segmentation approach, a DCNN model based on the U-Net++
framework has been used. The segmented parts are then fed
into Faster RCNN that itself is composed of two modules. The
first module is a DCNN used for region proposal, and the second
module is a detector that generates output by using the
proposed regions. ResNet 101 is used for feature extraction, and
subsequently, the feature map so generated is considered by the
region proposal network (RPN) to generate multiple
coordinates, each of them being graded according to an
objectness score. In this way, tumors having a diameter of less
than 48 pixels are targeted and identified by the
aforementioned methodology.

3.4.5 Thermography-Based

3.4.5.1 Machine Learning-Based Thermography Analysis

Venkatachalam et al. [11] have proposed effective ROI

segmentation from the relatively complex thermogram that
suffers from issues like intensity inhomogeneity, ROI
overlapping, poor contrast, and low SNR. The image
preprocessing involves the conversion of pseudocolored images
to gray scale followed by the segmentation of non-breast
regions. The left and right breasts are thereafter separated and
undergo binary histogram difference thresholding (BHDT) for
segmenting the contours exhibiting high-intensity
inhomogeneity. The Multiscale Local Global Image Fitted
(MLGFI) model is used for getting more clear segmentation
results from the outputs produced after the application of
BHDT. This procedure makes use of multiscale median filters,
and according to the mathematical derivations presented, we
get the image devoid of intensity inhomogeneity as a result. To
prove the efficiency of the proposed segmentation approach, its
performance is compared with state-of-art ML methods like
Fuzzy C Means (FCM), Chan-Vese (CV-ACM), and K-means.
Benign and malignant tissues exhibit significant differences
corresponding to the area-based features (AFs) and average
intensity-based features (AIFs), and therefore, the work
proposes the computation of the absolute difference of the
aforementioned features corresponding to the inflamed region
and residual region. GLCM is used for extracting the textural
features, and for classification, all of these features are fed into
a support vector machine (SVM). The proposed work exhibits
an accuracy of 91.5%.

Figure 3.2 CNN architecture as illustrated by Mohamed et al. in [29].

3.4.5.2 Deep Learning-Based Thermography Analysis

Mohamed et al. [29] have proposed an approach that can be

systematically segmented into three phases, namely, image
resizing, breast area segmentation, and classification. The
thermal images are available of the size 640 × 480 pixels that
can prove to be computationally intensive for processing, and
therefore, the first phase involves resizing the input image to
the dimension 280 × 280 pixels. This is followed by segmenting
the breast area from other body parts that act as noise, and this
is achieved by making use of the U-Net network structure. U-Net
over a series of works has been proven effective for biomedical
segmentation. The two-class classification concerning abnormal
and normal areas is done using a CNN model that is
indigenously developed and trained from scratch. The novelty,
therefore, lies in the architecture of the proposed CNN model.
The architecture diagram is shown in Figure 3.2.

Here, the use of transfer learning is not done as opposed to

existing works and therefore indicates the novelty in the work.
The proposed system is evaluated considering the measures of
accuracy, sensitivity, and specificity and registers the figures
99.33%, 100%, and 98.67%, respectively. The figures put up are
indeed impressive and commendable.

3.5 Conclusion

The paper discussed the characteristics of some of the most

prominent imaging modalities that are used for breast cancer
detection. It also presented a review of the characteristics of
popular public datasets corresponding to each imaging
modality so that the readers can infer and make informed
choices concerning the dataset to be used for research.
Additionally, some very notable works systematically
segmented into sections concerning machine learning and deep
learning have been presented as a source of reference so that
the readers can imbibe and implement hybrid formulations
from the aforementioned proposed methodologies.

There are still some relatively unexplored techniques and

subjects that can open up new avenues for future research.
Concerning histopathological WSI, the research can be
extended for the formulation of efficient patch extraction
techniques, thereby reducing the rate of misclassifications.
Mammographic scans are majorly used for predicting whether
a given mass is malignant or benign. However, the scope of
multiclass classification is evident and must be explored.
Research works concerning the use of ultrasound scans are
mostly involved in mass detection, and the dataset used is also
standard 2D scans. With the introduction of high-resolution US
and 3D scans as rightly pointed out by Cheng et al. [20],
research works implementable on these latest version datasets
can be developed for the detection of microcalcifications. Lastly,
in recent years, there has been a trend of using complementary
imaging modalities (mammography with ultrasound scans for
instance), and therefore, there arises a need for developing CAD
systems implementable on mixed modality datasets.

From the survey, it has been noticed that most of the works
based on medical image processing have preferred the use of
deep learning techniques over machine learning. However,
machine learning too can be employed over datasets for
representing diagnostic parameters in numerical form as has
been represented in the dataset titled “Breast Cancer Wisconsin
(Diagnostic) Data set” in Kaggle. It has the diagnosis based on 10
real-valued features like radius, texture, perimeter, concavity,
and symmetry to name a few. This, therefore, opens up a new
avenue for research concerning classification based on
numerical diagnostic parameters.

3.6 Acknowledgment

I would like to extend my gratitude towards Dr. Girish N. Bedre

(MD Radiation Oncology) of Shree Guruji Hospital, Nashik for
giving his valuable insights and sharing information
concerning the technical nuances of the imaging modalities
discussed in this work. I would also like to thank my guide and
co-author Dr. Leninisha Shanmugam for propelling my research
efforts in the correct direction. Lastly, I would like to thank
Vellore Institute of Technology, Chennai for providing me with
the opportunity and knowledge resources to carry forward my
research.

References

1. Chaudhury, S., Rakhra, M., Memon, N., Sau, K., Ayana, M.T.,
Breast cancer calcifications: Identification using a novel
segmentation approach. Comput. Math. Methods Med., 2021,
9905808, Oct. 2021, doi: 10.1155/2021/9905808.
2. Loizidou, K., Skouroumouni, G., Nikolaou, C., Pitris, C., An
automated breast micro-calcification detection and
classification technique using temporal subtraction of
mammograms. IEEE Access, 8, 52785–52795, 2020, doi:
10.1109/ACCESS.2020.2980616.
3. Jeong, J.-W., Yu, D., Lee, S., Chang, J.M., Automated detection
algorithm of breast masses in three-dimensional ultrasound
images. Healthc. Inform. Res., 22, 4, 293–298, Oct. 2016, doi:
10.4258/hir.2016.22.4.293.
4. Hamed, G., Marey, M., Amin, S.E., Tolba, M.F., Automated
breast cancer detection and classification in full field digital
mammograms using two full and cropped detection paths
approach. IEEE Access, 9, 116898–116913, 2021, doi:
10.1109/ACCESS.2021.3105924.
5. Celik, Y., Talo, M., Yildirim, O., Karabatak, M., Acharya, U.R.,
Automated invasive ductal carcinoma detection based using
deep transfer learning with whole-slide images. Pattern
Recognit. Lett., 133, 232–239, May 2020, doi:
10.1016/j.patrec.2020.03.011.
6. Ehteshami Bejnordi, B. et al., Automated detection of DCIS in
whole-slide H&E stained breast histopathology images. IEEE
Trans. Med. Imaging, 35, 9, 2141–2150, Sep. 2016, doi:
10.1109/TMI.2016.2550620.
7. Debelee, T.G., Schwenker, F., Ibenthal, A., Yohannes, D., Survey
of deep learning in breast cancer image analysis. Evol. Syst.,
11, 1, 143–163, Mar. 2020, doi: 10.1007/s12530-019-09297-2.
8. Mahmood, T., Li, J., Pei, Y., Akhtar, F., Imran, A., Rehman, K.U.,
A brief survey on breast cancer diagnostic with deep
learning schemes using multi-image modalities. IEEE Access,
8, 165779–165809, 2020, doi: 10.1109/ ACCESS.2020.3021343.
9. Al-Dhabyani, W., Gomaa, M., Khaled, H., Fahmy, A., Dataset of
breast ultrasound images. Data Brief, 28, 104863, Feb. 2020,
doi: 10.1016/j. dib.2019.104863.
10. Witowski, J. et al., The NYU breast MRI dataset v1.0.
11. Venkatachalam, N., Shanmugam, L., Heltin, G.C.,
Govindarajan, G., Sasipriya, P., Enhanced segmentation of
inflamed ROI to improve the accuracy of identifying benign
 and malignant cases in breast thermogram. J. Oncol., 2021,
5566853, 2021, doi: 10.1155/2021/5566853.
12. Roy, S.D., Das, S., Kar, D., Schwenker, F., Sarkar, R., Computer
aided breast cancer detection using ensembling of texture
and statistical image features. Sensors, 21, 11, 3628, Art. no.
11, Jan. 2021, doi: 10.3390/s21113628.
13. Sanyal, R., Kar, D., Sarkar, R., Carcinoma type classification
from high-resolution breast microscopy images using a
hybrid ensemble of deep convolutional features and gradient
boosting trees classifiers. IEEE/ACM Trans. Comput. Biol.
Bioinform., 19, 4, 2124–2136, Jul. 2022, doi: 10.1109/
TCBB.2021.3071022.
14. Das, K., Conjeti, S., Chatterjee, J., Sheet, D., Detection of breast
cancer from whole slide histopathological images using deep
multiple instance CNN. IEEE Access, 8, 213502–213511, 2020,
doi: 10.1109/ACCESS.2020.3040106.
15. Han, Z., Wei, B., Zheng, Y., Yin, Y., Li, K., Li, S., Breast cancer
multi-classification from histopathological images with
structured deep learning model. Sci. Rep., 7, 1, 4172, Jun.
2017, doi: 10.1038/s41598-017-04075-z.
16. Mohiyuddin, A. et al., Breast tumor detection and
classification in mammogram images using modified YOLOv5
network. Comput. Math. Methods Med., 2022, 1359019, Jan.
2022, doi: 10.1155/2022/1359019.
17. Darweesh, M.S. et al., Early breast cancer diagnostics based
on hierarchical machine learning classification for
mammography images. Cogent Eng., 8, 1, 1968324, Jan. 2021,
doi: 10.1080/23311916.2021.1968324.
18. Tsochatzidis, L., Koutla, P., Costaridou, L., Pratikakis, I.,
Integrating segmentation information into CNN for breast
cancer diagnosis of mammographic masses. Comput. Methods
Programs Biomed., 200, 105913, Mar. 2021, doi:
10.1016/j.cmpb.2020.105913.
19. Mohapatra, S., Muduly, S., Mohanty, S., Ravindra, J.V.R.,
Mohanty, S.N., Evaluation of deep learning models for
detecting breast cancer using histopathological
mammograms images. Sustainable Oper. Comput., 3, 296–302,
Jan. 2022, doi: 10.1016/j.susoc.2022.06.001.
20. Cheng, H.D., Shan, J., Ju, W., Guo, Y., Zhang, L., Automated
breast cancer detection and classification using ultrasound
images: A survey. Pattern Recognit., 43, 1, 299–317, Jan. 2010,
doi: 10.1016/j.patcog.2009.05.012.
21. Sadad, T., Hussain, A., Munir, A., Habib, M., Ali Khan, S.,
Hussain, S., Yang, S., Alawairdhi, M., Identification of breast
malignancy by marker-controlled watershed transformation
and hybrid feature set for healthcare. Appl. Sci., 10, 6, 1900,
2020, https://doi.org/10.3390/app10061900.
22. Byra, M., Breast mass classification with transfer learning
based on scaling of deep representations. Biomed. Signal
Process. Control, 69, 102828, Aug. 2021, doi:
10.1016/j.bspc.2021.102828.
23. Li, Y., Gu, H., Wang, H., Qin, P., Wang, J., BUSnet: A deep
learning model of breast tumor lesion detection for
ultrasound images. Front. Oncol., 12, 848271, 2022, Accessed:
Dec. 17, 2022. [Online]. Available:
https://www.frontiersin.org/articles/10.3389/fonc.2022.848271
.
24. Dewangan, K.K., Dewangan, D.K., Sahu, S.P., Janghel, R.,
Breast cancer diagnosis in an early stage using novel deep
learning with hybrid optimization technique. Multimed. Tools
Appl., 81, 10, 13935–13960, 2022, doi: 10.1007/ s11042-022-
12385-2.
25. Hu, Q., Whitney, H.M., Giger, M.L., A deep learning
methodology for improved breast cancer diagnosis using
multiparametric MRI. Sci. Rep., 10, 10536, 2020,
https://doi.org/10.1038/s41598-020-67441-4.
26. Ha, W. and Vahedi, Z., Automatic breast tumor diagnosis in
MRI based on a hybrid CNN and feature-based method using
improved deer hunting optimization algorithm. Comput.
Intell. Neurosci., 2021, 5396327, Jul. 2021, doi:
10.1155/2021/5396327.
27. Ma, D., Shang, L., Tang, J., Bao, Y., Fu, J., Yin, J., Classifying
breast cancer tissue by Raman spectroscopy with one-
dimensional convolutional neural network. Spectrochim.
Acta A Mol. Biomol. Spectrosc., 256, 119732, Jul. 2021, doi:
10.1016/j.saa.2021.119732.
28. Jiao, H., Jiang, X., Pang, Z., Lin, X., Huang, Y., Li, L., Deep
convolutional neural networks-based automatic breast
segmentation and mass detection in DCE-MRI. Comput. Math.
Methods Med., 2020, 1–12, May 2020, doi:
10.1155/2020/2413706.
29. Mohamed, E.A., Rashed, E.A., Gaber, T., Karam, O., Deep
learning model for fully automated breast cancer detection
system from thermograms. PLoS One, 17, 1, e0262349, Jan.
2022, doi: 10.1371/journal.pone.0262349.

Note

1. *Corresponding author:
[email protected]
4
Enhancing Feature Selection
Through Metaheuristic Hybrid
Cuckoo Search and Harris Hawks
Optimization for Cancer
Classification
Abrar Yaqoob1*, Navneet Kumar Verma1, Rabia Musheer
Aziz2 and Akash Saxena3

1
School of Advanced Science and Language, VIT Bhopal
University, Kothrikalan, Sehore, India

2
State Planning Institute (New Division), Planning Department
Lucknow, Utter Pradesh, India

3
School of Engineering and Technology, Central University of
Haryana, Mahendergarh, India

Abstract
Gene expression platforms offer vast amounts of data that can
be utilized for investigating diverse biological processes.
However, due to the existence of redundant and irrelevant
genes, it remains challenging to identify crucial genes from
high-dimensional biological data. To overcome this obstacle,
researchers have introduced different feature selection (FS)
methods. Developing more efficient and accurate FS techniques
is essential to select important genes for the classification of
complex biological information with multiple dimensions for
many purposes. To tackle the difficulty of selecting genes in
high-dimensional biological datasets, a novel approach called
the Harris hawks optimization and cuckoo search algorithm
(HHOCSA) is proposed for commonly used machine learning
classifiers such as K-nearest neighbors (KNN), support vector
machine (SVM), and naive Bayes (NB). The effectiveness of the
hybrid gene selection algorithm was assessed using six
commonly used datasets and compared to other techniques for
selecting features. The experimental findings demonstrate that
the HHOCSA outperforms alternative methods when
considering performance metrics of accuracy measures such as
precision, sensitivity, and specificity. Furthermore, the HHOCSA
approach proposed in this study is both computationally
efficient and consistent in terms of variability when compared
to the other methods. Therefore, the proposed hybrid algorithm
can be used as a useful instrument for the classification of
cancer datasets and can help medical professionals make
better-informed decisions in cancer diagnosis.

Keywords: Harris hawks optimization and cuckoo search

algorithm (HHOCSA), K-nearest neighbors (KNN), support
vector machine (SVM), naive Bayes (NB)

4.1 Introduction

Cancer is a significant global health issue that affects

populations across the globe, with millions of new cases
diagnosed every year. While advances in medical technology
have significantly improved cancer diagnosis and treatment,
the complexity of the disease remains a challenge. Early and
accurate diagnosis is crucial for successful treatment outcomes,
and the potential of machine learning techniques is enormous
in this area. In particular, gene expression microarray data
analysis using machine learning has appeared as a promising
tool for identifying cancer [1]. Machine learning models have
been developed during the last 10 years as a result of the study
of publicly accessible cancer data. Recently, models for
classifying cancer types and diagnosing cancer in its early
stages based on machine learning have been published [2].
Machine learning has greatly helped genetic data analysis. It
has been put into use in several fields of biological sciences,
mostly in the last 20 years. Making prediction models for
identifying malignant growths from tumor tissue is among the
most promising uses of machine learning.

Gene expression microarray data analysis involves the

discovery of genes important for cancer categorization and
diagnosis. Because of the vast number of genes involved and
the complex nature of the data, which lead to a high-
dimensional space, it becomes a challenging task. Selecting a
group of genes that are most pertinent to the condition will
increase the precision of cancer categorization. This process is
known as feature selection, and it involves identifying a smaller
set of characteristics that can accurately discriminate between
cancer and non-cancer sample. The use of arrays to detect
cancer is a newer area of intelligence analysis. The capacity of
gene chip technology to examine hundreds of genes’ expression
patterns concurrently results in the development of unique
profiles of biological processes. Researchers are examining the
properties of microarray technology for classification of cancer,
which may help to provide a deeper understanding of the
disease.

Cancer classification issues are resolved by differentiating

healthy and unhealthy samples using microarray organic
phenomenon information. Microarray technology has the
potential to provide a substantial technique for prognosis and
diagnosis, which will also expand knowledge of the primary
origins of cancer and open the door to the creation of
innovative treatments. The majority of the alternatives (genes)
are useless even though there are more possibilities (genes) in
these data than there are samples, which results in an oversized
visit classification accuracy. The method of factor selection was
first used as a technique for categorization, comprises choosing
the smallest group of genetic variants that are most relevant to
the associated cluster. As a result, the classifier’s ability to
categorize samples precisely is improved. As a result, using
heuristic methods to solve this problem is more effective. The
two key concerns driving microarray research are the
knowledge is limited or deficient, meaning there are often just a
few dozen samples, and it is robust, meaning it encompasses
many thousands of genes (features). Second, because genes
seem to be either directly or indirectly associated, there is a
great level of richness in the understanding of organic
phenomena. These methods did not work effectively since
conventional machine learning procedures function best when
there are more samples than attributes. As they utilize various
technical advancements in the fields of deep learning, machine
learning, and artificial intelligence to aid healthcare
professionals in quickly spotting abnormalities, these image-
based medical systems can be highly helpful for the early
identification of cancer. Sadly, this system’s architecture
demands a training set containing useful characteristics, yet the
actual features acquired may be completely useless.
Consequently, before a consequence, prior to categorization,
relevant features with a high potential for discrimination are
selected [3].

Feature selection encompasses various techniques, such as

filter, wrapper, and embedded approaches. Filter methods
prioritize characteristics determined by their significance to the
desired outcome variable, while wrapper methods assess
feature subsets using machine learning algorithms. Embedded
methods integrate feature selection directly into the learning
algorithm. All of these techniques possess their own advantages
and limitations, and the selection of an appropriate method
primarily relies on the characteristics of the dataset and the
specific problem under consideration. Recently, various novel
hybrid optimization algorithms have been proposed for the task
of selecting relevant features in cancer classification. Hybrid
algorithms combine two or more optimization techniques to
achieve better performance than each individual technique.
The suggested method reduced the computing complexity of the
Harris hawks optimization (HHO) exploitation process by using
the cuckoo search (CS). Although the CS algorithm has a strong
exploitation process, it is less successful than the HHO
algorithm at discovering new ideal solutions. As a result, the CS
algorithm’s main shortcomings are its convergence issue and
early discovery of the local optima. CS is an algorithm designed
for optimization purposes that draws inspiration from the
behavior of cuckoo birds. It operates by generating a population
of potential solutions and creating new solutions by imitating
the cuckoo’s habit of deposition of eggs within the nests of
different bird species. The algorithm incorporates a random
walk to generate fresh candidate solutions, and the top-
performing solutions are retained for the subsequent
generation. On the other hand, HHO is another optimization
algorithm rooted in the hunting behavior of Harris hawks. It
involves a collective effort by a group of hawks to search for
prey, with each hawk employing its own distinct strategy. The
algorithm employs an integration of both local and global
search methods, techniques to uncover the optimal solution.
The combination of CS and HHO has been shown to be effective
for feature selection in cancer classification. This hybrid
algorithm combines the strengths of both CS and HHO to
achieve better performance than either algorithm alone. The
algorithm uses CS to generate candidate solutions, and HHO is
used to refine the solutions and determine the most suitable set
of features. The CSAHHO algorithm combines the HHO and CS
algorithms to preserve the equilibrium between the period
following the exploitation and exploration stages [3].

The recommended work contributes considerably to the

following:

1. A novel hybrid algorithm, known as the HHOCSA (cuckoo

search and Harris hawks optimizations), has been developed
to specifically target cancer identification.
2. The suggested approach fixes the issue of minimal
convergence of HHO.
3. The combination of HHO and CS algorithms is employed to
keep the processes of extraction and exploration in balance.

This article will follow a structured approach to explore a

specific topic. The Introduction will provide an overview of the
research purpose and structure. Section 4.2 will cover related
work, while Section 4.3 will introduce the unique research
approach used. Section 4.4 will present the overview of the
classifiers used in the proposed algorithm. Section 4.5 will
provide an analysis and interpretation of the results and their
potential consequences, while Section 4.6 will conclude the
article by summarizing the key findings, discussing their
significance, and providing recommendations for future
research. Figure 4.1 shows the proposed methodological
approach in broad strokes.
Figure 4.1 The proposed research methodology.

4.2 Related Work

Nanglia et al. developed a technique for categorizing cancer
that included three basic components. The data were initially
preprocessed in the first block, next using the SURF approach,
features were retrieved from the second block. In the third
block, the model was enhanced using a genetic algorithm, and
using a neural network employing feed-forward and back-
propagation techniques, the tumor was classified. The overall
classification correctness achieved by this hybrid algorithm was
an amazing 98.08% [4].

Sahu et al. created a prediction model by fusing learning

strategies based on artificial intelligence with multivariate
statistical methods. This study suggests using artificial neural
networks (ANNs) with principal component analysis (PCA) to
choose the characteristics. Data are preprocessed using PCA to
determine the most crucial properties. Utilizing the WBC
Dataset with 10-fold cross-validation, the suggested method
demonstrated improved accuracy with respect to sensitivity [5].

A hybrid metaheuristic technique for classifying cancer was

proposed by Motieghader et al. using the genetic algorithm and
learning automata (GALA). When assessed using the selected
method, GALA outperformed certain newly proposed
algorithms on each dataset [6].
The benefits associated with the artificial bee colony algorithm
and ant colony algorithm are combined in a unique method
known as AC-ABC hybrid, created by Shunmugapriya and
Kanmani to enhance feature selection. According to the findings
of the experiments, the suggested method seeks to increase
feature selection and classification accuracy [7].

A unique two-stage method for choosing pertinent genes for

cancer classification, called MI-GA Gene Selection, has been
proposed by Jansi Rani M. and Devaraj. First-stage analysis
revealed the presence of genes with a high degree of mutual
information that are connected to cancer. In the second phase, a
genetic algorithm is employed to select a set of genes capable of
accurately classifying cancer data. The classification process
utilized a support vector machine (SVM). The findings reveal
that the MI-GA method, as proposed, surpasses current
techniques in terms of classification accuracy for ovarian, lung,
and colon cancer datasets [8].

The mRMR-COA-HS is a two-step technique of gene selection

that was developed by Elyasigomari et al. To find a significant
set of genes, they first use the lowest redundancy and
maximum relevance strategy. Following that, they employ a
wrapper configuration made consisting of a support vector
machine classifier and a brand-new cuckoo optimization
technique called the harmony search (COA-HS). The maximum
degree of categorization accuracy is guaranteed by this
algorithm [9].

Jain et al. devised a hybrid approach for cancer classification

that combines correlation-based feature selection with an
improved binary particle swarm optimization. The model
utilizes the naive Bayes classifier alongside stratified 10-fold
cross-validation to classify biological samples with binary and
multiclass malignancies. Consistently, the outcomes exhibit the
superior performance of this model over seven other well-
established strategies, regarding both classification precision
and the quantity of selected genes [10].

Shahbeig et al. introduced a novel hybrid method for

identifying crucial genes linked to the development of breast
cancer. Their objective was to uncover the most concise gene set
capable of achieving optimal levels of classification accuracy,
sensitivity, and specificity. To achieve this, they merged the
teaching learning-based optimization (TLBO) technique with a
recently introduced mutant fuzzy adaptive particle swarm
optimization method. Their approach produced encouraging
outcomes, with a detection accuracy of 91.88%, sensitivity of
90.55%, and specificity of 93.33% when applied to microarray
data for breast cancer detection [11].
Lu et al. utilized a hybrid method for feature selection, merging
mutual information maximization (MIM) with an adaptive
evolutionary algorithm. Their experimental results
demonstrate that this approach effectively reduces the
dimensionality of gene expression data and eliminates
redundancies in the classification process. When compared to
traditional feature selection methods, their generated gene
expression dataset achieved the highest classification accuracy
[12].

Naeem et al. demonstrated the precise diagnosis of liver cancer

using machine learning algorithms. They integrated a dataset
consisting of two-dimensional computed tomography images
and magnetic resonance imaging. Through the application of
10-fold cross-validation, they assessed the performance of
multilayer perceptron (MLP), support vector machine, random
forest (RF), and J48 classifiers using an optimized hybrid-
feature dataset. Notably, the MLP classifier showcased
remarkable success, achieving an impressive accuracy of 99%
[13].

Kharrat and Neji employed simulated annealing and genetic

algorithm (SA-GA) techniques to generate the most optimal
subset of features. The efficacy of this approach was assessed by
employing two real datasets of magnetic resonance imaging for
brain tumor analysis. Advanced methods like simulated
annealing and genetic algorithms were compared to the
suggested methodology. The outcomes showed that SAGA
outperforms and computes more quickly than these approaches
when used separately. With segmentation correctness rates of
97.82% and 0.74 for glioma tumors and 95.12% and 3.21% for
pituitary malignancies, the analysis showed that the suggested
method is superior to cutting-edge methodologies [14].

Zheng et al. presented a hybrid method for selecting feature

subsets known as the maximum Pearson maximum distance
enhanced whale optimization approach. This technique
combines Pearson’s correlation coefficient with a correlation
distance-based MPMD filtering algorithm. Maximum Pearson
Maximum Distance (MPMD) utilizes two parameters to adjust
the weights of relevance and redundancy. The effectiveness of
the approach was assessed using 10 benchmark datasets from
the UCI machine learning databases. The results highlighted its
superiority over three existing wrapper strategies and one
hybrid strategy in terms of classification accuracy [15].

Stephan et al. developed a hybridization of the artificial bee

colony algorithm that blends the whale optimization’s bubble
net attacking strategy with the ABC’s exploitative employee bee
phase hybridization of artificial bee colony with whale
optimisation (HAW). The exploratory phase of the HAW
algorithm consists of a mutative initialization step, which
improves upon the limited exploration capability of the
standard ABC method. In the context of an ANN model, the
HAW technique is utilized to simultaneously select features and
optimize parameters. To create hybrid versions of HAW, three
backpropagation learning algorithms (Levenberg–Marquart,
robust backpropagation, and momentum-based gradient
descent) are incorporated, resulting in HAW-LM, HAW-RP, and
HAW-GD. The effectiveness of these hybrid variations is
assessed using datasets related to breast cancer, including
WBCD, WDBC, WPBC, DDSM, MIAS, and IN-breast, based on
accuracy, complexity, and processing time. Among these, the
HAW-RP variant, utilizing a low-complexity ANN model,
achieves the highest accuracy with values of 99.2%, 98.5%,
96.3%, 98.8%, 98.7%, and 99.1% for the respective datasets [16].

In their examination of breast cancer, Khamparia et al. focus a

lot of stress on the process of transfer learning. They used a
modified version of the visual geometry group (VGG) (MVGG)
on datasets that included 2D and 3D mammograms. According
to the findings of their studies, the suggested hybrid transfer
learning model’s accuracy, which combines MVGG with
ImageNet, is 94.3%. The MVGG design, which was the only one
studied, had an accuracy rate of 89.8%. The recommended
hybrid pretrained network is therefore superior to previous
convolutional neural networks (CNNs) of a similar design. The
authors give thorough explanations of these findings [17].

Shukla et al. have developed a unique hybrid framework

named CMIMAGA to extract important indications from gene
expression data. This method combines adaptive genetic
algorithm and conditional mutual information maximization
(CMIM). The method employs a wrapper approach as a
straightforward filter, followed by the CMIM approach as a
second filter, to exclude extraneous genes. The next step is to
choose highly discriminating genes that can reliably diagnose
tumors and malignancies using an adaptive genetic algorithm.
As a fitness function, the system uses classifiers to choose these
genes. In order to evaluate the effectiveness of the suggested
method, six commonly employed microarray datasets were
employed, alongside three classifiers: Extreme learning
machine, support vector machine, and K-nearest neighbors
(KNN) were employed for comparison purposes. The suggested
method using ELM outperformed earlier filter and wrapper
techniques, according to experimental findings, and achieved
improved classification accuracy with fewer genes [18].

In order to classify microarray data with high dimensionality

using artificial neural networks, Aziz R. et al. have developed a
unique method for feature selection and extraction. The
strategy uses artificial bee colonies (ABCs) as an optimization
technique and as an extraction method, independent
component analysis (ICA). By contrasting the ICA + ABC method
with ICA and ABC alone and running rigorous testing on six
gene expression microarray datasets, the study assesses the
effectiveness of the procedure. Comparing the suggested
approach to previous ICA-based bio-inspired algorithms and
well-known filtering strategies, in terms of classification
accuracy and the number of selected genes, it outperforms the
current technique. The studies’ findings show that the ICA +
ABC algorithm improves an ANN classifier’s classification
accuracy rate, making it a potential method for resolving
concerns using microarray data in the context of gene selection
and the classification of cancer [19].

Kilicarslan et al. carried out an investigation utilizing three

micro-array datasets: ovarian, leukemia, and central nervous
system (CNS). They employed autoencoder methods to reduce
the dimensionality and employed support vector machines and
convolutional neural networks for analysis. The leukemia
dataset comprised 72 samples with 7,129 genes and two classes,
while the ovarian and CNS datasets also contained 60 samples,
7,129 genes, and two classes. With bracket accuracy of 96.14 for
ovarian datasets, 94.83 for leukemia datasets, and 65 for CNS
datasets without dimension reduction, SVM shown good
performance. The ovarian, leukemia, and CNS datasets’
respective bracket accuracy for the proposed mixed system
Relief-F CNN was 98.6, 99.86, and 83.95 [20].

Melekoodappattu and Subbian proposed utilizing extreme

learning machines (ELMs) in conjunction with the fruit fly
optimization algorithm (ELM-FOA) to modify the input weight.
The objective is to achieve the optimal arrangement at the
hidden layer of the ELM, enabling the analytical acquisition of
the output. During ELM testing, FOA’s perception and perfection
are 97.5 and 100, respectively. The cutting-edge technology can
describe calcium deposits and cancers with 99.04 delicacies.
Through the utilization of advanced preprocessing and
segmentation techniques, incorporating features from various
point filters, and harnessing a robust classifier algorithm, this
approach showcases outstanding performance [21].

Two novel CNN-SVM hybrid models have been developed by

Keerthana et al. to classify dermoscopy pictures as benign or
malignant lesions. Features are extracted by the primary and
secondary CNN models, which are then merged and sent into
the SVM classifier for classification. The efficacy of the proposed
methodology is evaluated using expert dermatologist labeling.
Using the ISBI 2016 dataset, the suggested models outperformed
current CNN models and obtained an accuracy of 88.02% and
87.43%, respectively, outperforming traditional CNN models
[22].

The literature survey brings to the conclusion that CSAHHO, a

hybrid cuckoo search method with Harris hawks optimization,
has not yet been used for cancer classification. Filling this gap is
the main objective of the suggested inquiry. It is worth noting
that despite the cuckoo search algorithm being an effective
method, it has sluggish convergence and local minima
stagnation. While HHO is better for obtaining global optimum,
in some cases, it may have a low convergence problem.
Moreover, no one solution can solve all optimization problems,
according to the no free lunch (NFL) concept. The proposed
hybrid algorithm was developed with this goal in mind.

4.3 Proposed Methodology

The methodology introduced in this paper involves several key

steps, including data collection, data preprocessing, feature
selection, training the model, evaluation, refinement, and final
testing. The first step is to collect relevant data from various
sources, followed by cleaning, normalizing, and transforming
the data to make them suitable for analysis. Then, feature
selection is conducted to identify the most crucial
characteristics (features) that should be incorporated into the
model. Once this is done, a hybrid algorithm is selected to use
the data to train a machine learning model. The trained model
is then assessed for its performance on a separate validation
set, and any necessary changes are made based on the results
obtained. To assess the model’s generalization ability, it is tested
on a different test set. By following this proposed methodology,
the paper aims to build a robust and accurate model that can
effectively address the problem.

4.3.1 Cuckoo Search Algorithm

The cuckoo bird’s reproductive habits served as the inspiration

for the optimization method used in computer science. The
strategy makes use of a tactic known as Levy flying in place of
random movements with a wide range and unexpected pattern,
which would require a significant shift from the existing
position [23, 24]. In order to ensure the survival of its young,
the cuckoo lays its eggs in other birds’ nests, causing the host
birds to raise the cuckoo chicks as if they were their own [25,
26].

Brood Parasitism: Uses another bird’s nest as its own, hides its
eggs there, and then relies on the host bird to raise the young.
Host Bird: Owner of the host nest, responsible for raising and
feeding the cuckoo chicks.

Parasite Bird: Birds that deposit eggs within the host bird nest.

Levy Flight: A levy flight is a physical phenomenon whose

phase lengths follow the probability distribution with a
pronounced tail known as the Levy distribution. Levy
distribution is used to regulate the Levy flights’ transition
opportunity in the manner described below:

For 1 < λ ≤ 3

(4.1)
The process of using Levy flights to generate random numbers
comprises two parts from the perspective of implementation.

1. Uniform distribution-based random direction selection

2. Making decisions that follow the selected Levy distribution

Random Walk: Random process that describes a path that

comprises a sequence of discrete constant length steps in
random directions on some mathematical space (1D, 2D, or nD)
such as integers.

Random Walk Equation:

 (4.2)
where = new solution, = existing solution, α = walk
length, = entry considerate multiplication, L(λ) = Levy
expositor.

The fundamental cuckoo search algorithm in pseudocode is

developed based on these criteria. It is crucial to control the
situation while seeking new solutions. Lévy flights,
characterized by random walks, help prevent large movements
that could lead to escaping the search area. The size of each step
is determined by the scale of the problem, and this can be
achieved by incorporating interest. Exploring the use of Lévy
flight for optimization proves to be an intriguing topic for
further investigation. For ease of use, as recommended by the
authors, a common step size factor of 0.01 is employed.
PSEUDOCODE FOR THE CUCKOO SEARCH

1. Receive the initial assignment as the input.

2. Establish a baseline population based on assignment.
3. While (stop requirement not satisfied)
1. Select a cuckoo at random from Levy flights
2. Examine its effectiveness/quality Fi
3. Randomly select one nest out of n (let us say j).
4. If (Fi > Fj)
5. Substitute the new solution for j
6. Unfavorable nests are discarded, and new ones are
constructed
7. Maintain the top solutions (or nests of excellent
solutions)
8. Order the answers to determine which is currently
best.
4. End while
5. Report the best solution

4.3.2 Harris Hawks Algorithm

Native to the southern Arizona region of the United States, the

Harris hawk, is considered one of the most intelligent bird
species largely due to its feeding habits. This raptor is known
for its strong family ties, and research on animal behavior has
revealed that Harris hawks actively support their group. In
contrast to most other bird species, which hunt, attack, and eat
prey individually, Harris hawks participate in group meals that
gradually increase in size during nonbreeding months. Every
member of a Harris hawk family takes part in the nightly feast
[27].

Heidari et al. created a labeling algorithm called the HHO

method, which imitates the behavior of Harris hawks in
catching prey (2019). The Harris hawks have exceptional
behaviors that help them detect and capture prey quickly. They
use a technique called “7 killings” to weaken their prey before
capturing it. This technique involves several tasks, including
striking quickly to startle the prey, engaging in fast and
identical dives to confuse the prey, and having a skilled hawk
take the prepared victim. The prey is then distributed among
the team. The HHO method uses similar strategies in searching
for and attacking targets, combining exploratory and
exploitative functions to achieve success [28].

Assume Trabbit(H) indicates where the prey is and T(H) indicates

where the hawks are. Equation 4.3, which considers the
probability p, is used to calculate the positioning of the hawks
during the exploration phase of the (H + 1) iteration. The
locations of the other hawks and the prey are used to determine
the position of the hawks for the (H + 1) iteration if p is less than
0.5 (rabbit). An estimated location is given for the (H + 1)-
iteration. Similarly, for p >= 0.5 by randomly selecting hawks, as
suggested by Trabbit(H).

(4.3)

where rd1, rd2, rd3, rd4, and p are the random variables and lb,
ub represents lower and upr bounds, and Tm(H) denotes the
mean position and is calculated as follows:

(4.4)

When the energy level of the escaping rabbit (E) is considered,

the HHO algorithm transitions from the exploration phase to
the exploitation phase. This transition is symbolized as follows:

(4.5)
where E0 is the rabbit’s (the prey’s) initial state of energy,
randomly produced using a [-1, 1] range, and Max IT is the most
iterations allowed. When |E| > 1, the hawks are looking in
more areas to find where their prey is. According to this, the
exploitation phase of the HHO discovery process starts at |E| <
1.

HHO is used to plan the rabbit’s escape before the hawks’

unexpected approach. Let n represent the possibility of a rabbit
escaping successfully (n 0.5) or unsuccessfully (n >= 0.5) prior to
the hawk’s unexpected pounce. The hawks use a soft or harsh
siege depending on whether the victim is able to escape
successfully or not. The gentle siege is depicted in Equations
4.6–4.8 and took place when |E| > = 0.5.

(4.6)
(4.7)
(4.8)
where J is the rabbit’s power of chance leaps, and rd is a
random variable. Equation 4.9 illustrates a harsh siege that took
place for |E|< 0.5.

(4.9)
The HHO method models the prey jumping frog for increasing
fast diving using the Levy flight (FL) principle, which occurs for
|E|>= 0 and n < 0.5. The hawks’ ability to dive successfully or
unsuccessfully is measured by the following:

(4.10)
When the dive in Equation 4.10 fails, hawks perform a dive
based on Levy flying based on Equation 4.11.

(4.11)
where W is the random variable, and Pdim stands for the issue
dimension. The Levy flight functn Ly is calculated as follows:

(4.12)

(4.13)

where β is a constant with a value of 1.5, u and v are random

variables with a range of [0, 1]. Equations 4.10 and 4.11 are used
to calculate the final step in updating the hawks’ position for
this dive.

(4.14)

Similarly, Equation 4.15 computes the position of the hawks

during the hard besiege phase with successive dives, and it does
so for |E|< 0.5 and n < 0.5.
PSEUDOCODE FOR HARRIS HAWK

Inputs: population size h and Max_iter

1. Initialize a population of h random hawks xi (i=1, ….h)

2. Calculate the fitness of each hawk
3. Xrabbit = best position of minimum fitness
4. T=1
5. while(t<=Max_iter)
1. Update E
2. If (|E| >= 1)
1. Update x(t+1)
3. If (|E| < 1)
4. If (p >= 0.5 && |E| >= 0.5)
1. Update x(t+1)
5. If (p >= 0.5 && |E| < 0.5)
1. Update x(t+1)
6. If (p < 0.5 && |E| >= 0.5)
1. Update x(t+1) using soft besiege progressive rapid
dives
7. If (p >= 0.5 && |E| >= 0.5)
1. Update x(t+1) using hard besiege progressive
rapid dives
8. If (fitness(x(t+1)) < fitness(xrabbit))
1. Update xrabbit = x(t+1)
6. t++
Output: rabbit position xrabbit

4.3.3 The Proposed Hybrid Algorithm

The suggested method for improving the convergence and

effectiveness of the Harris hawks employs a combined
algorithm called CSAHHO, which combines the cuckoo search
algorithm and the Harris hawks optimization algorithm. By
combining the process of updating the population mechanism
of CSA with the integrated HHO’s population update structure,
the algorithm gains more flexibility to explore the population
and achieve greater diversity. As a result, the CSAHHO
approach enhances the capacity to explore and achieve a
broader range of solutions. Figure 4.2 shows a hybrid flowchart
of the HHO and CSA.

The following lists the CSAHHO implementation steps:

1. To specify the parameters, we need to assign values to the

following: maximum iterations (Max IT), population size
(Pop-size), number of feature variables, awareness
probability, flight duration, and crossover rate.
2. Population Encoding: A binary encoding strategy was used to
encode the population. In this approach, a binary string
 comprising of “0s” and “1s” is used to depict the population.
When a value is “1,” the feature is included, whereas it is
excluded when the value is “0.”
3. Fitness Function: By removing superfluous attributes, the
suggested CSAHHO algorithm chooses an ideal feature set.
With the aid of the fitness function, the algorithm
accomplishes this task. The effectiveness of the classifier and
the quantity of characteristics chosen are two competing
goals that are integrated by the fitness function (FS). The
convolutional neural network evaluates the performance of
each chosen feature set. The feature set that the algorithm
chooses offers maximum accuracy with the fewest features.
Using Equation 4.15, the fitness Xi(t) of each individual
within the population vector is computed.

(4.15)

where cost(xi(t)) stands in for Xi(t) fitness value. E(xi(t))

represents the performance rating assigned by the classifier.
The term “FS/N” stands for the ratio of “selected features” to
“total features.”

Depending on the probability value (Pi) determined, it is

necessary to decide whether to utilize CSA or HHO when
updating the population for the hybrid CSAHHO model. The
population update process is described in Equation 4.16, and
the algorithm below provides the implementation stages.

(4.16)
Figure 4.2 Hybrid flowchart of the HHO and CSA.
HYBRID CSAHHO ALGORITHM

Inputs:

population size h
Max_iter
Harris Hawks Optimization pa
Cuckoo Search Optimization n
Initial assignment

1. Initialize a population of h random hawks xi (i=1, ….h)

2. Calculate the fitness of each hawk
3. Xrabbit = best position of minimum fitness
4. T=1
5. While(t<=Max_iter)
6. Update E using Eq. (16)
7. If (|E| >= 1)
8. Update x(t+1) using Harris Hawks Optimization
9. If (|E| < 1)
10. If (p >= 0.5 && |E| >= 0.5)
11. Update x(t+1) using Harris Hawks Optimization
12. If (p >= 0.5 && |E| < 0.5)
13. Update x(t+1) using Harris Hawks Optimization
14. If (p < 0.5 && |E| >= 0.5)
15. Update x(t+1) using soft besiege progressive rapid dives
16. If (p >= 0.5 && |E| >= 0.5)
17. Update x(t+1) using hard besiege progressive rapid dives
18. If (fitness(x(t+1)) < fitness(xrabbit))
19. Update xrabbit = x(t+1)
Apply the Cuckoo Search Optimization
20. - Generate an initial population of n solutions based on
assignment
21. - While (stop criterion has not been met)
22. - Get a cuckoo randomly by Levy flights
23. - Evaluate its quality/fitness Fi
24. - Choose a nest among n (say j), randomly
25. - If (Fi > Fj)
26. - Replace j with the new solution
27. - A fraction (pa) of worse nests are abandoned and new
ones are built
28. - Keep the best solutions (or nests with quality solutions)
29. - Rank the solutions and find the current best
30. - End while
31. t++

Output: rabbit position

4.3.4 Classifiers Used

4.3.4.1 KNN Classifier

The K-nearest neighbor is an uncomplicated yet effective
classification algorithm used for both tasks of binary and
multiclass classification [29]. The functioning of the KNN
algorithm involves identifying the k closest data points to the
input point within the feature space and assigning the input
point the class label that appears most frequently among those
k points. The value of k is a hyperparameter that must be
adjusted to achieve optimal performance. One notable
advantage of KNN is its nonparametric nature, which means it
does not rely on assumptions about the underlying data
distribution. However, KNN is challenged by the curse of
dimensionality, as its computational complexity increases with
higher numbers of features [30, 31].

4.3.4.2 SVM Classifier

The support vector machine is a widely used and influential

classification algorithm in machine learning. SVM works by
finding the best hyperplane that separates data points into
different classes [32]. This hyperplane is selected to maximize
the margin, which represents the distance between the
hyperplane and the nearest points of each class. By employing
the kernel trick, SVM is able to transform the input data into a
higher-dimensional feature space, enabling the linear
separation of data that are not linearly separable in the original
space. Additionally, SVM exhibits robustness against overfitting
and exhibits efficient performance even with high-dimensional
data [33–35].

4.3.4.3 NB Classifier

When it comes to text classification scenarios, naive Bayes (NB)

is a simple and speedy algorithm that can be employed. The
naive Bayes algorithm is based on the principles of Bayes’
theorem and assumes that the features are conditionally
independent when given the class label. NB gives the class that
possesses the highest probability to the input data after
calculating the posterior probabilities of each class given the
input data. NB is a computationally efficient algorithm that
performs well with little training data. In contrast, naive Bayes
assumes feature independence, which may not be valid in
reality and can result in suboptimal performance on specific
datasets [36].

4.3.4.4. mRMR

The feature selection technique known as the mRMR, or

minimal redundancy maximum relevance, is frequently
employed in machine learning. In addition to minimizing
redundancy among the chosen features. The objective of
feature selection is to identify a subset of features that exhibit
high relevance to the target variable. Peng et al. [49] introduced
the mRMR algorithm in 2005, and it has since gained popularity
as a feature selection technique. Relevance and redundancy are
the two key requirements of the algorithm. While the
redundancy criterion evaluates the degree of overlap between
the chosen characteristics, the relevance criterion assesses the
amount of information that a feature holds regarding the target
variable [37]. The mRMR algorithm utilizes relevance and
redundancy scores to rank features. The relevance score of a
feature is determined by calculating its mutual information
with the target variable. In this situation, it evaluates how much
knowledge the characteristic offers about the target variable.
The mutual information quantifies the reliance between two
variables. A feature’s redundancy score is calculated by
measuring the average mutual information between the
specific feature and all other chosen features. The redundancy
score calculates how much the chosen features overlap one
another. The algorithm chooses the characteristics with the
highest relevance scores while minimizing redundancy among
them after the features are sorted according to their relevance
and redundancy scores. This is accomplished by gradually
adding features to the chosen set while periodically assessing
the set’s level of redundancy. The algorithm stops accumulating
features and returns the chosen set if the redundancy score is
greater than a given threshold [38]. It has been demonstrated
that the mRMR algorithm exhibits superior performance
compared to alternative feature selection techniques in a
number of applications, including text categorization and gene
expression analysis. However, it is crucial to remember that the
algorithm’s success depends on the particular application and
the selection of parameters, such as the redundancy score
cutoff [39].

In conclusion, the mRMR stands as a widely recognized and

popular approach for feature selection that balances the
relevance and redundancy of the selected features. The
algorithm assigns rankings to the features based on their
relevance and redundancy scores and selects the features that
maximize the relevance while minimizing the redundancy.

4.4 Experimental Setup

In this work, we sought to investigate the potential of this

system for the analysis of complicated biological data using an
HP laptop with an Intel Core i7 processor, 8 GB RAM, and a 512-
GB SSD. Six cancer microarray datasets were analyzed in this
study to see how effective the recommended approach was. The
study, which focused on gene expression, employed six high-
dimensional biological datasets that were collected from the
online database Kent Ridge. The datasets contained information
on high-grade gliomas for binary classification, colon cancer,
acute leukemia, prostate cancer, lung cancer-II, and leukemia 2
for multiclassification
(http://datam.i2r.astar.edu.sg/datasets/krbd/index.html). All of
the details and characteristics of these six datasets are listed in
Table 4.1.
Table 4.1 Information regarding the six cancer microarray data.

Class No. of
Dataset Classes Genes S
balance samples

Colon 2 2,000 (22\40) 62 C

cancer re
fo
ca
e
re
ca
re
ra
in
p
th
in
m
lo
fa
 Class No. of
Dataset Classes Genes S
balance samples

Acute 2 7,129 (47\25) 72 A

leukemia g
sy
a
d
b
b
m
ly
th
A
th

Prostate 2 12,600 (50\52) 102 P

tumor m
th
sl
a
re
o
 Class No. of
Dataset Classes Genes S
balance samples

High- 2 12,625 (28\22) 50 H

grade a
glioma b
a
a
g
a
o

Lung 2 12,533 (31\150) 181 T

cancer II cl
lu
c
m
cm
lu
c
p
o
ca
sp
 Class No. of
Dataset Classes Genes S
balance samples

p
a
la
m
th
S
m
m
lu
a
ti
fr
ca

Leukemia 3 7,129 (28\24\20) 72 L

2 o
a
a
th
a
c
 Class No. of
Dataset Classes Genes S
balance samples

g
o
m
re
p
th
b
c
im
u
w
T
d
c
sa
A
A
sa

4.4.1 The Compared Algorithms

The CSAHHO algorithm, proposed in this study, has been
compared to three other methods: mRMR, mRMR+CSA, and
mRMR+HHO. Furthermore, a hybrid approach that combines
both methods was also examined, the mRMR+CSAHHO, and has
also been tested for comparison. The comparison was
conducted based on the characteristics of the CSAHHO
algorithm. To assess the effectiveness of each method, the
classification outcomes were obtained by utilizing the genes
selected through the respective methods.

mRMR: The classification outcomes are achieved using genes

that have been chosen through the mRMR selection method.

mRMR+CSA: The classification outcomes are derived from

genes that have been chosen through the mRMR+CSA selection
process.

mRMR+HHO: The classification outcomes are derived from

genes that have been selected using the mRMR+HHO method.

mRMR+CSAHHO: The classification outcomes are derived from

genes selected using the hybrid method mRMR+CSAHHO.

4.4.2 Parameter Setting

The following table (Table 4.2) present the parameters used in
our experimental implementation of the proposed technique,
HHOCSA. In order to ensure a fair comparison, the gene
selection process involved selecting the same top-ranked genes
identified by the mRMR for each gene selection approach.

Fitness function: the proposed approach’s effectiveness is

evaluated by evaluating the accuracy of classification using
SVM, KNN, and NB classifiers. When the current fitness value
exceeds the previous value, the previous result is discarded and
replaced with the current one. However, if the current fitness
value is not better, the previous solution is retained. Ultimately,
the gene subset with the highest fitness value is chosen as the
optimal solution for predictive purposes. The fitness function
(fit) is defined as follows:
Table 4.2 Parameter settings of the proposed algorithm.

Parameter Value

Dimension Number of genes selected with

the mRMR algorithm

Maximum iterations 100

Lower bound (LB) and 0 and 1

upper bound (UB)

Global phase fitness Harris hawks optimization

modification method (HHO)

No. Harris hawks 10 and (K=5 and 10)

(SearchAgents_no)

Number of iterations 30 runs

without improvement

Local phase fitness Cuckoo search algorithm (CSA)

modification method
 Parameter Value

Cuckoo search a step size 1.0 and

Limit 5 iterations

4.5 Results and Discussion

We applied the proposed algorithm with the three most popular

classifiers for the cancer classification problem section 4.5.1
Section 4.5.2 presents the experimental results using the SVM
classifier, Section 4.5.3 showcases the experimental results
using the NB classifier, and Section 4.5.4 demonstrates the
experimental results using the KNN classifier.

4.5.1 Experimental Results of the Proposed Algorithm

With the SVM Classifier

Table 4.3 provided information presenting a comprehensive

analysis of various datasets and feature selection algorithms,
employing both 5-fold and 10-fold cross-validation techniques.
The mean accuracy values, accompanied by their respective
standard deviations, illuminate on the performance of each
combination. Notably, the results demonstrate that the
utilization of feature selection algorithms yields higher
accuracy rates compared to relying solely on the mRMR
algorithm. Among the feature selection approaches, the
mRMR+CSAHHO combination consistently stands out as the
most effective, attaining remarkably high mean accuracy values
ranging from 92.32% to a perfect 100% across different datasets.
This compelling evidence suggests that integrating feature
selection techniques into the classification models significantly
enhances their predictive capabilities, offering valuable insights
into the underlying patterns of the data. Provided information
presents a comprehensive analysis of various datasets and
feature selection algorithms, employing both 5-fold and 10-fold
cross-validation techniques. The mean accuracy values,
accompanied by their respective standard deviations, shed light
on the performance of each combination. Notably, the results
demonstrate that the utilization of feature selection algorithms
yields higher accuracy rates compared to relying solely on the
mRMR algorithm. Among the feature selection approaches, the
mRMR+CSAHHO combination consistently stands out as the
most effective, attaining remarkably high mean accuracy values
ranging from 92.32% to a perfect 100% across different datasets.
This compelling evidence suggests that integrating feature
selection techniques into the classification models significantly
enhances their predictive capabilities, offering valuable insights
into the underlying patterns of the data.
Table 4.3 Accuracies of the proposed algorithm with the mRMR, mRMR+CSA, and
mRMR+HHO for the SVM classifier with 5-fold cross-validation and 10-fold cross-
validation techniques.

5-fold cross-validation

mRM
mRMR+CSA mRMR+HHO
CSA

Datasets mRMR Mean(±STD) Mean(±STD) Mea

Colon 83.2± 95.5 ±4.90 93.6 ± 3.80 97.7

5.19

Acute 82.2 ± 83.5 ± 5.21 88.5 ± 3.65 91.0

Leukemia 6.29

Prostate 82.03 ± 91.71 ± 2.09 93.4 ± 1.4 95.3

Tumor 4.11

High 83.2 ± 84.71 ± 3.99 88.71 ± 3.19 92.32

Grade 4.77
Glioma
 Lung 84.72 ± 92.53 ± 2.11 92.53 ± 2.11 96.52
Cancer II 5.98

Leukemia 84.45 ± 94.54 ± 2.17 94.54 ± 2.17 97.42

2 3.99

Figure 4.3 presents a comprehensive comparison of errors

obtained from different analysis approaches applied to
microarray cancer datasets. The primary focus is on evaluating
the performance of the mRMR algorithm and its comparison
with two popular cross-validation techniques: 5-fold cross-
validation and 10-fold cross-validation. Additionally, the
diagram also includes the performance of three distinct
algorithms utilized for analysis. It is clear from the diagram that
10-fold validation has the least error with the SVM classifier.

The diagram presents a comprehensive comparison of errors

obtained from different analysis approaches applied to
microarray cancer datasets. The primary focus is on evaluating
the performance of the mRMR algorithm and its comparison
with two popular cross-validation techniques: 5-fold cross-
validation and 10-fold cross-validation. Additionally, the
diagram also includes the performance of three distinct
algorithms utilized for analysis. It is clear from the diagram that
10-fold cross-validation has the least error with the SVM
classifier.

The radar graph presented in this analysis showcases the

variance observed in the three algorithms: mRMR+CSA,
mRMR+HHO, and mRM-R+CSAHHO. These algorithms were
evaluated using six microarray cancer datasets. The radar
curve illustrates the variance-based performance of each
algorithm. In the graph, the magnitude of the area under the
curve indicates the algorithm’s performance level. A larger area
under the curve indicates an inferior ability to accurately detect
and classify different classes. Figure 4.4 in the graph focuses on
the variance achieved using the SVM classifier for all three
algorithms. The mRMR+CSA algorithm is represented by the
captivating blue color, while the mRMR+HHO algorithm is
depicted in a vibrant orange shade. The mRMR+CSAHHO
algorithm stands out with its green color. Notably, the algorithm
with the largest area under the curve demonstrates less
performance.
Figure 4.3 Error comparison with the SVM classifier.

Figure 4.4 The variance observed in the proposed algorithm (mRMR+CSAHHO)

compared to the mRMR+CSA and mRMR+HHO algorithms with the SVM classifier.
4.5.2 Experimental Results of the Proposed Algorithm
With the KNN Classifier

Table 4.4 represents the performance of various machine

learning models on different datasets using 5-fold and 10-fold
cross-validation techniques. The datasets include colon, acute
leukemia, prostate tumor, high-grade glioma, lung cancer II, and
leukemia 2, while the models include the mRMR, mRMR+CSA,
mRMR+HHO, and mRMR+CSAHHO. The mean values with
standard deviations are presented in the table. The results show
that the combination of the mRMR with the CSAHHO provides
better results in most cases than using the mRMR alone or with
other models. For instance, in the colon dataset, the
combination of the mRMR with the CSAHHO achieves the
highest accuracy of 97.6% with a standard deviation of 2.3%,
while the mRMR alone achieves 74.9% with a standard
deviation of 5.31%. Similarly, in the leukemia 2 dataset, the
combination of the mRMR with the CSA achieves the highest
accuracy of 96.76% with a standard deviation of 1.59%, while
the mRMR alone achieves 76.82% with a standard deviation of
4.56%. Furthermore, the use of 10-fold cross-validation
generally results in higher mean values than using 5-fold cross-
validation. For instance, using 10-fold cross-validation, the
combination of the mRMR and the CSAHHO in the lung cancer
II dataset produces a mean accuracy of 97.2% with a standard
deviation of 1.23%, but with 5-fold cross-validation, the
accuracy is 92.53% with a standard deviation of 2.11%. The
results show that the mRMR and the CSAHHO work better
together than other models, and that using 10-fold cross-
validation typically yields higher mean values than using 5-fold
cross-validation.
Table 4.4 Accuracies of the proposed algorithm with the mRMR, mRMR+CSA, and
mRMR+HHO for the KNN classifier with 5-fold cross-validation and 10-fold cross-
validation techniques.

5-fold cross-validation

mRMR+CSA mRMR+HHO mRM

Datasets mRMR Mean(±STD) Mean(±STD) Mea

Colon 74.9 ± 95.5 ±4.90 97.6 ± 2.3 98.7

5.31

Acute 77.2 ± 83.5 ± 5.21 87.4 ± 2.02 91.0

leukemia 6.75

Prostate 77.7 ± 91.71 ± 2.09 93.7 ± 1.02 95.3

tumor 4.56

High- 78.4 ± 84.71 ± 3.99 87.81 ± 2.02 92.32

grade 5.37
glioma

Lung 77.02 ± 92.53 ± 2.11 94.01 ± 2.02 96.52

cancer II 8.09
 5-fold cross-validation

mRMR+CSA mRMR+HHO mRM

Datasets mRMR Mean(±STD) Mean(±STD) Mea

Leukemia 76.82 ± 94.54 ± 2.17 96.76 ± 1.59 97.42

2 4.56

Figure 4.5 shows a thorough comparison of the errors caused

by various analytic techniques applied to microarray cancer
datasets. The main objective is to evaluate and compare the
performance of the mRMR algorithm with two commonly used
cross-validation methods: 5-fold and 10-fold cross-validation.
Figure 4.5 also demonstrates the performance of three different
algorithms utilized in the analysis. Notably, the diagram reveals
that the approach employing 10-fold cross-validation and a
KNN classifier achieves the lowest error rate.
Figure 4.5 Error comparison with the KNN classifier.

Figure 4.6 The variance observed in the proposed algorithm (mRMR+CSAHHO)

compared to the mRMR+CSA and mRMR+HHO algorithms with the KNN classifier.
Figure 4.6 illustrates the variability of the three algorithms
utilizing the KNN classifier. The graph offers valuable insights
into the performance and disparities among these algorithms.
Specifically, the proposed algorithm using the KNN classifier
demonstrates a smaller coverage area compared to the other
algorithms, indicating less variance in the achieved results of
classification performance.

4.5.3 Experimental Results of the Proposed Algorithm

With the NB Classifier

Table 4.5 shows how various feature selection techniques and

classification algorithms performed on diverse datasets. The
datasets utilized in the study are represented by the first
column, and the effectiveness of the mRMR feature selection
approach is displayed in the second column. The performance
of various combinations of the mRMR with the cuckoo search
algorithm and Harris hawks optimization feature selection
methods is shown in the next six columns. Two popular cross-
validation methods, 5-fold and 10-fold cross-validation, are used
to assess performance. Each column contains the mean and
standard deviation for the accuracy of each method. With
regard to the colon dataset, the mRMR and CSAHHO
combination obtains the maximum accuracy, with a 5-fold
cross-validation accuracy of 93.8% and a 10-fold cross-
validation accuracy of 100%. Similar to this, using 5-fold and 10-
fold cross-validation, the combination of the mRMR and the
CSAHHO yields the greatest accuracy for the prostate tumor
dataset, with 88.4% and 96.3%, respectively. The combination of
the mRMR and the CSAHHO yields the maximum accuracy for
both 5-fold and 10-fold cross-validation for the datasets for
acute leukemia and high-grade glioma. The best accuracy is
obtained when the mRMR is combined with the HHO or CSA for
the lung cancer II and leukemia 2 datasets, respectively. Overall,
the findings imply that the mRMR and the CSAHHO can greatly
enhance the performance of classification systems.
Table 4.5 Accuracies of the proposed algorithm with the mRMR, mRMR+CSA, and
mRMR+HHO for the NB classifier with 5-fold cross-validation and 10-fold cross-
validation techniques.

5-fold cross-validation

mRMR+CSA mRMR+HHO mRM

Datasets mRMR Mean(±STD) Mean(±STD) Mea

Colon 88.4 ± 90.16 ± 2.02 93.8 ± 2.56 95.7

5.19

Acute 84.4 ± 84.16 ± 2.19 86.4 ± 6.34 89.3

leukemia 5.22

Prostate 88.3 ± 86.5 ± 1.76 88.4 ± 2.33 91 ±

tumor 3.23

High- 85.6 ± 85.31 ± 3.24 88.34 ± 3.76 90.37

grade 5.22
glioma

Lung 90.72 ± 92.33 ± 2.33 94.45 ± 4.23 95.72

cancer II 4.56
 5-fold cross-validation

mRMR+CSA mRMR+HHO mRM

Datasets mRMR Mean(±STD) Mean(±STD) Mea

Leukemia 90.38 ± 91.34 ± 2.11 94.56 ± 2.71 96.62

2 5.32

Figure 4.7 provides a comprehensive comparison of error rates

obtained from different analysis methods applied to microarray
cancer datasets. The primary emphasis is on assessing the
performance of the mRMR algorithm in comparison to two
commonly employed cross-validation techniques, 5-fold and 10-
fold cross-validation. The diagram also presents the
performance results of three distinct algorithms used in the
analysis. Notably, it is clear from the diagram that the 10-fold
cross-validation approach with the NB classifier demonstrates
the lowest error rate.

Figure 4.8 portrays the variance of all three algorithms using

the NB classifier. This depiction further highlights the distinct
performance characteristics of each algorithm and helps in
understanding its strengths and limitations. Overall, the radar
graph and the numbers that go with it provide an appealing
visual depiction of the variation found in the mRMR+CSA,
mRMR+HHO, and mRMR+CSAHHO algorithms when applied to
various classifiers. In order to enable successful detection and
classification in cancer research and diagnosis, it is essential to
have this information when choosing the best algorithm for
certain cancer datasets.

Figure 4.7 Error comparison with the NB classifier.

Figure 4.8 The variance observed in the proposed algorithm (mRMR+CSAHHO)
compared to the mRMR+CSA and mRMR+HHO algorithms with the SVM classifier.

4.5.4 Comparison of the Proposed Algorithm

Compared to Other Recently Published and Popular
Algorithms for Cancer Classification

Table 4.6 shows the accuracy comparison of the published

algorithms and the proposed algorithm. According to the
findings, the CSAHHO algorithm is the best and most efficient
method for the given problem out of the three. The CSA’s mean
performance was the lowest, indicating that it did not
successfully optimize the problem. Although the HHO algorithm
outperformed CSA, it was not as effective as the CSAHHO. It is
worth highlighting that the performance of all three methods
exhibited variations across different tests, indicating that the
performance of an algorithm can be influenced by the specific
instance of the problem being addressed. It is essential to
compare algorithms across various issue contexts in order to
get a fuller view of their performance. The standard deviation
values suggest that the performance of each algorithm varied
significantly across the different experiments. This indicates
that the performance of an algorithm may be sensitive to
changes in the problem instance or input parameters.
Therefore, it is important to carefully tune the input parameters
of an algorithm to ensure that it is able to effectively optimize
the problem.
Table 4.6 Comparison of the different published methods with the proposed method
for cancer classification.

High-
Acute Prostate
Methods Colon grade
leukemia tumor
glioma

GWO + 95.935% – – 94.33%

SVM [40]

DL + DIP 96.33% – – –
[41]

SCLB+ CLB 99.58% – 89.32% –

[42]

NN +CT – 97.63 – –
[43]

ICA + ABC 98.14% 98.68% 98.88% 94.22%

[44, 45]

miRNA – – – –
[46]
 High-
Acute Prostate
Methods Colon grade
leukemia tumor
glioma

FHTWSVM – – 90.48% –
[47]

CNN + ECA – 94.34% – 96.33%

+VGG16
[48]

Proposed 99.2% 98.7% 99% 97.34%

method

Overall, the results and comments demonstrate that the

CSAHHO approach is the most successful and efficient for the
specified task out of the three methods. To confirm the
generalizability and scalability of these algorithms to other
issues and contexts, additional research and experimentation
are necessary.

4.6 Conclusion
The study investigated the effectiveness of three optimization
methods for identifying and classifying cancer features. The
findings revealed that, for the given problem, the CSAHHO
algorithm demonstrated superior efficiency and effectiveness
compared to the other two algorithms. However, the study
highlighted the need for further experimentation and
evaluation on different cancer datasets and scenarios to verify
the generalizability and scalability of the suggested algorithm.
The findings emphasize the importance of carefully tuning
algorithm input parameters and evaluating algorithms across
multiple problem instances to obtain a comprehensive
understanding of their performance. The work gives knowledge
for more research in this area and aids in the development of
hybrid optimization algorithms for cancer feature selection and
classification.

The generalizability and scalability of the suggested approach

might be established through additional testing and evaluation
on various cancer datasets in the future directions of this study.
The employment of additional hybrid optimization algorithms
or machine learning techniques applied to the selection and
classification of cancer feature sets can offer additional
benefits. In order to enhance the precision and efficiency of the
categorization procedure, this can involve the incorporation of
deep learning models or ensemble learning techniques.
Researchers could also look into how the suggested approach
might be used in sectors other than cancer feature selection and
classification. Finally, the development of user-friendly
software tools for implementing the proposed algorithm could
make it more accessible to researchers and practitioners in the
medical community.

References

1. Tefferi, A. et al., The 2008 World Health Organization

classification system for myeloproliferative neoplasms: Order
out of chaos. Cancer, 115, 17, 3842–3847, 2009, doi:
10.1002/cncr.24440.
2. Kourou, K. et al., Machine learning applications in cancer
prognosis and prediction. Comput. Struct. Biotechnol. J., 13, 8–
17, 2015, doi: 10.1016/j. csbj.2014.11.005.
3. Nakariyakul, S., A hybrid gene selection algorithm based on
interaction information for microarray-based cancer
classification. PLoS One, 14, 2, e0212333, 2019, doi:
10.1371/journal.pone.0212333.
4. Nanglia, P. et al., A hybrid algorithm for lung cancer
classification using SVM and Neural Networks. ICT Express, 7,
3, 335–341, 2021, doi: 10.1016/j. icte.2020.06.007.
5. Sahu, B., Mohanty, S., Rout, S., A hybrid approach for breast
cancer classification and diagnosis. EAI Endorsed Trans.
Scalable Inf. Syst., 6, e2, 2018.
6. Motieghader, H. et al., A hybrid gene selection algorithm for
microarray cancer classification using genetic algorithm and
learning automata. Inform. Med. Unlocked, 9, 246–254, 2017,
doi: 10.1016/j.imu.2017.10.004.
7. Shunmugapriya, P. and Kanmani, S., A hybrid algorithm using
ant and bee colony optimization for feature selection and
classification (AC-ABC Hybrid). Swarm Evol. Comput., 36, 27–
36, 2017, doi: 10.1016/j.swevo.2017.04.002.
8. Jansi Rani, M.J. and Devaraj, D., Two-stage hybrid gene
selection using mutual information and genetic algorithm for
cancer data classification. J. Med. Syst., 43, 8, 235, 2019, doi:
10.1007/s10916-019-1372-8.
9. Elyasigomari, V. et al., Development of a two-stage gene
selection method that incorporates a novel hybrid approach
using the cuckoo optimization algorithm and harmony
search for cancer classification. J. Biomed. Inform., 67, 11–20,
2017, doi: 10.1016/j.jbi.2017.01.016.
10. Jain, I. et al., Correlation feature selection based improved-
Binary Particle Swarm Optimization for gene selection and
cancer classification. Appl. Soft Comput., 62, 203–215, 2018,
doi: 10.1016/j.asoc.2017.09.038.
11. Shahbeig, S. et al., A fuzzy multi-objective hybrid TLBO-PSO
approach to select the associated genes with breast cancer.
Signal Process., 131, 58–65, 2017, doi:
10.1016/j.sigpro.2016.07.035.
12. Lu, H. et al., A hybrid feature selection algorithm for gene
expression data classification. Neurocomputing, 256, 56–62,
2017, doi: 10.1016/j. neucom.2016.07.080.
13. Naeem, S. et al., Machine-learning based hybrid-feature
analysis for liver cancer classification using fused (MR and
CT) images. Appl. Sci., 10, 9, 2020, doi: 10.3390/app10093134.
14. Kharrat, A. and Neji, M., Feature selection based on hybrid
optimization for magnetic resonance imaging brain tumor
classification and segmentation. Appl. Med. Inform., 41, 9–23,
2019, Available at:
https://ami.info.umfcluj.ro/index.php/AMI/article/view/648.
15. Zheng, Y. et al., A novel hybrid algorithm for feature
selection based on whale optimization algorithm. IEEE
Access, 7, 14908–14923, 2019, doi: 10.1109/
ACCESS.2018.2879848.
16. Stephan, P. et al., A hybrid artificial bee colony with whale
optimization algorithm for improved breast cancer
diagnosis. Neural Comput. Appl., 33, 20, 13667–13691, 2021,
doi: 10.1007/s00521-021-05997-6.
17. Khamparia, A. et al., Diagnosis of breast cancer based on
modern mammography using hybrid transfer learning.
Multidimens. Syst. Signal Process., 32, 2, 747–765, 2021, doi:
10.1007/s11045-020-00756-7.
18. Shukla, A.K. et al., A two-stage gene selection method for
biomarker discovery from microarray data for cancer
classification. Chemom. Intell. Lab. Syst., 183, 47–58, 2018, doi:
10.1016/j.chemolab.2018.10.009.
19. Aziz, R.M., Application of nature inspired soft computing
techniques for gene selection: A novel frame work for
classification of cancer. Soft Comput., 26, 22, 12179–12196,
2022, doi: 10.1007/s00500-022-07032-9.
20. Kilicarslan, S. et al., Diagnosis and classification of cancer
using hybrid model based on ReliefF and convolutional
neural network. Med. Hypotheses, 137, 109577, 2020, doi:
10.1016/j.mehy.2020.109577.
21. Melekoodappattu, J.G. and Subbian, P.S., Automated breast
cancer detection using hybrid extreme learning machine
classifier. J. Ambient Intell. Hum. Comput., 14, 5489–5498,
2023, https://doi.org/10.1007/s12652-020-02359-3.
22. Keerthana, D. et al., Hybrid convolutional neural networks
with SVM classifier for classification of skin cancer. Biomed.
Eng. Adv., 5, 2023, doi: 10.1016/j. bea.2022.100069.
23. Dara, S. and Tumma, P., Feature extraction by using deep
learning: A survey. Second Int. Conf. Electron. Commun.
Aerosp. Technol., vol. 2018, 2018, doi:
10.1109/ICECA.2018.8474912.
24. Fister, I. et al., Cuckoo search: A brief literature review. Stud.
Comput. Intell., 516, 49–62, 2014, doi: 10.1007/978-3-319-
02141-6_3.
25. Kalita, K. et al., Optimizing frequencies of skew composite
laminates with metaheuristic algorithms. Eng. Comput., 36, 2,
741–761, 2020, doi: 10.1007/ s00366-019-00728-x.
26. Kalita, K. et al., A comparative study on the metaheuristic-
based optimization of skew composite laminates. Eng.
Comput. Springer, 38, 4, 3549–3566, 2022, doi: 10.1007/s00366-
021-01401-y.
27. Alabool, H.M. et al., Harris hawks optimization: A
comprehensive review of recent variants and applications,
in: Neural Comput. & Applic., vol. 33, no. 15, pp. 8939–8980,
Springer, London, 2021, doi:10.1007/s00521-021-05720-5.
28. Heidari, A.A. et al., Harris hawks optimization: Algorithm
and applications. Future Gener. Comput. Syst., 97, 849–872,
2019, doi: 10.1016/j. future.2019.02.028.
29. Shukla, S. and Naganna, S., A review on K-means data
clustering approach. Int. J. Inf. Comput. Technol., 4, 17, 1847–
1860, 2014.
30. Priyadarshini, J. et al., Analyzing physics-inspired
metaheuristic algorithms in feature selection with K-nearest-
neighbor. Appl. Sci., 13, 2, 906, 2023, doi:
10.3390/app13020906.
31. Ganesh, N. et al., Efficient feature selection using weighted
superposition attraction optimization algorithm. Appl. Sci.,
13, 5, 3223, 2023, doi: 10.3390/ app13053223.
32. Reddy, G.T. et al., Analysis of dimensionality reduction
techniques on big data. IEEE Access, 8, 54776–54788, 2020,
doi: 10.1109/ACCESS.2020.2980942.
33. Bhattacharya, S. et al., A comparative analysis on prediction
performance of regression models during machining of
composite materials. Mater. (Basel), 14, 21, 6689, 2021, doi:
10.3390/ma14216689.
34. Gayathri, R. et al., A comparative analysis of machine
learning models in prediction of mortar compressive
strength. Processes, 10, 7, 1387, 2022, doi:
10.3390/pr10071387.
35. Narayanan, G. et al., PSO-tuned support vector machine
metamodels for assessment of turbulent flows in pipe bends.
Eng. Comput., 37, 3, 981–1001, 2019, doi: 10.1108/EC-05-2019-
0244.
36. Aziz, R.M., Application of nature inspired soft computing
techniques for gene selection: A novel frame work for
 classification of cancer. Soft Comput., 26, 22, 12179–12196,
2022, doi: 10.1007/s00500-022-07032-9.
37. Lv, J. et al., A multi-objective heuristic algorithm for gene
expression microarray data classification. Expert Syst. Appl.,
59, 13–19, 2016, doi: 10.1016/j. eswa.2016.04.020.
38. Alshamlan, H. et al., mRMR-ABC: A hybrid gene selection
algorithm for cancer classification using microarray gene
expression profiling. Biomed. Res. Int., 2015, 604910, 2015,
doi: 10.1155/2015/604910.
39. Mohamed, N.S. et al., Metaheuristic approach for an
enhanced mRMR filter method for classification using drug
response microarray data. Expert Syst. Appl., 90, 224–231,
2017, doi: 10.1016/j.eswa.2017.08.026.
40. AbdElNabi, M.L.R. et al., Breast and colon cancer
classification from gene expression profiles using data
mining techniques. Symmetry, 12, 3, 408, 2020, doi:
10.3390/sym12030408.
41. Masud, M. et al., A machine learning approach to diagnosing
lung and colon cancer using a deep learning-based
classification framework. Sensors (Basel), 21, 3, 1–21, 2021,
doi: 10.3390/s21030748.
42. Ali, M. and Ali, R., Multi-input dual-stream capsule network
for improved lung and colon cancer classification.
 Diagnostics (Basel), 11, 8, 1–18, 2021, doi:
10.3390/diagnostics11081485.
43. Kuruvilla, J. and Gunavathi, K., Lung cancer classification
using neural networks for CT images. Comput. Methods
Programs Biomed., 113, 1, 202–209, 2014, doi:
10.1016/j.cmpb.2013.10.011.
44. Aziz, R. et al., A fuzzy based feature selection from
independent component subspace for machine learning
classification of microarray data. Genomics Data, 8, 4–15,
2016, doi: 10.1016/j.gdata.2016.02.012.
45. Aziz, R. et al., A novel approach for dimension reduction of
microarray. Comput. Biol. Chem., 71, 161–169, 2017, doi:
10.1016/j.compbiolchem.2017. 10.009.
46. Yanaihara, N. et al., Unique microRNA molecular profiles in
lung cancer diagnosis and prognosis. Cancer Cell, 9, 3, 189–
198, 2006, doi: 10.1016/j. ccr.2006.01.025.
47. Hua, D. et al., Tumor classification of gene expression data
by fuzzy hybrid twin SVM. Chin. J. Electron., 31, 1, 99–106,
2022.
48. Zakir Ullah, M. et al., An attention-based convolutional
neural network for acute lymphoblastic leukemia
classification. Appl. Sci., 11, 22, 10662, 2021, doi:
10.3390/app112210662.
49. Peng, H., Long, F., Ding, C., Feature selection based on
mutual information criteria of max-dependency, max-
relevance, and min-redundancy. IEEE Trans. Pattern Anal.
Mach. Intell., 27, 8, 1226–1238, 2005.

Note

1. *Corresponding author: [email protected]

5
Anomaly Identification in
Surveillance Video Using
Regressive Bidirectional LSTM
with Hyperparameter
Optimization
Rajendran Shankar1* and Narayanan Ganesh2

1
Department of Computer Science and Engineering, Koneru
Lakshmaiah Education Foundation, Vaddeswaram, India

2
School of Computer Science & Engineering, Vellore Institute of
Technology, Chennai, India

Abstract

Urban planners and academics are influenced by the idea of

smart cities to develop sustainable, modern, and reliable
infrastructure that offers their citizens a respectable standard
of living. To meet this demand, there have been video
monitoring devices installed to improve public security and
welfare. Despite scientific advancements, it is difficult and
labor-intensive to identify odd events in surveillance video
systems. In this research, we concentrate on the improvement
of anomaly detection in intelligent video surveillance using
regressive bidirectional Long Short-Term Memory (LSTM)
(RBLSTM) with hyperparameter optimization (HO). The
suggested framework is tested on a real-time dataset, the
ShanghaiTech Campus dataset, and it outperforms state-of-the-
art techniques in terms of performance. It is important to take
advantage of higher-quality features from accessible videos.
This work uses the Video Swin Transformer model to extract
features. As a consequence, anomaly detection in video
surveillance applications provides reliable outcomes for real-
time situations. In this study, an abnormality was correctly
identified in videos with a 98.5% accuracy rate. Future research
might explore further experiments by investigating other
methods to reduce the noise present in the positive bag.

Keywords: Regressive bidirectional LSTM (RBLSTM),

hyperparameter optimization (HO), video anomaly detection,
Video Swin, ShanghaiTech

5.1 Introduction
The use of video anomaly detection in surveillance systems has
received a lot of interest. Even if the amount of installing
monitoring devices has decreased dramatically in recent years,
aberrant activities like fighting, abuse, and theft still need to be
detected by humans. The development of efficient video
anomaly detection algorithms is necessary due to the increased
expense of human labor and the loss of productive time [1].
Anomaly identification is challenging because anomalies are ill-
defined and annotated data are not readily available. Protection
of ourselves and one’s property is more vital nowadays. The
real-time video surveillance serves a beneficial purpose. As a
result of these demands, cameras are installed at every turn,
and video surveillance systems can comprehend the situation
and recognize unusual activity right away [2]. The primary
component recognizes the activity and automatically alerts the
operator or users when an unexpected occurrence occurs.
When it comes to managing personal safety and security as well
as public safety, video surveillance works more effectively than
conventional approaches. To reduce the burden of an observer,
as well as build an autonomous surveillance device to replace
observer-oriented services provided by humans. Cameras are
used to gather information on various events that indicate the
behavior of anomalies in a surveillance-enforced environment
during the anomaly identification process. The system’s
functionality conducts feature extraction on the gathered data,
processes it, and then converts the generated features into
various inputs for the designated algorithm [3]. Contextual,
local, and communal anomalies are all possible. Data instances
that are examined against a specific context connected with the
data instance are referred to as contextual anomalies. Single
data examples that vary from others are referred to as point
anomalies. Intelligent video surveillance systems detect
suspicious activity and sound alarms without human
involvement [4]. Data are gathered by visual sensors in the
surveillance environment throughout the detecting phase. The
next step is feature extraction and preprocessing of these raw
visual data. Video surveillance anomaly detection is a delay-
sensitive application that needs minimal latency [5].

In this study, we aim to provide readers with a comprehensive

overview of various anomaly detection strategies and
viewpoints. We will also discuss how different surveillance
targets are selected based on the usage of the device. Previous
evaluations of anomaly detection methods have varied in terms
of surveillance targets, and we will consider recent methods for
anomaly detection. A key contribution of this study is the
introduction of regressive bidirectional LSTM (RBLSTM) with
hyperparameter optimization (HO) for anomaly detection in
intelligent video surveillance. Additionally, we will address the
challenges and prospects of anomaly detection in video
surveillance. Notably, this study is the first to discuss how
anomaly detection in video surveillance has converged.

The following are the scientific contributions of this research

article:

1. The research that is being presented attempts to establish

connections between the various issue formulations and
offered solutions for anomaly identification.
2. It provides a comprehensive classification of anomaly
detection algorithms based on the aims and techniques of
monitoring.
3. Anomaly detection methods are discussed in the application
domain, surveillance targets, learning strategies, and
modeling approaches.
4. A more recent transformer-based feature extraction
approach called Video Swin Transformer is utilized to
enhance comprehension of the videos.
5. The suggested model’s impacts on the open-source
ShanghaiTech dataset are investigated via a comparison to
existing state-of-the-art methodologies.

The remaining sections of the article are organized as follows:

Section 5.2 reviews the literature on anomaly detection, while
Section 5.3 introduces the proposed approach. In Section 5.4, we
present the results of our experimental analysis, and Section 5.5
concludes the article with our observations.

5.2 Literature Survey

To identify actions in sporting events, researchers developed an

attention-based LSTM. To enhance the spatial characteristics,
convolution block attention was used [6]. To find abnormalities
in complicated situations, they suggested an ensemble of deep
models. A human position estimate model that recognizes the
human joints is integrated into the initial stage. The second
phase involves treating the discovered joints as features and
providing them with a densely linked complete CNN for
anomaly detection [7]. They introduced an unsupervised
method for finding anomalies in films utilizing time-varying
sparse coding, online query signals, and sparse reconstruction
skills developed from a learned event-specific terminology. It
has remained difficult to acquire the capacity to spot an
anomaly promptly, which has drawn the attention of various
academics [8]. The researchers created a deep auto-encoder
based on clustering to effectively extract data from routine
occurrences. To learn spatial–temporal feature regularity, two
modules are employed, with the first module’s spatial auto-
encoder taking care of the last individual video frame. In
contrast, the second module’s temporal auto-encoder works and
generates the RGB difference between the frames. Additionally,
abnormalities in videos are found using generative models [9].
The researchers suggested a tube extraction method that makes
use of coordinates to create an abnormality regression model.
Before transferring the data to the regression model, the
medium pooling layer mixes the temporal aspects of the ideal
flow with the spatial variables from the inception block [10].
The researchers created a method for spotting irregularities
with limited supervision and a supervised system for action
classification with noisy labels. Because the anomalous
occurrences were unexpected, the novelty consisted of merely
keeping the anomaly video labels loud. In addition, a graph
CNN was trained to clean up these noisy labels, and an action
classifier was used to categorize the actions [11].

Following this context, the researchers proposed the creation of

anomaly detection in intelligent video surveillance using
RBLSTM with HO, which covers both short- and long-range
temporal dependencies and requires less computational power
than alternative approaches. The suggested approach enhances
the performance of the model overall by extracting features
using a more efficient transformer-based model.
5.3 Proposed Methodology

The suggested approach assumes that although clips taken from

anomaly videos would contain at least one anomaly, clips taken
from regular videos would only contain normal clips. A
normalization procedure is used for preprocessing in stage 1.
Stage 2 of the Video Swin model uses the extracted features
from the clips. Stage 3 involves applying RBLSTM and HO to the
characteristics to cover relevant long- and short-range temporal
relationships. The final step involves performance analysis by
employing optimization to find anomalies. Figure 5.1 depicts
the suggested model in this article.

Figure 5.1 Schematic architecture of our proposed system.

5.3.1 Dataset
The ShanghaiTech Campus dataset [12] is a large-scale video
anomaly detection dataset used in this article. It comprises
videos captured by street surveillance cameras with a fixed
perspective. The dataset includes 337 videos from 13 different
settings, of which 80 are abnormal and 257 are normal. This
dataset is a standard benchmark used in anomaly detection
tasks and contains data that are both anomalous and typical in
nature.

Figure 5.2 Normal and abnormal clips from the ShanghaiTech dataset.

Figure 5.2 displays several samples of normal and abnormal

clips taken from the dataset. The dataset was reorganized into a
weakly supervised training set by selecting examples of
abnormalities testing films and converting them into training
images. This ensured that each of the 14 background clips
would be included in both the testing set and the training set.

5.3.2 Preprocessing Using Normalization

The identification of anomalies is crucial in an intelligent

transportation system. Data undergo early processing to
prepare information for either primary processing or further
analysis. When several procedures are required to prepare data
for the user during the first or preliminary processing phase,
the term “data normalization” may be employed. Data
normalization refers to the process of improving,
standardizing, and preserving the integrity of data. The process
reduces the need for storage space by removing any redundant
or duplicate data from the database. Normalization can be a
scaling technique, a mapping approach, or a preprocessing step.
It can be useful for making forecasts or predictions whenever a
new range can be inferred from an old range. As there are
various ways to anticipate or make predictions, we took
measures to standardize the data during preprocessing.
Equation 5.1 is applied to the data as part of the normalization
procedure.
 (5.1)

where Pnor = Normalized value

Pmax = Maximum value and

Pmin = Minimum value.

5.3.3 Feature Extraction Using Video Swin

Transformer

The Video Swin Transformer model is utilized in this work to

extract features. When using a pretrained model, we can
extract higher-quality features. In the case of images, standard
transformer models compute self-attention for all components,
but this operation can be computationally expensive. The Swin
Transformer addresses this issue by dividing images into
windows and computing self-attention only within the current
window. The window is then shifted across the images to obtain
the self-attention value for all image collections in a more time-
efficient manner.

To begin the process of feature extraction, the videos must first

be split into individual frames (of size C × N, for example). The
collection of these frames, denoted by the letter S, is the video
that will be used for feature extraction, with each image
containing RGB channels. This provides us with the following
formula for determining the input dimension: M × H × S × C × N,
where M is the batch size and H is the number of channels.

5.3.4 Anomaly Detection

The suggested RBLSTM with HO model is used for anomaly

detection. In this model, normal video clips are represented by
minimum-magnitude features, while anomaly video clips are
represented by maximum-magnitude features. The suggested
method is based on the assumption that abnormal clips have a
higher mean feature magnitude than typical clips.

5.3.4.1 Regressive Bidirectional LSTM

The suggested approach carries out the task of predicting a

dependent variable (target) based on the provided independent
variable(s). Therefore, this regressive approach identifies a
linear connection between a dependent variable and other
provided independent variables. Unlike a standard LSTM, an
RBLSTM’s output relies on both the next and preceding frames
in the sequence. For this study, we utilized an RBLSTM with two
cells per layer, one for the ahead pass and one for the reverse
pass. The RBLSTM receives the characteristics extracted by
Video Swin in the form of chunks to categorize the condition
supplied as either normal or abnormal for an abnormality
detection conclusion. The input to the RBLSTM at a time t is
composed of the first section of the thousand features from the
first frame of the video, followed by the second section of the
following feature at a time t+1, and so on. Algorithm 1
represents the proposed RBLSTM approach.
ALGORITHM 1: RBLSTM

Procedure anomaly_dectector (data)

Model, detector<-CREATE ()

Nn, Vn, Tn, < − SPLIT (data)

For n epochs and batch_size do

Model train (Nn, Vn,)

End for

N_predictionn <-model. predict (Nn)

N_errorn<-compute the loss function

Thresholds<-detector.SET_THERSHOLD as given in Equation

(5.2)

(5.2)

Detector.APPLY (thresholds)

For every Tn do

T_predictionn <-model. predict (Tn)

T_eroorn<-compute the loss function

If (T_eroorn<(thresholds) then

Tn is a normal point

Else

An<-Tn is anomalous

Return [Tn]

Stop for

Stop procedure

5.3.4.2 Hyperparameter Optimization

Optimizing hyperparameters to enhance algorithm

performance is crucial, since almost all machine learning (ML)
methods include them [13, 14]. With the help of these
hyperparameters, performance is optimized, requiring less
work from the user and leading to a quicker, easier solution.
The success of machine learning strongly depends on humans
choosing the right hyperparameters. This is a difficult and time-
consuming operation that becomes much more crucial when it
must be performed by ML nonexperts rather than
professionals, which occurs fairly often. The process of finding
the ideal combination of hyperparameter values to obtain the
best performance on the data in the shortest possible time is
called hyperparameter optimization. It may provide you with
hyperparameter optimized settings that increase the predicted
accuracy of your model [15]. The hyperparameter optimization
process is represented by Algorithm 2, which may be used to
make predictions once the model has been trained. A test point
must have all of its hash values generated before it can be
checked. The precomputed counts in the model, which indicate
the characteristics of the training data distribution, are used to
build an estimate.
ALGORITHM 2: HYPERPARAMETER OPTIMIZATION

Input: p← Test point,

hasher← Hasher of the trained model,

estPerHash← Estimator dictionary,

threshold← Estimator threshold, Output: Boolean value

indicating whether p is an anomaly hashes

←hasher.HASH(p);

estimator ←0;

For every h ϵ hashes do

estimator←estimator + estPerHash [h];

Stop

Output: estimator < threshold

5.4 Result and Discussion

The dataset from ShanghaiTech Campus is utilized in this study

for analysis. Video Swin serves as the backbone model for
feature extraction. Comparison and result evaluation provide a
full assessment of the proposed RBLSTM with HO technique.
When compared to other methodologies such as the graph
convolutional network (GCN) [16], margin learning embedded
prediction (MLEP) [17], and 3-dimensional convolutional
networks (3D ConvNets) [18], the suggested methodology proves
powerful and effective in identifying anomalies in video
surveillance. Accuracy, precision, recall, and error rate are
performance metrics.

Equation 5.3 used to calculate accuracy is defined as the

proportion of all forecasts to the corrected predictions.

(5.3)

where tn is true negative, tp is true positive, fn is false negative,

and fp is false positive.
Figure 5.3 Accuracy comparison between the suggested and current techniques.

Figure 5.3 displays the accuracy of the proposed system, along

with the accuracy prediction of existing systems. The suggested
method is denoted as well. The GCN has achieved 74%, the
MLEP has acquired 66%, and the 3D ConvNets has reached 85%.
In contrast, the suggested method achieves 98.5% accuracy. This
demonstrates that the proposed approach is more effective than
the existing ones.

Precision can be calculated using Equation 5.4 below.

(5.4)
Figure 5.4 illustrates the precision of the proposed system as
well as the precision of current systems’ consumption
predictions. The GCN achieves 60%, the MLEP achieves 71%, the
3D ConvNets achieves 81%, and the suggested system achieves
98% precision. This demonstrates that the proposed approach is
more effective than the current ones.

As seen here in Equation 5.5, we can calculate the recall as

follows:

(5.5)

Figure 5.4 Precision comparison between the suggested and current techniques.
Figure 5.5 displays the recall of the proposed system as well as
the recall of the consumption estimates from existing systems.
The suggested system achieves 90% recall, while the GCN
achieves 74%, the MLEP achieves 83%, and the 3D ConvNets
achieves 62%. This proves that the suggested technique is more
efficient than the current ones.

The error rate is the ratio of invalid data units to all other data
units delivered.

Figure 5.5 Recall comparison between the suggested and current techniques.
Figure 5.6 Error rate comparison between the suggested and current techniques.

(5.6)

Figure 5.6 illustrates the error rate of the proposed system, the
prediction of consumption of error rate in existing systems, and
the suggested technique. The GCN has attained 0.15%, the MLEP
has acquired 0.14%, the 3D ConvNets has reached 0.12%, and
the suggested technique has an error rate of 0.07%. This shows
that the proposed approach has a lower error rate than those of
the existing ones.

5.5 Conclusion
Due to several variables influencing outcomes, such as video
noise, outliers, and resolution, the detection of anomalies in
video surveillance is a tough task. In this study, we suggested an
RBLSTM with HO to improve anomaly detection in intelligent
video surveillance. It makes use of higher-quality features taken
from the Video Swin Transformer model. In this investigation, a
video of an anomaly was accurately detected with a 98.5%
accuracy rate, 98% precision, 90% recall, and 0.07% error rate.
According to experimental results that show the importance of
anomaly detection, locating anomalies in the frame helps to
provide outstanding results throughout extensive video
surveillance, and this method is highly robust. Due to the rare
incidence of anomalous occurrences, inconsistent behavior of
various types of anomalies, and an uneven availability of data
for normal and abnormal settings, anomaly identification in
videos is a difficult challenge. In the future, further tests may be
conducted by experimenting with various methods to reduce
the noise present in the positive bag.

References

1. Franklin, R.J. et al., Anomaly detection in videos for video

surveillance applications using neural networks, in: Fourth
 International Conference on Inventive Systems and Control.
(ICISC), 2020, doi: 10.1109/ICISC47916.2020.9171212.
2. Zhu, S. et al., Video anomaly detection for smart surveillance,
in: Computer Vision: A Reference Guide, pp. 1–8, Springer
International Publishing, Cham, 2020, doi: 10.1007/978-3-030-
03243-2_845-1.
3. Nasaruddin, N. et al., Deep anomaly detection through visual
attention in surveillance videos. J. Big Data, 7, 1, 1–17, 2020,
doi: 10.1186/ s40537-020-00365-y.
4. Tang, Y. et al., Integrating prediction and reconstruction for
anomaly detection. Pattern Recognit. Lett., 129, 123–130, 2020,
doi: 10.1016/j. patrec.2019.11.024.
5. Pranav, M. and Zhenggang, L., A day on campus-an anomaly
detection dataset for events in a single camera, in: Proc.
Asian Conference on Computer Vision, 2020.
6. Ullah, M. et al., Attention-based LSTM network for action
recognition in sports. J. Electron. Imaging, 2021, 302–301,
2021, doi: 10.2352/ ISSN.2470-1173.2021.6.IRIACV-302.
7. Riaz, H. et al., Anomalous human action detection using a
cascade of deep learning models, in: 9th European Workshop
on Visual Information Processing (EUVIP), 2021, doi:
10.1109/EUVIP50544.2021.9484062.
8. Wu, C., Shao, S., Tunc, C., Satam, P., Hariri, S., An explainable
and efficient deep learning framework for video anomaly
 detection. Clust. Comput., 25, 2715–2737, 2021.
9. Chang, Y. et al., Clustering-driven deep autoencoder for video
anomaly detection, in: Computer Vision-ECCV, 16th European
Conference, Proceedings, Part XV, pp. 329–345, 2020, doi:
10.1007/978-3-030-58555-6_20.
10. Landi, F. et al., Anomaly locality in video surveillance, arXiv
preprint arXiv:1901.10364, 2019.
11. Madan, N. et al., Temporal cues from socially unacceptable
trajectories for anomaly detection, in: Proc. IEEE/CVF
International Conference on Computer Vision, pp. 2150–2158,
2021, doi: 10.1109/ICCVW54120.2021.00244.
12. Deshpande, K. et al., Anomaly detection in surveillance
videos using transformer-based attention model, 2022, arXiv
Preprint ArXiv:2206.01524, doi: 10.1007/978-981-99-1648-
1_17.
13. Ganesh, N. et al., Efficient feature selection using weighted
superposition attraction optimization algorithm. Appl. Sci.,
13, 5, 3223, 2023, doi: 10.3390/ app13053223.
14. Priyadarshini, J. et al., Analyzing physics-inspired
metaheuristic algorithms in feature selection with K-nearest-
neighbor. Appl. Sci., 13, 2, 906, 2023, doi:
10.3390/app13020906.
15. Shaik, K. et al., Big data analytics framework using squirrel
search optimized gradient boosted decision tree for heart
 disease diagnosis. Appl. Sci., 13, 9, 5236, 2023, doi:
10.3390/app13095236.
16. Zhong, J.-X. et al., Graph convolutional label noise cleaner:
Train a plug-and-play action classifier for anomaly detection,
in: Proc. IEEE/CVF Conference on Computer Vision and Pattern
Recognition, pp. 1237–1246, 2019, doi:
10.1109/CVPR.2019.00133.
17. Liu, W. et al., Margin learning embedded prediction for
video anomaly detection with A few anomalies, in: IJCAI, pp.
3023–3030, 2019, doi: 10.24963/ ijcai.2019/419.
18. Maqsood, R. et al., Anomaly recognition from surveillance
videos using 3D convolution neural network. Multimedia
Tools Appl., 80, 12, 18693–18716, 2021, doi: 10.1007/s11042-
021-10570-3.

Note

1. *Corresponding author: [email protected]

6
Ensemble Machine Learning-
Based Botnet Attack Detection for
IoT Applications
Suchithra M.

Department of Computing Technologies, School of Computing,

SRM Institute of Science and Technology, Kattankulathur,
Chennai, India

Abstract

The Internet of Things (IoT) systems are currently the subject of

a rising prevalence of attacks as a result of their fast
development and widespread adoption. It has been claimed
that attacks using botnets account for the majority of those in
IoT networks. Because the majority of IoT systems possess the
memory and processing power necessary for effective security
methods, there are still several security vulnerabilities in these
systems. Additionally, attackers are capable of circumventing
many of the present rule-based detection algorithms. Even
though the machine learning (ML)-driven method can identify
the variations of the various types of attacks, periodically
special new types of attacks can be initiated. As a result, this
paper suggests a novel ML technique termed random tree-
adaptive artificial neural network (RT-ANN) in detecting botnet
attacks. Initially, we collect the raw data samples and these
samples are normalized. Principal component analysis (PCA) is
used to retrieve key aspects of the RT-ANN. Our proposed RT-
ANN method’s performance is analyzed, and comparative
analysis is also provided. The experimental results demonstrate
that the proposed method achieves the highest detection rate in
the botnet attack detection.

Keywords: Internet of Things (IoT), machine learning (ML),

botnet attacks, principal component analysis (PCA), random
tree-adaptive artificial neural network (RT-ANN)

6.1 Introduction

The fast development of Internet of Things (IoT) devices in

recent years has improved the convenience of our everyday
routines. IoT devices averaged one assault every 2 minutes [1].
The IoT collects, processes, and analyzes vast volumes of data
from a variety of heterogeneous IoT devices to seamlessly link
millions of smart objects and devices. The IoT is one of the top
10 key technology developments identified by Gartner in 2020.
According to this forecast, in 2023, there will be 20 times as
many smart devices as traditional IT equipment. It is
anticipated that the IoT will be used more widely across a
variety of industries, including utilities, healthcare,
government, physical security, and automobiles. However, the
Internet of Things infrastructure is being built quickly at the
expense of more assaults and security vulnerabilities.
According to Symantec, an IoT device gets attacked every 2
minutes [2]. The IoT is one of the networking industry’s newest
concepts. It may be characterized as the interconnection of
objects with limited computing resources. It allows for the
transmission and reception of data via the Internet without the
need for computer-to-computer or person-to-person
communication. The networked IP-enabled objects are referred
to as things (both physical and virtual). Along with the regular
IoT, it is important to discuss resource-constrained IoT devices.
Although these gadgets use IoT applications, they are battery-
operated, compact, low-power gadgets with a variety of design
trade-offs. They also have restricted computational and storage
capacities. Simply put, resource-constrained devices are sensors
on endpoints that are utilized for a certain application. These
gadgets have limited storage and processing power, restricted
energy owing to energy-vulnerable batteries, and other
resource limitations. Additionally, they are linked via weak
radio circumstances, low-power lossy communications, and
without human involvement [3]. The proliferation of IoT
devices in recent years has brought forth new dangers like
botnet assaults. Such assaults may be organized and seem to
compromise the target devices. A botnet refers to an
unauthorized remote control of a host. Attackers are in charge
of hacked IoT devices, using them to carry out nefarious deeds.
Botnets may cause significant harm in the real world, as shown
in the 2016 Mirai attack that impacted approximately 1,000
closed-circuit television (CCTV) cameras [4]. A botnet is a group
of Internet-connected computers that have been infiltrated and
are now being remotely managed by an attacker using bots, a
kind of malicious software. Malicious software is often used by
an attacker. Among many other hazards, botnets will benefit the
most from IoT security weaknesses. Botnets are networks made
up of nodes that have been infected with malware that turns
them into attacking bots that follow orders from a botmaster.
The Internet of Things provides a great setting for botnets for
two main reasons. First, the absence of security mechanisms in
IoT devices makes virus transmission and installation easier.
Second, the sheer number of linked devices that are anticipated
in the next year will enable huge attacks, providing
unprecedented quantities of susceptible resources to attackers
[5]. In this article, we propose a random tree-adaptive artificial
neural network (RT-ANN) for detecting botnet attacks.

The organizational structure of the paper can be understood by

referring to the following outline. Section 6.2 provides a
thorough analysis of relevant work. In Section 6.3, we outline
the formulation of the proposed work. Section 6.4 presents and
discusses the results. Finally, in Section 6.5, we offer the
conclusions and recommendations for further research.

6.2 Literature Survey

Sriram et al. [6] proposed a botnet detection mechanism based

on network traffic flows and deep learning (DL). The platform
uses a DL model to capture network traffic flows, convert them
into connection records, and identify attacks originating from
compromised IoT devices. Popoola et al. [7] presented the
federated deep learning (FDL) strategy for preventing data
privacy leaks in IoT edge devices by detecting zero-day botnet
attacks. This approach uses an ideal deep neural network (DNN)
architecture to classify network traffic. Al-Duwairi et al. [8]
proposed an IoT botnet DDoS attack detection and mitigation
system based on Security Information and Event Management.
Several IoT companies offer a wide range of heterogeneous IoT
solutions, each implementing their own network and
communication protocol. Hussain et al. [9] provided a
comprehensive set of characteristics to more effectively identify
botnet attacks, regardless of the underlying dataset. The
experimental findings showed that the norms and principles
feature set effectively supports machine algorithms to better
identify various forms of botnet attacks. Zagrouba and Alhajri
[10] proposed ML strategies with low-power consumption for
identifying IoT botnet attacks, utilizing random forest as an ML-
based detection method and detailing IoT typical assaults with
their countermeasures. Hikal and Elgayar [11] developed a
lightweight anomaly-based machine learning intrusion
detection system based on the use of an ensemble data
preparation step in advance. Al-Sarem et al. [12] offered a
methodology for feature selection based on aggregate mutual
information and ML techniques to improve the detection of
botnet attacks.

This research analyzed actual traffic data from nine

commercial IoT devices to analyze the N-BaIoT benchmark
dataset to identify different botnet attack types. The
experimental findings demonstrated that, when a binary
dataset was employed, the mutual information filter-based
approach produced the greatest score.
McDermott et al. [13] suggested that a bidirectional strategy
lengthens processing times and adds cost to each epoch, but
eventually turns out to be a more effective progressive model.
Alissa et al. [14] suggested machine learning techniques for
categorizing binary classes using the freely accessible dataset
UNSW-NB15. As a result, cybercriminals are always searching
for fresh ways to identify and take advantage of security gaps.
Vishwakarma and Jain [15] offered a honeypot-based
methodology that uses machine learning methods to identify
malware. Abu Khurma et al. [16] proposed a wrapper feature
selection model by hybridizing the salp swarm algorithm and
ant lion optimization. The findings indicate that by employing
these pertinent subsets of characteristics, the ant lion
optimization achieved the greatest outcomes in terms of the
metrics under study. Vinayakumar et al. [17] proposed a botnet
detection system based on a two-level deep learning framework
for semantically discriminating botnets and legitimate
behaviors at the application layer of the domain name system
services. Nõmm and Bahşi [18] created models for
unsupervised learning that are very accurate and have reduced
feature set lengths, allowing for a reduction in the amount of
CPU power needed.

6.3 Proposed System

In the following section, we discuss in detail an enhanced
model for predicting botnet attacks and explain the proposed
method of the random tree-adaptive artificial neural network.
Figure 6.1 illustrates the proposed methodology. Initially, raw
data samples are collected and normalized. The dataset is then
prepared, and data processing techniques, such as
normalization and principal component analysis (PCA), are
used to extract the key features for the RT-ANN model.

6.3.1 Dataset

The studies discussed in this section utilized the N-BaIoT

dataset, which was obtained by executing hazardous IoT
malware such as Mirai and BASHLITE. The dataset comprises a
total of 849,233 instances, including 17,935 benign and 831,298
assault instances. For the tests, the ANN was used with a variety
of common activation mechanisms. This information, collected
from IoT environments using well-known IoT devices, is
publicly accessible. The dataset includes 115 traffic
characteristics, eight separate attack classes, and two botnets
and one benign botnet. Figure 6.2 shows the distribution of
cases in the dataset.
Figure 6.1 The proposed methodology.
Figure 6.2 The dataset’s distribution.

6.3.2 Data Processing Using Normalization

Z-score normalization, also known as zero-mean normalization,

normalizes each input feature vector by calculating the mean
(M) and standard deviation (SD) of each feature across a
training dataset and dividing it by the dataset size. The average
and standard deviation for each attribute are computed, and
the transformation is necessary as per the general formula.

(6.1)

The property n has a mean of μ and a standard deviation of σ.

Before training can commence, all features of the dataset are z-
score normalized. After collecting the training data, the mean
and standard deviation of each characteristic should be saved
to use them as algorithm weights.
6.3.3 Feature Extraction Using Principal Component
Analysis

To enhance our model’s performance, we employed

dimensionality reduction and feature selection techniques to
reduce the variability while retaining the important features.
Principal component analysis is one method used to decrease
the number of dimensions. The data are transformed into a
new feature space using PCA, where the first coordinate
accounts for the majority of the variation, the second principal
component accounts for the second most variance, and so on.
The traits in this research that exhibit the most variance are
used as input by the classifiers. This process ensures that only
relevant features are selected, improving the computational
efficiency and simplicity of the machine learning models.
Moreover, by utilizing all of the variables, this method can
potentially reduce any possible overfitting.

6.3.4 Random Tree-Adaptive Artificial Neural

Network

The random tree algorithm is comparable to the random forest

approach and shares many of its advantages, as it uses some
decision trees. It can effectively manage data with high-
dimensional features without the need for feature selection,
and it has a strong classification impact and high accuracy. The
execution process can be parallelized to achieve efficiency.
Unlike the random forest method, which generates training
samples via bootstrap sampling in the integrated learning
approach, the extreme random tree strategy employs all of the
original data for each decision tree. Moreover, when the
extreme random tree divides at a node, the splitting node is
chosen at random, and the ideal splitting threshold or feature is
not used. Figure 6.3 illustrates the architecture of the ANN.

The artificial neural network is a fundamental building block of

the neural network technique. The paradigm is useful for
teaching vector-valued functions. The components, which are
also known as neurons, operate in parallel and are
interconnected. The layer is a subgroup that deals with the
components that are linked together. The hidden, output, and
input layers are essentially the layers in the neural network.
Although the corresponding weights are present in each
connecting connection, they must be adjusted during the
backward calculation stages to have the best weight values.
Figure 6.3 Architecture of the ANN.

(6.2)

(6.3)

(6.4)
Therefore, the activation function assumes a crucial role in
obtaining these ideal weight vectors. Equations 6.1–6.3 show
some well-known activations, such as the sigmoid, ReLU, and
tanh functions.

6.4 Results and Discussion

This study explores the use of ensemble ML-based botnet attack
detection for IoT applications. Raw data samples are collected
and normalized, and principal component analysis is used to
retrieve the key aspects of the RT-ANN. The performance of our
proposed RT-ANN method is analyzed, and a comparative
analysis is provided. Accuracy, recall, precision, and F-measure
metrics are used to test the system’s ability to identify botnet
attacks. The equations used to define these metrics are as
follows:

6.4.1 Accuracy

The percentage of samples for which the suggested approach

accurately predicted results is used to assess how accurate the
system is. The accuracy is calculated using Equation 6.5.

(6.5)

Figure 6.4 displays comparable values for the accuracy metrics,

and it is evident that the suggested technique can generate
better performance results than the current research
approaches. The suggested method’s accuracy of 97%
outperforms the results of existing methods, such as the CNN
[19] with a 91.6% accuracy rate, the LGBA-NN [20] at 90%, and
the PSO-NN [20] at 85.2%. When classifying the data, the
suggested technique outperforms the existing methods.
Figure 6.4 Results of accuracy.

Figure 6.5 Results of precision.

6.4.2 Precision

Precision is one of the most crucial standards for accuracy, and

it is well-defined as the ratio of correctly classified cases to all
instances of predictively positive data. The precision is
calculated using Equation 6.6.

(6.6)

Figure 6.5 displays comparable values for the precision

measures. With a precision of 98%, the suggested method
outperforms the currently used methods, such as the CNN with
a precision of 90.9%, the LGBA-NN with 85%, and the PSO-NN
with 81%. The proposed method achieved better data
categorization precision than other currently used methods.

6.4.3 Recall

Recall is the potential of a model to identify each important

sample within a data collection. It is statistically defined as the
percentage of true positives (TPs) divided by the sum of TPs and
false negatives (FNs). The recall is calculated using Equation 6.7.

(6.7)

Comparative data for the recall metrics are shown in Figure 6.6.
Recall rates for the CNN are 97.6%; for the LGBA-NN, 90%; for
the PSO-NN, 85%; and the RT-ANN, 99%.

6.4.4 F-Measure

The harmonic mean of the proposed model is computed to

merge recall and precision into a single component called the F-
measure. Equation 6.8 is used to compute the F-measure.

Figure 6.6 Results of recall.

(6.8)

In Figure 6.7, the RT-ANN scored 98% on the F-measure,

followed by the CNN (scored 94.1%), LGBA-NN (86%), and PSO-
NN (86.23%).
Figure 6.7 Results of the F-measure.

6.5 Conclusion

The IoT botnet attacks can be devastating due to various

factors, including the rapid growth of connected IoT devices
and their potential security vulnerabilities, as well as the
possibility that the affected devices may not display any signs of
danger. In this paper, the use of emerging ML techniques for
detecting botnet attacks was proposed. We collected raw data
samples and normalized them. Principal component analysis
was used to extract the key features of the RT-ANN. The
performance of our proposed RT-ANN method was analyzed,
and comparative analysis was provided. The results of the
experiments demonstrate that the suggested approach has the
highest detection rate for botnet attacks. This approach may be
capable of detecting various types of botnet attacks and other
forms of disruptive network activity. To confirm the findings,
future studies using the complete dataset will need to be
evaluated using the same methodology. Additionally, to evaluate
the performance of the algorithms with other botnet attacks,
this experiment can potentially be expanded to include more
datasets, such as the Bot-IoT dataset and the CTU-13, which are
more current.

References

1. Soe, Y.N. et al., Machine learning-based IoT-botnet attack

detection with sequential architecture. Sensors (Basel), 20, 16,
4372, 2020, doi: 10.3390/ s20164372.
2. Alkahtani, H. and Aldhyani, T.H.H., Botnet attack detection by
using the CNN-LSTM model for the Internet of Things
applications. Secur. Commun. Netw., 2021, 1–23, 2021, doi:
10.1155/2021/3806459.
3. Abu Al-Haija, Q. and Al-Dala’ien, M.A.-M., ELBA-IoT: An
ensemble learning model for botnet attack detection in IoT
networks. J. Sens. Actuator Netw., 11, 1, 18, 2022, doi:
10.3390/jsan11010018.
4. Alqahtani, M. et al., IoT botnet attack detection is based on
optimized extreme gradient boosting and feature selection.
Sensors (Basel), 20, 21, 6336, 2020, doi: 10.3390/s20216336.
5. Rethinavalli, S. and Gopinath, R., Botnet attack detection in
the Internet of Things using optimization techniques. Int. J.
Electr. Eng. Technol., 11, 412–420, 2020.
6. Sriram, S. et al., Network flow-based IoT botnet attack
detection using deep learning, in: IEEE Conference on
Computer Communications Workshops (INFOCOM WKSHPS),
IEEE Infocom, 2020, doi: 10.1109/ INFOCOMWKSHPS50562.
2020.9162668.
7. Popoola, S.I., et al., Federated deep learning for zero-day
botnet attack detection in IoT-edge devices. IEEE Internet
Things J., 9, 5, 3930–3944, 2021, doi:
10.1109/JIOT.2021.3100755.
8. Al-Duwairi, B. et al., SIEM-based detection and mitigation of
IoT-botnet DDoS attacks. Int. J. Electr. Comput. Eng., 10, 2,
2182, 2020, doi: 10.11591/ ijece.v10i2.pp2182-2191.
9. Hussain, F. et al., Towards universal features set for IoT
botnet attack detection, in: 23rd International Multitopic
Conference (INMIC), IEEE, 2020, doi:
10.1109/INMIC50486.2020.9318106.
10. Zagrouba, R. and Alhajri, R., Machine learning-based attacks
detection and countermeasures in IoT. Int. J. Commun. Netw.
 Inf. Secur., 13, 2, 158–167, 2021, doi: 10.17762/ijcnis.v13i2.4943.
11. Hikal, N.A. and Elgayar, M.M., Enhancing IoT botnets attack
detection using machine learning-IDS and ensemble data
preprocessing technique, in: The Internet Things-Appl. Future
Proc. ITAF, vol. 2020, 2019, doi: 10.1007/978-981-15-3075-3_6.
12. Al-Sarem, M. et al., An aggregated mutual information-based
feature selection with machine learning methods for
enhancing IoT botnet attack detection. Sensors (Basel), 22, 1,
185, 2021, doi: 10.3390/s22010185.
13. McDermott, C.D. et al., Botnet detection in the Internet of
Things using deep learning approaches, in: International
Joint Conference on Neural Networks (IJCNN), vol. 2018, 2018,
doi: 10.1109/IJCNN.2018.8489489.
14. Alissa, K. et al., Botnet attack detection in IoT using machine
learning. Comput. Intell. Neurosci., 2022, 4515642, 2022, doi:
10.1155/2022/4515642.
15. Vishwakarma, R. and Jain, A.K., A honeypot with a machine
learning-based detection framework for defending IoT-based
botnet DDoS attacks, in: 3rd International Conference on
Trends in Electronics and Informatics (ICOEI), vol. 2019, 2019,
doi: 10.1109/ICOEI.2019.8862720.
16. Abu Khurma, R. et al., IoT botnet detection using salp swarm
and antlion hybrid optimization model. Symmetry, 13, 8,
1377, 2021, doi: 10.3390/ sym13081377.
17. Vinayakumar, R. et al., A visualized botnet detection system
based on deep learning for the Internet of Things networks
of smart cities. IEEE Trans. Ind. Appl., 56, 4, 4436–4456, 2020,
doi: 10.1109/TIA.2020.2971952.
18. Nõmm, S. and Bahşi, H., Unsupervised anomaly-based
botnet detection in IoT networks. 17th IEEE International
Conference on Machine Learning and Applications (ICMLA),
vol. 2018, 2018, doi: 10.1109/ICMLA.2018.00171.
19. Nguyen, H.T. et al., IoT botnet detection approach based on
PSI graph and DGCNN classifier, in: IEEE International
Conference on Information Communication and Signal
Processing (ICICSP), vol. 2018, 2018, doi: 10.1109/
ICICSP.2018.8549713.
20. Alharbi, A. et al., Botnet attack detection using local global
best bat algorithm for the industrial internet of things.
Electronics, 10, 11, 1341, 2021, doi:
10.3390/electronics10111341.

Note

1. Email: [email protected]
7
Machine Learning-Based
Intrusion Detection System with
Tuned Spider Monkey
Optimization for Wireless Sensor
Networks
Ilavendhan Anandaraj1* and Kaviarasan Ramu2

1
School of Computer Science and Engineering, Vellore Institute of
Technology, Chennai, India

2
Department of Computer Science and Engineering, RGM College
of Engineering and Technology, Nandyal, India

Abstract

An intrusion detection system is used to look into the problem

that occurs in a wireless sensor network (WSN) or system.
Intrusion detection is software or a piece of hardware that
looks for suspicious activity on a system or network. As
computers become more connected to each other, intrusion
detection becomes more important for network security.
Different types of intrusion detection systems (IDSs) have been
made to protect wireless sensor networks with the help of
machine learning and statistical methods. Accuracy is the most
important factor in how well an IDS works. The accuracy of
intrusion detection needs to be improved to cut down on false
alarms and raise the rate of detection. The main job of an IDS is
to look at huge amounts of network traffic data. To solve this
problem, a well-organized way to classify intrusion is needed.
This issue is taken into account in this research. In this article,
we suggest machine learning techniques such as the support
vector machine (SVM) with the tuned spider monkey
optimization (TSMO) for wireless sensor networks. The NSL-
KDD knowledge discovery dataset is used to measure how well
an intrusion detection system works. The results show that the
suggested model works better than the methods that are
already in use. The suggested model is assessed on a number of
performance criteria, and it is compared to other models to
figure out how well it works overall.

Keywords: Intrusion detection, support vector machine (SVM),

spider monkey optimization (SMO), wireless sensor networks,
machine learning
7.1 Introduction

The fundamental structure of a wireless sensor network (WSN)

consists of sensor nodes. These nodes are often dispersed
thickly to facilitate data processing, data sensing, transmission,
and computation. WSNs consist of a variety of tasks, operations,
and capacities. WSNs may be deployed in parks, highways,
battlefields, certain machines, commercial structures, and even
on the human body [1]. Distributed, infrastructure-free, failure,
extensible, and dynamical are all characteristics of wireless
sensor networks. Multiple security flaws exist in WSNs, which
may reduce their efficiency. Despite providing secure data
transmission through access control, strong authentication, and
secure network algorithms, these methods are unable to
guarantee the timely delivery of messages. In plenty of other
words, these safeguards function to prevent malicious outsiders
from accessing the network but fall short when faced with
malicious insiders. Data privacy, identification, and data
integrity are all goals. These methods protect sensitive data
from prying eyes during an external assault. When an insider
attacks a sensor system, a node inside the network begins
acting maliciously without first attempting to get access to the
data in the messages it has received. Computer and wireless
sensor network intrusions may be identified with the help of an
intrusion detection system (IDS). There are a few main methods
to categorize IDSs, the most prevalent of which are misuse-
based and anomaly-based. Misuse-based intrusion detection
systems are effective at spotting common attacks like a snort.
When compared to other IDSs, this one generates less false
positives. As it does not modify any of the predefined
instructions in its database, it cannot identify novel threats.
Anomaly-based intrusion detection systems first create a
baseline of normal behavior, then identify and flag any
noteworthy departures from that baseline as potential security
threats. There is a significant false alarm rate with this IDS;
however, it can identify both known and unknown assaults.
False positives are reduced by the use of a number of machine
learning (ML) methods [2].

7.1.1 Intrusion Detection System

The term intrusion poses one of the most significant threats to

online security based on the assumption that a single error or
incursion may quickly take over or destroy data from a
computer or a system’s architecture. System security flaws can
cause system harm and result in massive economic losses and
actual hardware operations, leading to inaccurate data in the
cyber digital war. To avoid failures, an intrusion detection
system and error identification architecture are required [3].
In a matter of minutes, data stored on computers or wireless
sensor networks can be accessed or lost if an attack occurs,
making infiltration one of the most pressing issues in network
safety. Intruders can cause damage to system hardware.
Various IDSs are used to combat intrusions; however, accuracy
is a key concern. Detection rate and false alarm rate are critical
for accuracy analysis. To reduce false alarms and improve the
detection rate, intrusion detection must be enhanced.

7.1.2 Machine Learning

Analytical models can be automatically built with the assistance

of machine learning, a method of data analysis. Artificial
intelligence (AI) operates on the premise that computers can
learn to solve problems, make decisions, and identify patterns
with little human intervention. The two main types of ML are
supervised and unsupervised learning. Training algorithms
often require labeled instances, such as input and desired
output. Unsupervised learning is used to train new instances in
the absence of such labels. The primary objectives of
unsupervised learning are to (1) detect patterns in the data and
(2) learn more about the data. Semi-supervised learning and
reinforcement learning are two other strategies utilized.
Semi-supervised learning employs less labeled data and vast
volumes of unlabeled data for training purposes.
Reinforcement learning adopts a trial-and-error approach to
determine which activities offer the greatest rewards. The
methods of categorization, regression, and prediction are used.
In this learning approach, agent, environment, and actions are
the three key components. The goal is for the agent to select
activities that maximize the predicted payoff. By using a sound
strategy, the agent may achieve the objective much faster [4].
Therefore, in this research, the support vector machine (SVM)
with the tuned spider monkey optimization (TSMO) is used.
These algorithms can handle classification. Additionally,
normalization and feature reduction are used to conduct a
comparative analysis.

This work is divided into five sections, starting with the

introduction. The second section describes related works.
Additionally, the methodology was outlined in section 7.3.
Performance evaluations were presented in section 7.4. In
conclusion, section 7.5 concludes the article.

7.2 Literature Review

A digital test bed with a star topology and hosts and servers
connected to the OpenFlow OVS-switch mimics the actual
network environment. The signature-based Snort IDS monitors
and detects attacks by mirroring server traffic. The risk
assessment demonstrates that the Snort IDS sufficiently protects
the network topology against threats, except for a few that
require mitigation. A flow-based IDS model is created for
architecture-scaled threat detection. The flow-based outlier
detection using machine learning defeats the signature-based
IDS [5]. Verma and Ranga [6] proposed ML classification
techniques that are used to secure the Internet of Things (IoT)
against DoS attacks. Classifiers that enhance anomaly-based
IDSs are thoroughly studied. Classifiers are evaluated using
popular metrics and validation techniques. CIDDS-001, UNSW-
NB15, and NSL-KDD datasets are used to benchmark classifiers.
Friedman and Nemenyi tests compare classifiers quantitatively.
Raspberry Pi tests classifier response time on IoT frameworks.
The primary focus of the study is on the latest most
authoritative literature available on the subject of machine
learning techniques and their implementation in the IoT and
the IDS for the purpose of protecting computer networks.
Therefore, the aim of the study is to conduct a comprehensive
literature review of current publications that address the
intersection of the IoT and ML in the IDS for computer
networks [7]. The IDSs have been using deep learning (DL),
along with other new approaches, to improve their
effectiveness in safeguarding computer networks and hosts.
Lansky et al. [8] proposed a comprehensive study and
categorization of DL-based intrusion detection techniques. It
begins with a presentation of the fundamental ideas of IDS
designs and the different DL techniques. Then, it groups the
designs according to the specific DL techniques they employ. It
explains how the use of deep learning networks can enhance
the accuracy of intrusion detection. The expansion of
cybersecurity research projects, such as those focusing on
intrusion detection and prevention systems, is of critical
importance. The use of an intrusion detection system, often
known as an IDS, is an efficient method of guarding against
harmful assaults. During the course of research conducted by
Alrowaily et al. [9], the CICIDS2017 intrusion detection dataset
was used in a variety of tests carried out on seven different
machine learning algorithms. The most effective method of
preventive control against assaults originating from IoT devices
is the use of ML and DL techniques, as Asharf et al. [10]
provided an overview of intrusion detection techniques and a
detailed assessment of the technologies, protocols, design, and
dangers connected to IoT systems that arise from compromised
IoT devices. That study also includes an evaluation of several
machine learning and deep learning-based strategies that may
be used to spot Internet of Things devices affected by
intrusions. Rahman et al. [11] proposed two approaches, semi-
distributed and distributed, that combine high-performing
feature extraction and selection with the possibility of fog-edge
synchronized analytics to overcome the limitations of the
centralized IDS for devices with limited resources. To divide up
the work, they each create their simultaneous ML algorithms
for a split intrusion dataset. On the semi-distributed side,
parallel methods are used for feature choices on both the edge
and fog sides, and then a single multilayer perceptron is used
for categorization. In the distributed manner, feature extraction
and multilayer perceptron categorization are carried out
independently by the parallel systems, and then the results are
aggregated by a coordinating edge or fog. Sarker et al. [12]
described an intrusion detection tree machine learning-based
security model that first ranks security mechanisms by
relevance and then builds a tree-based generalized intrusion
detection algorithm based on the chosen critical features. The
model predicts unseen test cases accurately and reduces
computation costs by lowering feature dimensionality. Dang
[13] analyzed the existing research by explicating the choices
made by IDSs constructed in earlier research. Users’ confidence
in using the system will increase if they have a stronger grasp
of the IDS, while programmers may depend on the explanation
to tweak the system. Based on the model’s interpretation, the
experimental findings indicate that they may greatly lower the
computing power requirements of the IDS. Idrissi et al. [14]
analyzed well-known security risks in accordance with the
Cisco IoT reference model architecture to examine the current
state of the art in IoT security threats and vulnerabilities.
Existing studies in the subject of IoT security are also reviewed,
with a focus on those that use DL methods to detect intrusions.
DL techniques are on the rise and are becoming more
important in many sectors, including cybersecurity. Sethi et al.
[15] presented a deep reinforcement learning-based adaptive
cloud IDS framework that overcomes the restrictions
mentioned above while performing accurate determination and
fine-grained categorization of novel and sophisticated assaults.
They conducted thorough testing using the benchmark UNSW-
NB15 dataset, which demonstrates higher accuracy and lower
False Positive Rate (FPR) when matched to state-of-the-art IDSs.
Network entities are susceptible to serious assaults due to a lack
of security measures. An effective security framework for
dealing with threats and detecting assaults is required to solve
that. To address the massive volume of intrusion data
categorization difficulties and enhance detection accuracy by
lowering false alarm rates, Ethala and Kumarappan [16]
proposed a hybrid optimization technique that combines the
spider monkey optimization and hierarchical particle swarm
optimization. The random forest classifier was used to classify
assaults from the NSL-KDD and UNSW-NB 15 datasets after
identifying the most optimal values. Wang [17] used the NSL-
KDD dataset to analyze the effectiveness of several attack
techniques against deep learning-based intrusion detection. It is
experimentally shown that neural networks used by IDSs are
vulnerable in certain ways. The function of various
characteristics in producing hostile data is investigated, and
they analyze the practicality and viability of the assault
strategies in light of their results.

7.3 Proposed Methodology

7.3.1 Problem Statement

The detection accuracy also decreases, and the number of false

alarm reports of network intrusion rises as a result of the
massive amounts of data. This is a significant problem if the
system is attacked by something it has never seen before.
Improving accuracy while reducing false positives is a top
priority. The SVM with the TSMO is proposed to tackle the
aforementioned problems.

7.3.2 Methodology
The investigation used the NSL-KDD dataset, and data
preparation was conducted to remove undesirable data from
the dataset. In the SVM and the TSMO, the characteristics from
which the leader selects whether to split or merge are
discussed, resulting in the selection of the optimal feature
value. Additionally, the TSMO hierarchically arranges the
particles based on their optimal function and categorizes the
attacks using the SVM. The flowchart of the SVM-TSMO is
illustrated in Figure 7.1.

7.3.3 Data Collection

The NSL-KDD dataset is built on the KDD Cup 99 datasets, with

duplicate and redundant data removed. The NSL-KDD dataset
contains 41 labeled features, such as assault kinds. It is
suggested as a solution to the problems with the KDD’99
datasets. The training and testing sets both provide enough data
to be useful, eliminating the need to randomly pick a small
subset of the collection for testing. This advantage enables
results from different studies to be compared and contrasted
consistently [16].
Figure 7.1 Flowchart of the proposed SVM-TSMO model.
7.3.4 Data Preprocessing

Data cleansing, standardization, and integration are among the

preprocessing techniques employed. The data cleaning process
includes removing empty fields, duplicate records, syntax
mistakes, and missing codes. The datasets’ data cleansing
procedures are standardized to facilitate data analysis, and the
cleaned data may be quickly retrieved for study. To accomplish
normalization, new data vectors must be created. One of the
main benefits of normalizing data is a decreased potential for
data inconsistency. The Min-Max normalization paradigm is
crucial when it comes to integrating and normalizing data. The
values of each feature range from 0 to 1, with 0 representing
the lowest and 1 representing the maximum.

Equation 7.1 expresses the normalization process.

(7.1)

where Zr represents a data point, Zmin and Zmax denote the

minimum and maximum values when working with batches.
Recursive feature elimination (RFE) ranks the most essential
characteristics and returns their relative relevance as a value. It
is hypothesized that reducing redundant data would result in
smaller feature subsets. The ranking of characteristics may be
derived by sorting the features from most important to least
important.

7.3.5 Support Vector Machine

The supervised learning approach from which the support

vector machine is derived involves training a model using data
from a wide range of sources and topics [18, 19]. The SVM is
capable of generating a hyperplane or multiple hyperplanes in
a high-dimensional space. The ideal hyperplane is the one that
effectively partitions the input data into distinct categories. A
nonlinear classifier employs a variety of kernel functions to
evaluate the distances between hyperplanes. The primary
objective of kernel functions, such as linear, polynomial, radial
basis, and stochastic, is to maximize the distance between
hyperplanes. The growing interest in the SVMs has resulted in
the development and exploration of various notable
applications. The SVM plays a significant role in image analysis
and pattern recognition. Figure 7.2 illustrates the SVM basic
diagram.
Figure 7.2 The support vector machine.

In most cases, the primary aspect of a classification task

involves splitting data into two distinct groups: the training
datasets and the testing datasets. Within the context of this
class, labels will be referred to as target variables, while
attributes will be understood as features or observed variables.

7.3.6 Tuned Spider Monkey Optimization

The TSMO is a swarm intelligence-based system that consists of

a cluster of social media (SM) whose foraging activity inspires
them. The decision to fission or join together is based on the
number of monkeys in one group. The algorithm depends on
the social structure of a set of characteristics determined by the
leader, who decides whether to pool resources or move in
separate directions in the quest for food. The global leader is in
charge of all groups, while the local leaders are in charge of
their respective groups. One idea is that different parts of the
brain have different control over different aspects of the mind’s
thought processes, and the similarities between these
operations and the monkey brain are well recognized. A fission-
fusion society is a community observed in ethology where
animals undergo constant transformations. They are split into
smaller groups during the day, which have the fission status,
and mixed back together at night with the fusion status. Spider
monkeys provide a good example of this kind of civilization. In
this community, female monkeys are viewed as leaders over
their male counterparts, while male monkeys are at a lower
status (i.e., they are not dominating) despite their greater
activity.

This society’s customs are based on the same social

arrangement as that of a spider monkey, which allows for a
unique blend of old and new. Several research studies have
been conducted on the TSMO technique. These efforts show that
the TSMO performs well in exploration and exploitation, but
there are still certain details to consider in order to achieve
optimal performance. Mini-groups of 40 or 50 partners are
formed inside the larger female-governed monkey
communities, with locating the food supply being a top priority
for these smaller groupings. If one female-led group does not
achieve its goal, another female-led subgroup will attempt, and
so on, until the mission is accomplished. In an effort to keep up
with the times, several changes have been made, such as
analyzing search volumes and selecting the most relevant
results. The TSMO approach is segmented and built on
population repetitive methodology, with each stage detailed
below. Figure 7.3 shows the proposed model block diagram.

Phase 1: Initialization

One starting point involves finding N completely random

answers. The spider monkey population, m, is divided into n
equal halves. Then, a local leader is appointed to each group,
with a global leader overseeing all groups.

(7.2)
where Jmin m and Jmax m are limits of Jr in the mth vector. I(0,1) is
an arbitrary number (0, 1).

Phase 2: Local Leader

At this stage, the SMs will be aware of what their local leader
and their nearby neighbor know, and they will update their
location as a result. The following are the steps involved:

(7.3)
s
Phase 3: Global Leader

During this phase, the SMs have another opportunity to update

their positions using fitness and reach the global optimum.
They can draw inspiration from their perseverance, their
neighbors, and their global leaders. The updating behavior at
this stage is more precisely described by the following equation:

(7.2)
Phase 4: Local Leader Learning

Since the global optimum has already been identified, the

method determines the leader of the local subgroups and the
local optimal solutions. This step involves determining whether
or not local leaders have updated themselves by assessing the
threshold’s number.

Phase 5: Global Leader Learning

The phase’s name alone suggests that the global leader is

present within the pack. It also verifies whether the leader has
updated its position to a specific limit, allowing for an increase
in activity.

Phase 6: Local Leader Decision When this happens, the social

media managers in the group will upgrade the platform
entirely either by following the instructions of the global leader
or by initializing randomly based on the perturbation rate. This
will happen regardless of whether the local leaders have
already been updated to a certain threshold or not.

(7.5)

Phase 7: Global Leader Decision

During this phase, the group undergoes both fission and fusion,
particularly regarding the concept of a limit global leader,
which suggests that global leaders do not need to meet a
specific threshold to be considered as such.

7.4 Result and Discussion

After constructing the SVM-TSMO model, it is important to

adjust its parameters to meet the prediction requirements. The
key elements to consider include precision, accuracy, recall, and
F1-measure [20, 21].

7.4.1 Accuracy

The accuracy of the recommended intrusion detection scheme

is a measure of how effectively it detects intrusions. Figure 7.3
shows a comparison between the accuracy of the suggested
model and the current model, indicating that the suggested
system is more precise than the current one.

Figure 7.3 Accuracy of the existing and proposed methods.

(7.6)

7.4.2 Precision

Precision, also known as positive predictive value, is the

proportion of relevant concepts among the recoverable
occurrences. It serves as a benchmark for quality and
represents the average likelihood of recovery. In comparison to
existing methods, the proposed research is much more precise.
Figure 7.4 provides a comparison between the precision of the
recommended and existing methods.

(7.7)

7.4.3 Recall

Recall, also known as sensitivity or true positive rate,

represents the number of instances that were searched for and
recovered. The proposed approach has the highest recall rate
when compared to conventional approaches. Figure 7.5
provides a comparison between the suggested method and the
existing one in terms of recall.
Figure 7.4 Precision of the existing and proposed methods.

(7.8)

7.4.4 F1 Score

The F1-measure is a unified metric that considers both the

frequency and amplitude of a system’s clarity and memory
harmonic means. Its primary purpose is to evaluate the relative
merits of two approaches, with higher F1 scores indicating a
more efficient system. In comparison to the old approach, the
proposed approach has a better F1 score. Figure 7.6 presents a
comparison of the F1-measure for both new and old
approaches.
 (7.9)

Figure 7.5 Recall % of the existing and proposed methods.

Figure 7.6 F1-measure of the existing and proposed methods.

7.5 Conclusion

Staying up to date with the latest security developments

involves both detecting and preventing intrusions. Given that
most activities rely heavily on networks and information
systems, both intrusion detection and prevention are crucial.
Various methodologies have been used to develop intrusion
detection systems, with ML algorithms like the SVM with the
TSMO playing a crucial role. The comparison results between
the recommended approach and the other three techniques
demonstrate that the suggested method performs much better
in terms of intrusion detection according to the same criteria.
References

1. Jabbar, A.H. and Alshawi, I.S., Spider monkey optimization

routing protocol for wireless sensor networks. Int. J. Electr.
Comput. Eng., 11, 3, 2432–2442, 2021, doi:
10.11591/ijece.v11i3.pp2432-2442, (2088-8708).
2. Halimaa, A. and Sundarakantham, K., Machine learning
based intrusion detection system, in: 3rd International
Conference on Trends in Electronics and Informatics (ICOEI),
vol. 2019, 2019, doi: 10.1109/ICOEI.2019.8862784.
3. Azizan, A.H. et al., A machine learning approach for
improving the performance of network intrusion detection
systems. Ann. Emerg. Technol. Comput., 5, 5, 201–208, 2021,
doi: 10.33166/AETiC.2021.05.025.
4. Disha, R.A. and Waheed, S., Performance analysis of machine
learning models for intrusion detection system using Gini
Impurity-based Weighted Random Forest (GIWRF) feature
selection technique. Cybersecurity, 5, 1, 1, 2022, doi:
10.1186/s42400-021-00103-8.
5. Abubakar, A. and Pranggono, B., Machine learning based
intrusion detection system for software defined networks, in:
Seventh International Conference on Emerging Security
Technologies (EST), 2017, doi: 10.1109/EST.2017.8090413.
6. Verma, A. and Ranga, V., Machine learning based intrusion
detection systems for IoT applications. Wirel. Pers. Commun.,
111, 4, 2287–2310, 2020, doi: 10.1007/s11277-019-06986-8.
7. Da Costa, K.A.P. et al., Internet of Things: A survey on machine
learning-based intrusion detection approaches. Comput.
Netw., 151, 147–157, 2019, doi: 10.1016/j.comnet.2019.01.023.
8. Lansky, J. et al., Deep learning-based intrusion detection
systems: A systematic review. IEEE Access, 9, 101574–101599,
2021, doi: 10.1109/ ACCESS.2021.3097247.
9. Alrowaily, M. et al., Effectiveness of machine learning based
intrusion detection systems, in: Security, Privacy, and
Anonymity in Computation, Communication, and Storage,
Proceedings: 12th International Conference, SpaCCS 2019,
Atlanta, GA, USA, July 14-17, 2019, 2019, doi: 10.1007/978-3-
030-24907-6_21.
10. Asharf, J. et al., A review of intrusion detection systems
using machine and deep learning in internet of things:
Challenges, solutions and future directions. Electronics, 9, 7,
1177, 2020, doi: 10.3390/electronics9071177.
11. Rahman, M.A. et al., Scalable machine learning-based
intrusion detection system for IoT-enabled smart cities.
Sustain. Cities Soc., 61, 102324, 2020, doi:
10.1016/j.scs.2020.102324.
12. Sarker, I.H. et al., Intrudtree: A machine learning based
cyber security intrusion detection model. Symmetry, 12, 5,
754, 2020, doi: 10.3390/sym12050754.
13. Dang, Q.-V., Understanding the decision of machine learning
based intrusion detection systems, in: Future Data and
Security Engineering, pp. 379–396, 2020, doi: 10.1007/978-3-
030-63924-2_22.
14. Idrissi, I. et al., IoT security with deep learning-based
intrusion detection systems: A systematic literature review,
in: Fourth International Conference on Intelligent Computing
in Data Sciences (ICDS), vol. 2020, 2020, doi: 10.1109/
ICDS50568.2020.9268713.
15. Sethi, K. et al., Deep reinforcement learning based intrusion
detection system for cloud infrastructure, in: International
Conference on Communication Systems & NetworkS
(COMSNETS), vol. 2020, 2020, doi: 10.1109/
COMSNETS48256.2020.9027452.
16. Ethala, S. and Kumarappan, A., A hybrid spider monkey and
hierarchical particle swarm optimization approach for
intrusion detection on Internet of Things. Sensors (Basel), 22,
21, 8566, 2022, doi: 10.3390/s22218566.
17. Wang, Z., Deep learning-based intrusion detection with
adversaries. IEEE Access, 6, 38367–38384, 2018, doi:
10.1109/ACCESS.2018.2854599.
18. Kumar, G. et al., MLEsIDSs: Machine learning-based
ensembles for intrusion detection systems-a review. J.
Supercomput., 76, 11, 8938–8971, 2020, doi: 10.1007/s11227-
020-03196-z.
19. Narayanan, G. et al., PSO-tuned support vector machine
metamodels for assessment of turbulent flows in pipe bends.
Eng. Comput., 37, 3, 981–1001, 2019, doi: 10.1108/EC-05-2019-
0244.
20. Bhattacharya, S. et al., A comparative analysis on prediction
performance of regression models during machining of
composite materials. Materials (Basel), 14, 21, 6689, 2021, doi:
10.3390/ma14216689.
21. Shaik, K. et al., Big data analytics framework using squirrel
search optimized gradient boosted decision tree for heart
disease diagnosis. Appl. Sci., 13, 9, 5236, 2023, doi:
10.3390/app13095236.

Note

1. *Corresponding author: [email protected]

8
Security Enhancement in
IoMT‑Assisted Smart Healthcare
System Using the Machine
Learning Approach
Jayalakshmi Sambandan1*, Bharanidharan Gurumurthy2
and Syed Jamalullah R.3

1
Department of Master of Computer Applications, MEASI
Institute of Information Technology, Royapettah, Chennai, India

2
Department of BCA, The New College, Chennai, India

3
Department of Computer Science and Engineering, Sathyabama
Institute of Science and Technology, Chennai, India

Abstract

The Internet of Medical Things (IoMT) is an emerging area that

offers numerous valuable benefits to patients and healthcare
providers in the identification and management of various
illnesses. Despite its advantages, security remains a concern.
IoMT devices contain a vast amount of clinical data, including
personal information such as names, addresses, and medical
histories, making data privacy and security challenging to
maintain. The IoMT is highly susceptible to security breaches
due to inexperienced users’ weak security practices and the
prevalence of various intermediate attacks that may
compromise sensitive medical data. To address these concerns,
we propose utilizing the support vector machine with
multilayer particle swarm optimization (SVM-MPSO) to
enhance the IoMT data security. In the experimental analysis,
the suggested techniques outperform the current system. The
proposed method can be evaluated using a variety of criteria,
including accuracy, precision, security, and sensitivity. The
findings demonstrate that the recommended approach
enhances security measures.

Keywords: Internet of Medical Things, security and privacy,

machine learning, optimization, smart healthcare system

8.1 Introduction

The Internet of Medical Things (IoMT) has emerged as the new

technique for handling patient health data. It primarily serves a
variety of purposes in homes, hospitals, and body sensors. The
IoMT environment is made up of a vast array of medical
equipment, sensors, and other things. Additionally, these
effective technologies have the capacity to cover a specific body
surface. The medical business can simply manage medical
identification, where the practitioner assists in the early
identification of illnesses and effective diagnosis, thanks to the
IoMT system’s quick development [1]. Even while IoMT
applications have been growing steadily over the last several
years, security and privacy remain major issues. Given that the
health data stored through the IoMT architecture comprises
many pieces of medical information, its management is
essential. The illustrative nature of the IoMT network makes it
vulnerable to a variety of assaults that may severely impact its
efficiency. As a result, it loses all the crucial information. When
medical equipment or sensors are not properly authorized or
authenticated, when users do not understand the need of data
encryption, when device interfaces are not properly protected,
etc., the whole system is put at risk. For an IoMT system to be
useful, it must include authentication procedures to keep
patient information safe and private [2]. IoMT-smart
healthcare-based systems are shown in Figure 8.1.
Figure 8.1 The IoMT-smart healthcare system.

The economy has benefited immensely from smart healthcare

systems (SHSs), which are now an indispensable part of it.
Telehealth, smart medication, and local and remote monitoring
of medical resources are just a few examples of how the IoMT is
essential to the growth of SHSs. The degree to which a patient
follows their doctor’s advise and modifies their behavior in
response to therapy is important for the success of the process.
The IoMT refers to the network of sensor-enabled healthcare
equipment that is linked to one another. By reducing the
number of needless hospital visits, the IoMT has the potential to
lessen the burden on healthcare facilities. Additionally, it offers
a secure data transfer environment for the exchange of private
medical data across various medical sectors. The practical uses
of the IoMT have improved people’s quality of life. Rapid
development in issues related to the safety, privacy, and
trustworthiness of the IoMT has occured in recent times.
Recently, the research community has been paying increasing
attention to security, privacy, and trust [3]. Data security
involves taking precautions during data storage and transfer to
protect the data’s veracity, accuracy, and completeness. In
addition, it guarantees that only authorized individuals have
access to the data and may make any necessary changes. When
developing an SHS, it is also important to keep privacy
protection (PP) in mind. When information is sent via an
unsecured route, its severity and sensitivity are taken into
consideration. There are both subject matter and environment
constraints in PP. Content privacy prevents sensitive medical
data from being leaked, but protecting patients’ identities may
be challenging because an assailant may readily determine the
health state of a patient by learning the identity of the treating
doctor. Contextual privacy must also be protected. The concept
of “contextual privacy” pertains to the safeguarding of the
communication’s background details. Multiple asymmetric and
symmetric encryption techniques are utilized in the IoMT-
enabled SHS to protect user data [4]. Figure 8.2 shows the
systematic block of security enhancement in the healthcare
system.
Figure 8.2 A systematic diagram of security enhancement in the IoMT using machine
learning.

The rest of the paper will be presented below. In section 8.2, we

will examine the existing works. The theory behind the
suggested procedures is detailed in section 8.3. In section 8.4,
we will discuss the metrics we used and the outcomes we
obtained. The last section of the report will summarize the
findings.

8.2 Literature Review

Ahmed et al. [5] presented an overview of several federated

learning (FL) architectures and future IoMT-based FL
architectural techniques. Furthermore, they explored how
physical layer security may be used for the efficient protection
of data privacy in the FL-based IoMT. They discussed current
research in this field and important research problems
associated with Physical Layer Security (PLS)-assisted FL in the
IoMT. In addition, they demonstrated a case study indicating
that clustering IoMT devices improves the secrecy performance
of the FL-based IoMT network in comparison to its non-
clustering equivalent. Awotunde et al. [6] discussed the
research hurdles involved in deploying an artificial intelligence
based IoMT (AiIoMT) system and the potential benefits of
implementing such a system in medical systems. An AiIoMT-
based architecture for continuous patient monitoring and
diagnosis was also presented in that chapter. The model was
validated using a collection of cytology images and scored on
several metrics, including several parameters. An improved
diagnostic accuracy of 99.5% is shown, indicating that the
artificial intelligence (AI) method is a potential approach for the
diagnosis of illnesses in an IoMT-related model. Kavitha et al.
[7] showed how the IoMT and deep learning (DL) algorithms
may be used to observe the physical behaviors of the elderly for
faster evaluation and better therapy options. Wearable sensors
attached to the individual’s chest, left ankle, and right arm
collect and transmit data to the consolidated cloud and data
analysis layer through IoMT devices. For the purpose of
determining which parts of the body are engaging in a
particular action, the MapReduce framework is put to good use
in tandem with a complex-valued deep Convolution Neural
Network (CNN) with an improved political optimizer, a classifier
that boasts both increased flexibility and superior performance
relative to its contemporaries. Rehman et al. [8] examined the
literature on the intersection of blockchain technology and
federated learning in the medical field in depth. The aim of that
research is to build a blockchain-based IDS-enabled healthcare
5.0 system that can monitor patients with the help of sensors
and provide physicians with the data they need to make
periodic diagnoses and treatments based on the predicted
course of the disease.

The results of this study show that the suggested system is a

successful optimization of the method for healthcare
monitoring. The framework relies on edge computing for its
design, which protects user privacy using hybrid encryption
methods and lightens the burden on the main server as a
bonus. Security is further strengthened by using differential
privacy based on Laplacian noise on the common
characteristics, ensuring that the transmitted data remain
private even in the face of an adversary. The experimental
findings on benchmark datasets reveal that the variation in key
system performance metrics is not proportional to the change
in the volume of gradients communicated and the proportion of
people involved [9]. Machine learning has been effectively used
for the resolution of a wide range of networking issues,
including routing, traffic engineering, resource allocation, and
security. There has been a recent uptick in the use of ML-based
approaches for enhancing several Internet of Things (IoT)
applications. While research on big data analytics and ML is
abundant, research on the development of ML-based
methodologies for big data analysis in the IoT healthcare
industry is lacking. Li et al. [10] provided an in-depth look at
how machine learning may be used to analyze large amounts of
healthcare data. Abbas et al. [11] proposed a blockchain-
assisted secure data management architecture for health
information based on the IoMT to facilitate the safe transfer of
patient data and improve the scalability and accessibility of the
healthcare system. The suggested blockchain-assisted secured
data management framework (BSDMF) allows for the safe
transfer of information between local and remote servers, such
as those located in the cloud and those located on a user’s local
network, for the sake of medical treatment. To provide secure
data management and data transfer across connected nodes,
the IoMT-related security architecture makes use of blockchain
technology. Artificial intelligence is a new and dynamic concept
that will revolutionize IoMT-related applications and
innovations.

The use of AI in IoMT-based diagnostics and monitoring may

speed up the process and reduce the workload involved. Thus,
the scope of use for IoMT-related systems that may use AI is
explored in that section. Several outstanding potential and
research problems brought about by AI-enabled IoMT are also
discussed in that chapter. In addition to decreasing the
workload on healthcare systems by shortening the assessment
time required for traditional approach detection methods, that
technology has also been employed to achieve reasonably
accurate identification accuracy. Human intervention in
healthcare will be reduced as a result of the continuing use of
AI using IoMT-based techniques, which will greatly enhance
healthcare monitoring, diagnosis, analysis, forecasting, and the
creation of drugs and vaccines [12]. Concerns about patient
privacy are heightened by IoMT’s lax security measures
regarding sensitive data. Fresh information sent out by the
sensor node is not guaranteed to be genuine. For this reason, a
blockchain-based system is essential. A healthcare professional
would be able to get access to a patient’s medical data using a
more secure authentication-based method regardless of the
network connection used. Al-Otaibi [13] offered a novel method
of safe authentication that makes use of machine learning. In
that study, they use K-nearest neighbor and ML through smart
contracts to detect and authenticate dynamic temporal attacks
in an IoMT setting (KNN-MLSC). It increases safety for both
doctors and patients, decreases waiting time, and protects
personal information in health records.

8.3 Proposed Methodology

There is a need to boost the safety of IoMT applications in

healthcare settings. As a result, we recommend protecting the
data using a support vector machine and multilayer particle
swarm optimization (SVM-MPSO). The suggested method is
described in great depth below. The steps of the suggested
technique are displayed in Figure 8.3.

Figure 8.3 Diagrammatic representation of the proposed method.

8.3.1 Data Collection

The Information Security Centre of Excellence 2012 dataset was
used to run the experiments and estimate how well the
proposed way of finding anomalies worked. The whole set of
data includes nearly 1.5 million network traffic packets with 20
features that cover 7 days of activity on the network. There are
more details about the dataset that can be found. A random
part of the original dataset has been taken out. The training
data consists of 30,814 “regular” lines and 15,375 “attack” links.
There are 13,154 regular links and 6,580 more attack lines in
the testing data [14].

8.3.2 Data Preprocessing

A new dataset with 14 characteristics was constructed from the

specified dataset’s original format. Since the vast majority of
packets included nothing, we removed the payload feature,
which included the packet itself. In its stead, we implemented a
duration feature that combines the functionality of the
previously separate start time and end time fields. For the
purpose of eliminating the bias of higher-valued features, the
attributes were scaled between [0,1] using the Min-Max
approach during the data normalization step, with the
corresponding math looking like this:
 (8.1)

Categorical characteristics are not accepted by most classifiers;

hence, the data mapping approach was employed to convert
them to numerical values.

8.3.3 Support Vector Machine

A supervised machine learning classification method, the SVM

algorithm finds the greatest separation hyperplane across
classes to distinguish between positive and negative samples.
Since the kernel controls the form of the separating hyperplane,
multiple kernels may be employed as part of the SVM approach
depending on the nature of the dataset. When the data can be
separated along a straight line, a linear kernel may be used to
describe the hyperplane as a linear equation. However, other
kernels are required when the data are not linearly separable.
The Gaussian kernel is one example of a suitable kernel that
can be used in this case. This kernel transforms the input data
into a greater feature space [15].

(8.2)
The basic goal of the SVM model is to choose the hyperplane
that optimally divides classes and maximizes the lowest
distance between any two data points (margin). In this sense, all
boundary points are support vectors. Figure 8.4 shows a
simplified representation of the basic variant of this system, the
linear SVM model.

The optimal hyperplane equation is found below.

(8.3)

where

Y: is the normal vector of the decision hyperplane

Data pointi is denoted by the symbol Yi.

xi: classifies data point i as belonging to either the positive or

negative set
Figure 8.4 The linear SVM model.

The formula for the classifier is as follows: f (yi) = signUTyi + p

The formula xi = UTyi + p describes the functional margin of xi.

Algorithm: Support vector machine algorithm

N←choice of input features

Sv←the support vector

ft ← characteristic of support vector machines

out ←The Results of the Decision Function

for j ← 1 ti M do

L = 0;

for i ← 1 to Sv do

dist = 0

for r ← 1 to ft do

g ← Sv [i]mul l(r)

GG ← M[i] mull d(r)

t ← g sub gg

dist ← t2

end
r ← −x mul dist

LL ← e mul r

L ← LL add LL

end

out ← L mul p

end

For classification purposes, the Multi Support Vector Machine

(MSVM) creates a hyperplane in a high-dimensional feature
space to divide the data into two groups. The search for the best
hyperplane fulfills the classification requirement and employs a
tailored method to achieve a nearly ideal level of margin
separation. In the theory of learning, the margin scale is an
illustration of the capacity to generalize to very large settings,
and the maximum generalization power of the SVM is required
for the largest margin scale.

8.3.4 Multilayer Particle Swarm Optimization

The MLPSO algorithm is a method of SI global random search
that simulates the swarming and migrating behavior of feeding
insects. The usual strategy toward each part of the swarm
aggregation method is to guard every piece of information,
accomplish each information rate in the nearby area, and
change the information center regardless of where they are
going in the case of MPSO. The multilayer particle swarm
optimization, commonly known as the MLPSO, locates a particle
in the search space that best addresses each optimization
challenge. The fitness value of each particle is determined by
the optimum function, and the distance traveled by each
particle is determined by its velocity. Particles follow the
optimum particle, which leads the way, and move across the
subspace. Figure 8.5 depicts the flow diagram for the most
fundamental form of the MLPSO algorithm. The MLPSO is
integrated with a specified analytical model to increase the
level of security of the data transfer process. As a result, the
various types of assaults that are currently active inside the
system are uncovered. The iteration values are determined by a
set of population matrices, in which each individual is assigned
a particular set of fitness values ranging from 0.5 to 1 because
the MLPSO is being used. The variation in these two values
generates a binary matrix that yields two optimal values for
individuals represented by the variables p and g. The optimal
values, indicated above, shift between 10 and 100 with each
new iteration, with a step variation of 20 between each number.
After identifying the optimal value location for low-security
elements, the associated rapidity rates are measured as the
output of the MLPSO. This is done when the speed of the search
space is boosted with security measures.

The phrase bbest [j], also known as the individual extremum, is

used to refer to the position that particle j has determined to be
the best possible option [16]. The global ideal point that was
found by the comprehensive particle swarm search is denoted
by the abbreviation sbest [j]. The following set of random values
is used to update the locations and velocities of the particles in
accordance with Equation 8.4 so that the next generation may
be generated correctly.

(8.4)
(8.5)
where iter denotes the j th− generational iteration, Randj
denotes ranges of random integers between 0 and 1 that are
dispersed equally, zj denotes the particle’s unique velocity
value, and h denotes the inertia weight that established the
particle’s speed before the present speed and acts as a hybrid of
the global search algorithm and the local search capability.
Figure 8.5 The MLPSO algorithm flowchart.

Convergence and accuracy for the MLPSO approach both

increase with decreasing inertia weight. In order to improve
the MLPSO algorithm’s precision, the inertia weight factor may
be computed.

8.3.5 Performance Evaluation

Several performance criteria, such as accuracy, precision,

sensitivity, security, and specificity, are evaluated to find out
how well the proposed method works.

8.3.5.1 Accuracy

The standard method for measuring accuracy is the weighted

arithmetic mean, which is also known as the inverse precision
of the precision value. Accuracy measures the precise
classification of secured data. Table 8.1 illustrates the accuracy
of the proposed and existing methods.

Accuracy is determined by the following:

(8.6)

8.3.5.2 Precision

In the field of information retrieval, precision is the number of

relevant observation results out of all the ones that were found.
It is the percentage of documents that are actually true among
the reported to be true, also known as the positive predictive
rate. Table 8.2 presents the precision of the existing and
recommended methods.

(8.7)

Table 8.1 Comparison of the accuracy.

Methods Accuracy (%)

Deep CNN (Kavitha et al., 2022) 91

Blockchain (Rehman et al., 2022) 85

KNN (Al-Otaibi, 2022) 73

SVM-MLPSO (Proposed) 98
Table 8.2 Comparison of the precision.

Methods Precision (%)

Deep CNN (Kavitha et al., 2022) 83

Blockchain (Rehman et al., 2022) 75

KNN (Al-Otaibi, 2022) 85

SVM-MLPSO (Proposed) 97.5

8.3.5.3. Sensitivity

Sensitivity is the ratio of true positives to the sum of true

positives and false negatives, which determines how well a test
can detect positive instances. It is calculated as the positive
classification of secured data and can also be referred to as true
positive or false negative. Table 8.3 presents the sensitivity of
the old and new strategies.

The sensitivity is computed by using the following:

(8.8)
8.3.5.4 Specificity

Specificity measures the negative classification of secured data

when the actual condition is not present. It is expressed as a
false positive or true negative rate. Table 8.4 presents the
specificity of the existing and proposed methods.

Table 8.3 Comparison of the sensitivity.

Methods Sensitivity (%)

Deep CNN (Kavitha et al., 2022) 70

Blockchain (Rehman et al., 2022) 87

KNN (Al-Otaibi, 2022) 79

SVM-MLPSO (Proposed) 96
Table 8.4 Comparison of the specificity.

Methods Specificity (%)

Deep CNN (Kavitha et al., 2022) 70

Blockchain (Rehman et al., 2022) 80

KNN (Al-Otaibi, 2022) 85

SVM-MLPSO (Proposed) 95

Table 8.5 Comparison of the security.

Methods Security (%)

Deep CNN (Kavitha et al., 2022) 85

Blockchain (Rehman et al., 2022) 73

KNN (Al-Otaibi, 2022) 90

SVM-MLPSO (Proposed) 97.6

It is determined by the following:

(8.9)

8.3.5.5 Security

A security metric is a set of interconnected dimensions

(measured against a standard) that allows the measurement of
the degree of protection against potential harm or loss from
malicious attacks. Table 8.5 illustrates the security comparison
between the current and proposed techniques.

8.4 Conclusion

The SVM-MLPSO methods are used in this research to protect

the data. While the suggested technique offers excellent
security characteristics for IoMT applications, some of its
shortcomings become apparent when it is used in actual
situations. The particle swarm optimization approach yields
more accurate results as compared to other methods. The
required charts demonstrate that without utilizing any feature
extraction techniques, the accuracy of the suggested strategy
employing the MLPSO improves to 99%. In contrast to the
observed values in the current approach, the suggested method
offers 98% accuracy, which is substantially greater. Additionally,
the MLPSO iteration values provide optimum outcomes that
raise the prediction and quantifiable score throughout the
process. The suggested study employing the MLPSO may be
expanded in the future to include a variety of cloud computing
platforms, which will further enrich the complete dataset with
high-security characteristics. The expansion is also made
feasible by taking into account the distinction between internal
and external assaults in situations where all users may send
and receive various pieces of data using a machine learning
method.

References

1. Khatiwada, P. and Yang, B., An overview on security and

privacy of data in IoMT devices: Performance metrics,
merits, demerits, and challenges, in: pHealth, vol. 2022, pp.
126–136, 2022, doi:10.3233/SHTI220970.
2. Khan, I.A. et al., XSRU-IoMT: Explainable simple recurrent
units for threat detection in Internet of Medical Things
networks. Future Gener. Comput. Syst., 127, 181–193, 2022,
doi: 10.1016/j.future.2021.09.010.
3. Vaiyapuri, T., Binbusayyis, A., Varadarajan, V., Security,
privacy and trust in IoMT enabled smart healthcare system:
A systematic review of current and future trends. Int. J. Adv.
 Comput. Sci. Appl. (IJACSA), 12, 2, 1–7, 2021,
http://dx.doi.org/10.14569/IJACSA.2021.0120291.
4. Mawgoud, A.A. et al., A secure authentication technique in
internet of medical things through machine learning, arXiv
Preprint ArXiv:1912.12143, 2019.
5. Ahmed, J. et al., On the physical layer security of federated
learning based IoMT networks. IEEE J. Biomed. Health
Inform., 27, 2, 691–697, 2023, doi: 10.1109/JBHI.2022.3173947.
6. Awotunde, J.B., Folorunso, S.O., Ajagbe, S.A., Garg, J., Ajamu,
G.J., AiIoMT: IoMT-based system-enabled artificial
intelligence for enhanced smart healthcare systems. In
Machine Learning for Critical Internet of Medical Things, F. Al-
Turjman, A. Nayyar (eds.), Springer, Cham, USA, 2022,
https://doi.org/10.1007/978-3-030-80928-7_10
7. Kavitha, D. et al., An efficient IoMT based health monitoring
using complex valued deep CNN and political optimizer.
Trans. Emerg. Telecommun. Technol., 33, 12, e4610, 2022, doi:
10.1002/ett.4610.
8. Rehman, A. et al., A secure healthcare 5.0 system based on
blockchain technology entangled with federated learning
technique. Comput. Biol. Med., 150, 106019, 2022, doi:
10.1016/j.compbiomed.2022.106019.
9. Nair, A.K. et al., Privacy preserving Federated Learning
framework for IoMT based big data analysis using edge
 computing. Comput. Stand. Interfaces, 86, 103720, 2023, doi:
10.1016/j.csi.2023.103720.
10. Li, W. et al., A comprehensive survey on machine learning-
based big data analytics for IoT-enabled smart healthcare
system. Mob. Netw. Appl., 26, 1, 234–252, 2021, doi:
10.1007/s11036-020-01700-6.
11. Abbas, A. et al., Blockchain-assisted secured data
management framework for health information analysis
based on Internet of Medical Things. Pers. Ubiquitous
Comput., 1–14, 2021, doi: 10.1007/s00779-021-01583-8.
12. Awotunde, J.B., Chakraborty, C., AbdulRaheem, M., Jimoh,
R.G., Oladipo, I.D., Bhoi, A.K., Chapter 1 - Internet of medical
things for enhanced smart healthcare systems, in: Intelligent
Data-Centric Systems, Implementation of Smart Healthcare
Systems Using AI, IoT, and Blockchain, C. Chakraborty, S.
Kumar Pani, A.A. Mohd, X. Qin (eds.), pp. 1–28, Academic
Press, USA, 2023, https://doi.org/10.1016/B978-0-323-91916-
6.00009-6.
13. Al-Otaibi, Y.D., K-nearest neighbour-based smart contract for
internet of medical things security using blockchain. Comput.
Electr. Eng., 101, 108129, 2022, doi:
10.1016/j.compeleceng.2022.108129.
14. Injadat, M. et al., Bayesian optimization with machine
learning algorithms towards anomaly detection, in: IEEE
 Global Communications Conference (GLOBECOM), vol. 2018,
2018, doi: 10.1109/GLOCOM.2018.8647714.
15. Narayanan, G. et al., PSO-tuned support vector machine
metamodels for assessment of turbulent flows in pipe bends.
Eng. Comput., 37, 3, 981–1001, 2019, doi: 10.1108/EC-05-2019-
0244.
16. Shankar, R. et al., Hybridized particle swarm-gravitational
search algorithm for process optimization. Processes, 10, 3,
616, 2022, doi: 10.3390/pr10030616.

Note

1. *Corresponding author: [email protected]

9
Building Sustainable
Communication: A Game-
Theoretic Approach in 5G and 6G
Cellular Networks
Puppala Ramya*, Tulasidhar Mulakaluri, Chebrolu
Yasmina, Pandi Bindu Madhavi and Vijay Guru Balaji K. S.

Department of Computer Science and Engineering, Koneru

Lakshmaiah Education Foundation, Vaddeswaram, Guntur, India

Abstract

High data volume is added to the currently used digital

technologies as a result of cellular networks’ technical
development. These new digital technologies must maintain
trade-offs between clients and service providers, which present
issues in terms of quality of service (QoS). Real-time, audio,
visual, and text data types, each of which has specific needs, are
handled by cellular service providers in an effort to satisfy the
demands of end customers. There were many potential
incentive-based algorithms to achieve an effective QoS. The
mobile network operator and access points are not guaranteed
a fair share of the profits under these programs, however.
Furthermore, these schemes are unable to guarantee an
equitable channel distribution during crowning time and do
not offer the end user efficient QoS. In order to offload the
method of decision-making in a diverse network, we present a
motivation-based strategy employing game theory and a two-
stage Stackelberg technique for integrated data. Our suggested
design models an individual mobile base station and many
built-in access points in a congested urban region. Here is
competition among access sites for incentives when traffic is
unloaded at this station, which gives financial incentives
depending on the sorts of traffic. To simulate and evaluate a
real-world event, a mathematical game is developed. By
comparing the findings with those from other models, the
experimental technique is used to verify the numerical results.

Keywords: Cellular networks, digital technologies, game theory,

Stackelberg approach, Nash equilibrium

9.1 Introduction

Due to the growing need for fifth-generation (5G) wireless

networks to have more coverage and capacity, ultra-dense
networks (UDNs) have been used to address some of the most
critical issues. In order to overcome the basic impasses of
beyond, the deployment of UDNs on a very wide scale has been
envisioned [1]. The sixth-generation (6G) wireless networks will
bring several orders of magnitude greater improvements than
current technology. However, it is generally acknowledged to be
a formidable challenge to use a mathematical tool to optimize
the system performance under the strict radio resource
constraints [2]. Current UDNs typically do system-level
performance optimization by relying on numerical simulations,
which are often time-consuming and have proven increasingly
challenging in the setting of 6G with exceptionally high density
[3]. Therefore, it is crucial to create a practical mathematical
model for optimizing the 6G UDNs. In this work, we present
difficulties and problems that must be carefully taken into
account while implementing UDNs in a practical setting. In
order to improve the performance of UDNs, we review effective
mathematical approaches including game theory and real-time
optimization [4]. Additionally, emerging technologies that can
be used in UDNs are also covered. While some of them are
already being investigated, others are anticipated to improve
UDN performance to meet 6G requirements. Some of them have
already been used in UDNs with high efficiency. The combined
optimum technique of real-time optimization and game theory
(ROG), which is a powerful tool to handle the optimization
issues of large-scale UDNs with minimal complexity, is
important, and we present it for the first time [5]. Then, we go
over two methods for utilizing ROG in UDNs. Finally, a few ROG
case studies are provided to show how to use ROG to address
issues with various applications in UDNs.

9.2 Related Works

The research [6] examined physical layer security issues on

device-to-device (D2D) communications in light of various
attack scenarios. The hierarchical and diverse interactions
between authorized users and active intruders in 5G Internet of
Things (IoT) networks are described using game theory, and
then several scalable machine learning techniques are
suggested to find situations of equilibrium between various
agents. The study [7] presented historical discussion of the
many concepts used to reduce or eliminate the interference
issue in 5G communication systems. It has also been suggested
where future research should go. The study [8] provided a
brand-new gamification framework using utilizing geometric
modeling and evaluate cache-enabled small cell base stations
(SBSs) with the shared framework for regional 5G open
platforms. The study [9] provided thorough analysis of the area
of wireless communication resource distribution. The survey
starts out by providing a general resource distribution
summary issue in current linked wireless network papers on
the assessment of resource distribution strategies. The article
[10] provided a reduced-constraints method for finding
solutions for two non-convex mixed-integer problems with
combined allocation of power and channel assignment that are
not ideal nonlinear programming (MINLP). The research [11]
examined the issue of the offloading choice-making systems
engineering for the intelligent secure mobile edge computing
(MEC) system with a unmanned aerial vehicle (UAV)
eavesdropper, where the eavesdropper may be hearing the safe
computational job from the user to the computational access
point (CAP). The research [12] suggested an adaptive VR
architecture that enables high-quality wireless VR in mobile
edge computing-enabled wireless networks. Tasks related to
instant VR rendering may be adaptively offloaded to MEC
servers, and these servers’ cache-related abilities allow further
performance enhancement.

9.3 Methodology

Due of its portability, wireless cellular networks are widely

used, and cellular communication is presently the most
populated sector. The amount of data traffic through cellular
networks has dramatically expanded along with technical
advancement. Mobile network operators (MNOs) implement
4/5G LTE, which offers the quickest cellular connection, to
accommodate customers and their data needs. Although
network service providers are implementing the most recent
technology, the constant volume of motionless traffic reduces
their output by impacting throughput, reaction time, and other
quality of service (QoS) parameters. Costly technologies have
been shown to be far less able to satisfy end user needs. The
traffic flow is increasingly crowded, and the response is
becoming worse every day due to end users’ growing data
needs and the increased bandwidth being used for a variety of
activities through networks, particularly cellular networks.
Customers are given services at a QoS-level via the suggested
incentive-based strategy. Figure 9.1 depicts a densely populated
area where the majority of residents use smart devices, such as
phones and tablets, for example various Internet access. Due to
the overuse of data services, the quality of service was
degraded, and many new customers were allowed. Particularly
during peak hours, a very packed atmosphere was created;
nevertheless, things were different during off-peak hours. The
demand for video streaming, which uses a lot of bandwidth,
was strong during peak hours. In order to experience the
situation actually, an example of an unbalanced network
simulation made up of a mobile base station (MBS) and a few
access points (APs) was created taking these aspects into
account and unloading while implanting through Wi-Fi was
implemented as the solution to the issue.

Figure 9.1 IDO based on a game model.

MBS has a greater coverage area, but APs within its range with
lesser coverage areas have nonoverlapping zones of coverage.
For this offloading process in this network that is diverse, we
take into account one MBS and many APs in order to simulate it
accurately.
 (9.1)

where

MBS coverage area equals AMBS, while AP coverage area

equals AAP.

The MBS area without an offload option

The following formulae may be used to calculate the physical

areas of the areas mentioned above.

(9.2)
(9.3)
The 5G and 6G communication models are shown in the above
graphic. Every user is online and working on downloading and
uploading tasks. We use the integrated data offloading (IDO)
using a game concept to prevent the congestion problem.

Adding these numbers to Equation 9.1, we acquire the MBS

coverage area, which no AP is able to access. In Equation 9.1,
MMBS depicts the portion of the MBS without offloading
circumstance, EMBS reflects the region that MBS covers, EAP
displays the region that APs cover. The following categories are
used to distinguish distinct traffic kinds:
 (9.4)
For Equation 9.4, DD shows different forms of traffic, K reflects
current information, C is used for video, E for sound, whereas
the set of APs is provided by represents text data

(9.5)
P denotes a team of players, and EB indicates the entry point in
Equation 9.5 from 1 to m [2]. For each AP, the different sorts of
data are represented by the Cartesian product of the two sets
previously indicated:

(9.6)

In Equation 9.6 [1], EB DD demonstrates how in an unloading

scenario, various types of traffic are offloaded by the access
point, including real, audio, video, and text the total traffic per

EB, is given as:

(9.7)

In Equation 9.6 [1], EBDD demonstrates in an unloading

scenario the entry point.offloads many traffic kinds, including
text, audio, video, and real. The total of function f ’s
components, where any specific EB when this relationship is
taken into account, the total traffic per
EB, which is stated to be:

(9.8)

In Equation 9.7, fm reveals the entire amount of traffic. Among

these, the traffic that a cellular network offloads includes the
following:

(9.9)

In Equation 9.7, the variable fm stands for traffic that a mobile

phone network has offloaded, where fm displays the location of
the download and EBrf demonstrates the discharge posture. Our
simulation is built based on a game-theoretic process that aims
to achieve the highest offloading EBs; however, a traffic-wise
offloading mechanism will be used to control offloading
percentage during peak and off-peak hours. For this, the
present rate of EB is considered to be its time, which is
symbolized by Ka, in which the subscript makes a distinction
EB. Utilizing this immediate rate, we are able to display the
throughput of certain EB as follows:

(9.10)
In Equation 9.10, Kr demonstrates the immediate pace of
EB, where Thro stands for throughput. The sum of the two
nodes represents the heterogonous network’s entire amount.

(9.11)
In Equation 9.11, ThroughputMBS shows the throughput of MBS,
and ThroughputEBs represents throughput of EBs autonomous
decision-making from a variety of methods is used to study the
construction of real-world models and how they interact. Three
things are involved: players, strategy, and payout. It defines
how players react and make decisions, as well as how they
adopt different positions and work together. The payoff
function is given by the following:

(9.12)

The various functions are represented by the variables in

Equation 9.12: Br is the payoff function in action, βm displays
the reward being given, and σrm highlights the price paid for
the incentive that was provided. The payback function is
intended to assess how well the model based on spurs provided
by MBS and the volume that was by offloading EBs performs.
When AP competes against its rival to increase its revenue, the
game is said to be noncooperative. The following items are
included in the game set.

(9.13)
In Equation 9.13, r ∈ B, S exemplifies the game, Bf symbolizes
MBS participants in the game, and EBs and Gr represent a
group of strategies that players can use to obtain the highest
payout, and Br is the reward or a scoring mechanism that
depicts the players’ earnings.

In Equation 9.14 [5], fr demonstrates the existence of a

quadrilateral space for traffic load. Q, R, S, T are subspaces. To
indicate a particular instance data load, the ordered pair below
is used:

(9.14)
where lO = {r1, r2,…,r0},lK = {r1, r2,…,rk},lG = {r1, r2,…,rg}, and lD =
{r1, r2,…,rd}

9.3.1 Nash Equilibrium

In the case of the aforementioned game set, Nash equilibrium

exists if, for each r ∈ B, Br (fr, f_r) ≥ Br (fr, f_r) aimed at wholly fr
≥ 0 , somewhere, Br (fr, f_r) continues to be the next bid for the r
participant, presuming that the results of the extra troupes’
tenders f−r [6]. Br is concave in fr , while Gr in the specified
spaces is convex.

(9.15)
Data are offloaded using a kind-based method that takes into
account a clearly determined importance; for example, video
data may stream with a higher priority than text data. We
remove the second-order derivative pertaining to km after
confirming its wrap and seeing that Utlm continues to l. The
following equation makes a leading derivative assumption:

(9.16)

The partial derivative rule is used by Equation 9.16 [7] to

demonstrate its concavity. The element ∂ demonstrates partial
derivative, βm displays the reward provided for Wi-Fi, fn
demonstrates redirected traffic, and σrm displays the amount
spent on that unloading traffic.

Regarding the second derivative for fn , we have the following:

(9.17)

The second-order derivative in Equation 9.15 [8] displays a

negative multiplier, and it turns out that in each instance, the
outcome will be unfavorable, demonstrating that Br is a
concave function. Um is concave, as may be seen in fn, while the
Gr is convex; thus, in the game, the Nash equilibrium is present.
In Equation 9.15, the variables have previously been explained.

9.3.2 Unique Nash Equilibrium

The game contains a Nash equilibrium, as has been shown

before vector fn. To prove its originality, to ensure consistency,
the Best Response function K must be given, which requires
understanding. Scalability is one of the three prerequisites,
positivity, and monotonicity. Here, R(fn) depicts the Best
Response feature.

The optimum result for the ideal reaction would be for all
traffic downloaded by any EB is limited to the EB’s full
capability. In any case, the former would not be as good and
could not match the later. This limitation may thus be
mathematically represented as indicated below [9]:

(9.18)

The variables in Equation 9.18 are σr, the unloading fee, which
provides a driving incentive, and K (fr), which depicts the Best
Response feature. The aforementioned connection meets the
game’s positive attributes. The immediate rate of AP, which
includes the component specified in Equation 9.15, would be
approached by the Best Response function in order to increase
revenue. Adding to the Best Response relation is the following:

(9.19)

Equation 9.19 demonstrates [10], a monotonic growth;

nevertheless, to demonstrate its scalability, we take into account
the following scenario:

(9.20)

(9.21)

In Equations 9.20 and 9.21, the variables Kr (αl) show the Best
Respone function for the cost paid for offloading. fe shows the
offloaded traffic. Combining Equations 9.17 and 9.18 and
simplifying them, we obtain as change parameter,
which is always greater than 0; therefore, using the Best
Response feature is also scalable. By satisfying all three criteria,
it is ensured that the Best Response function is a standard
function and satisfies the Nash equation.

9.3.3 Total Unique Equilibrium

The game’s traffic type set’s overall unique equilibrium may be

calculated as [10] , where

(9.22)

where for which the Best Response value exists f in cases

when β is dependent on meeting both of the requirements in
Equations 9.18 and 9.19, and going over either side will cause it
to approach the limitations. This brings about the game’s
complete unique balance.

The MBS return may be calculated as follows:

(9.23)

where r ∈ DD, δ is unit spectrum made clear of congestion

using EBs for offloading, and β is a financial inducement for r-
type traffic. The MBS will optimize its usefulness in this
competitive game by optimizing the β incentive in its broadest
sense. Equation 9.16 is substituted, and the highest utility
benefit is as follows:

(9.24)

Subject to β ≤ βmax.

The utilitarian feature of the MBS is convex in shape because it

can be maximized. Equity and justice must be offloaded in
between EBs. The logarithmic goal will be accomplished. The
MBS utility may be calculated in several ways as [11]:

(9.25)

Substituting Equation 9.16, we obtain:

(9.26)

By replacing Γ and φ, we obtain:

 (9.27)

The inference implies that the app’s MBS feature is concave.

The right first-order status is related to determine the overall
economic advantage.

(9.28)

Considering the problem’s Lagrangian with consideration λ

being a Lagrange multiplie [12], the outcome is as follows:

(9.29)

The dual function 0 is the topic of the maximum state of the

aforementioned equation, which may be solved via sub-
gradient retrieval of the Lagrange multiplier as shown below:

(9.30)

wherever r(g) is a step-by-step size in repetition s. The step size

is always positive of the form,[x]+ = max{0, X} total economic
stimulation β* can be acquired using Karush–Kuhn–Tucker
(KKT), given that the first-order derivative of the Lagrangian
function of βr with respect to β is 0. The whole financial gain
solution has been simplified as follows [13]:

(9.31)

According to the demand-based traffic categorization, the

suggested strategy in independent games rewards to the
secondary participants. Using comprehensive data offloading,
which is based on the philosophy of Stackelberg, it will enhance
QoS to prevent the congestion problem. The Stackelberg model,
a noncooperative game theory model, is what we have used to
address the congestion problem faced by mobile consumers.
Nash equilibrium in the Stackelberg model was established by
the point where the Best Response intersects functions from
each player. The MBS information is the input of the algorithm,
which efficiently provides the greatest output reward and
offloading to benefit the cellular network. The spectrum, the
unit electricity cost, and the proper kind of incentives are all
represented by different symbols. Temporary variables like fm
are added, and lm is utilized as AP’s condition, in order to
condense more complex equations with recurring components.
Numerous calculations are carried out, including those for total
cost, individual costs, economic incentives, progressive balance,
proportional fairness, and others. MBS then determines the net
profit gain before making a decision. To calculate optimum
rewards and traffic conditions, which are often utilized in
various forms, two functions are needed: ideal incentives and
cutoff information.

9.4 Result

The data flow rate offloading each unit of time is represented by

the offload ratio. We collect data in chunks of 10 MB and
measure time in seconds. Our model’s first accomplishment is
maximized offload, which has a significant impact on the other
elements shown in the data. According to the graph in Figure
9.2, offloading begins as soon as at least one AP is deployed and
grows as additional APs are added. The suggested approach
successfully offloaded 70 MB/s of data onto 20 APs in seven
units. As seen in Figures 9.2 and 9.3, this demonstrates a
sufficient improvement over the models that are already in use.
Based on the incentive structure each model offers, the lines are
distinguished. Comparatively speaking, our model did well.

Overload traffic causes a drop in QoS, which causes network

congestion. Five picocells have been found to be adequate to
manage range value 2 congestion; whereas IMDO and IFPC
models perform the same function with 12 and 17 picocells,
correspondingly. This is because our model proposes a plan for
resolving the problem that results in maximal offloading and
special integrated Wi-Fi management. This demonstrates that,
in contrast to previous models, our approach improved the
management of congestion, as illustrated in Figure 9.3. The
suggested model’s output is superior to those of the other
models. The method of making decisions in a diverse network
has been improved by our idea of a motivation-based system
based on a two-stage process and game theory Stackelberg
technique for integrated data offloading. To mimic our
suggested system, an individual mobile base station, a few
cross-platform access points in a populated urban area are
used. Access points face rivalry against one another to gain
rewards for unloading traffic at this station, which gives
financial incentives depending on the sorts of traffic. Through
simulation, a game using math is derived by examination of
real-world situations. The quantity of data that is sent from a
source to a destination in a certain period of time is known as
throughput. While seconds are the unit of time and the amount
of data is expressed in bytes, throughput is their benchmark
and is frequently expressed as a percentage. The graph displays
throughput results in relation to the chosen spectrum. The
utilized spectrum is between 0 and 2, and its effect on each of
all three models’ throughput is investigated. Spectrum 0
indicates no throughput, whereas spectrum 1 indicates
improved throughput. Congestion causes throughput to
decrease; however, when congestion is reduced by the models,
throughput is increased. According to Figure 9.2’s spectrum
value 2, the suggested model achieves 100% throughput,
whereas the other two models only manage 48% and 62%,
respectively.

Figure 9.2 Spectrum use ratio.

Figure 9.3 Offload ratio.

Based on the different traffic kinds, we determine the best

incentives using the Stackelberg-based game model. The main
player will provide the secondary player the ideal incentive,
and the supporting character will unload optimally based on
the ideal incentive. The throughput graph demonstrates that we
have great offloading throughput. The suggested technique
shown in Figure 9.4 has demonstrated the absolute outcome.

The term “response delay” refers to the length of time required

for data to switch between host and a client. The lag is seen in
the graph as shown measured in milliseconds on the y-axis, and
the range is used to assess the reaction using the data’s x-axis as
the load on the spectrum increases. However, despite its impact,
the increase in load for each of the three models lowered the
spectrum’s burden. Overall, IMDO fared the best in this
outcome, although as can be seen in Figure 9.5, our model
advanced more quickly.

IDO is used in the suggested approach to address the congestion

problem. According to the kind of traffic, it will move
movement from the major auxiliary channel to the main
channel, reducing the delay and reducing the likelihood of
congestion. The spectrum was under increasing demand, which
decreased offloading latency and cleared the channels of
congestion. Since it will make use of an independent game
paradigm to equally channelize both users’ use, IMDO fared the
best overall. Figure 9.5 displays the fair outcomes of our
approach.
Figure 9.4 Throughput analysis.

Figure 9.5 Response delay analysis.

Energy use is thought to have a significant impact on the QoS of
mobile devices. Modern wireless communication systems
concentrate on reducing their energy needs. In consideration of
the growing number of picocells, the use of energy is measured
offloading megabytes per watt. The greatest performance came
from the IMDO model, albeit with a little change; the suggested
model displayed its results. The suggested plan is superior, but it
is not perfect because of the integrated design, which requires
extensive sublevel administration. Figure 9.6 illustrates that the
outcome is excellent when compared to alternative approaches.

Figure 9.6 Energy consumption analysis.

9.5 Conclusion
The usage of data-hungry services has greatly increased
network traffic as a result of the widespread adoption of the
newest smart gadgets and fast Internet connections. Mobile
network operators must respond to the scenario by meeting the
substantial requirements, the volume, scope, and superior QoS
to secure the longevity of their connection with users and to
profit in a more sophisticated manner. All providers can afford
mobile data offloading, which improves QoS by lightening the
strain on the network. Offloading data off the cellular network,
many technologies are utilized, including Wi-Fi, mobile
networks using WiMAX, femtocells, opportunistic
communication, and IP. Internet offloading, however, is
common among them since it is cost-effective for both service
suppliers and customers. In order to test the linkage of Wi-Fi
employing cellular networks to offload, we developed many
scenarios and investigated offloading through Wi-Fi may be a
choice. Our major goal was to use the Stackelberg technique to
get the most offloading over Wi-Fi. We investigated the
Stackelberg approach’s effects on the commercial market. A
model for a game with two levels has been presented. In the
first stage, the MBS gives the APs economic stimulation
depending on the kinds of traffic they utilize, and in the second
phase, the APs take advantage of these impulses by releasing
other forms of traffic. To analyze our efforts, a mathematical
model is created and put through simulation and
experimentation. In order to investigate the impact of
incentives on offloading, we presented simulations for a range
of incentive levels using the simulation program MATLAB. The
performance of the QoS parameter was also examined and
contrasted with that of the current model. The trials
demonstrated showing the suggested model outperforms the
current model, and the proposed approach enabled Nash
equilibria to be reached using the Best Response potential
function. The suggested traffic type-based game also attained a
fully unique equilibrium. The suggested traffic type-based game
is deemed to be objectively better in all significant structures
and operations as a result of these numerical results. Our
methodology achieves maximum offloading and increased
throughput based on the best incentives. In such a two-way
offloading procedure, privateness and security is a contentious
subject for investigation, and research into numerous base
stations and many connected APs with criteria for efficiency
and delivery disparities is the unresolved problem that needs
more study.

References
1. Buffa, S. et al., 5th generation district heating and cooling
systems: A review of existing cases in Europe. Renew. Sustain.
Energy Rev., 104, 504–522, 2019, doi:
10.1016/j.rser.2018.12.059.
2. Stergiou, C.L. et al., IoT-based big data secure management in
the fog over a 6G wireless network. IEEE Internet Things J., 8,
7, 5164–5171, 2020, doi: 10.1109/JIOT.2020.3033131.
3. Youn, J.Y. et al., High-density proximity mapping reveals the
subcellular organization of mRNA-associated granules and
bodies. Mol. Cell, 69, 3, 517–532.e11, 2018, doi:
10.1016/j.molcel.2017.12.020.
4. Chen, S. et al., A survey on user-centric cell-free massive
MIMO systems. Digit. Commun. Netw., 8, 5, 695–719, 2022, doi:
10.1016/j.dcan.2021.12.005.
5. Enkavi, G. et al., Multiscale simulations of biological
membranes: The challenge to understand biological
phenomena in a living substance. Chem. Rev., 119, 9, 5607–
5774, 2019, doi: 10.1021/acs.chemrev.8b00538.
6. Luo, Y. et al., Secure and reliable D2D communications with
active attackers: A game-theoretic perspective and machine
learning approaches, in: Machine Learning and Intelligent
Communications, Proc. 5: 5th International Conference,
MLICOM 2020, Shenzhen, China, September 26-27, 2020,
 Springer International Publishing, pp. 519–533, 2021, doi:
10.1007/978-3-030-66785-6_55.
7. Sarma, S.S. and Hazra, R., Interference mitigation methods
for D2D communication in 5G network, in: Cognitive
Informatics and Soft Computing: Proceeding of CISC, Springer,
Singapore, pp. 521–530, 2020, doi:10.1007/978-981-15-1451-
7_54.
8. Sanguanpuak, T. et al., Radio resource sharing and edge
caching with latency constraint for local 5G operator:
Geometric programming meets Stackelberg game. IEEE
Trans. Mob. Comput., 20, 2, 707–721, 2019, doi: 10.1109/
TMC.2019.2948630.
9. Liang, H. and Zhang, W., A survey and taxonomy of resource
allocation methods in wireless networks. J. Commun. Inf.
Netw., 6, 4, 372–384, 2021, doi: 10.23919/JCIN.2021.9663102.
10. Hashad, O. et al., Resources allocation in underlay device-to-
device communications networks: A reduced-constraints
approach. IEEE Access, 8, 228891–228904, 2020, doi:
10.1109/ACCESS.2020.3046417.
11. Lai, S. et al., Intelligent secure mobile edge computing for
beyond 5G wireless networks. Phys. Commun., 45, 101283,
2021, doi: 10.1016/j. phycom.2021.101283.
12. Cite Guo, F., Yu, F.R., Zhang, H., Ji, H., Leung, V.C., Li, X., An
adaptive wireless virtual reality framework in future
 wireless networks: A distributed learning approach. IEEE
Trans. Veh. Technol., 69, 8, 8514–8528, 2020.
13. Kadam, S., Bhargao, K.S., Kasbekar, G.S., Node cardinality
estimation in a heterogeneous wireless network deployed
over a large region using a mobile base station. J. Network
Comput. Appl., 221, 103779, 2024.

Note

1. *Corresponding author: [email protected]

10
Autonomous Vehicle Optimization:
Striking a Balance Between Cost-
Effectiveness and Sustainability
Vamsidhar Talasila1*, Sagi Venkata Lakshmi
Narasimharaju1, Neeli Veda Vyshnavi1, Saketh Naga
Sreenivas Kondaveeti1, Garimella Surya Siva Teja1 and
Kiran Kumar Kaveti2

1
Department of Computer Science and Engineering, Koneru
Lakshmaiah Education Foundation, Vaddeswaram, Guntur, India

2
Department of Computer Science and Engineering, Vignan’s
Foundation for Science Technology and Research Deemed to be
University, Vadlamudi, Guntur, India

Abstract

Transportation, energy use, and the environment are all

projected to be significantly impacted by the mainstream
deployment of autonomous vehicles (AVs). Even though AVs
promise to minimize collisions, traffic, and pollution, they also
spark questions about their affordability, sustainability, and
societal effects. In this study, we offer a methodology for AV
optimization that strikes a compromise between performance
that is both cost-effective and sustainable. The suggested
strategy minimizes trade-offs in sustainability and cost-effective
effectiveness via offering adaptability, redundant operation,
and adaptive capabilities for handling the logistic cluster.
Analyses of the logistic network showed a 23% reduction in
emissions of carbon dioxide by using the AV technique
combined with consolidation facilities using actual information
through interactions with experts in the food business. Due to
the use of autonomous cars in the supply chain, the trip did not
require any rest stops, which reduced the cost of the entire
logistical network. Legislation and infrastructural concerns,
however, impede the practical application of the established
distribution strategy, delaying the development of a cost-
effective supply chain.

Keywords: Autonomous vehicles (AVs), cost-effective

performance, sustainability, emissions

10.1 Introduction

Transportation is one of the sectors that is being severely

impacted as the globe pushes toward more sustainable
practices. Finding solutions for decreasing these emissions is
essential for reaching sustainability goals because the
transportation sector contributes significantly to greenhouse
gas emissions. Utilizing autonomous vehicles (AVs), which are
designed to strike a balance between sustainable performance
and cost-effectiveness, is one way to address this issue. The
term “autonomous vehicle” refers to a class of self-driving
vehicles that can navigate and perform tasks on roads without
a driver’s input by using sensors, cameras, and sophisticated
algorithms [1]. Being electrically powered, requiring less
maintenance, and being able to maximize routes and speeds to
save fuel consumption, AVs have the potential to significantly
reduce the environmental impact of the transportation sector.
But AVs must strike a balance between sustainability and cost-
effective performance if they are to be a real solution to the
sustainability issue facing the transportation sector. As a result,
AVs must be adjusted to strike a balance between efficiency and
long-term viability, which can be accomplished via several
techniques [2].

Route optimization is one approach to AV performance and

sustainability optimization for cost-effectiveness. AVs can
customize their routes to use less fuel and emit fewer
greenhouse gases by analyzing the current information on
traffic, weather, and road conditions. Due to the decreased trip
time and distance, AVs are now more cost-effective.
Optimization can also lead to cost savings. Energy-efficient
driving is another technique for AV performance and
sustainability optimization. To maximize energy economy, AVs
can be configured to accelerate and decelerate at the best
possible rates and speeds. By doing this, AVs can increase the
performance and longevity of the vehicle while lowering the
consumption of fuel and carbon emissions [3].

The battery management systems of AVs can also be used to

optimize them for sustainability. The longevity of the battery
can be increased and the battery’s need for regular charging
decreased by having AVs keep track of the battery’s level of
charge and rate of discharge. AVs become more cost-effective
because of this optimization, which also lowers their carbon
footprint and maintenance expenses [4]. Vehicle-to-vehicle
communication is an additional way that AVs might strike a
balance between cost-effective performance and sustainability.
AVs can adjust their speed and routes to save on fuel and cut
down on carbon emissions by interacting with other vehicles on
the road. Additionally, this communication can assist AVs in
avoiding traffic jams, which can significantly cut down on fuel
use and carbon emissions [5]. The production and disposal
processes for AVs can also be improved for sustainability. To
lessen their carbon footprint, AV producers should employ
environmentally friendly resources and manufacturing
techniques, such as recycling garbage and reducing waste.
Similar to AVs, AVs may be made to be simple to disassemble
and recycle, minimizing the negative environmental effects of
their disposal [6]. A crucial answer to the sustainability issue
facing the transportation sector is the employment of
autonomous cars that are designed for cost-effective
performance and sustainability. AVs may balance performance
and sustainability through optimization of routes, energy-
effective driving, charge administration, communication
between vehicles, and environmentally friendly production and
disposal methods. To achieve a more sustainable future, it will
be crucial to optimize AVs for sustainability as the globe moves
toward more sustainable practices [7].

Through information sharing, automated and connected cars

(CAVs) have a great potential to increase traffic safety and
efficiency. This study deals with the issue of longitudinal
coordination between two strings of automobiles at highway
on-ramps. For cars entering from the ramp and coming from
the mainline, a rule-based adjusting algorithm is suggested to
achieve a close-to-ideal merging sequence [8]. For the
realization of an entirely autonomous system (vehicle and
charger), wireless charging technology is a great fit for self-
driving electric vehicles. This article presents a planned
optimization study for a pilot concept of in-route mobile
charging infrastructure servicing corrected-route immediately
shared automated electric shuttles (SAESs) in Greenville, South
Carolina, USA. The formulation of a single-objective nonlinear
combined integers system planning optimization problem was
presented. The whole induction-charged SAES system is
represented by a detailed cost model that takes into account the
costs of the installation, traction battery, power electronics, and
road construction [9]. By enhancing travel mobility,
autonomous vehicles are opening up new possibilities for the
management and operation of transportation options in cities.
They suggest using AVs in this study to address the first-mile
(FM) problem, which entails getting people from their houses to
metro stations. A fleet operator dispatches AVs and sets up
rideshare in a rolling horizons framework once passengers
submit their travel requests in advance. One AV is assigned to
fulfill several requests following the vehicle’s capacity, its
maximum trip time, and its accessibility restrictions [10].

As more and more massive Internet applications are employed,

energy costs in data center networks have skyrocketed. The
interaction latency between two automobiles and vehicles and
the controller in technology-defined vehicular networks
(SDVNs) will sharply increase as the number of vehicles rises.
To reduce latency, additional controllers are needed to offer
communication services. High energy expenses are a result of
more controllers. As a result, the issue of controller placement
(CPP), or the quantity and positioning of controllers, needs to be
addressed [11]. The DC nano grid with efficient energy
management has recently received a lot of interest. In the
future decades, it is anticipated that the DC nano grid could
improve energy effectiveness and lessen carbon footprint due
to its propensity to incorporate developing power electronics
burdens, such as solar energy and electric vehicles (EVs), as
well as renewable energy sources. In this study, a leadership
structure based on a smart charging point is introduced that
might be utilized to develop a small, inexpensive, and
straightforward energy administration system for a nano grid
[12]. Enhancing energy efficiency has been acknowledged as a
successful strategy for promoting energy conservation,
reducing emissions, and realizing sustainable development.
Numerous and dispersed methods of enhancing energy
efficiency toward environmentally friendly production are
inadequate in terms of overall approaches and system
integration. Following a thorough review of academic
literature, 166 research papers specifically focused on
enhancing energy performance are examined. To understand
the pattern of energy use and implement efficient energy-
saving measures, a thorough examination and analysis of
energy surveillance, assessment, optimization, and
benchmarking are conducted [13]. Autonomous vehicles need
real-time performance to reach their present position in time
for decision-making, while real-time localization is crucial for
AVs to achieve safe, effective driving. There has not been a
review paper that compares the real-time performance of
various localization approaches using a variety of hardware
systems and programming languages or that examines the
connections between localization methodologies, real-time
performance, and correctness. As a result, this study examines
the most recent localization methods and evaluates how well
they function in AV applications. This paper first proposes an
equivalent comparison method based on the localization
method operations capability (LAOC) to evaluate and analyze in
depth the links between methodologies, computational
difficulty, and accuracy after evaluating the relative
computational cost of several localization approaches [14]. The
development and implementation of automated and connected
cars (CAVs) must include testing and assessment, but there is no
organized structure for creating testing scenario libraries. A
broad architecture for the testing scenario library generation
(TSLG) issue with various operating design domains (ODDs),
CAV models, as well as performance metrics is what this paper
intends to give. The evaluation scenario library is described as a
crucial collection of situations that can be used for CAV testing,
given an ODD [15]. For systems that assist drivers to be
effective, traffic participants must be detected. The bulk of
pedestrian fatalities that could have been avoided happened at
night, according to data on traffic safety. Sensors like cameras
may malfunction at night due to a lack of light [16]. The
performance of several industries and enterprises, including
the transportation industry, is being improved—thanks to the
rapid pace of artificial intelligence (AI) development. The
improvements brought forward by AI include extremely
sophisticated computational techniques that resemble how the
human brain functions [17]. Researchers introduce L3-Net, a
new starting-to-learn-based LiDAR localization system with
centimeter-level localization accuracy, which is similar to
earlier state-of-the-art systems with manually built pipelines.
Instead of depending solely on these manually created modules,
they creatively implement the utilization of multiple deep
neural network topologies to provide a learning-based strategy
[18]. Today, it is thought of as a highly strategic goal to create
effective fuel cell hybrid vehicles (FCHVs), to have them
deployed as widely as possible across the entire transportation
sector, and to fully satisfy well-known ecological and regulatory
constraints on a global scale [19]. The growing volume of traffic
is destroying cities all around the world. Smart cities are a
recent development that is viewed as a successful approach to
dealing with several serious urban issues like traffic, pollution,
energy usage, and waste treatment [20].

10.2 Methods

A depth analysis was done to maximize the advantages of

autonomous vehicle distribution. Unfortunately, the current
industry does not employ autonomous vehicles, making it
impossible to gather empirical data through the use of
established methods like surveys or interviews. Real-world data
obtained from the most successful distribution strategies in
conjunction with cutting-edge technologies must be thoroughly
analyzed. A computer simulation was run using the data
collected to determine the potential levels of efficiency for the
suggested competitiveness approach. The research was adopted
to produce quantitative proof to support the suggested
framework. Because simulations offer precise constraints and
assumptions, using a contribution to the scientific literature
was still legitimate—thanks to a computation method created
by a computer. Additionally, the chosen method allowed for the
evaluation of multiple scenarios, which was not possible with
qualitative methods. The information was used to model a
dynamic logistic network. The Monte Carlo (MC) simulated
annealing process was used to simulate partial goods logistics,
choosing the optimal routes and suppliers at random. The
possible advantages of autonomous vehicles were learned from
corporate sources and other publications. The final paragraph
of the method reveals a more thorough step-by-step
breakdown.

10.2.1 Competition and Sustainable Supply Chains at

Odds

The introduction of manufacturing lines to the sector is linked

to the industrial age, which began in Britain. The need for a
labor force has altered as a result of the invention of machines
and motors. Job workshops and batch manufacturing were the
two main methods of production employed in the past. The
laborer must be skilled and specialized in a variety of activities
during that time. The working population was forced to become
an expert in particular objectives and devote more attention to
single jobs as a result of the industrial revolution’s increased
usage of mass manufacturing on the market. The fourth
industrial revolution, often known as Industry 4.0, is taking
place right now. It is known as the digital revolution. However,
it is crucial to comprehend what brought about the industry’s
drastic change before analyzing the unique characteristics of
the modern industrial revolution. Because manufacturing costs
were reduced by the development of new inventive
technologies, consumers began to demand more individualized
production. It made it feasible for people to buy a variety of
goods for a low price, which was before impossible. Due to the
abundance of supply for various products, customers adopted a
demanding rather than suppliant stance. Customers began to
request more customized products for their use, thus the
manufacturing process had to change once more. In earlier
times, when customers did not have the option of choosing
from a wide range of manufacturers, mass production was
advantageous. As a result, the current industrial revolution
offers means of producing more customized goods. On the one
hand, Industry 4.0 empowers producers to provide consumers
with what they want. Cycle times are getting shorter and
customer order amounts are dropping. Short-notice orders are
being sent to production in increasing numbers. On the
opposite hand, investments are necessary to match the
technology with what businesses are already doing. The price of
the product may rise as a result of the incorporation of cutting-
edge technologies into business operations, which could result
in the loss of the competitive advantage. The cost of production
makes up a very small portion of the product’s total cost.
Marketing, sales, distribution, and customer service are further
factors. The distribution issue is the primary emphasis of this
essay. Another important component of current consumer
demand is distribution. Internet access and e-commerce have
made acquiring goods even simpler, and it is typical to see
customers buying goods from different continents. As a result,
the local market must now contend with the global market.
There is just one market in the globe, according to some, and a
regional or global market no longer exists. Naturally, not every
industry is covered by this. The worldwide distribution of the
quickly paced consumer goods sector serves as the best
illustration. Low distribution costs are necessary to sustain a
competitive edge, but product demand additionally relies on
the price of the product. In addition, the manufacturing and
delivery lead times are highly important. Present-day
customers want their items immediately; therefore, careful
demand planning is necessary. Furthermore, due to trends,
shifting consumer demand, or expiration dates, many items
cannot be kept on hand for an extended period. Small
production orders are a significant problem in the quickly
evolving consumer goods sector.

10.2.2 Industry’s Effect Application of the 4.0 Idea to

Business Operations

The emergence of the fourth industrial revolution has

significantly altered business operations. To maintain
sustainable and low management, an effective competitiveness
strategy must be developed while simultaneously implementing
the Industry 4.0 idea. Businesses that fail to adapt to the
changing environment will go out of business. The notion of
Industry 4.0 has multiple components. Fundamentally, it
implies that everything is interconnected and that total
connection and openness of data and information are crucial.
More sophisticated business decisions have been made as a
result of this type of technical progress, including flexible
business operation adaptation and flexible human resource
planning. Only via a creative integration of all these
technologies can Industry 4.0’s advantages become apparent.
The Internet of Things (IoT) technology is one of the
cornerstones of Industry 4.0 principles. The IoT is a
communication innovation that connects everything using
several technologies and makes it easier for communication
devices to connect. The concept of the Internet of Things opens
significant opportunities for visualizing supply chain operation
procedures. With the use of this method, real-time business
activity scheduling, monitoring, and control are all possible for
food enterprises. It can help them reduce food waste, manage
risk from unforeseen deviations, and keep up with food safety
laws. The Internet of Things idea can be used when planning a
transit system. Big data analytics is the flip side of the Industry
4.0 notion. To find new patterns or eliminate weaknesses, data
analysis is important; simply collecting data from the
environment is not sufficient. Utilizing the data that are
currently available effectively provides a thorough overview of
business operations. Future items can be completely suited to
the needs of the customer—thanks to the study of the client’s
wants. Descriptive, predictive, diagnostic, and prescriptive
statistics are some of the disciplines that make up big data
analytics. Making smarter decisions can only be assisted by
real-time information. Given that agricultural products have a
limited shelf life, this is especially useful. Diagnose statistics is
used to model different business processes and provide a
statistical justification for how the market could influence the
choices made. Finally, everything needs to be condensed and
translated into business terms. The notion of Industry 4.0 needs
to be modified for operational and commercial strategy. Big
data models for process control, for instance, can be used to
build the most effective strategies for enhancing the
environmental effectiveness of sustainable supply chain
management and for managing natural resources sustainably.
Even larger advancements have resulted from the concepts of
big data analysis and IoT. It is now possible to use autonomous
vehicles in the industry. Figure 10.1 depicts the overall
competitiveness approach. There are four steps in the method.
Operations of those involved in the supply chain must be
defined and visualized during the initial phase. The activities
include several distinct components, including purchasing,
warehousing, and several modes of transportation. A more
thorough organization study should be carried out during the
second stage to identify the willing businesses that would be
most suited to create a logistical cluster. The entire supply chain
should have its activities computerized during the third phase.
The management strategy’s main considerations should be the
potential for gathering, processing, and utilizing information.
Key components of the cluster should be continuously assessed
and watched during the final phase. The cluster can achieve
higher levels of cost-effective efficiency and long-term viability
by incorporating self-driving cars and big data analysis into
organizational procedures. Big data analytics and autonomous
vehicles both can improve flexibility and redundancy in the
management process. Continuous analysis of the strategy would
develop adaptive capabilities, which would aid in making
better decisions in the present. Since autonomous vehicles can
sense their navigation without input from humans, they are
more environmentally friendly than traditional cars. As a
result, costs, emission levels, and working hours can all be
significantly decreased. Aside from long distances, autonomous
vehicles can also be used at the operational level of the supply
chain. Some nations have acknowledged the invention of
autonomous vehicles as a significant economic growth factor.
Recently, regulators in the USA said that the artificial
intelligence system used by Google Cars would qualify as a
driver under federal law. The introduction of driverless
vehicles has many potential advantages. Transport makes up a
sizable portion of the supply chain. There is a chance that the
cargo will be harmed during shipment. Therefore, the use of
self-driving cars can lower the amount of damaged goods.
According to statistics, drivers are to blame for up to 90% of
traffic accidents. On the other hand, because less fuel is used,
there is less of an influence on the environment. Furthermore,
by consistently driving short distances, which is not always
achievable with regular ground vehicles, the amount of fuel
used can be further decreased. It is difficult to quantify the
possible environmental advantages, though. There is currently
no method of fuel economy testing that can accurately assess
fuel economy ratings, the energy consumption and
environmental damage caused by light automobiles. When
traveling in convoys, preliminary fuel economy studies have
shown that fuel efficiency can increase by up to 10%. The
current economy testing process lacks the means to assess how
autonomous vehicle technologies would affect fuel efficiency
ratings. As a result, it is necessary to perform more precise
calculations for the exact autonomous vehicles. The evaluation
of the literature, however, can show that autonomous vehicles
have positive effects on the economy and the environment. The
utilization of transport is a key factor in autonomous vehicle
development. The trucks are typically not utilized well. After
all, the rest period is also deleted when the driver is removed.
Additionally, there is a chance for the logistic cluster members
to share the autonomous vehicles. Shared autonomous vehicles
may offer affordable transportation options. A logistic cluster
could use the sharing technique mentioned in the paper as well.
Additionally, combining autonomous drive technologies with
big data analysis, cloud computing, and the Internet of Things
can increase safety on the roads while making transport more
easily accessible to customers and suppliers. The cloud-based
option enables the storage of far more past information than an
independent system and provides sufficient accuracy to
comprehend trends in fuel usage and to help users make deft
judgments. Analysis tools for cloud computing allow for the
conversion of fuel use to expenses or CO2 emissions, the
analysis of energy use per tonne used in production, and the
monitoring of energy consumption by each vehicle.
Figure 10.1 Framework for logistic clusters that limits supply chain management
trade-offs.

10.3 Results

The data used to create the logistic network model for the food
industry were obtained through professional interviews. Fifty
Lithuanian businesses involved in the continent’s food industry
took part in the interview. Invoices and other significant data
have been acquired, and several supply chain-related questions
have been raised. A significant portion of the interview’s results
were previously published. Table 10.1 shows a sample of the
collected data. The primary finding from the conversation was
that tiny markets frequently use partial goods. Typically,
businesses have between 5 and 20 pallets of cargo that need to
be imported or exported. The cost per pallet when shipping a
full truck is roughly 50 euros; however, when transporting
lower quantities, the cost per pallet might increase three to four
times. These findings compel the establishment of logistic
clusters of small- and medium-sized businesses, which will
share resources like transportation and information. For some
sectors, partial goods transit might be a challenge. In this
instance, we emphasized the results in light of the food
business. The food sector was selected as a reference sector
because it frequently uses temperature-controlled
transportation methods, which have a greater negative impact
on the environment than would other industries Figure 10.2.
Table 10.1 Example of demand and supply data from expert interviews with the food
industry.

Quantity
Loading of cargo, Delivery Required
country-city Euro countrycity temperature
pallets

Belgium-Geel 32 Lithuania- Chilled

Šiauliai

Belgium-Aalst 15 Lithuania- None

Vilnius

Belgium-Alken 8 Lithuania- None

Vilnius

Belgium-Alken 11 Lithuania- None

Vilnius

Belgium- 9 Lithuania- Chilled

Kortrijk Kaunas
 Quantity
Loading of cargo, Delivery Required
country-city Euro countrycity temperature
pallets

Belgium- 13 Lithuania- Chilled

Antwerp Šiauliai

Germany-Kiel 23 Lithuania- None

Šiauliai

Germany-Löhne 11 Lithuania- None

Šiauliai

Germany- 18 Lithuania- Chilled

Bielefeld Šiauliai

Germany- 17 Lithuania- None

External Šiauliai

Germany-Melle 16 Lithuania- Chilled

Šiauliai
 Quantity
Loading of cargo, Delivery Required
country-city Euro countrycity temperature
pallets

Germany-Lügde 12 Lithuania- None

Šiauliai

Netherlands- 16 Lithuania- Frozen

DeKwakel Šiauliai

Netherlands- 12 Lithuania- None

Veenendaal Šiauliai

Netherlands- 32 Lithuania- Frozen

IJsselmuiden Šiauliai

Netherlands- 5 Lithuania- Frozen

Purmerend Šiauliai

Netherlands- 6 Lithuania- None

Veldhoven Šiauliai
 Quantity
Loading of cargo, Delivery Required
country-city Euro countrycity temperature
pallets

Netherlands- 12 Lithuania- Chilled

Helmond Šiauliai

UnitedKingdom- 15 Lithuania- Chilled

London Vilnius

UnitedKingdom- 16 Lithuania- Frozen

London Vilnius

UnitedKingdom- 12 Lithuania- Chilled

Aylesford Kaunas

UnitedKingdom- 16 Lithuania- None

London Kaunas

UnitedKingdom- 16 Lithuania- None

Bristol Kaunas
 Quantity
Loading of cargo, Delivery Required
country-city Euro countrycity temperature
pallets

UnitedKingdom- 16 Lithuania- None

Cambridge Kaunas

Figure 10.2 Illustrates the assumptions of the logistic network model.

Only the immediate consolidated storage facility and partial

freight collecting distribution procedures are considered in this
model, which starts with the first warehouse sites. The initial
one is timely, which mimics direct distribution from suppliers
to the end destination by removing warehouses from the
model. The second tactic is known as consolidation warehouses,
in which merchandise is collected from the area and taken to
the closest warehouse before being transported from the
warehouse to a central warehouse in Lithuania. Last but not
least, the simulated annealing principle is used to program the
partial freight optimization strategy. The CO2 emission level
computation was the optimization objective; transportation
costs, warehouse management costs, and warehouse building
costs were also taken into account. Based on the journey
distances, in this study, the amount of emissions brought on by
supply chain transportation has been independently
determined.

The model is best expressed as follows:

(10.1)

where UDD stands for warehouse building costs, and SMD

stands for overall network expenses. UGD is the cost of
managing the warehouse, and SD is the cost of transportation.
Given the variety of situations, L is the warehouse index (L = 1,
2, 3) and OR is the number of pallets. EDQ stands for energy
consumption rate, C for distance, and S for a type of
temperature.
 (10.2)
where FK stands for emission level, BF for average CO2
emission, C for distance, FEDQ for fuel consumption of engines,
SEDQ for fuel consumption of heat, and SS for travel time.

(10.3)

where BT denotes the mean speed, and CJX denotes an operator

that only leaves whole numbers.

The obtained emission level data were then updated on the

assumption that the cargo was consolidated onto one truck. The
path with the lowest CO2 emissions has been chosen. The
optimal route that was chosen has been eliminated from the list
of suppliers. The procedure was carried out repeatedly until the
best routes were allocated to all of the suppliers in the supplier
set. The suppliers have been individually added to the end
objective functions without being combined, with the
understanding that they have been withdrawn without being
assigned to specific routes. Due to optimization limits, some
suppliers cannot be consolidated. For instance, there are
restrictions on distance, a truck’s maximum load capacity is 32
containers, the transit time cannot be longer than 30% of the
time required for direct delivery, etc. In both situations, the
traditional system of distribution and the independent
distribution system, the optimum paths for every supplier have
been chosen, after this method has been used 50,000 times
(Figure 10.3).

The usual consumption rates and working hours have been

used to examine these three situations and two options. During
the first approach, a traditional system of distribution has been
simulated. The second method involved modeling an
autonomous distribution system. The distribution model for
autonomous vehicles has been simulated by varying fuel
consumption and travel times. The simulation’s findings
demonstrated that shorter lead times and more effective
utilization of fuel rates were made possible by convoy driving.
Convoy driving also makes better use of available road space
and minimizes the demand for parking. Instead of 50 meters
between each connected vehicle in a platoon, only 15 meters
are needed. The aerodynamic drag is significantly reduced by
this much shorter distance. Although it is difficult to predict the
extent to which fuel consumption and CO2 emissions might be
reduced by up to 10% by deploying autonomous vehicles in
actual use.
Figure 10.3 Distribution model simulations with simulated annealing.

The outcomes of simulations based on selected scenarios are

displayed in Table 10.2. In essence, the findings indicate that
distribution in a timely approach has the highest transportation
expenses, but there are no costs associated with building or
managing warehouses. Thus, this kind of strategy, which is
more focused on individual businesses than on clusters, is
suitable if there is a shortage of available land or starting
money. In comparison to the just-in-time case, the
transportation costs for the consolidation warehouse are
reduced by 19%, and the costs associated with the partial goods
gathering strategy are reduced by 41%. The partial goods
scenario has the greatest CO2 emission level, whereas the
consolidation warehouse case has the lowest. This is a result of
the driver’s rest period, which must be at least 10 hours long
after each 90 minutes of driving. CO2 is constantly generated
during the period of rest because the caravan’s temperature
control must be maintained so that the amount of time left to
the destination remains constant. This scenario can be
explained by the longer trip duration.
Table 10.2 Provides an analysis of the logistic network situation.

Independent distribution system C

Type Just-in- Distribution A Ju

sequence(JIS) center portion s
of the
cargo

Transport 142 006 110 151 102 1

expenses, € 555

CO2 level, kg 141 856 91 182 3 911 1

Warehouse 0 341 550 207 0

construction 942
costs, €

Travel time, 1958 1 074 3081 3

hours
 Independent distribution system C

Warehouse 0 10 263 3 904 0

management
costs, €

10.4 Conclusions

According to research, using consolidation warehouses in

supply chains can lead to a more environmentally friendly
system with just-in-time distribution, which has CO2 emissions
that are 55% higher and transportation costs that are 19%
lower. While larger markets with larger order quantities should
concentrate on the just-in-time distribution strategy to optimize
their distribution systems, this approach is more appropriate
for smaller cargo quantities. Although the cost of transport may
be reduced by 18.4% as a result of this strategy, CO2 emissions
may rise by 43%. By establishing logistics clusters to share
transportation and use autonomous vehicles, this problem can
be solved while cutting transportation costs and CO2 emissions
by 5% and 23%, respectively, in comparison to conventional
distribution models. Industry 4.0, big data analytics, and the
Internet of Things implementation can aid in demand
forecasting, cargo optimization, and food waste reduction while
boosting the cluster’s adaptability and competitiveness.
Autonomous cars can be utilized to distribute goods in huge
volumes across vast distances using consolidation warehouses
and regionally gathered partial goods. The study notes several
theoretical contributions, such as the significance of a
methodical approach to technology use, the requirement for
information utilization skills in management, and the
contribution of organizational factors to technical advantages.
The study’s quantitative evidence, however, is restricted to the
system’s complexity and does not take into account outside
variables like policies or economic trends. Therefore, to
validate the suggested strategy, further study should adopt a
complexity theory perspective.

References

1. Zhang, S. et al., Self-sustaining caching stations: Toward cost-

effective 5G-enabled vehicular networks. IEEE Commun.
Mag., 55, 11, 202–208, 2017, doi: 10.1109/MCOM.2017.1700129.
2. Zhong, Z. et al., Multiobjective optimization framework for
cooperative adaptive cruise control vehicles in the
automated vehicle platooning environment. Transp. Res. Rec.,
2625, 1, 32–42, 2017, doi: 10.3141/2625-04.
3. Shahian Jahromi, B. et al., Real-time hybrid multi-sensor
fusion framework for perception in autonomous vehicles.
Sensors (Basel), 19, 20, 4357, 2019, doi: 10.3390/s19204357.
4. Jones, E.C. and Leibowicz, B.D., Contributions of shared
autonomous vehicles to climate change mitigation. Transp.
Res. D, 72, 279–298, 2019, doi: 10.1016/j.trd.2019.05.005.
5. Feng, S. et al., Safety assessment of highly automated driving
systems in test tracks: A new framework. Accid. Anal. Prev.,
144, 105664, 2020, doi: 10.1016/j.aap.2020.105664.
6. Hao, M. and Yamamoto, T., Shared autonomous vehicles: A
review considering car sharing and autonomous vehicles.
Asian Transp. Stud., 5, 1, 47–63, 2018.
7. Phan, D. et al., Intelligent energy management system for
conventional autonomous vehicles. Energy, 191, 116476,
2020, doi: 10.1016/j.energy.2019.116476.
8. Ding, J. et al., A rule-based cooperative merging strategy for
connected and automated vehicles. IEEE Trans. Intell. Transp.
Syst., 21, 8, 3436–3446, 2019, doi: 10.1109/TITS.2019.2928969.
9. Mohamed, A.A.S. et al., Planning optimization for inductively
charged on-de-mand automated electric shuttles project at
Greenville, South Carolina. IEEE Trans. Ind. Appl., 56, 2, 1010–
1020, 2019, doi: 10.1109/TIA.2019.2958566.
10. Chen, S. et al., Solving the first-mile ridesharing problem
using autonomous vehicles. Comput. Aided Civ. Infrastruct.
 Eng., 35, 1, 45–60, 2020, doi: 10.1111/mice.12461.
11. Lin, N., Zhao, Q., Zhao, L., Hawbani, A., Liu, L., Min, G., A
novel cost-effective controller placement scheme for
software-defined vehicular networks. IEEE Internet Things J.,
8, 18, 14080–14093, 2021.
12. Yu, H. et al., Towards energy-efficient and cost-effective DC
nanaogrid: A novel pseudo hierarchical architecture
incorporating V2G technology for both autonomous
coordination and regulated power dispatching. Appl. Energy,
313, 118838, 2022, doi: 10.1016/j.apenergy.2022.118838.
13. Cai, W. et al., A review on methods of energy performance
improvement towards sustainable manufacturing from
perspectives of energy monitoring, evaluation, optimization
and benchmarking. Renew. Sustain. Energy Rev., 159, 112227,
2022, doi: 10.1016/j.rser.2022.112227.
14. Lu, Y. et al., Real-time performance-focused localization
techniques for autonomous vehicle: A review. IEEE Trans.
Intell. Transp. Syst., 23, 7, 6082–6100, 2021, doi:
10.1109/TITS.2021.3077800.
15. Feng, S. et al., Testing scenario library generation for
connected and automated vehicles, part I: Methodology. IEEE
Trans. Intell. Transp. Syst., 22, 3, 1573–1582, 2020, doi:
10.1109/TITS.2020.2972211.
16. Li, G. et al., Detection of road traffic participants using cost-
effective arrayed ultrasonic sensors in low-speed traffic
situations. Mech. Syst. Signal Process., 132, 535–545, 2019, doi:
10.1016/j.ymssp.2019.07.009.
17. Abduljabbar, R. et al., Applications of artificial intelligence in
transport: An overview. Sustainability, 11, 1, 189, 2019, doi:
10.3390/su11010189.
18. Lu, W. et al., L3-net: Towards learning based lidar
localization for autonomous driving, in: Proc. IEEE/CVF
Conference on Computer Vision and Pattern Recognition, pp.
6382–6391, 2019, doi: 10.1109/CVPR.2019.00655.
19. Sorrentino, M. et al., Development of flexible procedures for
co-optimizing design and control of fuel cell hybrid vehicles.
Energy Convers. Manage., 185, 537–551, 2019, doi:
10.1016/j.enconman.2019.02.009.
20. Seuwou, P. et al., The future of mobility with connected and
autonomous vehicles in smart cities, in: Digital Twin
Technologies and Smart Cities, pp. 37–52, 2020,
doi:10.1007/978-3-030-18732-3_3.

Note

1. *Corresponding author: [email protected]

11
Adapting Underground Parking
for the Future: Sustainability and
Shared Autonomous Vehicles
Vamsidhar Talasila1*, Madala Pavan Pranav Sai1, Gade Sri
Raja Gopala Reddy1, Vempati Pavan Kashyap1, Gunda
Karthik1 and K. V. Panduranga Rao2

1
Department of Computer Science and Engineering, Koneru
Lakshmaiah Education Foundation, Vaddeswaram, Guntur, India

2
Department of Computer Science and Engineering, Sree Vahini
Institute of Science and Technology, Tiruvuru, India

Abstract

A significant field of worldwide science and technical

innovation is the creation of shared autonomous vehicles
(SAVs). SAVs’ ability to drastically minimize the requirement for
parking is one of its most coveted qualities in urban areas.
Particularly in densely populated metropolitan areas, the
town’s underground parking space (UPS) is presently the
biggest fixed-traffic area. The requirement for UPS renewals
will become more urgent in the SAV situation. However, due to
its unique form attributes, which are significantly constrained
by outside factors, the UPS is challenging to renew, requiring
aimed techniques and tactics. To gather information for this
study, fieldwork was initially done to examine the geographical
morphology and operating circumstances of typical UPSs in
various parts of Ningbo. The time series assessment technique
and functional substitution choice models for the sustainable
renewal of subterranean parking were developed dependent on
the driver status reaction and the multiple-purpose feature
models. The study also analyzes the best design approaches for
combining common examples’ spatial features with their
operational substitution objectives. The findings will offer
scientific direction for builders’ and city designers’ upcoming
designs in SAVs.

Keywords: Area of underground parking, autonomous vehicle

shared, revitalize methods, design techniques

11.1 Introduction

The number of automobiles on our roadways is growing in

tandem with city expansion and urbanization, which causes
traffic jams, air pollution, and other environmental and social
issues. Parking issues in crowded metropolitan areas have been
successfully solved by underground parking buildings.
However, underground parking facilities are now facing novel
possibilities as well as obstacles as the transportation sector
changes in response to the introduction of discussed
autonomous vehicles (AVs) [1]. We must modify subterranean
parking infrastructure to meet the demands of shareable and
driverless vehicles if we are to create a sustainable future. By
integrating technology and designing parking structures with
flexibility in mind, we can create parking facilities that are
more efficient, cost-effective, and sustainable [2].

Implementing dynamic parking systems, which enable many

people to utilize a single parking space at various times, is a
way to adjust to the demands of shared vehicles. Automatic
parking structures that arrange and shift cars as required or
shared parking applications that let users book and share
parking spaces can also be used to do this. Designing parking
lots with the adaptability to accept the various sizes and forms
of these cars is another option to satisfy the demands of
autonomous vehicles [3]. This can be done by establishing
modular, readily reconfigurable parking spaces or by
constructing parking spaces with flexible proportions. In
addition to these modifications, modifying subterranean
parking facilities for shared and driverless cars has positive
financial and ecological consequences. We can lessen the
negative effects of urban mobility on the natural world and
save money on infrastructural expenditures by minimizing the
demand for parking spaces and maximizing the use of existing
facilities [4, 5].

In this article, we will look at how subterranean parking

garages may be revitalized to better meet the demands of
shared and autonomous vehicles and, eventually, build an
environmentally friendly transport network for cities. We will
look at how technologies may be incorporated into these
facilities, the benefit of multiuser parking, and the possible
financial and ecological advantages of making these changes.
Underground parking garages have a chance to play a
significant role in the development of a more sustainable and
effective urban transportation system by adjusting to the
requirements of shared and self-driving vehicles. The way we
view urban transportation has the potential to change as a
result of shared and autonomous cars. These cars can assist in
lowering the need for private automobile possession, cutting
down on emissions of carbon dioxide, and improving the
atmosphere. However, we also need to reconsider how we
approach parking if we want to fully realize these advantages.
The structure of the paper is as follows. The associated work is
shown in Section 11.2. Algorithms for the methodology are
described in Section 11.3. The analysis produced is shown in
Section 11.4, and the conclusion is presented in Section 11.5.

11.2 Related Works

The paper [6] examines the current literature on the benefits of

the renewal of urban public parking spaces and suggests the
four primary problems entailed: the amount to get renewed
(i.e., forecasting demand assessment), when to renew (i.e.,
update time series evaluation), what to renew (i.e., function
replacement decision), as well as how to keep up (i.e., design
empirical research). Additionally, it presents a rough idea for
and develops a research framework for beneficial renewal
techniques for parking spaces within the shared autonomous
vehicle (SAV) situation. To make cities smarter and to shape the
future of urban travel, this presentation will present a cutting-
edge examination of several connectivity projects. These are
both newly developed processes with promoting acceptance
thus far and a far greater ability to assist in a shift to an
improved transportation framework or they remain in an early
phase of growth and have not yet been accepted as potent
processes that could change travel behavior norms [7]. The
paper [8] provides an overview of the historically cars-oriented
design and the world’s century-long capitulation to vehicle
culture. It then goes into further detail about various self-
driving technology and how they could affect how cities look.
The approach is used in a case study on Regensburg, a
historically significant German city. Findings indicate that the
dedicated lane for buses causes a pollution decrease of 3.25% to
6.65% and a mode shift of approximately 1.6% from auto modes
to public transportation (PT). It has been determined that a
shuttle service enhances the current PT system by boosting the
volume of PT trips in the service region [9]. The study [10]
shows how policies have a significant impact on the
transportation sector’s supply, but human expectations also
have a role in determining the framework in which the sector’s
demand will operate. The efforts that have been examined
emphasize the need for multifaceted sustainable transportation
measures to reduce dependence on automobiles and
accomplish broader sustainability objectives. The study [11]
examines the interaction in Korea between the general
pollutants in the air and the behavior of Bike-Sharing System
(BSS) riders. They investigated the relationships between
exposure to airborne particles and BSS enrollment across time.
The study [12] aims to explain the current state of the European
cooperation programs that concentrate on the growth of the AV
industry and to illustrate in detail the difficulties faced by
urban planners and auto manufacturers. A competitive plan is
taken into account in the article, as it supports sustained and
effective cost control. The chain of supply has to change and
transition from conventional to more environmentally friendly
approaches. Long-term competitive edge may be maintained
with the use of the Internet of Things, big data analytics, online
computing, cyber-physical networks, autonomous cars, and
logistic clusters [13]. The paper [14] aims to give a mechanism
for future studies to employ realistic flow models to achieve
more precise forecasts on SAV benefits. In this study, a based-
on-events approach for including SAV actions in current traffic
simulation frameworks is presented. The study [15] uses
current taxonomy types from literature to describe the
generally defined AV fleet administration problem, adds new, or
more nuanced, dimensions to existing taxonomic categories,
and introduces new taxonomy types to categorize particular AV
fleet management obstacles.

11.3 Methodology

11.3.1 Framework for Research

First, a field study is used to gather information for the study on

the physical characteristics and operational circumstances of
underground parking spaces (UPSs). According to these
characteristics, the collected items were categorized. Following
that, a technique to evaluate the time frame of UPSs within the
SAV situation is developed for analyzing the creating time series
of various kinds of UPSs, and a library for operating substitutes
is developed by professional opinions and literature research.
The attribute-based multiple goals-making choices model
determines the future role of each type of UPS. The topic of
sustainability is finally brought up about renewal methods.

11.3.2 Area of Research

To speed up SAV in real-world applications, China is working to

create modern AI technologies. China’s east coast is home to the
sizable town of Ningbo, which has an advanced economy and a
dense population. Digital equipment is widely embraced in
Ningbo. Meanwhile, the use of SAVs will be more appealing
given the elevated level of automobile ownership and lack of
parking spaces.

11.3.3 Collection of Data

The geographical makeup, architectural kinds, and uses of

structures above the ground all have an immediate impact on
UPS’s renewal. Additionally, UPS service circumstances in
various parts of a city may vary. Consequently, we carried out
direct UPS inquiries across multiple Ningbo urban areas to
gather more complete and varied data. The final sample
information gathered includes several indicative regions in
each of Ningbo’s seven administrative regions, including
Haishu, Jiangbei, Beilun, Zhenhai, Yinzhou, Fenghua, and
Xiangshan. Industrial shop structures, workplaces, lodgings,
and other structures with above-ground uses are examples. In
Ningbo, 200 samples of UPSs were eventually collected.
Graphics were also used to assess the structure of their area.

11.3.4 Analysis of Clusters

Various UPS groups were autonomously created by two-step

clustering based on the characteristics of the UPS. Additionally,
the primary distinctions between the categories were made
clear. Each category’s renewal period and functionality
substitution are covered individually. The benefit of spreading
in two steps is that it allows for concurrent spreading using
continuous and categorized factors and can calculate an ending
amount of categories. The two phases of the clustering method
are the creation of a cluster features tree and the merging of the
leaf nodes using a clustering algorithm.

11.3.5 DSR Model

The DSR model includes solid principles for management and is
strongly tied to ecologically sustainable urban development
objectives. The study enhances the component arrangement in
the current DSR layout through the uniqueness of the UPS. The
updated model highlights the crucial part capital area renewal
plays in sustainable growth while maintaining the rational
connection among indices. One of them, “Drivers” (D) refers to
the requirement for the renovation of the parking space as a
result of external factors, such as city evolution; “Status” (S)
refers to the present state of the parking space construction,
surroundings, underground space, and natural surroundings,
expressing the area’s potential for revitalization; and
“Response” (R) refers to the UPS owners’ views on
reconstruction. The study improves and establishes the
requirements, elements, and signal levels associated with the
UPS renewal by the level of execution. The surroundings
(“location, use of the ground building, mix of uses, and spatial
connectivity”), the geometry arrangement (”number of floors,
the scale of UPS, structural span, space height, and coverage of
ground building”), and effective features (i.e., “whether it is a
civil air defense basement”) are all considered driver
indicators. The kind of UPS property owner is what defines the
response index. Table 11.1 contains a list of each indicator’s
nature and particular meanings.
Table 11.1 Index of the driver, status, and response model system.

Standards Nature Index Factors

Responses Unfavorable UPS’s rights Renewal

(RE) of ownership desire
and their
nature (RE1)

Status (ST) Unfavorable Number of Spatial

floors (ST1) structure

Favorable Size of UPS

(ST2)

Favorable Structure’s
length (ST3)
Standards Nature Index Factors
Favorable Area height
(ST4)

Favorable Proportion of
ground-to-
building
exposure
(ST5)

Favorable Underground Functional

defense from attributes
the air (ST6)

Drivers Unfavorable Parking Operation

s(DR) charge price status
(DR1)
Standards Nature Index Factors
Unfavorable Parking lot
rate of
change (DR2)

Favorable Location The

(DR3) atmosphere
outside

Favorable Use of
groundlevel
floor (DR4)
 Standards Nature Index Factors
Favorable Several
functions
(DR5)

Favorable Connection
(DR6)

Every index is a measure’s assessment level and is split into two

groups and given 1 or 2 points by the pertinent regional criteria
and the peculiarities of the UPS. The math is made simpler, and
two-step clustering is made easier by separating the groups into
only two. Study professionals present the criteria for evaluating
an index and, following several revisions, establish the
quantitative assessment criteria for every indicator. For more
information, see Table 11.2. Utilizing the conclusions of
professional evaluation, the “AHP” technique is also utilized to
evaluate the metric of the size of drivers and statuses. The
outcomes are then checked for accuracy. The driving forces and
the condition of the specimens were assessed in the study using
a weighted total approach. In this research, the driving force
stage and state stage assessment outcomes are coupled using
the four-quadrant technique, and renewal interest is divided
into three groups, ranging from elevated to inadequate: higher
driving forces with a high status indicates recent growth;
higher driving forces with a low status indicates the short-term
growth; and a weak driving force with a weak state indicates
long-term growth.
Table 11.2 The DSR indexes’ weights and value attributions.

Index evaluation standard and score

Index 2 1 Weight

ST1 1 Floor Greater than 1 0.182

floor

ST2 More than 100 Every parking 0.200

parking spaces space under 100
together

ST3 A mean column Less than 6 0.211

spacing of at meters between
least 6 meters columns on
average

ST4 Net height of at Less than 3 0.153

least 3 meters meters in net
height
 Index evaluation standard and score

Index 2 1 Weight

ST5 Less than 50% of Greater than 50% 0.141

a building’s of a building’s
area is on the area is on the
ground. ground

ST6 Basement for Basement civil air 0.113

non-civil air defense
defense

RE Mixed Single ownership 1.000

ownership

DR1 Charge an Charge a rate that 0.303

hourly rate of is less than 10
more than 10 yuan per hour.
yuan.
 Index evaluation standard and score

Index 2 1 Weight

DR2 More than two A daily turnover 0.215

times every day rate of no more
are changed. than two occurs.

DR3 Urban traffic Region of non- 0.139

node area urban traffic
nodes

DR4 Public Civil 0.096

DR5 Not combined Combined with 0.130

with other additional
activities operations

DR6 Connected Unconnected 0.117

11.3.6 Model with Multiple Objectives

A classic multiple-goals issue is choosing an appropriate

substitute for UPSs. Its traits include the incompatibility of the
objectives and the inconsistency of the numerous quality
markers. To thoroughly assess and rate the plans of a small
number of UPSs that need to be substituted, choice-makers
study the choice facts and example data that are already
available. The two components of multiple goals attribution
choice-making are options and choice characteristics. The
following formula describes the choice-making process:

(11.1)

(11.2)
bj is the case of No. j, vi is the attribute of No.i, and wji = e
(bj, vi) is the utility function value of case j for attribute i. CQ is
the decision criterion in this scenario. Equation 11.2 may be
used to evaluate and order different UPS service choices based
on the selection criteria. Various utility function matrices and
criteria for making decisions lead to matching mega-attribute
processes for making decisions. The function substitute
evaluation for UPS entails six steps, all of which are by the
general procedure of multiple goals and mega-attribute-making
choices: (i) setting up the selection matrices, (ii) creating a
recursive hierarchy, (iii) normalizing the values of the
attributes, (iv) determining the index’s weight, (v) calculating
the complete characteristic usefulness value, and (vi) ranking
the various possible schemes.
We first create substitute libraries for the UPS functionality. The
renewal purpose possibilities for city UPS are arranged and
described, presenting regional rules, industrial standards for
design, and possible uses in future urbanization.

By various function objectives and the DSR model, the study

then creates matching qualities and weights. The expert’s AHP
approach is used in the weighing technique to determine how
to determine whether the 10 functional objectives are
appropriate. The optimal substitution goal for every UPS case is
chosen after replacing the values of the attributes of the various
UPS cases (Figure 11.1) by contrasting the total amount size
replacements function. For instance, the following formula is
used to determine the importance of the operational goal
Commercial (F1):

(11.3)
Figure 11.1 Weight matrix of attributes.

11.3.7 Statistical Data

Radar graphs are utilized for contrasting the characteristics of

various UPS kinds, multiple regression analyses are utilized to
examine the connection between Driver indices and Status
indices for different types of UPSs, communication evaluation is
used to determine the connection between renewal time and
associated UPS characteristics, and intersect-tabs are utilized to
contrast the renewal times in instances of various substituted
purposes.
11.4 Analysis

11.4.1 Feature Types

The mean of the 13 indications was 1.52 for all 200 instances.
The mean value of 200 instances for each of the six Driver
indices was 1.64 (Dr1), 1.43 (Dr2), 1.8 (Dr3), 1.53 (Dr4), 1.32
(Dr5), and 1.13 (Dr6), accordingly. The mean values for all
instances for each of the six Status indices were
correspondingly 1.36 (St1), 1.52 (St2), 1.53 (St3), 1.67 (St4), 1.75
(St5), and 1.43 (St6). The reaction index had an average value of
1.87. The Drivers index had a weighted mean of 1.48. The
current state indices’ weighted mean was 1.52.

The two-step clustering algorithm separates 200 examples into

four groups based on 13 category factors. Figure 11.2 displays
the parameter ordering of categorization relevance. Parking lot
cost is the most crucial factor, then the spot, the level of parking
changeover, and floor level. According to the information, the
identifying factor of categorization accuracy was 0.26, which is
satisfactory. These kinds have unique characteristics. For
example, the initial group includes 40 cases where 100% of the
parking prices are under or exactly match 10 yuan/hour, where
45% of the destinations are close to traffic nodes, and where
65% have an elevated parking rate of turnover. The next group
consists of 50 cases, with 96% of the parking costs above 10
yuan/hour, 92% of the locations being close to traffic nodes, and
all of the parking rates of turnover falling into one of two
categories: either very high or very weak. There are 74
examples in the third group, of which 62.2% have a weak rate
of parking changeover, 89.2% are geographically located in
traffic node regions, and 91.9% have parking rates of more than
10 yuan/hour. There are 36 examples in the fourth category, of
which 72.2% have parking fees that are under or equal to 10
yuan/hour, 83.3% have sites that are outside of traffic nodes,
and 72.2% have significant parking turnover rates.

Figure 11.2 Ranking of the factors.

The mean values of the signals for the various UPS kinds are
displayed on the radar chart in Figure 11.3. The indicator
readings of D, S, and R for the four categories are quite distinct,
as shown in Figure 11.3. While the outcomes of the S indices of
type 3 are more comparable to those of other kinds, the
readings of the D indices of type 2 are more dissimilar than
those of the other kinds. The D, S, and R readings for the first
type are 1.31, 1.49, and 1.90, accordingly. The D, S, and R
readings for the second type are 1.48, 1.40, and 1.93; for the
third type, they are 1.60, 1.66, and 1.77; and for type 4, they are
1.47, 1.51, and 2.01, accordingly. It is clear that the third
category’s D and S are often greater than those of the other
categories.

Figure 11.3 UPS-type characteristics.

11.4.2 Timing of Renewal

A UPS will be considered recent expansion if D has a higher

value than 1.494 and S is bigger than 1.539; middle-term
expansion if Dr is bigger than 1.494 and St is bigger than 1.539;
and long-term expansion if D is lower than 1.494 and S is lower
than 1.539, as per Chapter 3.4. This usual procedure creates
four parts on the outside of Figure 11.4’s two-axis values.

The green region in the top right side represents the current
phase of growth, the yellow regions at the higher left and
bottom right sides indicate the medium-term phase of
expansion, and the red region in the lowered left corner
represents the long-term phase of expansion. Figure 11.4
reveals that a total of 92 instances have been categorized as
medium-term expansion, of which 30 are type 3 instances and
26 are type 2, 48 instances have been categorized as short-term
expansion, of which 18 are type 2 instances and 10 are each of
types 1, 3, and 4. Of the 60 instances categorized as recent
expansion, 34 are type 3 instances and 14 are type 4, while 30
are type 3 instances and 26 are type 2, and there are 92
instances categorized as medium-term expansion. Nearly half
of type 3 (44%) is growing rapidly compared to more than one-
third of type 2 (37%) and half of types 1, 2, and 3. Figure 11.4
also demonstrates the varied relationships between the Dr and
St. Type 3 Dr and St have a lesser positive association, whereas
Dr and St of other kinds have lesser opposing interactions.

Figure 11.4 Renewal time outcomes.

For the 13 DSR model indices and renewal time, numerous

connection types of research are used in this work. The
relationship research chart (Figure 11.5), which highlights the
association among the parameters, reflects the findings. The
level of each strewn point in Figure 11.5 corresponds to a
particular parameter. If the dispersed spots have a nearer level,
they are near together. The variable renewal timing features
are discovered and assigned to various regions in Figure 11.5
based on the description of the associated analysis graph. Long-
term growth and recent growth exhibit distinct features
because they diverge greatly from their points of origin.
Instances without a subterranean connection, a modest lower-
level location, and no additional integrated obligations are the
main characteristics that are ideal for medium-term expansion.
The top building’s usage for civil purposes, together with its
narrow essential spans and minimal height, is the main
characteristic of long-term expansion scenarios.

11.4.3 Replacement of Function

The mass matrix structures as shown in Figure 11.1 are the

basis for the weighing calculations and assessment of every
alternative function for every instance kind. The calculations
are done to correspond with the mean value of 13 sustainable
criteria for every instance kind. Table 11.3 is a list of the
outcomes. Various types have varying capacities for practical
adaptation. For the first type, business operations, educational
institutions, and libraries are a better option; for the second
type, clt & ent, logistics, and libraries are the best substitute
functions; and for type 4, data centers, cultural entertainment,
and libraries are the best substitute functions. Table 11.3
aggregates the many types of acceptable performances as well
as the suitability of each function for distinct sorts. The
commercial function is the least ideal of the 10 jobs, as can be
shown, while the screen is the most versatile.
Figure 11.5 Analyzing renewal timing and UPS properties.

One by one, we statistically analyzed the substitution of 200

functions to determine the renewal distribution of time ratio
for various function kinds. The exhibition is the most frequent
role for renewal in 46 cases, while commercial and office are
the least frequent (only four of each). This is seen in Figure 11.6.
Most of the business operations, together with everything
related to the school, library, and half of the office, are in the
short-term expansion phase. The office operation is nearing its
conclusion. Most of the scenarios that are ideal for libraries and
logistics centers are in the long-term expansion phase, whereas
libraries and data centers make up the largest and lowest
percentage of mid-term renewals.
Table 11.3 Rankings of function replacement for each UPS type.

Type Type Type Type

Functions Ranking
2 4 1 3

Culture & 4 9 19 2 4
Entertainment

Office 5 8 23 5 5

DataCenter 1 10 25 7 7

Laboratory 7 3 19 3 6

Exhibition 2 2 13 8 1

Logistics 9 1 27 9 8

Commercial 6 5 30 10 9

Stadium 8 4 16 1 3

School 3 6 17 6 2
 Type Type Type Type
Functions Ranking
2 4 1 3

Library 10 7 31 4 10

Figure 11.6 Distribution of renewal times.

11.5 Conclusion

This study performed a preliminary investigation on

sustainable renewal procedures and plans for urban
subterranean parking spaces based on the SAV situation. The
next findings were reached: The subterranean parking space
with the extension of a broad building and for public use of the
topmost growing is acceptable for early regrowth, but the
examples with a short architectural span and minimal
amenities and the above structure for civil purposes are largely
suitable for long-term expansion. Most of the situations have a
successful location for exhibition replacement, and certain
situations also have effective roles as educational institutions
and athletic facilities. Instances, where laboratories work well,
tend to be those that are short- and medium-term solutions. The
lighting and ventilation of underground rooms represent the
primary sustainable design inconsistency for the majority of
parking garages. Contradictory architectural elements can be
efficiently resolved with the use of inset lighting, ventilation
shafts, and squares.

A rising in groundwater owing to leaks or mishaps is one risky

technical and geological occurrence that might result from
replacing the UPS as a building. This study does not take into
account the technical and geographical features of UPSs, in
addition to the dependability of buildings in terms of
accountability and obligations, which are closely tied to how
the underground space operates. As a result, from a design
perspective, the practical adaptation of this research is simply a
reference.
Regardless of its shortcomings, this research has a significant
amount of value. It provided an evaluation of feasibility and
potential answers while exposing the new issues regarding
physical rejuvenation brought on by the shift toward electronic
towns by building an empirical foundation. To use the results of
science to free the process of design, further studies will be
connected to low-carbon, environmentally friendly, and other
aims, improve the structure of buildings with a variety of
reasons, and combine it into innovative ideas in a more
acceptable way. Future studies will also focus on the backdrop
of the digital city and the features of urban system spaces,
develop more efficient design techniques, and execute real-
world project presentations.

References

1. Tscharaktschiew, S. and Reimann, F., Less workplace parking

with fully autonomous vehicles? J. Intell. Connected Veh., 5, 3,
283–301, 2022, doi: 10.1108/JICV-07-2022-0029.
2. Papa, E. and Lauwers, D., Smart mobility: Opportunity or
threat to innovate places and cities. In 20th International
Conference on Urban Planning and Regional Development in
the Information Society (REAL CORP 2015), pp. 543–550, 2015.
3. Clark et al., The impacts of autonomous vehicles and e-
commerce on local government budgeting and finance.
Urbanism Next, Sustainable Cities Initiative University of
Oregon, 2017, https://cpb-us-
e1.wpmucdn.com/blogs.uoregon.edu/dist/f/13615/files/2017/0
7/Impacts-of-AV-Ecommerce-on-Local-Govt-Budget-and-
Finance-SCI-08-2017-2n8wgfg.pdf
4. Victor, T. et al., When autonomous vehicles are introduced on
a larger scale in the road transport system: The Drive Me
project, in: Autom. Driving Safer More Effic. Future Driving,
pp. 541–546, 2017, doi:10.1007/978-3-319-31895-0_24.
5. Baumgardner, W., Cassidy, C., Ruhl, M., The promise of
seamless mobility: Autonomous vehicles and the mobility-as-
a-service revolution. In Disruptive Transport Driverless Cars,
Transport Innovation and the Sustainable City of Tomorrow,
W. Riggs (Ed.), pp. 11–20, Routledge Taylor & Francis Group,
London, UK, New York, NY, USA, 2019.
6. Xia, B. et al., Sustainable renewal methods of urban public
parking spaces under the scenario of shared autonomous
vehicles (SAV): A review and a proposal. Sustainability, 13, 7,
3629, 2021, doi: 10.3390/su13073629.
7. Nikitas, A. et al., How can autonomous and connected
vehicles, electromobility, BRT, hyperloop, share-use mobility,
and mobility-as-a-service shape transport futures in the
 context of smart cities? Urban Sci., 1, 4, 36, 2017, doi:
10.3390/urbansci1040036.
8. Noyman, A. et al., Road map for autonomous cities:
Sustainable transformation of urban spaces, 2017.
9. Ceder, A., Urban mobility and public transport: Future
perspectives and review. Int. J. Urban Sci., 25, 4, 455–479,
2021.
10. Ortegon-Sanchez, A. et al., Car-free initiatives from around
the world: Concepts for moving to future sustainable
mobility, in: TRB 96th Annual Meeting Compendium of Papers,
Washington, DC, 2017, Jan.
11. Park, J. et al., Air pollution and public bike-sharing system
ridership in the context of sustainable development goals.
Sustainability, 14, 7, 3861, 2022, doi: 10.3390/su14073861.
12. Ágnes, S.P.F., Autonomous vehicles from the EU perspective,
2020.
13. Gružauskas, V. et al., Minimizing the trade-off between
sustainability and cost-effective performance by using
autonomous vehicles. J. Cleaner Prod., 184, 709–717, 2018, doi:
10.1016/j.jclepro.2018.02.302.
14. Levin, M.W. et al., A general framework for modeling shared
autonomous vehicles with dynamic network-loading and
dynamic ride-sharing applications. Comput. Environ. Urban
 Syst., 64, 373–383, 2017, doi: 10.1016/j.
compenvurbsys.2017.04.006.
15. Hyland, M.F. and Mahmassani, H.S., Taxonomy of shared
autonomous vehicle fleet management problems to inform
future transportation mobility. Transp. Res. Rec., 2653, 1, 26–
34, 2017, doi: 10.3141/2653-04.

Note

1. *Corresponding author: [email protected]

12
Big Data Analytics for a
Sustainable Competitive Edge: An
Impact Assessment
Rajyalakshmi K.1*, Padma A.2, Varalakshmi M.3, Suhasini A.4
and Chiranjeevi P.5

1
Department of Computer Science and Engineering, Koneru
Lakshmaiah Education Foundation, Vaddeswaram, Guntur, India

2
Chaitanya Bharathi Institute of Technology, Hyderabad, India

3
Department of Humanities and Sciences, Vardhaman College of
Engineering, Shamshabad, Hyderabad, India

4
Sri Durga Malleswara Siddhartha Mahila Kalasala, Vijayawada,
India 5Department of Mathematics and Humanities, R V R and J C
College of Engineering, Chowdavaram, Guntur, India

Abstract
One of Industry 4.0’s key components is big data analytics
(BDA), a cornerstone. It has developed into an effective
instrument for boosting businesses’ competitive advantages by
improving data-driven performance. While this is happening,
the global resource problem is forcing corporations to take a
commission based on sustainability seriously. According to the
literature, BDA and innovation can increase a company’s
productivity and provide a competitive advantage. According to
this study, a sustainable competitive advantage (SCA) might be
produced by combining big data analytics capabilities (BDACs)
with elevated levels of data availability (DA). Research from the
past and the dynamic capacity hypothesis support this claim.
PLS-SEM, a statistical method that uses partial least squares to
model structural equations, was utilized to analyze the results
of this study’s survey of 117 manufacturing companies.
According to the findings, BDAC significantly boosts internal
consistency (IC) and is highly dependent on DA. Additionally,
statistics show that compared to BDAC, IC directly impacts a
firm’s SCA. This study provides important theoretical and
practical insights into the challenges of BDA application to
achieve suitable procedures, promoting industrial
organizations to improve their sustainable business
performance (SBP).
Keywords: Sustainable competitive advantage (SCA), data
availability (DA), big data analytics (BDA), firm performance
(FPER)

12.1 Introduction

Big data (BD) research is growing in popularity in both

academic and business research. With the assistance of the
appropriate tools and systems, BD and predictive analytics
(BDPA) are analytical techniques that include accessing,
identifying, retrieving, storing, processing, visualizing, and
managing vast amounts of complex data [1]. When compared to
the current state of conventional data processing systems, this
form of data demonstrates a large quantity and speed of data
processing, which might provide frequent compensation and
settlement. The second is utilizing business analytics (BA) to
monitor and decide on organizations. BAs are prepared with
the knowledge and abilities to assess organizational strategy
efforts and create successful business planning [2]. E-commerce
companies can use BD to handle the massive volumes of data
accessible to identify market trends, anticipate customer needs,
and assess consumer buying habits. Undoubtedly, BD’s
contribution to the innovation process is becoming more visible
in practice and acknowledged in research [3]. The definition of
big data customer (BDC) activity may be concealed in BD, and
consumer analytics is the practical use of this information, even
though in the contemporary economy, BD is viewed as a new
type of capital. Businesses must allocate enough organizational,
human, and physical capital resources to BD to benefit from
this new type of capital [4]. By enhancing time, management,
and market transaction costs such as those associated with
online buyer–seller interactions, big data analytics (BDA) may
assist e-vendors. Especially in the case of e-commerce, business
intelligence (BD) enables businesses to monitor consumer
behavior and identify the best strategies for turning one-time
clients into devoted ones. Comparing a company’s value chain
to its rivals, one that employs BD analytics sees a 5%–6%
increase in productivity [5]. This study covers the impact of
BDA in supplying a long-lasting competitive edge.

The remaining portions of the paper are given below: Section

12.2 summarized the literature, Section 12.3 described the
suggested strategy, and Section 12.4 presented and discussed
the experimental data sets and simulation findings. The
analysis is concluded in Section 12.5, which also recommends
further study.

12.2 Related Works

The research [6] investigated the relationship between the
elements of internal consistency (IC) and the competitive
advantage (CA) of the hotel industry by mediating the influence
of service innovation and moderating the impact of BDA
capabilities. The study [7] developed a paradigm to examine
how exploitative, explorative, and ambidextrous innovation
techniques result in big data customer (BDC). Additionally, it
will look at how it can provide a company with a sustainable
competitive edge. By utilizing the dynamic capacity theory to
examine how data analytics capabilities (DACs) work as a
mediating element, the study [8] investigated how the
deployment of BDA influences the quality of judgments. The
article [9] introduced BDA to understand the interactions
between competitive advantage (CA) and supply chain
management practices (SCMPs) in Jordanian manufacturing
firms. The article [10] introduced the concept of sustainable
smart manufacturing by fusing the fundamental technologies of
intelligent manufacturing with pervasive servitization over the
whole life span. The research [11] intends to investigate how
wearable technology usage has affected Indian health
insurance companies. To gain a competitive edge, it emphasizes
the essential dynamic skills that health insurance companies
need to develop to handle massive data produced by wearable
technologies. The article [12] shows how BDA may give IT
businesses a competitive advantage and boost business
performance. China and India are the subjects of the research.
The study [13] enhanced by evaluating the direct effects of big
data analytics capabilities (BDACs) on firm performance (FPER)
and the mediating effects of process-oriented dynamic
capabilities (PODCs) on the relationship between BDAC and
FPER. Madhani [14] created a variety of frameworks to
demonstrate how BD and BDA are being used to alter the
supply chain further digitally. One of these frameworks
examines how BD-enabled supply chains’ competitive
objectives support the expansion of customer value and, as a
result, aid in creating competitive advantages. The article [15]
employed BDA to analyze thoroughly and research business
sustainability plans.

12.3 Hypothesis and Research Model

12.3.1 Theoretical Model

The proposed model designed for this study is exposed in

Figure 12.1, and it comprises the four reflecting components
data availability (DA), BDAC, IC, and sustainable competitive
advantage (SCA). The foundation for the idea is the dynamic
capabilities concept, which maintains that businesses must
create new strategic talents to keep up with the market’s growth
rate of technology advancements and shifting customer
preferences. This model, founded on the theory that DA, when
supported by BDAC and creative processes, may improve
organizations’ sustainable performance, illustrates the dynamic
potential of BDAC and IC. Finally, the characteristics above will
increase organizations’ SCA.

Figure 12.1 Suggested research design.

12.3.1.1 Hypothesis 1 (H1)

BDAC is a dynamic capability because it exhibits emotional

capability traits like repeatability, making an experience hard
to replicate and, consequently, providing valuable assets that
increase competitive advantage. However, data must be
accessible to use BDA; otherwise, overall performance and
competitive advantage may suffer. Furthermore, deleted data
cannot be recovered even after they are discovered. Therefore,
businesses strive to make their data as accessible as possible to
improve their BDAC by addressing relevant concerns, including
data source, size, and gathering technique. Massive volumes of
data may be delivered over the Internet of Things (IoT) through
cyber-physical systems, proximity sensors, radio-frequency
identification (RFID), and other advanced Industry 4.0
production technologies, which can then be connected to the
cloud to ensure DA. BDA might be linked to enterprise resource
planning (ERP) and other management information systems.
We thus suggest that DA is a BDAC facilitator, as higher DA
encourages improved data analysis. The following is the
situation that we offer:

H1: DA may improve BDAC.

12.3.1.2. BDAC as a Facilitator for IC (H2)

Businesses may boost their IC in items, processes, and markets

by integrating BDA with additional data sources like customer
research, strategy, and equipment monitoring. BDAC might
improve IC by simplifying routine processes and the work of an
innovation team. Additionally, they demonstrated how BDAC
positively affected IC. Consequently, we suggest the following:

H2: BDAC benefits IC in an optimistic way.

12.3.1.3 The Contribution of BDAC and IC to Raising SCA

(H3) and (H4)
BDAC is an important center for the creation of knowledge, to
start. This is accomplished using various data sources, such as
those about consumers, processes, and production, to help
organizations make better decisions, enhance performance,
and preserve a competitive edge. For instance, BDA has a
substantial influence on the social and environmental
sustainability of a corporation. In hypotheses 3 and 4, the
primary impacts of BDA on businesses’ financial performance
were examined in more detail. Second, the dynamic ability of
innovation substantially contributes to raising sales,
profitability, and performance excellence by creating new
products and processes.

Furthermore, an innovation that facilitates quicker market

change adaptation promotes competitive advantage
development. As a result, IC dramatically impacts a company’s
ability to operate profitably and competitively. Adaptable
talents by strengthening and maintaining their competitive
edge and overall performance, BDA and innovation both help
businesses overcome external obstacles and competencies. We
thus put forth the following hypotheses:

H3: Businesses’ SCA is positively impacted by BDAC’s dynamic

capacity.
H4: The SCA of an organization is positively affected by IC, since
it is an emotional competence.

12.3.1.4 Sustainable Competitive Advantage

A company gains a competitive edge when it outperforms its

present or future rivals in terms of success. Modern
technologies like BDA also provide sustained performance by
supporting data-driven performance. Finding an advantage in
competition usually requires contrasting the success of a
business to that of its rivals. In light of this, we operationalize
SCA as a standard for outperforming competitors regarding
environmental and social outcomes. This design is the first to
consider the three sustainability components at the business
level, according to prior research that forms the foundation of
the competitive advantage framework.

12.3.2 Proposed Method

Business decision analysis is the process of sifting through

enormous complicated information in search of patterns,
correlations, and insights that may be utilized to impact choices
and acquire valuable data. Large amounts of structured and
unstructured data must be subjected to the most up-to-date
analytics techniques and algorithms in order to extract useful
information and provide insights that may be put into practice.

12.3.2.1 Layout Assessment

The fundamental presumptions and methods of the study were

evaluated using a web-based questionnaire. Each of the 15
survey questions received a score from respondents on a 5-
point Likert scale, with 1 being a great deal of unhappiness and
5 a great deal of agreement. The previous material was used as
a guidance for creating the questionnaire. The four components
of the four multi-item reflecting measures for the research
model were created using the questions. The G*Power 3.1
program set the minimum sample size at 117 samples. As a
result, 5%, 90%, and 0.3 were chosen as the power, effect size,
and significant threshold parameters, respectively.

12.3.2.2 Analysis Method and Data Collection

This study concentrated on manufacturing firms from various

industrial sectors that were aware of and had some experience
with Industry 4.0 technology in order to test the hypothesis. The
service sectors were not included in the study, since several
survey questions were only intended to be addressed by
respondents from the industrial sectors. The secondary
objective of the questionnaire was to collect data to evaluate the
theories behind the structures. To strengthen the validity of the
data gathered, the members of management at the highest
levels were strongly encouraged to reply to the questionnaire.
This study employed PLS-SEM, since it is a statistical software
tool for DA that is effective at analyzing data from tiny sample
sizes, mediating factors, and indirect correlations. SmartPLS
may also directly assess reliability, discriminant, and
convergent validity. It delivers graphical outputs and a user
interface that is attractive to the eye. Compared to other
software packages, it has a more flexible manufacturing
process.

12.4 Results

The conceptual model is subjected to two analyses using PLS-

SEM software: first, the measurement model is assessed, and
then the structural model is taken into account. The two
following subsections go into considerable detail about the
results and rationale for both acts.

12.4.1 Evaluation of Validity and Reliability

Evaluation of the reflected measurement models requires the

conduct of validity and reliability tests for the constructs.
Passing both tests is necessary for the structural model exam.
By evaluating the values of all variables to those of their related
signals with the legitimacy of the questionnaire problem, these
evaluations attempt to determine the internal coherence of the
data provided by those who responded. While Cronbach’s
alpha, especially composite reliability (CR) testing, was used to
assess the IC or dependability, discriminate and convergent
validity tests were utilized to determine the reliability.

To ascertain an item’s internal coherence inside a construct, use

CA. Both the item and factor reliability scores must be at least
0.60. Additionally, for the combination of reliability, which is
used to verify the IC, values better than 0.8 are required. Due to
matching CA values above 0.8 with compound reliability scores
over 0.8, the results displayed in Table 12.1 indicate the
dependability of all structures. The outcomes confirm the
acceptability and reliability of the IC for each conception.
Table 12.1 Results of validity and reliability tests.

Construct R2 CA CR AVE

DA 0.639 0.804 0.580

BDAC 0.168 0.743 0.837 0.564

IC 0.241 0.711 0.818 0.531

SCA 0.440 0.743 0.839 0.567

Figure 12.2 shows Cronbach’s alpha. A scale or questionnaire’s

internal consistency or reliability is measured by the statistic
known as CA. Lee Cronbach, an American psychologist, is
credited with coming up with the concept. The range of the
Cronbach’s alpha coefficient (CAC) is 0 to 1. The alpha value is
nearer to 1, the more internally consistent the scale’s
components are. When the alpha is more significant, the items
on the scale are more closely related to one another, indicating
that they are evaluating the same underlying notion. The CA
(0.743) value will be bigger than CR and average variance
extracted (AVE).
In Figure 12.3, CR measures the proportion of actual score
variance to the construct’s overall variation. By dividing the
total squared loadings (factor loadings) of the items on the
construct by the total squared loadings plus the total residual
variances (measurement error) of the items, the coefficient of
determination (CR) is determined. Equation 12.1 may be written
as follows.

(12.1)

In structural equation modeling (SEM), the AVE statistic is

employed to evaluate the convergent validity of a construct or
latent variable. It gauges how much variation is captured by a
construct’s items compared to how much variance results from
measurement error in Figure 12.4. The average of the squared
correlations between each item and the other items in the same
construct determines the AVE, which is then divided by the total
variance of the things. Equation 12.2 may be written as follows:
Figure 12.2 Reliability and validity of the CA.

Figure 12.3 Reliability and validity of the CR.

Figure 12.4 Reliability and validity of the AVE.

(12.2)

The composite reliability requires values greater than 0.7 to

verify the internal consistency. The findings in Table 12.2
demonstrate the dependability of all constructs because
matching CA values and combined reliability values are higher
than 0.7 and 0.8, respectively. The conclusions attest to the
accuracy and acceptability of each idea’s internal coherence.
The data in Table 12.3 demonstrate that all AVE values are
much more than 0.6, supporting the convergent assertions. The
square root of the average variance of each notion is computed
to establish discriminating validity (DV), and the results are
compared to those of other pictures. It is asserted that DV exists
when there are more correlations between any two concepts
than there are between any other ideas. The outcomes in Table
12.2 show that the DV is sufficient.

These heterotrait-monotrait (HTMT) ratio tests, which assess the

intensity of distinctions among overlapping components,
provided additional evidence for the discriminate validity.
Acceptable values range from 0 to 1, with values under 1
indicating high dependability. The HTMT values for this
investigation are less than 0.90, which supports the
discriminate validity, as shown in Table 12.2.

Table 12.2 HTMT values.

BDAC DA IC SCA

BDAC

DA 0.56554

IC 0.62260 0.48970

SCA 0.53015 0.32422 0.87068

Table 12.3 Examine the legitimacy of differences.

1 2 3 4

1 BDAC 0.75

2 DA 0.41 0.76

3 IC 0.49 0.33 0.73

4 SCA 0.4 0.23 0.66 0.75

When the indication was most recently produced, the

acceptable values for external loadings had to be less than 0.7
but greater than 0.4. Otherwise, they must be more than or
equal to 0.7. To validate the significance of the outer loads in the
model, the t-statistics must be checked with values more than
2.00. The signals are therefore significant, according to the
findings of the exterior loading of the t-test.

12.4.2 Structural Model of Coefficient of

Determination (R2)
R2 is a mathematical indicator with permissible values between
0 and 1 that shows how much variance in the exogenous
constructions can be explained by the exogenous constructs.
According to Table 12.2, where the R2 value for SCA is 0.439, the
BDAC and IC constructs account for 45.0% of the conflict in the
SCA constructs. Table 12.4 demonstrates that the bootstrapping
method also used PLS-SEM to calculate the t-value and assess
the structural model.

Verifying the path coefficient values or the statistical

significance of the relationships between the latent variables in
the structural framework is crucial for assessing the offered
hypothesis. The t-value, significance level, and the path’s related
standardized coefficients (B-value), which must be at or higher
than 0.1 to be deemed significant, are calculated using the 0.05
threshold of importance.

DA considerably and positively impacts IC, while BDAC

significantly influences IC but not SCA, according to the model
fit findings in Table 12.4. SCA is significantly and favorably
affected by IC. However, the findings imply that IC has the
greatest influence on SCA. The findings in Table 12.4 are
consistent with H1, H2, and H4. The theory that BDAC and SCA
are connected, however, is unsupported by any data.
Table 12.4 The model’s fit outcomes.

t- B- p- T
STDEV
value value value statistics

H2 6.062 0.491 0.081 0 6.06

H1 5.256 0.41 0.078 0 5.238

H4 8.897 0.605 0.068 0 8.922

H3 1.178 0.106 0.09 0.237 1.183

12.4.3 Discussion

This study looked into the relationship between a firm’s SCA

and dynamic skills like BDA and innovation. The results of the
study suggest that BDAC has a major impact on IC and that IC
improves SCA capabilities but not BDA. The study’s findings add
to the body of knowledge in this area by presenting a number
of new viewpoints. This work is the first to assess the
relationship between the DA level and BDAC as well as the
contribution of BDAC and IC to the formation of SCA. Our
findings show that an increase in DA level favors BDAC, and the
results of the conceptual model are compatible with H1. Data
availability will considerably rise when a company’s system is
connected to BDA, since all components are connected to the
IoT and the backup mechanism is effective, boosting BDAC.
However, IoT technology enablers must be used to assist the DA
in industrial firms if the BDAC is to be enhanced. This paradigm
can be used, and it works incredibly well when cutting-edge
products with high levels of customization generate a lot of
data. BD will not aid manufacturers in proportion to the
amount of money invested in this industry, and conventional
mass manufacturing processes do not require highly
customized products. Thus, this approach could be more
effective in these situations. The technique is also useless when
organizations lack the IT infrastructure or agility to meet a
broad range of market demands or all client expectations. The
IC may not be impacted by the BD and BDAC in such
circumstances; therefore, the SCA may not be improved. The
same is true for businesses that operate in competitive markets.
The study also examines network needs as a precondition for
BDA when looking at DA concerns. Furthermore, by
emphasizing emotional skills as crucial facilitators for SCA at
the organizational level in the industrial sector, this research
assists managers’ decisions regarding allocating resources and
making investments.
12.5 Conclusion

The findings support the hypothesis in the study’s research

questions that DA significantly affects BDAC, substantially
improving the interaction with IC. However, BDAC plays a tiny
part in raising a company’s SCA. The findings also suggest that
IC significantly influences the association between BDAC and
SCA. Based on the outcomes of each hypothesis, a corporation
could make a decision. Businesses must digitalize their
manufacturing, operational, and corporate processes in order
to expand the data that are accessible to BDA. In H1, DA
significantly and favorably affects BDAC. By incorporating
goods, machinery, and other intelligent objects into the IoT
network and feeding a stream of data to BDA, businesses may
digitalize their production processes. In H3, it was
demonstrated that BDAC, a dynamic capability, influenced
enterprises’ SCA in a favorable way. According to our analysis,
BDAC has little impact on an organization’s SCA. The dynamic
capacity known as IC in H4 substantially and positively impacts
an enterprise’s SCA. This encourages individuals making
decisions to invest money in technology that supports
innovations. This study offers a helpful framework for
addressing various difficulties, including how DA, BDAC, and IC
support manufacturing businesses’ SCA. Since the methodology
of this study only permits generalizing findings to the service
industry, there is still room for improvement. Given this, future
research may employ a sector-based study design,
concentrating on the industrial or service sectors. This strategy
may target distinct qualities in each industry by incorporating
more particular criteria in the research questions, producing
more detailed findings and a more accurate sector
differentiation analysis. To achieve sustainability in processes,
this research provides significant theoretical and practical
insights into these problems. The long-term competitive
advantage of industrial enterprises may be strengthened using
these insights. But future research could benefit from this
study’s constraints and flaws.

References

1. Nilashi, M. et al., How can big data and predictive analytics

impact the performance and competitive advantage of the
food waste and recycling industry? Ann. Oper. Res., 1–42,
2023, doi:10.1007/s10479-023-05272-y.
2. Bag, S. et al., Big data analytics as an operational excellence
approach to enhance sustainable supply chain performance.
Resour. Conserv. Recycl., 153, 104559, 2020,
doi:10.1016/j.resconrec.2019.104559.
3. Sazu, M.H., Does big data drive innovation in e-commerce: A
global perspective? SEISENSE Bus. Rev., 2, 1, 55–66, 2022,
doi:10.33215/sbr.v2i1.797.
4. Erevelles, S. et al., Big data consumer analytics and the
transformation of marketing. J. Bus. Res., 69, 2, 897–904, 2016,
doi:10.1016/j.jbusres.2015.07.001.
5. Le, T.M. and Liaw, S.Y., Effects of pros and cons of applying big
data analytics to consumers’ responses in an E-commerce
context. Sustainability, 9, 5, 798, 2017, doi:10.3390/su9050798.
6. Alkhatib, A.W. and Valeri, M., Can intellectual capital promote
a competitive advantage? Service innovation and big data
analytics capabilities in a moderated mediation model. Eur. J.
Innov. Manag., 2022, doi:10.1108/ EJIM-04-2022-0186.
7. Zhang, Z. et al., Big data capability and sustainable
competitive advantage: The mediating role of ambidextrous
innovation strategy. Sustainability, 14, 14, 8249, 2022,
doi:10.3390/su14148249.
8. Li, L. et al., Evaluating the impact of big data analytics usage
on the decision-making quality of organizations. Technol.
Forecast. Soc. Change, 175, 121355, 2022, doi:
10.1016/j.techfore.2021.121355.
9. Baqleh, L.A. and Alateeq, M.M., The impact of supply chain
management practices on competitive advantage: The
moderating role of big data analytics. Int. J. Prof. Bus. Rev., 8,
 3, e0679–e0679, 2023,
doi:10.26668/businessreview/2023.v8i3.679.
10. Ren, S. et al., A comprehensive review of big data analytics
throughout product lifecycle to support sustainable smart
manufacturing: A framework, challenges, and future
research directions. J. Cleaner Prod., 210, 1343–1365, 2019,
doi:10.1016/j.jclepro.2018.11.025.
11. Nayak, B. et al., Integrating wearable technology products
and big data analytics in business strategy: A study of health
insurance firms. J. Syst. Inf. Technol., 21, 2, 255–275, 2019,
doi:10.1108/JSIT-08-2018-0109.
12. Behl, A., Antecedents to firm performance and
competitiveness using the lens of big data analytics: A cross-
cultural study. Manage. Decis., 60, 2, 368–398, 2022,
doi:10.1108/MD-01-2020-0121.
13. Wamba, S.F. et al., Big data analytics and firm performance:
Effects of dynamic capabilities. J. Bus. Res., 70, 356–365, 2017,
doi:10.1016/j.jbusres.2016.08.009.
14. Madhani, P.M., Big data usage and big data analytics in
supply chain: Leveraging competitive priorities for
enhancing competitive advantages. IUP J. Supply Chain
Manag., 19, 2, 1–35, 2022.
15. Zhang, W. et al., Research on the mechanism of the
sustainable development model of enterprises based on big
 data analysis model. Mob. Inf. Syst., 2021, 1–12, 2021,
doi:10.1155/2021/4469255.

Note

1. *Corresponding author: [email protected]

13
Sustainability and Technological
Innovation in Organizations: The
Mediating Role of Green Practices
Rajyalakshmi K.1*, Rajkumar G. V. S.2, Sulochana B.3, Rama
Devi V. N.4 and Padma A.5

1
Department of Computer Science and Engineering, Koneru
Lakshmaiah Education Foundation, Vaddeswaram, Guntur, India

2
Department of Computer Science and Engineering, GITAM
Deemed to be University, Visakhapatnam, India

3
Department of Humanities and Sciences, Chebrolu Hanumaiah
Institute of Pharmaceutical Sciences, Chowadaram, Guntur, India

4
Department of Mathematics, GRIET, Hyderabad, India

5
Chaitanya Bharathi Institute of Technology, Hyderabad, India

Abstract
A few of an organization’s goals are innovation, large profits,
and continued survival. Even so, sustainability has evolved into
a rigorous activity for both commercial and charity
organizations, since it pushes businesses to greater
effectiveness. Sustainability requires enough funds and
expertise. The factors that impact sustainability were examined
in previous studies, but innovation is rarely considered from
this angle. The present study investigates the connection among
organizational performance, technical innovation, and
management creativity, with sustainability acting as a
mediating element. We tested the model using structural
equation modeling in the analysis of moment structures
(AMOS) using actual information collected from 304 CEOs and
top managers in various countries. The findings suggest that
incorporating new management practices and technologies into
an organization may greatly enhance its effectiveness and
longevity. Sustainability acts as a mediator between
management innovation and organizational performance and
between organizational performance and technical innovation.
To improve long-term viability and success, we urge CEOs and
other senior managers to invest in management and technical
innovation.

Keywords: Organizational performance, technology

innovation, management innovation
13.1 Introduction

The connection among management and technical innovation

may be significantly impacted by sustainability. The goal of
sustainability is to provide for current needs without
compromising future generations’ ability to do the same. This
implies that technology advancements and sustainable
management techniques are created with consideration for the
environment, social justice, and economic viability [1]. A
mediating factor is a variable that accounts for the process by
which one variable affects the other to explain the connection
between the two other factors. A variable that aids in
explaining how management and technological innovation
impact organizational performance would be a mediating
factor in the context of the interaction between management
and technological innovation on organizational performance.
The working population’s expertise and ability capacity are one
potential intermediary influence. To be embraced and utilized
successfully, advances in technology often need novel abilities
and knowledge [2]. The adoption and application of emerging
technologies may thus be facilitated by management
approaches that encourage learning and growth, which can
boost organizational performance. The degree of interaction
and collaboration inside the company is yet another possible
mediating element. The discovery and implementation of novel
innovations that may improve the efficiency of an organization
can be facilitated through effective communication and
cooperation across individuals, groups, and divisions [3].

The magnitude and uncertain environments limit technology

innovation, companies now depend more on IT. Due to changes
in the corporate environment, to attain and retain productivity,
boost performance, and survive in today’s challenging markets.
The innovation-related activities that are technology-based and
aimed at enhancing performance are driven by these
developments. Numerous research surveys have shown that
with the increasing adoption of ICT by enterprises across a
number of industries, many firms fail to advance beyond the
company phase of the life cycle [4]. In general, businesses may
design strategies to improve the way they handle management
and technology innovations for optimal organizational
performance by recognizing intermediary elements in the
interaction among management and technical innovation on
organizational performance [5]. This research is one of the first
to take into account such a complicated framework for
investigating links among many diverse dimensions by
assessing the suggested model and testing the relevant
predictions. This connects acceptance, creation, and
routinization of big data with developments in sustainability
benefits in rivalry, economic and ecological results,
management and other belonging points of view, business
ethical conduct, consumer appetite for environmentally
friendly products, and human capital businesses [6]. Despite the
growing interest in information management and innovation,
only a small number of studies, particularly those that looked at
developing countries, have shown tangible proof connecting
the use of knowledge management with firm-level innovation.
For theoretical and strategic reasons, it is crucial to concentrate
about the connection between innovation and knowledge
management. First, it is significant because, as intangible assets
like intellectuals capital, as well as expertise, have become
increasingly accepted as being necessary to recognize, oversee,
and improve to promote and improve company value,
investigators have noted that there is a lack of scientific
evidence that guides how these intangible resources can be
managed [7]. In this research, we analyze and investigate the
impact of sustainability as a mediating factor in the
relationship between “management innovation (MI) and
technological innovation (TI)” on organizational performance.

13.2 Related Work

The study [8] examined how knowledge management (KM)
factors affect organizational sustainability. It compares the
results of companies that have earned the ISO 9001 standard
with those that have not. Data from 156 private sector
employees in Iraq’s Kurdistan Region was used in the study.
Utilizing a partial least squares method, we evaluate the
proposed model and investigate the variances among ISO 9001-
certified and noncertified organizations. The findings
demonstrate the importance of knowledge storage in the
creation, exchange, and use of knowledge. Additionally, it has
been shown that knowledge management significantly affects
an organization’s capacity to remain viable. The research [9]
analyzed that entrepreneur innovativeness affects the
relationship between technological innovation and company
success in Kenya. The investigation made use of structural
equation modeling (SEM) and a sample of 240 businesses. The
results show that technological innovation has a favorable
impact on business success. According to the survey,
entrepreneurs should create cutting-edge plans to improve
business performance. The impact of organizational innovation
(OI) and technical innovation capabilities (TICs) on business
performance is evaluated in this research.

The paper [10] determined that TICs mediated the effects of OI

on firm performance using a sample of 265 manufacturing
enterprises from the Pearl River Delta in China. Additionally,
they looked at how OI affected the link between TICs and firm
performance. The association between OI and company
performance was shown to be partly mediated by TICs
according to the findings of structural equation modeling
investigations. Similar to the way TICs and business
performance were correlated, OI somewhat mitigated this
association. The study [11] examined the root causes and effects
of open technology in small and medium-sized enterprises
(SMEs). To evaluate the assumptions, we gathered multisource
data from 404 SMEs and employed structural equation
modeling. The study [12] suggested an approach to clarify the
function that information creation plays as a connecting link
between knowledge management enablers and organizational
success. Cooperation, T-shaped skills, education, and technology
support are all factors that might affect an organization. The
research argues that the way decisions are made might
attenuate the correlation between knowledge creation and
business success. A collection of statements that represent an
empirically driven research aim and also identify the
relationships among the variables of interest is offered to aid
the audience in comprehending within an organization context.
To examine [13] the connection between an organization’s
culture, total quality management (TQM), and performance
within the context of (SMEs) is mentioned. More specifically, it
suggests a framework outlining potential connections between
organizational culture, TQM measures, and achievement within
the organization. Data from functionality representatives of
Indian automotive module SMEs that are members of the Auto
Component Manufacturer’s Association of India were collected
using a survey approach. Utilizing the path evaluation methods
of modeling structural equations with AMOS, estimation of
parameter data analysis, and the fit of goodness data have been
utilized for evaluating the suggested conceptual framework.
The study [14] used an approach built on data envelopment
analysis (DEA), and it is applied to a group of SMEs in the
nation’s Eastern region. Financial, functional, environmental,
and social factors are regarded as output criteria of the
suggested framework, whereas agile and sustainable
development-oriented inventions are viewed as input criteria.
The DEA separates ineffective SMEs and proposes
benchmarking at least one of them. The research then uses a
qualitative approach to identify methods of enhancement for
the underperforming SMEs. The survey [15] shows that eco-
conscious proprietors who improve the efficiency of the firm
are impacted by technological innovation. Effective businesses
that go above and beyond their financial responsibilities to
support sustainability projects and social well-being may
generate much more money. Management innovation and
employee participation in environmental protection efforts
have the potential to boost both the financial health and the
business’s position among its constituents. The study [16]
analyzed how corporate sustainability (CS) and total quality
management (TQM) are structurally related, as well as whether
TQM methods might help businesses accomplish CS goals. The
research investigates how knowledge management provides an
essential part in the link between TQM and CS and how KM
mediates that relationship.

13.2.1 Hypothesis Development

Hypothesis (H1) Creativity among executives enhances

organizational success.

Hypothesis (H2) The efficiency of organizations is enhanced by

advances in technology.

Hypothesis (H3) Creativity in leadership enhances

sustainability.

Hypothesis (H4) Advancement in technology enhances

organizational sustainability.
Hypothesis (H5) Sustainability enhances the effectiveness of
organizations.

Hypothesis (H6) The interaction of MI and organizational

performance is mediated by sustainability.

Hypothesis (H7) The interaction of TI and organizational

performance is mediated by sustainability.

13.3 Methodology

The study will use a quantitative approach by surveying a cross

section of businesses. Random sampling will be used to choose
the sample, and structural equation modeling will be used to
evaluate the collected data. The survey will include questions
designed to gauge administration, technological innovation,
sustainability, and the efficiency of businesses. Items pertaining
to their leadership, strategic planning, and decision-making
abilities will be evaluated. Items that capture the extent to
which a company invests in R&D, adopts new technology, and
implements innovative procedures will be used to gauge
technological innovation.

13.3.1 Sample and Population

Using the application of a standard survey, Companies with
activity in various countries submitted empirical information
for this research via the use of a conventional questionnaire. It
is challenging to find information about management,
innovation, and technical projects in different countries.
Because most of the businesses fail to report every one of their
expenditures in their annual mentions, there is a measurement
problem. Additionally, stated (subjective) measures are
advocated as offering more trustworthy statistics in developing
countries like China, India, and Pakistan. Key cities in various
countries were picked as targets since they are home to the
majority of enterprises. We collected data from companies that
engage in production, commerce, and services. The 700
randomly selected organizations that made up the distribution
of the 700 surveys each received one survey. Directors were
requested to answer the survey, since they have a better
understanding of the plans and policies of their companies. The
survey explicitly said the information was not going to be
published publicly and was only going to be utilized for
research. In this study, 304 valid responses or a response rate of
43.43% were received. A sample size of at least 300 is necessary
in the cited cities to produce valid results. Table 13.1 lists the
different types of firms that participated in the study.
13.3.2 Techniques

One variable that is dependent, organization performance and

sustainability, was employed as a mediating variable in this
study along with two separate independent variables, MI and
TI. Today’s most common products are unquestionably
available in a range of sizes and the six elements are used it is
divided into three groups: management processes (items 1 and
2), management procedures (items 3 and 4), and structure
(items 5 and 6).

Technology innovation also lacks a general description or set of

measures. We utilized the nine elements, one of the nine being
“We use the most recent innovations for fresh product
advancement” to handle the numerous aspects. “Our
competitors believe us to be a pioneer in the area of
sustainability” is one of the five statements from their prior
survey that we used as a case item for assessing sustainability.

The success of a company is frequently measured using metrics

like “return on assets (ROA),” “return on equity (ROE),” “return
on investment (ROI),” and sales growth. This survey frequently
uses accounting records in the evaluation of the efficiency of
businesses. In this study, organizational performance was
evaluated using the exact method (self-reported) as was used to
evaluate the other components of organizational performance.
On the basis of indicators like ROA, ROE, and ROI, CEOs and
other top executives were asked to assess the performance of
the company.
Table 13.1 Description of the companies.

Variables % Period

Organizations’ mean years

21 and above years 36.5 111

10 years and less 28.3 86

11–20 years 35.2 107

Education of CEOs/ managers

Master 41.8 127

Intermediate and below 22.7 69

Bachelor 28.0 85

PhD etc. 7.6 23

Nature of industry
 Variables % Period

Organizations’ mean years

Trading 32.89 100

Total 100.0 304

Manufacturing 34.21 104

Services 32.89 100

13.3.3 Control

We adjusted for the educational backgrounds of executives and

managers as well as the age of the businesses to reduce
erroneous findings. These variables have a big impact on
performance. According to the structural model, there is a
considerable relationship between CEOs’ and managers’
educational backgrounds and the success of organizations.

13.3.4 Data Analysis

To determine averages, standard deviation (SD), and
information normality, we used descriptive statistical methods.
The findings indicate that MI has the lowest mean value, and
organization performance has the greatest average level. Given
that all of the variables’ skewness and kurtosis values are
within the suggested ranges, our results appear fair ±2 ranges
per George and Mallery. The results are represented in Table
13.2.

13.3.5 Verifying Factor Evaluation

To verify the accuracy and dependability of the variables

utilized, confirmatory factor analysis (CFA) was carried out.
First, we evaluated an estimation method (Figure 13.1) that took
into account every factor. The average value of chi sq./df is
2.135, which is less than 3 as required. We discovered an
appropriate model fit. In accordance with other examinations
of alternative models that fit: Software tools: GFI, AGFI, CFI, TLI,
and NFI provided satisfactory results. These values indicate that
the model is valid: chi sq./df = 2.235; GFI = 0.87; AGFI = 0.84; CFI
= 0.94; TLI = 0.89; NFI = 0.89; root mean squared error (RMR) =
0.015; root mean squared error of approximation (RMSEA) =
0.061. Similarly, RMR and RMSEA also provided appropriate
values demonstrating a satisfactory model fit. On each of their
respective constructs, all of the variables were substantially
weighted (p = 0.01) (Table 13.3).

We examined the average variance extracted (AVE), also known

as convergent validity. To achieve adequate validity for
convergence, AVE value should be greater than 0.50. The reality
that all of the variables have respectable values (e.g., larger
than 0.5) confirms that the variables account for a substantial
amount of variation in the pertinent components. We also
assessed discriminant validity to see whether the items were
not overlapping and explained distinct variances in the specific
constructs. The value obtained from (Table 13.4) provide the
satisfactory and valid results.
Table 13.2 Presentation of illustrative information.

Peak
Parameters Threshold Average S
performance

Organization 1.24 4.95 3.7062

performance

Management 1.44 4.4 3.2684

innovation

Technology 1.38 5.05 3.6157

innovation

Sustainability 1.41 4.56 3.3183

Figure 13.1 Method of measuring model.

To determine if the things are trustworthy and in line with one

another, we also examined compound reliability. All of the
elements’ averages exceeded the cutoff value of 0.70, which was
supported by the findings.

13.3.6 Standard Technique Bias

The standard technique bias (STB) was examined using SPSS
and the Harman one-factor test. Four of the components we
discovered had eigenvalues higher than 1. Additionally, the first
component only described 33.97% of the variance, which was
under 50%. As a result, it was shown that typical bias in
methods is not a concern. Additionally, we examined the impact
of a usual implicit section of the measurement structure and
found that the STB in no way put our investigation’s
conclusions at risk.
Table 13.3 Impact on the various aspects.

Sl.
Variables and items Estimate
no.

Management innovation

mi5 Company system of internal and 0.863

external communication undergoes
frequent revisions.

mi6 We make constant changes to several 0.847

systemic components.

mi3 Modern systems for governance are 0.805

often implemented in our company.

mi4 In the span of three years, there have 0.757

been changes to the reimbursement
policy.

mi1 In our organization, policies and 0.847

processes are updated on occasion.
Sl.
Variables and items Estimate
no.

Management innovation

mi2 The duties and responsibilities of our 0.807

staff are often modified.

Technological innovation

ti8 We incorporate the most recent 0.648

technology breakthroughs quickly.

ti9 In our business, procedures, methods, 0.874

and software evolve quickly.

ti3 The rate at which novel items are 0.768

developed is sufficiently rapid and
viable.

ti4 There are plenty fresh items on sales 0.645

right now.
Sl.
Variables and items Estimate
no.

Management innovation

ti5 We are currently marketing items that 0.854

are initial business adopters.

ti1 We have the capacity to generate goods 0.772

having unique characteristics.

ti2 When creating new products, we use 0.674

the most recent technologies.

Sustainability

s2 Our competition observes us as a top 0.737

business in the environment sector.

s5 We provide social and moral 0.784

requirements a more detailed response
than our rivals.
Sl.
Variables and items Estimate
no.

Management innovation

s3 We create novel products and facilities 0.772

or enhance already existing ones that
are thought to be beneficial for the
environment and for community.

s4 Our track record for durability is 0.824

superior than that of other companies
in this area.

s1 We are the initial business in the sector 0.731

to provide goods and services that are
ecologically conscious.

Organization performance

op1 Profit from properties 0.835

op5 Trade percentage 0.746

 Sl.
Variables and items Estimate
no.

Management innovation

op6 Ultimately profitable 0.854

op2 Income from stocks 0.789

op3 Profit from capital 0.726

op4 Revenues rise 0.858

13.3.7 Correlation

The findings of the Pearson correlation in SPSS are shown in

Table 13.4. Both the association among MI and organizational
performance and the relationship between MI and
sustainability were determined to be significant. TI is
substantially connected to sustainability and organization
performance. It is a significant connection between
sustainability and organizational effectiveness. None of the
correlation coefficients above 0.80 indicated the lack of
multicollinearity.
Table 13.4 Credibility, dependability, and relevance.

Factors AVE CR I II I

Qualification - - 1

Years - - 0.074 1

Org. performance 0.64 0.91 0.202 0.440 0

** ** *

Sustainability 0.58 0.88 0.047 0.192 0

** *

Managementinnovation 0.64 0.92 0.071 0.110 0

Technologyinnovation 0.50 0.89 0.112 0.219 0

** *

13.3.8 Robustness Checks

We carried out the processes in SPSS to test the robustness.

First, we looked at how MI and TI affected sustainability, and
we discovered that both had a positive impact. The results
confirmed those of AMOS. We looked at how MI and TI affected
company efficiency and found that there was a critical direct
influence.

Additionally, the connections between TI and company

efficiency as well as between MI and organizational efficiency
were shown to be strengthened when viability served as an
intermediary. As a result, we contend that there was no
apparent distinction between the results of SPSS and AMOS.

13.3.9 Interaction Analysis

To verify that the relationship between sustainability in MI and

TI is represented, we added an extra model. The regression
techniques are showing the immediate indications and
significant interaction affects for both MI and TI, indicating that
the signs linked to these values propose the exact direction that
MI and TI are likely to affect organizational performance.

In the same way, TI also altered throughout the sustainability

interaction period, going from 0.255 to -0.831. Sustainability
was important as a mediator. Sustainability’s relationship term
with MI produced a poor outcome, while it produced a good
result with TI. The substantial values of both interaction terms
demonstrated that sustainability had an impact on and had the
power to alter efficiency. The scale and methods used to
quantify the components meant that the change was not
ensured.

Table 13.5 Inferential statistics.

Unreliable Standard
parameters parameters

Model Standard B Beta T S

Error

(Constant) 0.206 1.281 6.221 0

Tech. 0.049 0.359 0.388 7.377 0

innovation

Mgt. 0.054 0.226 0.220 4.171 0

innovation
Figure 13.2 Model of structure.

Although a formal hypothesis was not put out to test whether

MI causes TI, MI may cause TI. A structural model is used to
determine if the case for better TI as a result of MI efforts
should be made. The findings in Table 13.5 showed that MI
strongly contributed to TI. Therefore, we contend that there is a
good likelihood that MI will integrate TI into enterprises. Figure
13.2 shows the model of structure. The evaluation of
interactions is shown in Table 13.6.

13.4 Discussion

In this study, the impact of TI and MI on organizational

effectiveness was investigated. The goal of this study is to put
the model to the test using data in a developing country. We
provide evidence that sustainability considerably enhances
financial performance in the developing economies. Opposing
views put out by Yu and Zhao [17] may be adequately
addressed by our study, which concludes that sustainability and
financial success are significantly positively correlated. Our
results are also very supportive of the resource-based view of
the firm (RBV) theory, demonstrates how a company with
special and beneficial assets and abilities may create an edge in
order for competing better in a market. There is much to be
done to more completely conceive and then experimentally
analyze this field in order to discover the distinctive effect of
each invention. For the same reason, we only employed six
dimensions to assess organizational success. Future research
should take into account other factors to further explain the
findings. In order to thrive over the long term, business
organizations, particularly those operating in volatile markets,
require creative strategies. Countries like India, Pakistan has
many characteristics with other nations; therefore, the
consequences apply equally to developing and mature
economies.
Table 13.6 Evaluation of interactions.

Coefficients

Unreliable Standard
parameters parameters

Model B Std. Beta t

Error

Years 0.165 0.027 0.298 7.217

Qualification 0.047 0.095 0.137 2.988

Management 1.778 0.375 1.525 5.248

innovation

Technological -0.828 0.224 -0.856 -3.157

innovation

Sustainability 0.718 0.208 0.647 3.538

Coefficients

Management -0.489 0.100 -2.357 -4.864

innovation ×
Sustainability

Technological 0.377 0.028 1.859 4.125

innovation ×
Sustainability

Years 0.159 0.089 0.328 7.185

Qualification 0.047 0.059 0.157 3.056

Management 0.157 0.089 0.157 3.089

innovation

Technological 0.257 0.056 0.277 5.256

innovation

Sustainability 0.324 0.052 0.297 6.210

(Persistent) -0.689 0.625 -0.959

 Coefficients

Years 0.158 0.056 0.357 7.368

Qualification 0.057 0.026 0.147 2.756

Management 0.247 0.089 0.203 4.347

innovation

Technological 0.347 0.04 0.362 7.671

innovation

(Persistent) 0.715 0.199 3.589

(Persistent) 3.024 0.083 36.62

Years 0.236 0.028 0.427 8.386

Qualification 0.083 0.025 0.171 3.350

(Persistent) 1.127 0.199 5.658

13.5 Conclusions
This study offers a clear image of opposing hypotheses about
the connection between sustainability and financial success, as
well as the RBV hypothesis. Because of the inconsistent findings
of other research, this study aims to assess the impact of
management innovation and technology innovation on the
efficiency of companies with an intermediary function for
sustainability. To assess our strategy, we used a model made up
of structural equations in AMOS and SPSS to analyze
information gathered from 304 executives and other senior
executives. Our investigation revealed that TI is more crucial
than MI for the viability and success of businesses in
developing countries’ emerging economies. Firm leadership
should pay close attention to how MI is configured across
divisions, since doing so significantly improves performance
and longevity. Companies would do well to focus on MI and TI
rather than relying on just one, as they are the primary
markers of long-term viability and financial success. In
conclusion, we advocate for CEOs and executives to give MI and
TI implementation the time and care it needs in order to
succeed in the long run.

References
1. Cheng, Y. et al., How do technological innovation and fiscal
decentralization affect the environment? A story of the
fourth industrial revolution and sustainable growth. Technol.
Forecast. Soc. Change, 162, 120398, 2021, doi:
10.1016/j.techfore.2020.120398.
2. Pan, X. et al., Dynamic relationship among environmental
regulation, technological innovation and energy efficiency
based on large scale provincial panel data in China. Technol.
Forecast. Soc. Change, 144, 428–435, 2019, doi:
10.1016/j.techfore.2017.12.012.
3. Garousi Mokhtarzadeh, N. et al., Investigating the impact of
networking capability on firm innovation performance:
Using the resource-action performance framework. J.
Intellect. Cap., 21, 6, 1009–1034, 2020, doi: 10.1108/ JIC-01-
2020-0005.
4. Teece, D.J., A capability theory of the firm: An economics and
(strategic) management perspective. N. Z. Econ. Pap., 53, 1, 1–
43, 2019, doi: 10.1080/00779954.2017.1371208.
5. Parida, V. et al., Reviewing literature on digitalization,
business model innovation, and sustainable industry: Past
achievements and future promises. Sustainability, 11, 2, 391,
2019, doi: 10.3390/su11020391.
6. El-Kassar, A.N. and Singh, S.K., Green innovation and
organizational performance: The influence of big data and
 the moderating role of management commitment and HR
practices. Technol. Forecast. Soc. Change, 144, 483–498, 2019,
doi: 10.1016/j.techfore.2017.12.016.
7. Ode, E. and Ayavoo, R., The mediating role of knowledge
application in the relationship between knowledge
management practices and firm innovation. J. Innov. Knowl.,
5, 3, 210–218, 2020, doi: 10.1016/j.jik.2019.08.002.
8. Demir, A. et al., Links between knowledge management and
organizational sustainability: Does the ISO 9001 certification
have an effect? ISO 9001. Knowl. Manage. Res. Pract., 21, 1,
183–196, 2023, doi: 10.1080/14778238.2020.1860663.
9. Chege, S.M. et al., Impact of information technology
innovation on firm performance in Kenya. Inf. Technol. Dev.,
26, 2, 316–345, 2020, doi: 10.1080/02681102.2019.1573717.
10. Chen, Q. et al., Effects of organizational innovation and
technological innovation capabilities on firm performance:
Evidence from firms in China’s Pearl River Delta. Asia Pac.
Bus. Rev., 26, 1, 72–96, 2020, doi:
10.1080/13602381.2019.1592339.
11. Singh, S.K. et al., Top management knowledge value,
knowledge sharing practices, open innovation and
organizational performance. J. Bus. Res., 128, 788–798, 2021,
doi: 10.1016/j.jbusres.2019.04.040.
12. Abubakar, A.M. et al., Knowledge management, decision-
making style and organizational performance. J. Innov.
Knowl., 4, 2, 104–114, 2019, doi: 10.1016/j.jik.2017.07.003.
13. Sinha, N. and Dhall, N., Mediating effect of TQM on
relationship between organisational culture and
performance: Evidence from Indian SMEs. Total Qual.
Manage. Bus. Excell., 31, 15-16, 1841–1865, 2020, doi:
10.1080/14783363.2018.1511372.
14. De, D. et al., Impact of lean and sustainability oriented
innovation on sustainability performance of small and
medium sized enterprises: A data envelopment analysis-
based framework. Int. J. Prod. Econ., 219, 416–430, 2020, doi:
10.1016/j.ijpe.2018.07.003.
15. Chege, S.M. and Wang, D., The influence of technology
innovation on SME performance through environmental
sustainability practices in Kenya. Technol. Soc., 60, 101210,
2020, doi: 10.1016/j.techsoc.2019.101210.
16. Abbas, J., Impact of total quality management on corporate
sustainability through the mediating effect of knowledge
management. J. Cleaner Prod., 244, 118806, 2020, doi:
10.1016/j.jclepro.2019.118806.
17. Yu, M. and Zhao, R., Sustainability and firm valuation: An
international investigation. Int. J. Account. Inf. Manage., 23,
289–307, 2015, https://doi.org/10.1108/IJAIM-07-2014-0050.
Note

1. *Corresponding author: [email protected]

14
Optimal Cell Planning in Two Tier
Heterogeneous Network through
Meta-Heuristic Algorithms
Sanjoy Debnath1*, Amit Baran Dey2 and Wasim Arif2

1
Department of ECE, VelTech Rangarajan Dr. Sagunthala R&D
Institute of Science and Technology, Tamil Nadu, India

2
Department of ECE, National Institute of Technology Silchar,
Assam, India

Abstract

This work comprises optimal cell planning of a two-tier

heterogeneous network that consists of macro and small cells
with mobile and machine type user in emergency. The optimum
cell planning is a promising work to be solved for efficient
resource allocation of a wireless network. Cell planning deals
with the optimum placement of cells while maximizing the
overall network utility. Here, the development of utility
function comprises intercell interference and user throughput
satisfaction. This work deals with the maximization of network
utility of a two-tier heterogeneous network under the
constraints of minimum power consumption, user coverage,
and satisfaction fulfillment. The user satisfaction could be
thought of as user association with the optimal cell while
fulfilling the demanded data rate. The effectiveness of the
developed model is analyzed through optimal allocation of
resources among emergency and normal users, while
maximizing network utility.

Keywords: Heterogeneous network, network slicing, cell

planning, resource allocation, user association

14.1 Introduction

Heterogeneous network (HetNet) is a phrase that relates to the

use of numerous devices and techniques in a wireless
connection. In order to give network coverage in an area with a
broad variety of cellular service zones, such as an open
outdoors environment, office buildings, dwellings, and
subterranean places, a Wide Area Network can employ a
combination of macrocells, picocells, and femtocells. A HetNet
is a service that connects complicated interoperation between
macrocell, small cell, and, in some circumstances, Wi-Fi
network parts to produce a mosaic of coverage, with handoff
capabilities across network elements. The ARCchart. analysis
predicts that HetNets would help propel the worldwide mobile
infrastructure market to roughly US$57 billion in spending by
2017 [3]. Wireless connection is being used extensively in
networks. Whether it is mobile communications, fixed
telephony via WLL, or wireless personal communication, the
possibilities are endless. The accessible spectrum is limited,
making it a valuable resource. It must be utilized with caution
in order to achieve the network operator’s goals. The key
principles for cell planning in mobile networks, particularly the
LTE network, are discussed here.

In HetNet, a femtocell is considered as a low-power small cell

that is often used in a home or small company. Small cell is a
broader term used in the corporate world, with femtocell being
a subset. Current designs often allow 4–8 constantly operational
mobile phones in a home environment, depending solely on
version number and femtocell hardware, and 8–16
concurrently active mobile phones in corporate settings, again
depending solely on version number and femtocell hardware. A
femtocell aids the service provider in increasing access within
or even at the cell edge, particularly in areas where connection
would otherwise be usually absent. Although WCDMA has
gotten a lot of press, the concept is applicable to each and every
protocol, notably GSM, CDMA2000, TD-SCDMA, and WiMAX.
Cell planning in a wireless communication network is
considered as one of the challenging problems, which has to
consider for each and every type of network design. In [1], the
Gazzai et al., discussed the LTE cell planning problem and
solved the same with the metaheuristic particle swarm
optimization techniques. It was observed that the optimal
deployment of cells mostly depends on the user distribution in
a particular geographical area and their demanded data rate.
Optimal cell planning in a multi-RAT network is discussed in
[2], where Debnath et al., proposed an energy-efficient scheme
to find the optimal resource allocation in a multi-RAT network
under the constraint of user allocation and fulfillment of
demanded data rate. To meet the increasing user demand in the
cellular network, multi-RAT HetNet is one of the solutions. To
improve the energy efficiency of the network, a multi-objective
framework for the resource allocation in HetNet is analyzed
through Pareto optimal solution [3].

On the other hand, spectrum sharing is an important feature

for utilizing the wireless resource effectively [4]. In [4], Ye et al.,
analyzed the schemes of optimal allocation of resource slices in
the base station (BS) for quality-of-service (QoS) improvement.
As optimum cell planning also offers effective resource
allocation in wireless communication, thus the concept of
resource slicing is merged with optimization for effective
allocation of resources [5]. It was observed that network slicing
scheme maximizes the overall bandwidth resource utilization.

Providing the best algorithm for solving strategy policy, upper

subchannel allocation, uplink maximum transmission
allocation, and computational resource scheduling, Melendex et
al., in 2017 used a technique called joint computation offloading
and resource allocation optimization. The advantages in this
scheme is lower complexity, time saving, and reduction of
energy consumption [6]. However, the computational time and
accuracy both are equally important in optimization, one of the
ways are to offload the computation and maximize the overall
efficiency of the network management [6]. In [7], Wang et al.,
discuss on the computation offloading and intercell
interference minimization while managing effective resource
allocation. Whereas minimization of energy consumption in
offloading is also a part of concern in wireless communication
[8]. Thus, considering task offloading, the resource allocation
and spectrum efficiency improvement work is discussed in [9].

It is also observed that the NOMA-based HetNet could also be

considered for the maximization of spectrum efficiency and
resource allocation [10]. As both spectrum efficiency and
resource allocation are important for optimal network utility,
we utilize a metaheuristic algorithm for energy-efficient
spectrum allocation to serve emergency users. For a guaranteed
QoS, the strategy of NOMA is highly important to minimize the
energy consumption of all users. Thus, the application of a
heuristic algorithm in NOMA-based HetNet for the
maximization of overall network utility will provide balanced
energy and spectrum efficiency. The advantage of NOMA
includes optimal power distribution, QoS-oriented user
fairness, suitable user pairing, and good link adaptations. To
achieve the desired throughput in 5G communication
techniques, NOMA is one of the desirable techniques that meet
both energy and spectrum efficiency [11]. NOMA is basically of
two types, namely, power domain and code domain. Scheduling
and performing the power description adaptation in the
downlink, power domain NOMA is a suitable technique. In [12],
Liu et al., discussed and analyzed the power domain NOMA in
respect of user throughput and power allocation strategy. It was
observed that the proposed joint transmission-based NOMA
model offers improved user coverage and throughput [12].

For the maximization of the network utility that consists of both

spectrum efficiency and energy efficiency, metaheuristic
algorithms are getting high importance. For finding the near-
optimal solutions because of the non-convexities, a technique
called mixed-integer nonlinear programs can also be utilized,
but these techniques have some complexities [13]. The particle
swarm optimization (PSO) with less complexity and high
exploration capability of the search space is well suited to find
the optimal solution in the complex communication problems
[14]. Due to simplicity and ease of implementation, the PSO and
its supplementary algorithms like the gravitational search
algorithm (GSA), grey wolf algorithm (GWO), and whale
optimization algorithm (WOA) are very popular among the
researcher [14]. Afterward, to improve the balancing between
the exploration and exploitation rate, a hybrid version of the
algorithms is suggested [15]. To get the solution of the complex
problem accurately and effectively, a hybrid version of the
algorithms is preferred [16].

The application of hybrid meta-heuristic algorithm to maximize

the energy efficiency is found effective and promising in
wireless communication. Towards that a hybridize algorithm
based on PSO and GSA is proposed for the maximization of
energy efficiency in a cognitive radio network is proposed in
[17]. A computational efficient hybrid differential evolution
with learning (HDEL) is used for utility maximization in un-
manned aerial vehicle network [16]. Still the problem of
network planning for emergency user coverage and network
utility maximization is not addressed in the literature. Thus, we
are motivated to perform this work to solve the said problem
with the highly computational efficient DADE algorithm.
The rest of the paper is arranged as follows. Section 14.2 of this
paper includes proposed methodology, section 14.3 presents
results and discussion, followed by section 14.4 that draws the
conclusion of the paper.

14.2 System Model and Formulation of the

Problem

As shown in the Figure 14.1, consider a NOMA-assisted two-tier

HetNet comprises one macro base station (MBS) and K number
of SBS. Assume that all SBSs are NOMA assisted and connected
to one another through a fusion center (FC). The task of the FC
is to accumulate the channel state information and allocate the
resource accordingly. Consider that there are M numbers of
NUs and F numbers of EUs randomly distributed over the
coverage region of the MBS. It is also assumed that for w
allocated bandwidth, S numbers of channels are allocated.
Figure 14.1 System model.

The index sets of all of the assumed parameters are denoted as

follows for the SBS: k = {k ||1 ≤ k ≤ K}, for NU: M = {m |1 ≤ m
≤ M}, for EU:F = {f |1 ≤ f ≤ F}, and for S: S = {s |1 ≤ s ≤ S}. At
any particular instant of time, the possibility of one channel is
free or idle is as follows:

(14.1)

where is the average idle time and is the busy time

of sth channel.

The number of channels assigned to SC may not be sufficient

for allocating to all the EU of the SC zone. Assume that F
number of EU multiplexed over an idle block of spectrum, and
the EUs multiplexed over the same channel form a NOMA
cluster. Considered that the set of NOMA cluster under kth SC is
Ck. Thus for total N number of NOMA clusters 1 < n < N, one
cluster could be represented as .

One EU tries to communicate to the SBS, and the interchange

modulated symbol xk,i between the kth SBS and ith EU can be
represented as follows:

(14.2)

where Pk,i is the transmitted power from the BS to the EU.

The received modulated signal at the EU, when the channel gain
between the BS and the EU is hk,i, can be written as:

(14.3)
(14.4)

where N0 is the additive white Gaussian noise.

Thus, for a particular NOMA cluster , the signal-to-

interference-plus-noise ratio (SINR) γk,i can be computed as:

(14.5)
where |Hk,i|2 is the normalized channel response of the ith EU.
Ik is the interference received from the other EUs from the
same NOMA cluster and can be represented as:

(14.6)

Selection parameter αn,k for one EU to choose a particular SBS

and NOMA cluster is represented as:

(14.7)

Based on that, the throughput of ith EU of cluster can be

computed as:

(14.8)
where can be represented as:

(14.9)

Now, based on the Shannon capacity formula, the network

utility for the SB network can be written as:

(14.10)

(14.11)
Accordingly, the optimization problem for resource allocation
can be formulated as:

(14.12)

S.t.

(14.13)
(14.14)

(14.15)
(14.16)

(14.17)

(14.18)

s (14.19)
(14.20)

Description of the constraint:

Equation 14.13 represents the limitation of the selection

parameter αn,k.

Equations 14.14 and 14.15 discuss the bandwidth constraint,

where it says that the total number of channels should be
greater than the idle channel.
Equations 14.16 and 14.17 deal with the clustering constraint,
where one EU can associate with only one NOMA cluster in one
SBS. Equation 14.17 describes that the entire set of Ck NOMA
clusters is a subset of F.

Constraint 6 (Equation 14.18) deals with the successive

interference cancellation (SIC) implementation complexity.

Equation 14.19 states that the power allocated to all of the EUs
is less than the total budget of the power pk.

Lastly Equation 14.20 provides a guarantee that one EU

acquires minimum data rate.

Consider a large geographical area where we can deploy n

number of base stations in order to distribute data among the
users. So here, we have to ensure to check all of the parameters
before we deploy them into the geographical area. We can say
that parameters are nothing but the macro base stations, small
base stations, primary users, and emergency users. So, we need
to check where we have to deploy all of these parameters in the
large geographical area. We can randomly insert the user
position as our wish. Now, we should take a random value as an
iteration count. It denotes the number of iterations taken to
complete the process. Figure 14.2 shows the flowchart of the
proposed model.

After the initialization of user position and parameters, we

have to randomly deploy macro base stations and small base
stations into the large geographical area. Here, we have some
rules to follow that the macro base station should be placed at
the center of the large geographical area. After placing a macro
base station at the center of the large geographical area, we
have to cordially place the SBSs with respect to the MBS. It looks
like neighboring BSs regardless of both macro and small base
stations. After arranging both the base stations in their
respective positions, we have to evaluate network resources
and air traffic.
Figure 14.2 Flowchart of the proposed model.

Where the network resources are nothing but the power and
the bandwidth used in any communication. We should evaluate
about how much power is transmitted through the entire
process or how much power is there for the process of
transmission. Coming to bandwidth, we have to evaluate how
much bandwidth is generally utilized by one communication
node while utilizing channels in this entire process. Now,
coming to air traffic, it is the data transferred in the air from
one user to another user or may be from one base station to
another base station.

There will be a lot of disturbances while the information is

passed. For example, if there is a single BS in our surroundings
and there are total N numbers of users surrounded by that BS.
It is not possible for one BS to accommodate all of the users in
the concerned coverage area because of its resource capacity.
This problem can be solved by deploying more BSs in that area.
So, we have to evaluate both network resources and air traffic.
After completion of this step, we have to formulate the network
utility function. The condition of a network in availability of
resources and capability to cater to all of the users in the
concerned area in terms of end user perceived QoS is known as
utility function.

These utility functions are used to allocate resources and

manage data rates. The original notion is expanded by the
network utility function (NUF) concept presented here. As a
measure of network connections, the NUF reflects signal
absorption as well as the efficiency and utility of a service.
Network utility function is evaluated based on three steps. They
are as follows:

1. Evaluate the inter-cell interference

2. Guarantee user QoS satisfaction
3. Verify total geographical coverage

The rapid expansion of communication networks over the

previous decade has set new aspirational expectations about
future wireless connections to meet the ever-increasing
demand for user. However, the fundamental limitations of
communication channels in cellular operators make achieving
the goals impossible. Better spectral reuse performance
between cells is essential to deal with the huge cost and
shortage of acceptable wireless spectrum, which contributes to
increasing levels of interference. Thus, the concept of QoS
comes in the communication network for resource monitoring
and framework designed to prevent interference such as
information leakage, jitter, and latency. QoS also establishes
limits and targets for sources of data that travel as high
bandwidth traffic through networks.

After the above two steps, we have to check the total

geographical area for the association of deploying BSs and
users. After the completion of the above three steps, we can
formulate the network utility function. In order to assign each
user to the nearest BS optimally, we have to follow two steps.
They are as follows:

1. Search for optimal solution based on a metaheuristic

2. Update MBSs and SBSs locations

Utilize a metaheuristic algorithm for finding optimal location of

BSs (SBS and MBS) while maximizing network utility. The
proposed algorithm will check if the utility function is
maximum or not. If it is maximum, then it is processed to the
next step of process until all of the iteration has completed. If
the utility function is low, it will again come and check the
locations of the BS and the location is updated accordingly
under the constraint of user data rate satisfaction. Here, user
data demand is nothing but the amount of data rate a normal
user required when the user start communication. If it is full,
then it is processed to the next step. If it is not full, then it will
go back and the locations of the base stations are updated again
until it reaches the user demand. The next step is the iteration
count should be less than the maximum iteration count.

14.3 Result and Discussion

In simulations, assume the coordinates for the MBS as (2500,
2500), and SBSs are chosen randomly in the geographical area.
A total of 1,300 users including 900 mobile and 400 machine-
type devices (MTDs) were randomly deployed. Here, MTDs are
the devices generating constant data traffic for communication
like IoT devices. Analyze the resource allocation and user
association model in two data suits having different data rates
for NU/EU and MTD. The parameter assumptions for the
simulation are listed in Table 14.1.

In general, users within the coverage area of one SBS also come
under the coverage of the MBS. Thus, received power-based
association gives the first priority to SBS, whereas demand
fulfillment has to be taken care of with the SINR. Thus, user’s
association has to perform with SBS or MBS based on received
power and data rate. In case a device is free to choose its
association with a BS. The association with the SBSs is
established on a first-come, first-served basis. When the
threshold is reached, the users are connected to the MBS
regardless of their geographical location.

The analysis of the optimal user association to BS is presented

in Figures 14.3 and 14.4. Observe that the users are optimally
associated with the nearby BSs on the basis of their data rate
demand.
Table 14.1 Parameter values.

Parameters Value

QoS exponent 3

Frequency of operating carrier signal, f 1800 MHz

Total bandwidth resources, w 20 MHz

Noise power, -104 dBm

Packet data size, 9000 bits

Machine data packet size, 2000 bits

Transmit power of MBS, Pk 46 dBm

Transmit power of SBS, Pk 23 dBm

Data Suit-1

Data rate demand for NU/EU 128 Kbps

 Parameters Value

Data rate demand for MTD 56 Kbps

Data Suit-2

Data rate demand for NU/EU 256 Kbps

Data rate demand for MTD 128 Kbps

Figure 14.3 Optimal user association to BSs with data suit-1.

Figure 14.4 Optimal user association to BSs with data suit-2.

From Figure 14.3, it is observed that the users are associated

with the BSs according to their geolocation and data rate
demand. The practical capacity of one SBS is limited and not
have sufficient ability to cater much user. Thus, for the
maximization of network utility, users have to be optimally
associated with the BS having higher capacity and ability to
fulfill the demanded data rate. Thereafter, we analyze the same
with higher demanded data rate of users. Figure 14.4 shows the
association of optimal users following demanded data rate suit-
2.

From Figure 14.4, it is observed that the smaller number of

users is associated with the SBS as compared to the previous
case of data suit-1. Results observed are intuitively justified, as
the capacity of one SBS is limited and this limitation provides
the variation in network accessibility on the basis of demanded
user data rate. Thus, in this case, SBSs are optimally deployed to
the same place with less number of associated users.

In this work, we analyze the optimal user association to optimal

BS by three well-known metaheuristic optimization algorithms,
namely, the PSO, GWO, and DADE. Results of Figure 14.5 depicts
that the maximum network utility achieved with the GWO and
DADE is 1.65E+04 and the maximum network utility achieved
with the PSO is 1.66E+04. It is also observed that the
convergence characteristics of the DADE is superior to those of
the GWO and PSO.
Figure 14.5 Network utility maximization graph.

14.4 Conclusion

In this paper, we analyzed optimal cell planning in a two-tier

heterogeneous network keeping the emergency user on high
priority than the normal user. It was observed that optimal user
association to optimal cell is highly required for user data rate
satisfaction. It is also observed that optimal resource allocation
while balancing emergency user and normal user association
provides maximum network utility. It was observed that
prioritizing emergency users and finding the optimal location
of SBS for catering EU/NU provide satisfactory resource
utilization. In the future, considering network function
virtualization, the performance of the network utility is
analyzed under the constraint of user satisfaction.

References

1. Ghazzai, H., Yaacoub, E., Alouini, M. -S., Dawy, Z., Abu-Dayya,

A., Optimized LTE cell planning with varying spatial and
temporal user densities, in IEEE Transactions on Vehicular
Technology, vol. 65, no. 3, pp. 1575–1589, March 2016,
doi:10.1109/TVT.2015.2411579,
https://ieeexplore.ieee.org/abstract/document/7056465.
2. Debnath, S. et al., Energy efficient optimal resource allocation
in multi-RAT heterogeneous network. Appl. Artif. Intell.
Taylor & Francis, 35, 15, 2246–2275, 2021, doi:
10.1080/08839514.2021.1998300.
3. Yu, G. et al., Multi-objective energy-efficient resource
allocation for multi-RAT heterogeneous networks. IEEE J. Sel.
Areas Commun., 33, 10, 2118–2127, 2018, doi:
10.1109/JSAC.2015.2435374.
4. Ye, Q. et al., Dynamic radio resource slicing for a two-tier
heterogeneous wireless network. IEEE Trans. Veh. Technol.,
67, 10, 9896–9910, 2018, doi: 10.1109/TVT.2018.2859740.
5. Bahlke, F. et al., Optimized cell planning for network slicing in
heterogeneous wireless communication networks. IEEE
 Commun. Lett., 22, 8, 1676–1679, 2018, doi:
10.1109/LCOMM.2018.2841866.
6. Melendez, S. and McGarry, M.P., Computation offloading
decisions for reducing completion time, in: Proc. 14th IEEE
Annu. Consum. Comm. Netw. Conf. (CCNC), Jan. 2017, pp. 160–
164, doi: 10.1109/CCNC.2017.7983099.
7. Wang, C. et al., Joint computation offloading and interference
management in wireless cellular networks with mobile edge
computing. IEEE Trans. Veh. Technol., 66, 8, 7432–7445, Aug.
2017, doi: 10.1109/TVT.2017.2672701.
8. Zhang, K. et al., Energy-efficient offloading for mobile edge
computing in 5G heterogeneous networks. IEEE Access, 4,
5896–5907, 2016, doi: 10.1109/ ACCESS.2016.2597169.
9. Xu, C. et al., Energy-minimization task offloading and
resource allocation for mobile edge computing in NOMA
heterogeneous networks. IEEE Trans. Veh. Technol., 69, 12,
16001–16016, 2020, doi: 10.1109/TVT.2020.3040645.
10. Shin, W. et al., Non-orthogonal multiple access in multi-cell
networks: Theory, performance, and practical challenges.
IEEE Commun. Mag., 55, 10, 176–183, Oct. 2017, doi:
10.1109/MCOM.2017.1601065.
11. Islam, S.M.R. et al., Power-domainnon-orthogonal multiple
access (NOMA) in 5G systems: Potentials and challenges. IEEE
 Commun. Surv. Tutorials, 19, 2, 721–742, Apr.-Jun. 2017, doi:
10.1109/COMST.2016.2621116.
12. Liu, C. and Liang, D., Heterogeneous networks with power-
domain NOMA: Coverage, throughput, and power allocation
analysis. IEEE Trans. Wirel. Commun., 17, 5, 3524–3539, May
2018, doi: 10.1109/TWC.2018.2816923.
13. Burer, S. and Letchford, A.N., Non-convex mixed-integer
nonlinear programming: A survey. Surv. Oper. Res. Manage.
Sci., 17, 2, 97–106, Apr.-Jun. 2012, doi:
10.1016/j.sorms.2012.08.001.
14. Clerc, M. and Kennedy, J., The particle swarm-explosion,
stability, and convergence in a multidimensional complex
space. IEEE Trans. Evol. Comput., 6, 3, 58–73, Feb. 2008, doi:
10.1109/4235.985692.
15. Debnath, S. et al., A hybrid memory based dragonfly
algorithm with differential evolution for engineering
application. Eng. Comput., Springer, 37, 4, 2775–2802, 2021,
doi: 10.1007/s00366-020-00958-4.
16. Debnath, S. et al., Computationally efficient Hybrid
differential evolution with learning for engineering
application. Int. J. Bio, 19, 1, 29–39, 2022, doi:
10.1504/IJBIC.2022.120744.
17. Eappen, G. and Shankar, T., Hybrid PSO-GSA for energy
efficient spectrum sensing in cognitive radio network. Phys.
 Commun., 40, 101091, 2020, doi:
10.1016/j.phycom.2020.101091.

Note

1. *Corresponding author: [email protected]

15
Soil Aggregate Stability Prediction
Using a Hybrid Machine Learning
Algorithm
M. Balamurugan

Department of Computer Applications, Acharya Institute of

Graduate Studies Bengaluru, India

Abstract

The concept of soil aggregate stability (SAS) is utilized as a

criterion for soil composition because several soil ecosystem
functions depend on the presence of stable aggregates.
Predictive models have gained increased attention as a
replacement for direct measurements because aggregate
stability is rarely recorded in soil surveys. In order to anticipate
soil aggregate stability, this work aims to develop a novel
machine learning (ML) technique called the hybrid tree-based
twin-bounded support vector machine (HT-TBSVM).
Unexpected discoveries were made by examining soil records
that provided information on soil properties such as structure,
soil nutrient concentration, alkalinity, and moisture aggregates.
This collection includes 109 soil samples from Chile’s hyperarid,
arid, and semiarid regions, as well as humid areas, including
cultivated fields, grasslands, and tree plantations. The most
prevalent soil types in this dataset were clay loam, sandal loam,
and loam, and the values for each soil attribute ranged widely.
The mean absolute error (MAE), R2, normalized root mean
square error (nRMSE), and root mean square error (RMSE) are
used as different indicators to determine the effectiveness of
the given prediction. We evaluated these measurements in
comparison to conventional approaches to predicting SA
stability. The experimental results show that the proposed
approach outperforms conventional approaches in terms of
outcomes.

Keywords: Soil, soil aggregate stability (SAS), deep neural

network (DNN), support vector machine (SVM)

15.1 Introduction

The management of soil must be sustainable to protect it for

future generations, given the increasing environmental
pressures it is facing. This management cannot be carried out
without an in-depth knowledge of the diverse features and
properties of the soil. Aggregate stability, which influences
water storage and the stabilization of organic carbon (OC), is a
crucial component of soil conservation, along with the
preservation of its environmental functions. Increased soil
structural stability may also result in increased resistance to
corrosive agents and compaction. Stable soil aggregates (SAs)
create a solid soil structure that allows for the best potential
gas, water, and nutrient transport and storage. Therefore, soil
aggregate stability (SAS) may serve as a valuable indication for
assessing soil quality [1]. The challenging process of soil
aggregation involves the complex interplay of the soil’s organic
and mineral components, resulting in a microaggregate (0.25
mm) within a macroaggregate [2]. SAS is essential for soil
resilience to mechanical stresses such as those brought on by
surface runoff and rainfall, which, in turn, contributes to soil
resistance to water erosion. As SA degrades, finer particles are
created. Since these particles are effectively transported by
water and wind flow, they tend to clog soil pores and create
crusts when re-sedimented [3]. Soil aggregate stability
distribution (SASD) data are typically statistically quantified
using the mean weight diameter (MWD) and the geometric
mean diameter (GMD). Because these mass fractions (mi) of SA
depend on different (di) resemble probability distribution
functions (PDFs), higher-order moments, such as standard
deviation (σ) and skewness coefficient, are used to characterize
soil aggregate size distribution (SASD) data in addition to their
mean values [4]. However, the dynamic and complicated
relationship between SAS and soil factors cannot be captured
by the linear regression model, which typically produces
inferior prediction accuracy [5].

The following is the paper’s primary contribution:

1. The hybrid tree-based twin-bounded support vector machine

was proposed.

The following is the paper’s organization: Section 15.2 reviews

the relevant works, Section 15.3 explains the suggested
approach, Section 15.4 presents the results and discussions, and
Section 15.5 presents the conclusions.

15.2 Related Works

Cooke et al. [6] studied the value of digestates as organic soil

additions and how it varies based on process variables and
posttreatments. Pavlů et al. [7] investigated whether the same
approach can be applied to different types of soils, including
Haplic Luvisol, two Haplic Cambisols, and a Calcaric Leptosol.
Xu et al. [8] explored the interplay between fluoranthene’s
actions and soil aggregate structure, while Shi et al. [9] showed
the potential of hyperspectral imagers for assessing SAS and
aggregating size fractions on a massive scale. Liu et al. [10]
analyzed the impact of monovalent cations on aggregate
stability and splash erosion, and Zhou et al. [11] studied the
effect of soil binders on aggregate stability in urban forest soils
with organic mulch. Ma et al. [12] used Gaussian geostatistical
simulation and multivariate empirical mode decomposition
techniques to investigate the geographical distribution and
scale-specific implications of water-stable aggregates (WSA)
content. Xue et al. [13] conducted a multiyear investigation to
understand the variables that affect SAS in a rice-rape
agricultural system, and Tang et al. [14] examined how different
land use practices influence SA stability and aggregate-
associated OC and total nitrogen (TN) concentrations.

15.3 Proposed Methodology

Soil aggregation refers to the joining and holding together of

aggregates of various sizes by various organic and inorganic
components. It encompasses both the formation and stability
processes, which run continuously and even act simultaneously
in the field. As a result, it can be challenging to distinguish their
interrelated impacts. Some writers attribute the development of
soil aggregation primarily to physical forces, while others argue
that a variety of elements, including the quantity and quality of
inorganic and organic stabilizing agents, are responsible for its
stabilization. Figure 15.1 shows a block representation of the
proposed method.

Figure 15.1 Block diagram of soil aggregation.

15.3.1 Soil Samples and Characteristics

To collect a diverse set of data for the soil characteristics used in

this research, the sampling area was designed to include a
variety of soil types, land uses, and climate variables. The study
collected 109 soil measurements from the ground in Chile
between the latitudes of 27°43′S and 51°20′S at a depth of 0–17
cm. The research area is located in four of Chile’s eight soil
zones, including the Magellan, arid Mediterranean (37 sites),
and desert (19 sites), with 23 sites in total. The study area covers
a range of climates, from humid to hyperarid. The desert zone,
for example, has an annual precipitation of only 43 mm,
making it hyperarid, while the Mediterranean region has an
arid climate with an annual precipitation of no more than 440
mm. The Mediterranean and Magellan zones, on the other
hand, receive annual rainfall totals of 1,220 mm and 238 mm,
respectively, and are classified as such. The soil orders
examined in the study include Alfisols, Andisols, Histosols,
Inceptisols, Mollisols, and Ultisols. The study found an increase
in Andisol presence from the Mediterranean to the Magellan
zones. The majority of the samples taken from the desert, wet,
and moist Mediterranean zones were found in the middle
valley, which has the majority of cultivated soils and is where
most of the country’s productive land, grassland, and plantation
regions are concentrated. However, the Serrano River Basin,
with its Magellan zone soils, is home to native forests, shrubs,
and peaty scrubland, according to Luzio et al. [15].

15.3.2 Analyzing Soil Samples

At a sieving size of 2 mm, the soil was dried at 40°C. The pH,
quantity of organic matter (OM), and distribution of particle
sizes were measured in a suspension of soil and water prepared
from the sieved fraction (clay, silt, and sand). The proposed
method emphasizes that unstable aggregates degrade when
submerged in water and uses moisture screening equipment to
assess aggregate stability. The wet sieving methodology includes
the fundamental mechanisms of aggregate dispersion, slaking,
and mechanical dispersion, simulating the pressures associated
with water infiltration into SA in the field. A 0.25-mm filter
holding distilled water and 4.0 g of solid crystals between 1 and
2 mm is repeatedly raised and lowered for 3 minutes during the
experiment. Stable aggregates that pass through the sieve
during the 3-minute wet sifting procedure are collected in the
distilled water container below the sieve. After the sieves have
been cleaned until only sand and small root fragments remain,
the canisters are reinstalled and filled with a dispersing
solution. Both sets of cans are then dried at 110°C after
dewatering. Distilled water and dispersing solution are added
to the canisters, and the weights of the water-filled can and its
contents are calculated and subtracted. The weight of the stable
materials is calculated by subtracting the weight of the
dispersing solution and 0.2 g of that solution from the weight of
the container containing the aggregates. The weight of the
stable aggregates obtained in the cans of dispersing solution is
then divided by the sum of the weights of the stable and
unstable aggregates in Equation 15.1, where 100% indicates
that no aggregates were broken and 0% indicates that all
aggregates were broken. The WSA index is then calculated
using the percentage of stable aggregates.

(15.1)

where WSA is the number of particles evenly distributed in

distilled water (g), the total weight (%) dispersed in the
dispersing solvent wds along with the wdw is the aggregates’
weight(%) (g).

15.3.3 Hybrid Tree-Based Twin-Bounded Support

Vector Machine-Based Model

The process of learning hybrid trees (HTs) from training sets of

class labels is referred to as hybrid tree induction. HT is used in
this approach because the greedy strategy employed by C5.0
and the top-down recursive approach are only suitable for
decision tree construction.

15.3.3.1 Hybrid Tree Algorithm-C5.0

HT is often used to address data classification problems. The HT

algorithm typically involves two phases: the training phase,
which constructs the decision tree using the chosen training
data sample, and the prediction phase, which uses the HT
created during the training phase to forecast the target data.
Developing an HT is a key step in creating data classification
rules.

The following data formula uses the most popular decision tree
algorithm, C5.0:

(15.2)

(15.3)

The C5.0 algorithm is composed of the following steps (Figure

15.2):

1. Assuming the conditional qualities are as stated, determine

the original data’s information gain ratio.
2. Select the property Ai with the highest ratio value as the base
node of the decision tree.
3. Sort these actual data into groups based on the root node.
4. Calculate the information gain ratio for the starting data in
each category under the given circumstances, and then
choose a greater value as the decision node.
5. Recursively calling these aforementioned methods until all of
the original data are categorized.
Figure 15.2 C5.0’s algorithm flow.

15.3.3.2. Twin-Bounded Support Vector Machine-Based

Model

Assume that represents the data points in class +1 and

represents the data points in class 1, respectively. Each
row in both matrices corresponds to a data sample.

The nonparallel hyperplanes are located via the twin-bounded

support vector machine (TBSVM) in linear form by Equation
15.1.

(15.4)
Such that each hyperplane is closer to and further distant from
the data points of one class than the data points of the other. It
is given the class name +1 or 1 depending on how close together
the two hyperplanes’ data samples are. TBSVM resolves the two
quadratic programming problems (QPPs) by using the objective
function of one class and constraints related to the other class.

(15.5)

and

(15.6)

where the positive factors c1,c2,c3, and c4 are present in

Equations 15.5 and 15.6.

We must obtain the dual formulations as follows to solve

Equations 15.4 and 15.5.

The QPP (15.4) Lagrangian is provided by:

(15.7)

where the vectors with and

are the…w and bLagrange multipliers and for α, β, ξ, w1, and b1,
the Karush-Kuhn-Tucker (KKT) conditions are given as
Equations 15.7–15.13.

(15.8)
 (15.9)
(15.10)
(15.11)
(15.12)
(15.13)
Since β≥ 0, Equation 15.9 gives us:

Equations 15.8 and 15.9 give us:

(15.14)
Here is the identity matrix, etc. Equation 15.14 becomes true if
V1 = [w1,b1]T, H = [A e1] and G = [B e2]

(15.15)
The dual of the QPP can be found by combining Equation
15.15 with Equations 15.8 through 15.13 in the Lagrangian.

(15.16)

In a similar vein, QPPs dual is given as Equation 15.17:

(15.17)

where γ the Lagrange multiplier is used. If V2 = [w2,b2]T, then

the augmented vector can be written as Equation 15.18.
 (15.18)
The test sample is given one of the class labels i(i = +1,−1)
depending upon its distance from the two hyperplanes (w1,b1)
and (w2,b2) as the solutions to the QPPs and are found by
solving the QPPs (15.16) and (15.17).

(15.19)

The class i sample is derived in Equation 15.19.

where the absolute value is |.|.

15.4 Result and Discussion

The details of the soil prediction for SA stability findings are

covered in this section. The variables used to evaluate the
effectiveness of the suggested method are the mean absolute
error (MAE), normalized root mean square error (nRMSE), and
the coefficient of determination (R2). To assess the effectiveness
of the suggested method, the current random forest (RF), k-
nearest neighbors (kNN), support vector machine, and artificial
neural network (ANN) [16, 17] techniques are compared.

The RMSE represents the standard deviation of the residuals

(prediction errors). Residuals measure the distance between
data points and the regression line, and the RMSE measures the
spread of these residuals. Equation 15.20 illustrates the formula
for calculating the RMSE.

(15.20)

Figure 15.3 demonstrates that for the RMSE, the suggested

method (0.11%) outperformed the results of the current
methods of RF (0.17%), KNN (0.33%), SVM (0.21%), and ANN
(0.32%). It shows the suggested system’s good performance.

The coefficient of determination (R2), which has a range of 0 to

1, is employed to measure how accurately a statistical model
predicts an outcome. The R2 number is the proportion of
variation in the dependent variable that the statistical mode
forecasts. Equation 15.21 illustrates the R2 formula.
Figure 15.3 Comparative analysis of the RMSE.

(15.21)

Figure 15.4 demonstrates that for the R2, the suggested method
(0.82%) outperformed the results of the current methods of RF
(0.74%), KNN (0.21%), SVM (0.28%), and ANN (0.15%). It shows
the suggested system’s better performance.

Comparing the models of different scales using the root mean

square normalized error (nRMSE) is simplified by using the
RMSE. The normalized RMSE connects the observed range of
the variables with the RMSE. The nRMSE expresses the RMSE as
a percentage of the overall range that the model typically
resolves. Equation 15.22 illustrates the formula for calculating
the nRMSE.

(15.22)

Figure 15.5 demonstrates that for the nRMSE, the suggested

method (15.84%) is significantly lower than the results of the
current methods of RF (24.28%), KNN (47.1g1094%), SVM
(45.71%), and ANN (30.00%). It shows the suggested system’s
better performance.

Figure 15.4 Comparative analysis of the R2.

Figure 15.5 Comparative analysis of the nRMSE.

The MAE score is obtained by averaging the absolute error data.

Absolute is a mathematical operation that converts a negative
integer into a positive one. When calculating the MAE in
Equation 15.22, the difference between an expected value and a
forecasted value could be either positive or negative.
Figure 15.6 Comparative analysis of the MAE.

(15.23)

Figure 15.6 demonstrates that for the MAE, the suggested

method (0.10%) is lower than the results of the current methods
of RF (0.12%), KNN (0.25%), SVM (0.25%), and ANN (0.15%). It
shows the suggested system’s good performance.

15.5 Conclusion

The stability of the soil aggregate directly affects the

sustainability of the soil and crop output. This paper provides a
general overview of the mechanisms and forces involved in the
growth and stability of SA. Therefore, the goal of this research
was to create a novel model based on HT-TBSVM to predict SAS.
Measuring the amounts of soil organic carbon (SOC) and SAS
indices in various aggregate sizes takes a lot of time, money,
and labor. In this study, four machine learning models (RF, SVM,
kNN, and ANN) were evaluated for their ability to predict SAS
and SOC in variations in aggregate fractions. When dealing with
large datasets, hybrid tree-based twin-bounded SVM may not
perform well due to a longer required training time. In future
studies assessing the contributions of soil biota to soil
aggregation, species combinations should also be considered.

References

1. Bouslihim, Y. et al., Machine learning approaches for the

prediction of soil aggregate stability. Heliyon, 7, 3, e06480,
2021, doi: 10.1016/j.heliyon.2021. e06480.
2. Rivera, J.I., and Bonilla, C.A., Predicting soil aggregate stability
using readily available soil properties and machine learning
techniques. CATENA, 187, 104408, 2020, doi:
10.1016/j.catena.2019.104408.
3. Besalatpour, A.A. et al., Estimating wet soil aggregate stability
from easily available properties in a highly mountainous
watershed. CATENA, 111, 72–79, 2013, doi:
10.1016/j.catena.2013.07.001.
4. Purushothaman, N.K. et al., National-scale maps for soil
aggregate size distribution parameters using pedotransfer
functions and digital soil mapping data products. Geoderma,
424, 116006, 2022, doi: 10.1016/j. geoderma.2022.116006.
5. Blomquist, J. et al., Soil characteristics and tillage can predict
the effect of structure lime on soil aggregate stability. Soil
Res., 60, 4, 373–384, 2022, doi: 10.1071/SR21022.
6. Cooke, J., Girault, R., Busnot, S. et al., Characterising the effect
of raw and post-treated digestates on soil aggregate stability.
Waste Biomass Valor., 14, 2977–2995, 2023,
https://doi.org/10.1007/s12649-023-02045-3.
7. Pavlů, L. et al., Estimation of the stability of topsoil aggregates
in areas affected by water erosion using selected soil and
terrain properties. Soil Tillage Res., 219, 105348, 2022, doi:
10.1016/j.still.2022.105348.
8. Xu, C. et al., Forest conversion effects on soil organic carbon
are regulated by soil aggregate stability and not by
recalcitrance: Evidence from a reforestation experiment.
CATENA, 219, 106613, 2022, doi: 10.1016/j.catena.2022.106613.
9. Shi, P. et al., Vis-NIR spectroscopic assessment of soil
aggregate stability and aggregate size distribution in the
Belgian Loam Belt. Geoderma, 357, 113958, 2020, doi:
10.1016/j.geoderma.2019.113958.
10. Liu, J. et al., Specific ion effects on soil aggregate stability
and rainfall splash erosion. Int. Soil Water Conserv. Res., 10, 4,
557–564, 2022, doi: 10.1016/j. iswcr.2022.02.001.
11. Zhou, W., Sun, X., Li, S. et al., Effects of organic mulching on
soil aggregate stability and aggregate binding agents in an
urban forest in Beijing, China. J. For. Res., 33, 1083–1094, 2022,
https://doi.org/10.1007/s11676-021-01402-z.
12. Ma, R. et al., Spatial distribution and scale-specific controls
of soil water-stable aggregates in southeastern China. J.
Cleaner Prod., 369, 133305, 2022, doi:
10.1016/j.jclepro.2022.133305.
13. Xue, B. et al., Straw residue incorporation and potassium
fertilization enhance soil aggregate stability by altering the
soil content of iron oxide and organic carbon in a rice-rape
cropping system. Land Degrad. Dev., 33, 14, 2567–2584, 2022,
doi: 10.1002/ldr.4333.
14. Tang, X. et al., Responses of soil aggregate stability to organic
C and total N as controlled by land-use type in a region of
south China affected by sheet erosion. CATENA, 218, 106543,
2022, doi: 10.1016/j.catena.2022.106543.
15. Luzio, W. et al., Soils of Chile, University of Chile. Maval
Press, Santiago, 2010.
16. Zeraatpisheh, M. et al., Spatial prediction of soil aggregate
stability and soil organic carbon in aggregate fractions using
 machine learning algorithms and environmental variables.
Geoderma Reg., 27, e00440, 2021, doi: 10.1016/j.
geodrs.2021.e00440.
17. Mo, J., Feng, J., He, W., Liu, Y., Cao, N., Tang, Y., Gu, S., Effects
of polycyclic aromatic hydrocarbons fluoranthene on the soil
aggregate stability and the possible underlying mechanism.
Environ. Sci. Pollut. Res., 30, 4, 10245–10255, 2023.

Note

1. Email: [email protected]
Index

5G, 196–197, 199, 287

6G, 196, 199

ABC, 13, 46, 57–62, 100, 102–103, 129

ACO, 2, 8–10, 24, 38, 43, 45, 56, 59, 60, 62
Algorithmic efficiency, 22
Algorithmic performance, 22
Anomaly detection, 62, 135–147
Ant colony optimization, 1–2, 8–10, 43, 45, 57
Artificial bee colony, 13, 23, 46, 57, 100, 102
Autonomous vehicle, 215–220, 222–224, 228–229, 231, 235–
237

Bat algorithm, 15, 43, 48, 57

Bayesian optimization, 37, 39, 41, 50–53
Big data, 51, 183, 222–224, 231, 238, 253–255, 257, 259, 261,
263, 265, 269
Botnet attack, 149, 151–152, 156, 160

Cancer, 73–79, 81, 83, 85–86, 90–91, 96–104, 115–117, 119–130

Cat swarm optimization, 18–19, 32, 62
Cell planning, 286–288, 299
Cellular networks, 195–197, 211
Coefficient of determination, 260, 263, 309–310
Communication, 18–19, 23, 150–151, 182, 195–197, 199, 210–
211, 216–218, 221, 245, 268, 276, 286–288, 294–296, 300
Competitive advantage, 221, 253, 254, 255, 256, 257, 258, 265
Competitive edge, 221, 238, 253, 254, 255, 257
Computational complexity, 37, 65, 113
Cost-effectiveness, 216
Cuckoo search, 15, 16, 45, 95, 96, 98, 104, 105, 106, 112, 118,
125

Data collection, 81, 104, 117, 158, 184, 258, 303

Data normalization, 140, 185
Deep learning, 3, 61, 63, 73, 75, 76, 83, 84, 88, 91, 97, 130, 151,
152, 166, 167, 168, 174, 175, 176, 182
Differential evolution, 2, 12, 21, 39, 43, 44, 288

Ensemble learning, 83, 84, 130

Feature extraction, 23, 74, 85, 86, 87, 88, 89, 136, 137, 139, 141,
143, 153, 167, 192
Feature selection, 8, 14, 41, 61, 65, 79, 95, 96, 97, 98, 100, 101,
103, 104, 114, 119, 121, 125, 130, 151, 152, 154, 155
Firefly algorithm, 14, 59
Flower pollination algorithm, 16

Genetic algorithm, 1, 2, 3, 5, 23, 39, 43, 44, 49, 100, 101, 102
Global optimization, 2, 51, 62
Global search, 2, 50, 98, 188
Glowworm swarm optimization, 18
Grasshopper optimization algorithm, 19
Gray wolf optimizer, 14, 19, 45, 57

Harmony search, 1, 12, 39, 43, 48, 100

Heuristic optimization, 1
Hybrid algorithm, 22, 50, 87, 88, 97, 98, 100, 104
Hyperparameter optimization, 25, 37, 38, 55, 135, 137, 142,
143

Imaging, 73, 75, 76, 77, 78, 79, 80, 81, 82, 83, 90, 91, 101
Imaging modalities, 73, 76, 82, 83, 90, 91
Imperialist competitive algorithm, 14, 15
Intrusion detection, 151, 163–167, 177
IoT, 149–152, 154, 156, 160, 166–167, 183, 196, 221–222, 256,
264–265, 295

Krill herd algorithm, 17

Local search, 2, 10–11, 16, 43–48, 50, 56, 188
LSTM, 135, 137–138, 141–146

Machine learning, 1, 3, 14, 23, 37–40, 44, 50, 53, 61, 63–64, 73,
76, 83, 86, 91, 95–97, 101–102, 104, 113–114, 122, 130, 142, 149,
151–152, 154, 163–167, 179, 183–185, 193, 197, 301, 313
Magnetic resonance, 81, 101
Mammography, 23, 79, 80, 81, 82, 91
Metaheuristic algorithm, 1, 2, 3, 10, 21–26, 38–39, 42, 46, 49,
53–56, 59, 62–65, 287, 288, 295
Moth–flame optimization, 19–20, 57

Nash equilibrium, 196, 202–203, 207

Neural architecture search, 37, 39, 53–54
Nonlinear problems, 2

Optimization problem, 1, 3, 7, 8, 10, 12, 14–16, 19–21, 24–25,

38, 42, 49–50, 55–56, 64, 104, 217, 291
Optimization technique, 3, 11, 12, 14–18, 20, 22, 24, 37, 39, 41–
42, 44, 52, 64–65, 97, 100, 103, 168, 286

Parking, 229, 235–242, 245–246, 250

Particle swarm optimization, 1, 2, 7, 8, 22–23, 39, 43, 44, 49,
50, 57, 100, 101, 179, 184, 188, 192, 286, 288
Pathological, 75, 78, 51, 82, 91
Performance metrics, 40, 41, 95, 143, 183, 219
Principal component analysis, 100, 149, 152, 154, 156, 160

Random tree, 149, 151, 152, 155

Reinforcement learning, 53, 54, 62, 165, 167
Reliability, 73, 258–262, 275

Scalability, 21, 37, 54, 65, 129, 130, 183, 203

Security enhancement, 182
Simulated annealing, 1, 2, 5, 7, 20, 22, 39, 43, 47, 49, 101, 220,
227, 229
Smart healthcare system, 181
Soil aggregate stability, 301–302
Structural model, 259, 263, 273, 279
Supervised learning, 61–62, 152, 165, 170
Sustainability, 215–217, 233, 237–238, 253, 255, 257–258, 265,
267–269, 271, 242, 274, 276–281, 313
Swarm intelligence, 1, 2, 21, 46, 171

Tabu search, 1, 2, 10, 11, 39, 43, 47, 48

Technological innovation, 268–269, 271, 276, 280
Transformer, 135, 137–138, 140–141, 146

Ultrasound, 74–75, 79–82, 86–87, 91

Unsupervised learning, 62, 152, 165
Whale optimization algorithm, 17, 22, 46, 57, 288
Wireless sensor networks, 14, 50, 163–165
Also of Interest

Check out these published and

forthcoming titles in the “Artificial
Intelligence and Soft Computing for
Industrial Transformation” series from
Scrivener Publishing

Metaheuristics for Machine Learning Algorithms and

Applications
Edited by Kanak Kalita, Narayanan Ganesh and S. Balamurugan
Published 2024. ISBN 978-1-394-23392-2

Artificial Intelligence for Sustainable Applications

Edited by K. Umamaheswari, B. Vinoth Kumar and S. K.
Somasundaram
Published 2023. ISBN 978-1-394-17458-4

Convergence of Deep Learning in Cyber-IoT Systems and

Security
Edited by Rajdeep Chakraborty, Anupam Ghosh, Jyotsna Kumar
Mandal and S. Balamurugan
Published 2023. ISBN 978-1-119-85721-1
Cognitive Intelligence and Big Data in Healthcare
Edited by D. Sumathi, T. Poongodi, B. Balamurugan and
Lakshmana Kumar Ramasamy
Published 2022. ISBN 978-1-119-76888-3

The New Advanced Society

Artificial Intelligence and Industrial Internet of Things
Paradigm
Edited by Sandeep Kumar Panda, Ramesh Kumar Mohapatra,
Subhrakanta Panda and S. Balamurugan
Published 2022. ISBN 978-1-119-82447-3

Digitization of Healthcare Data using Blockchain

Edited by T. Poongodi, D. Sumathi, B. Balamurugan and K. S.
Savita
Published 2022. ISBN 978-1-119-79185-0

Tele-Healthcare
Applications of Artificial Intelligence and Soft Computing
Techniques
Edited by R. Nidhya, Manish Kumar and S. Balamurugan
Published 2022. ISBN 978-1-119-84176-0

Impact of Artificial Intelligence on Organizational

Transformation
Edited by S. Balamurugan, Sonal Pathak, Anupriya Jain, Sachin
Gupta, and Sachin Sharma and Sonia Duggal
Published 2022. ISBN 978-1-119-71017-2

Artificial Intelligence for Renewable Energy Systems

Edited by Ajay Kumar Vyas, S. Balamurugan, Kamal Kant Hiran
Harsh S. Dhiman
Published 2022. ISBN 978-1-119-76169-3

Artificial Intelligence Techniques for Wireless

Communication and Networking
Edited by Kanthavel R., K. Ananthajothi, S. Balamurugan and R.
Karthik Ganesh
Published 2022. ISBN 978-1-119-82127-4

Advanced Healthcare Systems

Empowering Physicians with IoT-Enabled Technologies
Edited by Rohit Tanwar, S. Balamurugan, R. K. Saini, Vishal
Bharti and Premkumar Chithaluru
Published 2022. ISBN 978-1-119-76886-9

Smart Systems for Industrial Applications

Edited by C. Venkatesh, N. Rengarajan, P. Ponmurugan and S.
Balamurugan
Published 2022. ISBN 978-1-119-76200-3
Intelligent Renewable Energy Systems
Edited by Neeraj Priyadarshi, Akash Kumar Bhoi,
Sanjeevikumar Padmanabam, S.Balamurugan, and Jens Bo
Holm-Nielson
Published 2022. ISBN 978-1-119-78627-6

Human Technology Communication

Internet of Robotic Things and Ubiquitous Computing
Edited by R. Anandan. G. Suseendran, S. Balamurugan, Ashish
Mishra and D. Balaganesh
Published 2021. ISBN 978-1-119-75059-8

Nature-Inspired Algorithms Applications

Edited by S. Balamurugan, Anupriya Jain, Sachin Sharma,
Dinesh Goyal, Sonia Duggal and Seema Sharma
Published 2021. ISBN 978-1-119-68174-8

Computation in Bioinformatics
Multidisciplinary Applications
Edited by S. Balamurugan, Anand Krishnan, Dinesh Goyal,
Balakumar Chandrasekaran and Boomi Pandi
Published 2021. ISBN 978-1-119-65471-1

Fuzzy Intelligent Systems

Methodologies, Techniques, and Applications
Edited by E. Chandrasekaran, R. Anandan, G. Suseendran, S.
Balamurugan and Hanaa Hachimi
Published 2021. ISBN 978-1-119-76045-0

Biomedical Data Mining for Information Retrieval

Methodologies, Techniques and Applications
Edited by Sujata Dash, Subhendu Kumar Pani, S. Balamurugan
and Ajith Abraham
Published 2021. ISBN 978-1-119-71124-7

Design and Analysis of Security Protocols for

Communication
Edited by Dinesh Goyal, S. Balamurugan, Sheng-Lung Peng and
O.P. Verma
Published 2020. ISBN 978-1-119-55564-3

www.scrivenerpublishing.com
WILEY END USER LICENSE
AGREEMENT
Go to www.wiley.com/go/eula to access Wiley’s ebook EULA.