Studies in Computational Intelligence 720

OmidBozorg-Haddad Editor

Advanced
Optimization by
Nature-Inspired
Algorithms
Studies in Computational Intelligence

Volume 720

Series editor
Janusz Kacprzyk, Polish Academy of Sciences, Warsaw, Poland
e-mail: [email protected]
About this Series

The series Studies in Computational Intelligence (SCI) publishes new develop-

ments and advances in the various areas of computational intelligencequickly and
with a high quality. The intent is to cover the theory, applications, and design
methods of computational intelligence, as embedded in the elds of engineering,
computer science, physics and life sciences, as well as the methodologies behind
them. The series contains monographs, lecture notes and edited volumes in
computational intelligence spanning the areas of neural networks, connectionist
systems, genetic algorithms, evolutionary computation, articial intelligence,
cellular automata, self-organizing systems, soft computing, fuzzy systems, and
hybrid intelligent systems. Of particular value to both the contributors and the
readership are the short publication timeframe and the worldwide distribution,
which enable both wide and rapid dissemination of research output.

More information about this series at http://www.springer.com/series/7092

Omid Bozorg-Haddad
Editor

Advanced Optimization
by Nature-Inspired
Algorithms

123
Editor
Omid Bozorg-Haddad
Department of Irrigation & Reclamation
Engineering
College of Agriculture & Natural Resources,
University of Tehran
Karaj
Iran

ISSN 1860-949X ISSN 1860-9503 (electronic)

Studies in Computational Intelligence
ISBN 978-981-10-5220-0 ISBN 978-981-10-5221-7 (eBook)
DOI 10.1007/978-981-10-5221-7
Library of Congress Control Number: 2017943829

Springer Nature Singapore Pte Ltd. 2018

This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part
of the material is concerned, specically the rights of translation, reprinting, reuse of illustrations,
recitation, broadcasting, reproduction on microlms or in any other physical way, and transmission
or information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar
methodology now known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this
publication does not imply, even in the absence of a specic statement, that such names are exempt from
the relevant protective laws and regulations and therefore free for general use.
The publisher, the authors and the editors are safe to assume that the advice and information in this
book are believed to be true and accurate at the date of publication. Neither the publisher nor the
authors or the editors give a warranty, express or implied, with respect to the material contained herein or
for any errors or omissions that may have been made. The publisher remains neutral with regard to
jurisdictional claims in published maps and institutional afliations.

Printed on acid-free paper

This Springer imprint is published by Springer Nature

The registered company is Springer Nature Singapore Pte Ltd.
The registered company address is: 152 Beach Road, #21-01/04 Gateway East, Singapore 189721, Singapore
I like to dedicate this book to my loving
parents.

Omid Bozorg-Haddad
Preface

From the early 1990s, the introduction of the term Computational Intelligence
(CI) highlighted the potential applicability of this eld. One of the preliminary
applications of the eld was in the realm of optimization. Undoubtedly, the tasks of
design and operation of systems can be approached systematically by the application
of optimization. And while in most real-life problems, including engineering
problems, application of the classical optimization techniques were limited due to the
complex nature of the decision space and numerous variables, and the CI-based
optimization techniques, which imitated the nature as a source of inspiration, have
proven quite useful. Consequently, during the last passing decades, a considerable
number of novel nature-based optimization algorithms have been proposed in the
literature. While most of these algorithms hold considerable promise, a majority
of them are still in their infancy. For such algorithms to bloom and reach their full
potential, they should be implemented in numerous optimization problems, so that not
only their most suitable sets of optimization problems are recognized, but also
adaptive strategies need to be introduced to make them more suitable for wider sets of
optimization problems. For that, this book specically aimed to introduce some
of these potential nature-based algorithms that could be useful for multidisciplinary
students including those in aeronautic engineering, mechanical engineering, indus-
trial engineering, electrical and electronic engineering, chemical engineering, civil
engineering, computer science, applied mathematics, physics, economy, biology, and
social science, and particularly those pursuing postgraduate studies in advanced
subjects. Chapter 1 of the book is a review of the basic principles of optimization and
nature-based optimization algorithms. Chapters 215 are respectively dedicated to
Cat Swarm Optimization (CSO), League Championship Algorithm (LCA),
Anarchies Society Optimization (ASO), Cuckoo Optimization Algorithm (COA),
Teacher-Learning-Based Optimization (TLBO), Flower Pollination Algorithm
(FPA), Krill Herd Algorithm (KHA), Grey Wolf Optimization (GWO), Shark Smell
Optimization (SSO), Ant Lion Optimization (ALO), Gradient Evolution (GE),
Moth-Flame Optimization (MFO), Crow Search Algorithm (CSA), and Dragonfly
Algorithm (DA). The order of the chapters corresponds to the order of chronological
appearance of these algorithms, from earlier algorithms to newly introduced ones.

vii
viii Preface

Each chapter describes a specic algorithm and starts with a brief literature review of
its development and subsequent modication since the time of inception. This is
followed by the presentation of the basic concept on which the algorithm is based and
the steps of the algorithm. Each chapter closes with a pseudocode of the algorithm.

Karaj, Iran Omid Bozorg-Haddad

Contents

1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
Babak Zolghadr-Asli, Omid Bozorg-Haddad and Xuefeng Chu
2 Cat Swarm Optimization (CSO) Algorithm. . . . . . . . . . . . . . . . . . . . 9
Mahdi Bahrami, Omid Bozorg-Haddad and Xuefeng Chu
3 League Championship Algorithm (LCA) . . . . . . . . . . . . . . . . . . . . . . 19
Hossein Rezaei, Omid Bozorg-Haddad and Xuefeng Chu
4 Anarchic Society Optimization (ASO) Algorithm . . . . . . . . . . . . . . . 31
Atiyeh Bozorgi, Omid Bozorg-Haddad and Xuefeng Chu
5 Cuckoo Optimization Algorithm (COA) . . . . . . . . . . . . . . . . . . . . . . 39
Saba Jafari, Omid Bozorg-Haddad and Xuefeng Chu
6 Teaching-Learning-Based Optimization (TLBO) Algorithm . . . . . . 51
Parisa Sarzaeim, Omid Bozorg-Haddad and Xuefeng Chu
7 Flower Pollination Algorithm (FPA) . . . . . . . . . . . . . . . . . . . . . . . . . 59
Marzie Azad, Omid Bozorg-Haddad and Xuefeng Chu
8 Krill Herd Algorithm (KHA) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
Babak Zolghadr-Asli, Omid Bozorg-Haddad and Xuefeng Chu
9 Grey Wolf Optimization (GWO) Algorithm . . . . . . . . . . . . . . . . . . . 81
Hossein Rezaei, Omid Bozorg-Haddad and Xuefeng Chu
10 Shark Smell Optimization (SSO) Algorithm . . . . . . . . . . . . . . . . . . . 93
Sahar Mohammad-Azari, Omid Bozorg-Haddad and Xuefeng Chu
11 Ant Lion Optimizer (ALO) Algorithm . . . . . . . . . . . . . . . . . . . . . . . . 105
Melika Mani, Omid Bozorg-Haddad and Xuefeng Chu
12 Gradient Evolution (GE) Algorithm. . . . . . . . . . . . . . . . . . . . . . . . . . 117
Mehri Abdi-Dehkordi, Omid Bozorg-Haddad and Xuefeng Chu

ix
x Contents

13 Moth-Flame Optimization (MFO) Algorithm . . . . . . . . . . . . . . . . . . 131

Mahdi Bahrami, Omid Bozorg-Haddad and Xuefeng Chu
14 Crow Search Algorithm (CSA) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143
Babak Zolghadr-Asli, Omid Bozorg-Haddad and Xuefeng Chu
15 Dragony Algorithm (DA) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151
Babak Zolghadr-Asli, Omid Bozorg-Haddad and Xuefeng Chu
About the Editor

Omid Bozorg-Haddad is Professor in the Department of Irrigation and

Reclamation Engineering at the University of Tehran, Iran. His teaching and
research interests include water resources and environmental systems analysis,
planning, and management as well as application of optimization algorithms in
water related systems. He has published more than 100 articles in peer-reviewed
journals and 100 papers in conference proceedings. He has also supervised more
than 50 M.Sc. and Ph.D. students.

xi
List of Figures

Fig. 2.1 Flowchart of the CSO algorithm . . . . . . . . . . . . . . . . . . . . . . . . 13

Fig. 3.1 Flowchart of the basic LCA . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
Fig. 3.2 A simple example of league championship scheduling . . . . . . . 24
Fig. 3.3 Procedure of the articial match analysis in LCA . . . . . . . . . . . 26
Fig. 4.1 Flowchart of the ASO algorithm . . . . . . . . . . . . . . . . . . . . . . . . 33
Fig. 5.1 Flowchart of the COA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
Fig. 5.2 Random egg laying in ELR (the black circle is the cuckoos
initial habitat with three eggs; and the white circles are the
eggs at new positions) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
Fig. 5.3 Immigration of a sample cuckoo to the target habitat . . . . . . . . 46
Fig. 6.1 Flowchart of the TLBO algorithm . . . . . . . . . . . . . . . . . . . . . . . 54
Fig. 7.1 Flowchart of the FPA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
Fig. 8.1 Schematic representation of the sensing ambit around
a krill individual . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
Fig. 8.2 Simplied flowchart of the KHA . . . . . . . . . . . . . . . . . . . . . . . . 77
Fig. 9.1 Social hierarchy of grey wolves . . . . . . . . . . . . . . . . . . . . . . . . . 82
Fig. 9.2 Attacking toward prey versus searching for prey . . . . . . . . . . . . 85
Fig. 9.3 Updating of positions in the GWO algorithm . . . . . . . . . . . . . . 86
Fig. 9.4 Attacking toward prey and searching for prey . . . . . . . . . . . . . . 86
Fig. 9.5 Flowchart of the GWO algorithm . . . . . . . . . . . . . . . . . . . . . . . 89
Fig. 10.1 Schematic of sharks movement toward the source
of the smell . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
Fig. 10.2 Rotational movement of a shark . . . . . . . . . . . . . . . . . . . . . . . . 99
Fig. 10.3 Flowchart of the SSO algorithm . . . . . . . . . . . . . . . . . . . . . . . . 100
Fig. 11.1 Antlion hunting behavior . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
Fig. 11.2 Flowchart of the ALO algorithm (It = iteration counter;
and IT = number of iterations) . . . . . . . . . . . . . . . . . . . . . . . . . . 108
Fig. 11.3 Three random walk curves in one dimension
started at zero . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112
Fig. 12.1 Gradient determination. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
Fig. 12.2 Search direction for the original gradient-based method . . . . . . 122

xiii
xiv List of Figures

Fig. 12.3 Search direction for the GE algorithm . . . . . . . . . . . . . . . . . . . . 123

Fig. 12.4 Gradient approximation method modied from
individual-based search to population-based search:
a individual-based, b population-based . . . . . . . . . . . . . . . . . . . 123
Fig. 12.5 Vector jumping operator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126
Fig. 12.6 Flowchart of the GE algorithm . . . . . . . . . . . . . . . . . . . . . . . . . 127
Fig. 13.1 Moths spiral flying path around a light source . . . . . . . . . . . . . 134
Fig. 13.2 Flowchart of the MFO algorithm . . . . . . . . . . . . . . . . . . . . . . . . 135
Fig. 14.1 Flowchart of the standard CSA . . . . . . . . . . . . . . . . . . . . . . . . . 147
Fig. 15.1 Primitive corrective patterns of dragonfly individuals
in a swarm: a Separation; b Alignment; c Cohesion;
d Food Attraction; and e Predator distraction . . . . . . . . . . . . . . 153
Fig. 15.2 Flowchart of the DA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 157
List of Tables

Table 2.1 Characteristics of the CSO algorithm . . . . . . . . . . . . . . . . . . . .. 12

Table 3.1 Characteristics of the LCA . . . . . . . . . . . . . . . . . . . . . . . . . . .. 21
Table 3.2 Hypothetical SWOT analysis derived from the articial
match analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
Table 4.1 Characteristics of the ASO algorithm. . . . . . . . . . . . . . . . . . . . . 34
Table 5.1 Characteristics of the COA . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
Table 6.1 Characteristics of the TLBO algorithm . . . . . . . . . . . . . . . . . . . 55
Table 7.1 Characteristics of the FPA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
Table 8.1 The characteristics of the KHA . . . . . . . . . . . . . . . . . . . . . . . . . 77
Table 9.1 Characteristics of the GWO algorithm . . . . . . . . . . . . . . . . . . . . 87
Table 10.1 Characteristics of the SSO algorithm . . . . . . . . . . . . . . . . . . . . . 101
Table 11.1 Characteristics of the ALO algorithm . . . . . . . . . . . . . . . . . . . . 108
Table 12.1 Characteristics of the GE algorithm . . . . . . . . . . . . . . . . . . . . . . 128
Table 13.1 Characteristics of the MFO algorithm . . . . . . . . . . . . . . . . . . . . 135
Table 14.1 Characteristics of the CSA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147
Table 15.1 Characteristics of the DA. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 157

xv
Chapter 1
Introduction

Babak Zolghadr-Asli , Omid Bozorg-Haddad and Xuefeng Chu

Abstract In this chapter, some general knowledge relative to the realm of

nature-inspired optimization algorithms (NIOA) is introduced. The desirable merits
of these intelligent algorithms and their initial successes in many elds have
inspired researchers to continuously develop such revolutionary algorithms and
implement them to solve various real-world problems. Such a truly interdisciplinary
environment of the research and development provides rewarding opportunities for
scientic breakthrough and technology innovation. After a brief introduction to
computational intelligence and its application in optimization problems, the history
of the NIOA was reviewed. The relevant algorithms were then categorized in
different manners. Finally, one the most groundbreaking theorems regarding the
nature-inspired optimization techniques was briefly discussed.

1.1 Introduction

Articial intelligence (AI) refers to any sort of intelligence that is exhibited by

machines. The term of computational intelligence (CI), a branch of AI, was coined
by Bezdek in the early 1990s (Bezdek 1992), which inspired the development of a

B. Zolghadr-Asli
O. Bozorg-Haddad (&)

Department of Irrigation and Reclamation Engineering, Faculty of Agricultural Engineering
and Technology, College of Agriculture and Natural Resources, University of Tehran,
3158777871 Karaj, Iran
e-mail: [email protected]
B. Zolghadr-Asli
e-mail: [email protected]
X. Chu
Department of Civil and Environmental Engineering, North Dakota State University,
Dept 2470, Fargo, ND 58108-6050, USA
e-mail: [email protected]

Springer Nature Singapore Pte Ltd. 2018 1

O. Bozorg-Haddad (ed.), Advanced Optimization by Nature-Inspired Algorithms,
Studies in Computational Intelligence 720, DOI 10.1007/978-981-10-5221-7_1
2 B. Zolghadr-Asli et al.

new eld dedicated to computer-based intelligence. In principle, CI consists of any

science-supported approaches and technologies for analyzing, creating, and
developing intelligent systems (Xing and Gao 2014). Unlike AI that depends upon
the knowledge derived from human expertise, CI, however, mostly relies on the
collection of numerical data and a set of nature-inspired computational paradigms
(Du and Swamy 2016). As evidenced by numerous studies in various elds, CI was
able to flourish and achieve an unprecedented popularity during the past several
decades. While, once the CIs primary focus was on articial neural network
(ANN), fuzzy logic (FL), multi-agent system (MAS), and few optimization algo-
rithms including: evolutionary algorithms (EA) [e.g., genetic algorithm (GA),
genetic programming (GP), evolutionary programming (EP), and evolutionary
strategy (ES)], articial immune systems (AIS), simulated annealing (SA), Tabu
search (TS), as well as two variants of swarm intelligence (SI) [i.e., ant colony
optimization (ACO) and particle swarm optimization (PSO)] (Xing and Gao 2014),
nowadays CI has expanded to a vast domain that covers almost every science and
engineering eld.
One of the major subjects of CI is, as previously mentioned, the study of
unorthodox optimization techniques. It is no exaggeration to say that optimization
has surrounded our everyday life, from complex engineering problems to business
or even the holiday planning. What bonds these activities together is that essentially
they aim to utilize a nite commodity (say, a limited resource, nance, or even
time) to achieve certain goals or objectives. Mathematically speaking, optimization
searches the decision space in pursuit of an array of decision variables that could
produce the maximum acceptability considering the initial goal (objective function).
Since its introduction, CI has been able to make a considerable contribution to the
solving techniques of such optimization problems due to the development of var-
ious efcient search algorithms.
Traditionally, optimization problems were solved either using calculus-based
methods, random-based search, or in some cases enumerative searching techniques
(Du and Swamy 2016). Broadly speaking, such optimization techniques can gen-
erally be divided into derivative and nonderivative methods, depending on whether
or not the derivatives of the objective function(s) are required in the optimization
process. Derivative methods are calculus-based methods, which are based on a
gradient search (also known as the steepest search method). The Newtons method,
GaussNewton method, quasi-Newton methods, the trust-region method, and
LevenbergMarquardt method are a few examples of such techniques. These
classical methods used to dominate the eld of optimization. While efcient in
solving a wide range of optimization problems, when the number of the decision
variables grow large enough, or even when the decision space is discreet, which is
the case in most if not all practical optimization problems, these techniques are
unable to illustrate acceptable performances. Indeed, the aforementioned describes
one of the major advantages of CI-based optimization techniques over traditional
ones.
1 Introduction 3

1.2 Optimization: Core Principles and Technical Terms

Assume that a group of amateur climbers decide to summit the highest mountain in
a previously unknown and hilly territory. Indeed, searching for the summit is not
unlike searching for the optimal solution, for, in fact, the landscape represents the
decision space, while the highest mountain embodies the global optimum. But how
does one even begin to search such a vast area? One initial answer would be to map
out the entire landscape. However, this would be a both time- and
energy-consuming task. Perhaps, an alternative would be to search the area in a
random-based manner. While it would denitely help to save both time and energy,
ultimately as one can expect, it would not be an efcient strategy as well, for there
would be no guarantee to reach the highest mountain. A more intelligent alternative
would be to directly climb up the steepest cliff, assuming that the highest mountain
is more likely to have a steeper slope. Essentially, this strategy represents the core
principle of classical optimization techniques. While efcient in many cases, if the
climbers path would be interrupted by cliffs (discrete decision space), for instance,
this strategy would not be efcient to locate the highest mountain. Additionally, in a
hilly landscape, climbers could deceitfully climb to the top of a mountain which in
essential stands above the neighboring mountains, while in fact, it is not the highest
mountain in the entire area. This problem is known in technical terms as trapping in
local optima.
Alternatively, the climbers could do a random walk in the area, while looking for
some clues. Such hybrid strategies are formed using a combination of random-
based searching and an adaptive strategy, which is usually inspired by nature. In
fact, that is the description of CI-based optimization algorithms. Subsequently, such
searches could be conducted while the group maintains to stick with one another
(perform as an individual climber). The group members can also spread out and
share their gained information with each other, while the searching proceeds further
on. Technically speaking, the former strategy is better known as single-point
optimization technique, while the latter strategy represents population-based opti-
mization algorithms. The single-point strategies are also known as trajectory opti-
mization algorithms, for the optimization process would provide a path that could
lead to the optimum point, in this case, the highest mountain in the area.
Additionally, the climbers, either as one group or separate individuals, could
investigate the area only using the information currently at hand. An alternative
would be to take a record and map out some previously encountered locations. In a
technical term, the second strategy is called memory using algorithms. While such a
strategy is more efcient in most cases, if the population of the climbers increases or
the landscape is vast enough, storing such massive information could potentially
turn into a major problem.
Ultimately, the core principle of all CI-based optimization algorithms, which are
better known as metaheuristic algorithms, is a way of trial and error to produce an
acceptable solution to a complex problem in a reasonably practical time (Yang
2010). While the complexity of the practical optimization problem is in favor of
4 B. Zolghadr-Asli et al.

implementing such optimization algorithms, there could also be no guarantee that

the best solution (global optimum) can be spotted using such techniques.

1.3 Brief History of CI-Based Optimization Algorithms

Despite their novel and ubiquitous nature, implementation of the CI-based opti-
mization algorithms is indeed a relatively new technique, though it is difcult to
pinpoint when the whole story began. Accordingly, Allen Turing was perhaps the
rst person to implement the CI-based optimization algorithms (Yang 2010).
Evidently, during World War II, while trying to break the German-designed
encrypting machine called Enigma, Turing developed an optimization technique
which he later named heuristic search as it could be expected to work most of the
times, while there was no actual guarantee for a successful performance in each
trial. Turing was later recruited to the national physical laboratory, UK, where he
set out his design for the automatic computing engine. Later on, he outlined his
innovative idea of machine intelligence and learning neural networks and evolu-
tionary algorithms in an NPL report on intelligent machinery (Yang 2010).
The CI-based optimization techniques bloomed during the 1960s and 1970s. In
the early 1960s, John Holland and his collaborators at the University of Michigan
developed the genetic algorithms (GA) (Goldberg and Holland 1988). In essence, a
GA is a search method based on the abstraction of Darwins theory of evolution and
natural selection of biological systems, which are represented in the mathematical
operators. Hollands preliminary studies were showing promising results, while he
continued to further develop his technique by introducing novel and efcient agents
to his algorithms which were named crossover and mutation, although his seminal
book summarizing the development of the genetic algorithm was not published
until 1975 (Yang 2010). Hollands work inspired many to further develop and
adopt similar methods in their research, which beneted from a similar basic
principle in numerous and colorful elds. For instance, while Holland and his
collogues were trying to develop their revolutionary method, Ingo Rechenberg and
Hans-Paul Schwefel at the Technical University of Berlin introduced another novel
optimization technique for solving aerospace engineering problems, which they
later named evolutionary strategy (Back et al. 1997). In 1966, (Fogel et al. 1966)
developed an evolutionary programming technique by representing the solution as
nite state machines and randomly mutating one of these machines. The above
innovative idea and method have evolved into a wider area that became known as
evolutionary algorithms (EAs) (Yang 2010).
In the early 1990s and in another great leap forward in the eld of CI-based
optimization algorithms, Marco Dorigo nished his Ph.D. thesis on optimization and
nature-inspired algorithms, in which he described his innovative work on ant colony
optimization (ACO) (Dorigo and Blum 2005). This search technique was inspired by
the swarm intelligence of social ants using the pheromone to trace the food sources.
Slightly later, in 1995, the particle swarm optimization (PSO) was proposed by an
1 Introduction 5

American social physiologist James Kennedy, and engineered by Russell C.

Eberhart, which shall be considered as another signicant progress in the eld of
nature-inspired optimization techniques (Poli et al. 2007). In essence, PSO is an
optimization algorithm inspired by the swarm intelligence of the school of sh or
birds. Multiple agents or particles swarm around the search space starting from an
initial random guess. The swarm members communicate to one another by sharing
the current best answer, which in turn, enables the algorithm to nd the optimum
solutions. Since then, many attempts have been made to mimic and imitate the
natural beings in the searching process of optimization techniques. During the past
decades, numerous novel nature-inspired optimization algorithms have been pro-
posed to advance the CI-based optimization techniques (Xing and Gao 2014).

1.4 Classication of CI-Based Optimization Algorithms

Perhaps, it could be benecial for studding reasons to categorize and classify the
CI-based optimization techniques, although depending on specic viewpoints many
classications are possible. While some of these classications are quite technical
and in some cases, even based upon vague characteristics, more general classi-
cations could help better understand the core principles behind such algorithms.
Intuitively, one can categorize the CI-based algorithms using the number of
searching agents. This characteristic would divide the algorithms into two major
categories: (1) population-based algorithms, and (2) single-point algorithms. The
algorithms working with a single agent are called trajectory methods. The population-
based algorithms, on the other hand, perform search processes using several agents
simultaneously.
As stated in previous sections, some CI-based algorithms can keep a record of
previously inspected arrays of decision variable in the search space. Such algo-
rithms are known as the memory using algorithms. On the contrary, the algorithms,
known as memory-less algorithms, do not memorize the previously encountered
locations in the search space. The methods of making use of memory in a CI-based
optimization algorithm can be further divided into short term and long-term
memory using algorithms. The former usually keeps track of recently performed
moves, visited solutions or, in general, decisions taken, while the latter is usually an
accumulation of synthetic parameters about the search. The use of memory is
nowadays recognized as one of the fundamental elements of a powerful CI-based
optimization algorithm (Blum and Roli 2003).
As mentioned earlier on, one of the challenges in complex optimization prob-
lems is to avoid trapping in local optima, which has been a major problem for
classical optimization techniques. To overcome this problem, some CI-based
optimization algorithms would modify their objective function during the opti-
mization process. Such optimization techniques are known as dynamic algorithms.
The alternative would be to keep the objective function as is during the optimization
process. This is the characteristic of static algorithms.
6 B. Zolghadr-Asli et al.

Finally, a more general, yet sometimes vague, classication of the CI-based

optimization algorithms is based on their sources of inspiration. Based on their ori-
gins, the CI-based optimization algorithms could be classied as nature-based and
non-nature-based optimization algorithms. While such a classication could be
confusing, for there are cases in which it is rather difcult to clearly attribute an
algorithm to one of the two classes, yet, such classication is the most widely accepted
classication for the CI-based algorithms. The subject of this book is nature-based
optimization algorithms. Generally speaking, any biological-, physical-, or even
chemical-based algorithm that somehow imitates the nature can be categorized as a
nature-based optimization algorithm. Also, it should be mentioned that such a cate-
gorizing method for the CI-based optimization algorithms has no contradiction with
any previously introduced categories. For instance, a nature-based optimization
algorithm could also be a population-based or single-point optimization algorithm.
It is advantageous to remember that two major components of any nature-based
algorithms are intensication and diversication. Diversication means to generate
diverse solutions so as to explore the search space, while intensication means to
focus on the search in a local region by exploiting the area surrounding the current
good solutions. While in most cases through a random-based component (diver-
sication), nature-based optimization algorithms attempt to thoroughly investigate
the decision space and divert the solutions to be trapped in local optima, the
intensication component ensures the convergence of the algorithm to what is more
likely to be the global optimum solution. The good combination of these two major
components will usually ensure that the global optimality is achievable (Yang
2010).

1.5 No Free Lunch Theorem: The Reason Behind New

Algorithms

In the late 1990s, (Wolpert and Macready 1997) proposed a controversial,

groundbreaking no free lunch theorem for optimization. Researchers intuitively
believed that there exist some universally benecial robust algorithms for opti-
mization. However, the no free lunch theorem clearly states that if an algorithm
outperforms others for some optimization functions, there will be other optimization
functions in which the other algorithms shall be better than the aforementioned
algorithm. In other words, if all possible function spaces are to be considered, on
average, all CI-based algorithms, and subsequently all nature-based algorithms
should perform equally well. And thus, ultimately, there could be no universally
better algorithm.
While no free lunch theorem clearly dismisses the myth of an ultimate opti-
mization algorithm, there is a silver lining sub-layer to this theorem, as well.
Though all algorithms are to perform equally on the average over all possible
1 Introduction 7

function space, yet as evidenced by numerous studies, such algorithms may and
will outperform one another in specic sets of optimization problems. In addition,
the algorithm developments are now focused on nding the best and efcient
algorithm for a specic set of optimization problems. Ultimately, instead of aiming
to design a perfect solver for all the problems, algorithms are developed to solve
most types of problems. As a result, during the last several decades, a considerable
number of novel nature-based optimization algorithms have been proposed. While
most of the algorithms hold considerable promise, a majority of them are still in
their infancy. For such algorithms to bloom and reach their full potential, they
should be applied to numerous optimization problems, so that not only they are
tested for their most suitable sets of optimization problems, but also adaptive
strategies are introduced to make them more suitable for wider sets of optimization
problems. For that, this book is specically aimed to introduce some of the potential
nature-based algorithms that can be useful in multidisciplinary studies including
those in aeronautic engineering, mechanical engineering, industrial engineering,
electrical and electronic engineering, chemical engineering, civil and environmental
engineering, computer science and engineering, applied mathematics, physics,
economy, biology, and social science, and particularly the postgraduate studies in
advanced subjects.

1.6 Conclusion

In this chapter, a brief background of CI, which consists of any science-supported

approaches and technologies for analyzing, creating, and developing intelligent
systems has been discussed from an introductory perspective. A major subject of CI
is to develop nature-based optimization techniques. For decades the nature-based
optimization methods have proved their ability to solve complex practical opti-
mization problems. Nowadays there are many novel nature-based optimization
techniques. However, according to the no free lunch theorem such algorithms do
not have particular advantages over one another, but rather could outperform each
other for particular sets of optimization problems. As a result, this book aims to
introduce some of the potentially useful algorithms so that the right audience could
check their authentic outperformances in applications and perhaps modify or select
the best algorithm for any particular optimization problem in hand.

References

Back, T., Hammel, U., & Schwefel, H. P. (1997). Evolutionary computation: Comments on the
history and current state. IEEE Transactions on Evolutionary Computation, 1(1), 317.
Bezdek, J. C. (1992). On the relationship between neural networks, pattern recognition and
intelligence. International Journal of Approximate Reasoning, 6(2), 85107.
8 B. Zolghadr-Asli et al.

Blum, C., & Roli, A. (2003). Metaheuristics in combinatorial optimization: Overview and
conceptual comparison. ACM Computing Surveys (CSUR), 35(3), 268308.
Dorigo, M., & Blum, C. (2005). Ant colony optimization theory: A survey. Theoretical Computer
Science, 344(23), 243278.
Du, K. L., & Swamy, M. N. S. (2016). Search and optimization by metaheuristics. Switzerland:
Springer Publication.
Fogel, L. J., Owens, A. J., & Walsh, M. J. (1966). Intelligent decision making through a simulation
of evolution. Behavioral Science, 11(4), 253272.
Goldberg, D. E., & Holland, J. H. (1988). Genetic algorithms and machine learning. Machine
Learning, 3(2), 9599.
Poli, R., Kennedy, J., & Blackwell, T. (2007). Particle swarm optimization. Swarm Intelligence,
1(1), 3357.
Wolpert, D. H., & Macready, W. G. (1997). No free lunch theorems for optimization. IEEE
Transactions on Evolutionary Computation, 1(1), 6782.
Xing, B., & Gao, W. J. (2014). Innovative computational intelligence: A rough guide to 134 clever
algorithms. Cham, Switzerland: Springer Publication.
Yang, X. S. (2010). Nature-inspired metaheuristic algorithms. Frome, UK: Luniver Press.
Chapter 2
Cat Swarm Optimization (CSO)
Algorithm

Mahdi Bahrami, Omid Bozorg-Haddad and Xuefeng Chu

Abstract In this chapter, a brief literature review of the Cat Swarm Optimization
(CSO) algorithm is presented. Then the natural process, the basic CSO algorithm
iteration procedure, and the computational steps of the algorithm are detailed.
Finally, a pseudo code of CSO algorithm is also presented to demonstrate the
implementation of this optimization technique.

2.1 Introduction

Optimization algorithms based on the Swarm Intelligence (SI) were developed for
simulating the intelligent behavior of animals. In these modeling systems, a pop-
ulation of organisms such as ants, bees, birds, and sh are interacting with one
another and with their environment through sharing information, resulting in use of
their environment and resources. One of the more recent SI-based optimization
algorithms is the Cat Swarm Optimization (CSO) algorithm which is based on the
behavior of cats. Developed by Chu and Tsai (2007), the CSO algorithm and its
varieties have been implemented for different optimization problems. Different
variations of the algorithm have been developed by researchers. Tsai et al. (2008)

M. Bahrami
O. Bozorg-Haddad (&)

Department of Irrigation and Reclamation Engineering, Faculty of Agricultural Engineering
and Technology, College of Agriculture and Natural Resources, University of Tehran, Karaj,
Tehran 31587-77871, Iran
e-mail: [email protected]
M. Bahrami
e-mail: [email protected]
X. Chu
Department of Civil and Environmental Engineering, North Dakota State University,
Dept 2470, Fargo, ND 58108-6050, USA
e-mail: [email protected]

Springer Nature Singapore Pte Ltd. 2018 9

O. Bozorg-Haddad (ed.), Advanced Optimization by Nature-Inspired Algorithms,
Studies in Computational Intelligence 720, DOI 10.1007/978-981-10-5221-7_2
10 M. Bahrami et al.

presented a parallel structure of the algorithm (i.e., parallel CSO or PCSO). They
further developed an enhanced version of their PCSO (EPCSO) by incorporating
the Taguchi method into the tracing mode process of the algorithm (Tsai et al.
2012). The binary version of CSO (BCSO) was developed by Shara et al. (2013)
and applied to a number of benchmark optimization problems and the zeroone
knapsack problem. The chaotic cat swarm algorithm (CCSA) was developed by
Yang et al. (2013a). Using different chaotic maps, the seeking mode step of the
algorithm was improved. Based on the concept of homotopy, Yang et al. (2013b),
proposed the homotopy-inspired cat swarm algorithm (HCSA) in order to improve
the search efciency. Lin et al. (2014a) proposed a method to improve CSO and
presented the Harmonious-CSO (HCSO). Lin et al. (2014b) introduced a modied
CSO (MCSO) algorithm capable of improving the search efciency within the
problem space. The basic CSO algorithm was also integrated with a local search
procedure as well as the feature selection of support vector machines (SVMs). This
method changed the concept of cat alert surroundings in the seeking mode of CSO
algorithm. By dynamically adjusting the mixture ratio (MR) parameter of the CSO
algorithm, Wang (2015) enhanced CSO algorithm with an adaptive parameter
control. A hybrid cat swarm optimization method was developed by Ojha and
Naidu (2015) through adding the invasive weed optimization (IWO) algorithm to
the tracing mode of the CSO algorithm.
Several other authors have used CSO algorithm in different elds of research on
optimization problems. Lin and Chien (2009) constructed the CSO algorithm + SVM
model for data classication through integrating cat swam optimization into the SVM
classier. Pradhan and Panda (2012) proposed a new multiobjective evolutionary
algorithm (MOEA) by extending CSO algorithm. The MOEA identied the non-
dominated solutions along the search process using the concept of Pareto dominance
and used an external archive for storing them. Xu and Hu (2012) presented a
CSO-based method for a resource-constrained project scheduling problem (RCPSP).
Saha et al. (2013) applied CSO algorithm to determine the optimal impulse response
coefcients of FIR low pass, high pass, bandpass, and band stop lters to meet the
respective ideal frequency response characteristics. So and Jenkins (2013) used CSO
for Innite Impulse Response (IIR) system identication on a few benchmarked IIR
plants. Kumar et al. (2014) optimized the placement and sizing of multiple distributed
generators using CSO. Mohamadeen et al. (2014) compared the binary CSO with the
binary PSO in selecting the best transformer tests that were utilized to classify
transformer health, and thus to improve the reliability of identifying the transformer
condition within the power system. Guo et al. (2015) proposed an improved cat
swarm optimization algorithm and redened some basic CSO concepts and opera-
tions according to the assembly sequence planning (ASP) characteristics. Bilgaiyan
et al. (2015) used the cat swarm-based multi-objective optimization approach to
schedule workflows in a cloud computing environment which showed better per-
formance, compared with the multi-objective particle swarm optimization (MOPSO)
2 Cat Swarm Optimization (CSO) Algorithm 11

technique. Amara et al. (2015) solved the problem of wind power system design
reliability optimization using CSO, under the performance and cost constraints.
Meziane et al. (2015) optimized the electric power distribution of a solar system by
determining the optimal topology among various alternatives using CSO. The results
showed a better performance than the binary CSO. Ram et al. (2015) studied a 9-ring
time-modulated concentric circular antenna array (TMCCAA) with isotropic ele-
ments based on CSO, for reduction of side lobe level and improvement in the
directivity. Crawford et al. (2016) solved a bi-objective set covering problem using
the binary cat swarm optimization algorithm. In order to achieve higher overall
system reliability for a large-scale primary distribution network, Majumder and Eldho
(2016) examined the effectiveness of CSO for groundwater management problems,
by coupling it with the analytic element method (AEM) and the reverse particle
tracking (RPT) approach. The AEM-CSO model was applied to a hypothetical
unconned aquifer considering two different objectives: maximization of the total
pumping of groundwater from the aquifer and minimization of the total pumping
costs. Mohapatra et al. (2016) used kernel ridge regression and a modied
CSO-based gene selection system for classication of microarray medical datasets.

2.2 Natural Process of the Cat Swarm

Optimization Algorithm

Despite spending most of their time in resting, cats have high alertness and curiosity
about their surroundings and moving objects in their environment. This behavior
helps cats in nding preys and hunting them down. Compared to the time dedicated
to their resting, they spend too little time on chasing preys to conserve their energy.
Inspired by this hunting pattern, Chu and Tsai (2007) developed CSO with two
modes: seeking mode for when cats are resting and tracing mode for when they
are chasing their prey. In CSO, a population of cats are created and randomly
distributed in the M-dimensional solution space, with each cat representing a
solution. This population is divided into two subgroups. The cats in the rst sub-
group are resting and keeping an eye on their surroundings (i.e., seeking mode),
while the cats in the second subgroup start moving around and chasing their preys
(i.e., tracing mode). The mixture of these two modes helps CSO to move toward the
global solution in the M-dimensional solution space. Since the cats spend too little
time in the tracing mode, the number of the cats in the tracing subgroup should be
small. This number is dened by using the mixture ratio (MR) which has a small
value. After sorting the cats into these two modes, new positions and tness
functions will be available, from which the cat with the best solution will be saved
in the memory. These steps are repeated until the stopping criteria are satised.
12 M. Bahrami et al.

Table 2.1 Characteristics of the CSO algorithm

General algorithm Cat swarm optimization
Decision variable Cats position in each dimension
Solution Cats position
Old solution Old position of cat
New solution New position of cat
Best solution Any cat with the best tness
Fitness function Distance between cat and prey
Initial solution Random positions of cats
Selection
Process of generating new solution Seeking and tracing a prey

Following Chu and Tsai (2007), the computational procedures of CSO can be
described as follows:
Step 1: Create the initial population of cats and disperse them into the
M-dimensional solution space (Xi,d) and randomly assign each cat a
velocity in range of the maximum velocity value (ti,d).
Step 2: According to the value of MR, assign each cat a flag to sort them into the
seeking or tracing mode process.
Step 3: Evaluate the tness value of each cat and save the cat with the best
tness function. The position of the best cat (Xbest) represents the best
solution so far.
Step 4: Based on their flags, apply the cats into the seeking or tracing mode
process as described below.
Step 5: If the termination criteria are satised, terminate the process. Otherwise
repeat steps 2 through 5.
Table 2.1 lists the characteristics of the CSO and Fig. 2.1 illustrates the detailed
computational steps of the CSO algorithm.

2.2.1 Seeking Mode (Resting)

During this mode the cat is resting while keeping an eye on its environment. In case
of sensing a prey or danger, the cat decides its next move. If the cat decides to
move, it does that slowly and cautiously. Just like while resting, in the seeking
mode the cat observes into the M-dimensional solution space in order to decide its
next move. In this situation, the cat is aware of its own situation, its environment,
and the choices it can make for its movement. These are represented in the CSO
algorithm by using four parameters: seeking memory pool (SMP), seeking range of
2 Cat Swarm Optimization (CSO) Algorithm 13

Start

Define the parameters of the algorithm

Generate initial cats and velocities randomly

Distribute the cats into tracing or seeking mode

No Yes
Is cat k in seeking mode?

Start tracing mode Start seeking mode

Re-evaluate fitness functions and keep the cat with the best solution in the memory

No Yes
Are termination criteria satisfied?

Report the best position among the cats

End

Fig. 2.1 Flowchart of the CSO algorithm

the selected dimension (SRD), counts of dimension to change (CDC), and

self-position consideration (SPC) (Chu and Tsai 2007). SMP is the number of the
copies made of each cat in the seeking process. SRD is the maximum difference
between the new and old values in the dimension selected for mutation. CDC tells
how many dimensions will be mutated. All these parameters dene the seeking
process of the algorithm. SPC is the Boolean variable which indicates the current
position of the cat as a candidate position for movement. SPC cannot affect the
value of SMP.
14 M. Bahrami et al.

Following Chu and Tsai (2007), the process of the seeking mode is described
below.
Step 1: Make SMP copies of each cati. If the value of SPC is true, SMP-1 copies
are made and the current position of the cat remains as one of the copies.
Step 2: For each copy, according to CDC calculate a new position by using
Eq. (2.1) (Majumder and Eldho 2016)

Xcn 1  SRD  R  Xc 2:1

in which
Xc current position;
Xcn new position; and
R a random number, which varies between 0 and 1.

Step 3: Compute the tness values (FS) for new positions. If all FS values are
exactly equal, set the selecting probability to 1 for all candidate points.
Otherwise calculate the selecting probability of each candidate point by
using Eq. (2.2).
Step 4: Using the roulette wheel, randomly pick the point to move to from the
candidate points, and replace the position of cati.

jFSi  FSb j
Pi ; where 0\i\j 2:2
jFSmax  FSmin j

where
Pi probability of current candidate cati;
FSi tness value of the cati;
FSmax maximum value of tness function;
FSmin minimum value of tness function; and
FSb = FSmax for minimization problems and
FSb = FSmin for maximization problems.

2.2.2 Tracing Mode (Movement)

The tracing mode simulates the cat chasing a prey. After nding a prey while
resting (seeking mode), the cat decides its movement speed and direction based on
2 Cat Swarm Optimization (CSO) Algorithm 15

the preys position and speed. In CSO, the velocity of cat k in dimension d is
given by

vk;d vk;d r1  c1 Xbest;d  Xk;d 2:3

in which, vk;d = velocity of cat k in dimension d; Xbest;d = position of the cat with
the best solution; Xk;d = position of the catk; c1 = a constant; and r1 = a random
value in the range of [0,1]. Using this velocity, the cat moves in the M-dimensional
decision space and reports every new position it takes. If the velocity of the cat is
greater than the maximum velocity, its velocity is set to the maximum velocity. The
new position of each cat is calculated by

Xk;d;new Xk;d;old vk;d 2:4

in which
Xk;d;new new position of cat k in dimension d; and
Xk;d;old current position of cat k in dimension d.

2.3 Termination Criteria

The termination criterion determines when the algorithm is terminated. Selecting a

good termination criterion has an important role to ensure a correct convergence of
the algorithm. The number of iterations, the amount of improvement, and the
running time are common termination criteria for the CSO.

2.4 Performance of the CSO Algorithm

Chu and Tsai (2007) used six test functions to evaluate the CSO performance and
compared the results with the particle swarm optimization (PSO) algorithm and the
PSO with weighting factor (PSO-WF). According to the results CSO outperformed
PSO and PSO-WF in nding the global best solutions.
16 M. Bahrami et al.

2.5 Pseudo Code of the CSO Algorithm

Begin
Input parameters of the algorithm and the initial data

Initialize the cat population Xi (i = 1, 2, ... , n), , and SPC

While (the stop criterion is not satisfied or I < Imax)

Calculate the fitness function values for all cats and sort them

Xg= cat with the best solution

For = 1:

If SPC = 1

Start seeking mode

Else

Start tracing mode

End if

End for i

End while

Post-processing the results and visualization

End

2.6 Conclusion

This chapter described cat swarm optimization (CSO) which is a new swarm-based
algorithm. CSO consists of two modes, seeking mode and tracing mode which
simulate the resting and hunting behaviors of cats. Each cat has a position in the
M-dimensional solution space. The cats movement toward the optimum solution is
based on a flag that sorts them into the seeking or tracing mode, the rst one being a
slow movement around their environment and the latter being a fast movement
toward the global best.
A literature review of CSO was presented, showing the success of the algorithm
for different optimization problems, along with different variations of the code
2 Cat Swarm Optimization (CSO) Algorithm 17

developed by other researchers. The flowchart of the CSO along with the pseudo
code was also presented in order to make different parts of the algorithm easier to
understand. These sources are a good reference point for further exploration of the
CSO algorithm.

References

Amara, M., Bouanane, A., Meziane, R., & Zeblah, A. (2015). Hybrid wind gas reliability
optimization using cat swarm approach under performance and cost constraints. 3rd
International Renewable and Sustainable Energy Conference (IRSEC), Marrakech and
Ouarzazate, Morocco, 1013 December.
Bilgaiyan, S., Sagnika, S., & Das, M. (2015). A multi-objective cat swarm optimization algorithm
for workflow scheduling in cloud computing environment. Intelligent Computing,
Communication and Devices (pp. 7384). New Delhi, India: Springer.
Chu, S. C., & Tsai, P. W. (2007). Computational intelligence based on the behavior of cats.
International Journal of Innovative Computing, Information and Control, 3(1), 163173.
Crawford, B., Soto, R., Caballero, H., Olgun, E., & Misra, S. (2016). Solving biobjective set
covering problem using binary cat swarm optimization algorithm. The 16th International
Conference on Computational Science and Its Applications, Beijing, China, 47 July.
Guo, J., Sun, Z., Tang, H., Yin, L., & Zhang, Z. (2015). Improved cat swarm optimization
algorithm for assembly sequence planning. Open Automation and Control Systems Journal, 7,
792799.
Kumar, D., Samantaray, S. R., Kamwa, I., & Sahoo, N. C. (2014). Reliability-constrained based
optimal placement and sizing of multiple distributed generators in power distribution network
using cat swarm optimization. Electric Power Components and Systems, 42(2), 149164.
Lin, K. C., & Chien, H. Y. (2009). CSO-based feature selection and parameter optimization for
support vector machine. Joint Conferences on Pervasive Computing (JCPC), Taipei, Taiwan,
35 December.
Lin, K. C., Zhang, K. Y., & Hung, J. C. (2014a). Feature selection of support vector machine
based on harmonious cat swarm optimization. Ubi-Media Computing and Workshops
(UMEDIA), Ulaanbaatar, Mongolia, 1214 July.
Lin, K. C., Huang, Y. H., Hung, J. C., & Lin, Y. T. (2014b). Modied cat swarm optimization
algorithm for feature selection of support vector machines. Frontier and Innovation in Future
Computing and Communications, 329336.
Majumder, P., & Eldho, T. I. (2016). A new groundwater management model by coupling analytic
element method and reverse particle tracking with cat swarm optimization. Water Resources
Management, 30(6), 19531972.
Meziane, R., Boufala, S., Amara, M., & Hamzi, A. (2015). Cat swarm algorithm constructive
method for hybrid solar gas power system reconguration. 3rd International Renewable and
Sustainable Energy Conference (IRSEC), Marrakech and Ouarzazate, Morocco, 1013
December.
Mohamadeen, K. I., Sharkawy, R. M., & Salama, M. M. (2014). Binary cat swarm optimization
versus binary particle swarm optimization for transformer health index determination. 2nd
International Conference on Engineering and Technology, Cairo, Egypt, 1920 April.
Mohapatra, P., Chakravarty, S., & Dash, P. K. (2016). Microarray medical data classication using
kernel ridge regression and modied cat swarm optimization based gene selection system.
Swarm and Evolutionary Computation, 28, 144160.
Naidu, Y. R., & Ojha, A. K. (2015). A hybrid version of invasive weed optimization with
quadratic approximation. Soft Computing, 19(12), 35813598.
18 M. Bahrami et al.

Pradhan, P. M., & Panda, G. (2012). Solving multiobjective problems using cat swarm optimization.
Expert Systems with Applications, 39(3), 29562964.
Ram, G., Mandal, D., Kar, R., & Ghoshal, S. P. (2015). Cat swarm optimization as applied to
time-modulated concentric circular antenna array: Analysis and comparison with other
stochastic optimization methods. IEEE Transactions on Antennas and Propagation, 63(9),
41804183.
Saha, S. K., Ghoshal, S. P., Kar, R., & Mandal, D. (2013). Cat swarm optimization algorithm for
optimal linear phase FIR lter design. ISA Transactions, 52(6), 781794.
Shara, Y., Khanesar, M. A., & Teshnehlab, M. (2013). Discrete binary cat swarm optimization
algorithm. In Computer, Control & Communication (IC4). 3rd IEEE International Conference
on Computer, Control & Communication (IC4), Karachi, Pakistan, 2526 September.
So, J., & Jenkins, W. K. (2013). Comparison of cat swarm optimization with particle swarm
optimization for IIR system identication. Asilomar Conference on Signals, Systems and
Computers, Pacic Grove, CA, USA, 69 November.
Tsai, P. W., Pan, J. S., Chen, S. M., Liao, B. Y., & Hao, S. P. (2008). Parallel cat swarm
optimization. International Conference on Machine Learning and Cybernetics, Kunming,
China, 1215 July.
Tsai, P. W., Pan, J. S., Chen, S. M., & Liao, B. Y. (2012). Enhanced parallel cat swarm
optimization based on the Taguchi method. Expert Systems with Applications, 39(7),
63096319.
Wang, J. (2015). A new cat swarm optimization with adaptive parameter control. Genetic and
Evolutionary Computing, 6978.
Xu, L., & Hu, W. B. (2012). Cat swarm optimization-based schemes for resource-constrained
project scheduling. Applied Mechanics and Materials, 220, 251258.
Yang, S. D., Yi, Y. L., & Shan, Z. Y. (2013a). Chaotic cat swarm algorithms for global numerical
optimization. Advanced Materials Research, 602, 17821786.
Yang, S. D., Yi, Y. L., & Lu, Y. P. (2013b). Homotopy-inspired cat swarm algorithm for global
optimization. Advanced Materials Research, 602, 17931797.
Chapter 3
League Championship Algorithm (LCA)

Hossein Rezaei, Omid Bozorg-Haddad and Xuefeng Chu

Abstract This chapter briefly describes the league championship algorithm

(LCA) as one of the new evolutionary algorithms. In this chapter, a brief literature
review of LCA is rst presented; and then the procedure of holding a common
league in sports and its rules are described. Finally, a pseudo code of LCA is
presented.

3.1 Introduction

The league championship algorithm (LCA) is one of the new evolutionary algo-
rithms (EA) for nding global optimum in a continuous search space rst proposed
by Kashan (2009, 2011) developed the basic LCA for solving a constrained opti-
mization benchmark problem. The results demonstrated that LCA is a very com-
petitive algorithm for constrained optimization problems. Kashan et al. (2012)
modied the basic LCA by using two halves analysis, instead of the post-match
SWOT analysis. The performance of the more realistic modied LCA (RLCA) was
compared with those of the particle swarm optimization (PSO) and the basic LCA
in nding the global solutions of different benchmark problems. The results indi-
cated the better performance of RLCA in terms of the quality of nal solutions and
the convergence speed. LCA has been applied to different optimization problems.

H. Rezaei
O. Bozorg-Haddad (&)

Department of Irrigation and Reclamation Engineering, Faculty of Agricultural Engineering
and Technology, College of Agriculture and Natural Resources, University of Tehran, Karaj,
Tehran 31587-77871, Iran
e-mail: [email protected]
H. Rezaei
e-mail: [email protected]
X. Chu
Department of Civil and Environmental Engineering, North Dakota State University,
Dept 2470, Fargo, ND 58108-6050, USA
e-mail: [email protected]

Springer Nature Singapore Pte Ltd. 2018 19

O. Bozorg-Haddad (ed.), Advanced Optimization by Nature-Inspired Algorithms,
Studies in Computational Intelligence 720, DOI 10.1007/978-981-10-5221-7_3
20 H. Rezaei et al.

Lenin et al. (2013) utilized LCA for solving a multi-objective dispatch problem.
Abdulhamid and Latiff (2014) used LCA based on job scheduling scheme for
optimization of infrastructure as a service cloud. Sajadi et al. (2014) applied LCA
for scheduling context. The scheduling considered in a permutation flow-shop
system with makespan criterion. Abdulhamid et al. (2015a, b, c) used LCA to
minimize makespan time scheduled tasks in IaaS cloud. Abdulhamid et al. (2015a,
b, c) proposed a job scheduling algorithm based on the enhanced LCA in order to
optimize the infrastructures as a service cloud. The performance of the proposed
algorithm was compared with three other algorithms. The results proved that the
LCA scheduling algorithm performed better than other algorithms. Xu et al. (2015)
presented an improved league championship algorithm with free search (LCAFS),
in which the novel match schedule was implemented to improve the capability of
competition for teams and by introducing the free search operation, the diversity of
league was also improved. The global search performance and convergence speed
of LCAFS were compared with those of the basic LCA, PSO, and genetic algorithm
(GA) in solving some benchmark functions. The results demonstrated that LCAFS
was able to describe complex relationships between key influence factors and
performance indexes. Jalili et al. (2016) introduced a new approach for optimizing
truss design based on LCA, which considered the concept of tie. The performance
of the proposed algorithm was evaluated by using ve typical truss design examples
with different constraints. The results illustrated the effectiveness and robustness of
the proposed algorithm.

3.2 Review of LCA and Its Terminology

The basic idea of LCA was inspired by the concept of league championship in sport
competitions. Following terms related to the league, team and its structure are
commonly used in LCA.
League: league means a group of sport teams that are organized to compete
with each other in a certain type of sport. A league championship can be held in
different ways. For example, the number of games that each team should play with
other teams may vary. At the end of the league championship, the champion is
determined based on the win, loss and tie records during the leagues competition
with other teams.
Formation: formation of a team refers to the specic structure of the team when
playing with other teams, such as the positions of players, and the rule of each player
during match. For any sport teams, coaches arrange their teams based on their
players abilities to achieve the best available formation to play with other teams.
Match analysis: match analysis refers to the examination of behavioral events
occurring during a match. The main goal of match analysis after determining the
performance of a team is to recognize the strengths and weaknesses and to improve.
The important part of the match analysis process is to send the feedback of last
matches (their own match and the opponents match) to players. The feedback
3 League Championship Algorithm (LCA) 21

should be given to the player, pre-match and post-match to build up team for next
match. One of these analyses is the strength/weakness/opportunity/threat (SWOT)
analysis which links the external (opportunities and threats) and internal (strengths
and weaknesses) factors of the teams performance. Identication of SWOTs is
necessary because next planning step is based on the results of the SWOT analysis
to achieve the main objective. The SWOT analysis evaluates the interrelationships
between the internal and external factors of match in four basic categories:
S/T matches illustrate the strengths of a team and the threats of competitors. The
team should use its strengths to defuse threats.
S/O matches illustrate the strengths and opportunities of a team. The team should
use its strengths to take opportunities.
W/T matches illustrate the weaknesses and threats of a team. The team should
try to minimize its weaknesses and defuse threats.
W/O matches show the weaknesses coupled with opportunities of a team. The
team should attempt to overcome weaknesses by use of opportunities.
The SWOT analysis provides a structure for conducting gap analysis. A gap
refers to the space between the place where we are and the place where we want to
be. The identication process of a teams gap contains an in-depth analysis of the
factors that express the current condition of the team and subsequently help to make
a plan for improvement of the team.

3.3 League Championship Algorithm

LCA is a population-based algorithm that is used for solving global optimization

problems with a continuous search space. Like other population-based algorithms,
LCA tries to move populations from possible areas to promising areas during
searching for the optimum in the whole decision space. Table 3.1 presents a list of
the characteristics of the LCA.

Table 3.1 Characteristics of the LCA

General algorithm League championship algorithm
Decision variable Players strength in each teams formation
Solution Teams formation
Old solution Old teams formation
New solution New teams formation
Best solution The winner of league championship
Fitness function Teams playing strength
Initial solution Random formation for each team
Selection Match analysis process
Process of generating new solution SWOT matrix
22 H. Rezaei et al.

In the optimization process of LCA, a set of L (an even number) solutions are
rst created randomly to build initial population. Then, they evolve gradually the
composition of the population in sequential iterations. In LCA, league refers to
population; formation of a team stands for solution; and week refers to iteration. So,
team i denotes the ith solution of the population. A tness value is then calculated
for each team based on the teams adaption to the objectives (determined by the
concepts of player strength and teams formation). In LCA, new solutions are
generated for next week by applying operators to each team based on the results of
match analysis which are used by coaches to improve their teams arrangements.
An evolutionary algorithm (EA) is a population-based one that uses the Darwins
evolution theory as selection mechanism. Based on a pseudo code of EA and
according to the selection process of LCA (greedy selection), in which the current
teams formation is replaced by the best teams formation, LCA can be classied as
an EA group of the population-based algorithms. LCA terminates after a certain
number of seasons (S), each of which is composed of L  1 weeks. Note that the
number of iterations in LCA is equal to SL  1.
LCA models an articial championship during the optimization process of the
algorithm based on some idealized rule that can be expressed as follows:
(1) The team with better playing strength (ability of the team to defeat competitors)
has more chances to win the game.
(2) The weaker team can win the game but its chance to win the game is very low
(the playing strength does not determine the nal outcome of the game exactly).
(3) The sum of the wins probabilities of both teams that participate in a match is
equal to one.
(4) The outcome of the game only can be win or loss (tie is not acceptable as the
outcome of the game in the basic version of LCA).
(5) When teams i and j compete with each other and eventually team i wins the
match, any strength helps team i to win and dual weakness causes team j to lose
the match (weakness is a lack of specic strength).
(6) The teams just focus on their forthcoming match without consideration of other
future matches and the formation of team arranged only by previous week
results.
Figure 3.1 shows the flowchart of LCA, which illustrates the optimization
process of the basic LCA. As shown in Fig. 3.1, rst of all a representation for
individuals must be chosen. Solutions (teams formation) are represented with n
decision variables of real numbers. Each element of the solutions depends on one of
the teams players and shows the corresponding values of the variables with the aim
of optimization. Changes in each value can be the effect of changes in the
responsibility of the corresponding player. f x1 ; x2 ; . . .; xn denotes an n variable
function to be minimized during the optimization running of LCA over a decision
space
(a subset of Rn ). The solution of team i at week t can be represented by
Xit xti1 ; xti2 ; . . .; xtin and the value of its tness function (player strength) is ff xti .


Bt1
i bt1i1 ; bi2 ; . . .; bin
t1 t1
i denote the best formation of team i before
and ff Bt1
3 League Championship Algorithm (LCA) 23

Fig. 3.1 Flowchart of the basic LCA

week t and its tness function, respectively. The greedy selection in LCA can be
made between ff xti and ff Bt1
i . The modules of LCA, generation of league
schedule, determination of the winner or loser, and setup of new formation are
detailed in the following section.
24 H. Rezaei et al.

3.4 Generating League Schedule

The common aspect of different sport leagues is the structure, in which teams can
compete with each other in a nonrandom schedule, named season. Therefore, the
rst and the most important step in LCA, is to determine the match schedule in each
season. A single-round robin schedule can be applied in LCA to determine the
teams schedule, in which each team competes against other teams just once in a
certain season. In a championship containing L teams with the single-round robin
schedule rule, there are LL  1=2 matches in a certain season. In each of L  1
weeks, L=2 matches will be held in parallel (if L is odd, in each week L  1=2
matches will be held and one team has to rest).
The procedure of scheduling of the algorithm can be illustrated by a simple
example of league championship of 8 teams. The teams have named from a to h.
In the rst week, the competitors are identied randomly. Figure 3.2a shows
the competitors in the rst week. For example, team a competes with team d and
team b competes with team g. In the second week, in order to identify the pairs of
competitors, one of the teams (team a) is xed in its own place and all other teams
turn round clockwise. Figure 3.2b indicates the procedure of identifying the pairs of
competitors for week 2. This process continues until the last week (week 7) shown
Fig. 3.2c. In LCA, the single-round robin tournament is applied for scheduling
L teams in SL  1 weeks.

3.5 Determining the Winner or Loser

During the league championship, teams compete with each other in every week.
The outcome of each match can be loss, tie, or win. The scoring rules for the
outcome of the matches can be different for different sports. For instance, in soccer
the winner gets three, and the loser gets a zero score. By the end of the match, both
teams get one if the outcome is tie. According to the idealized rule 1, the chance of
a stronger team to win the match is higher than its competitor, but occasionally a
weaker team may win the match. Therefore, the outcome of the match is associated
with different reasons. The most important one is the playing strengths of the teams.
So we can consider a linear relationship between the playing strengths and the
outcome of the match (idealized rule 2).

Fig. 3.2 A simple example of league championship scheduling

3 League Championship Algorithm (LCA) 25

In LCA, a stochastic criterion of playing strength, which depends on the degree

of t of a team, is utilized to determine the winner or loser of the match. Note that in
the basic version of LCA the outcome of the matches can only be win or loss
(no tie). The teams degree of t refers to the proportion of playing strength, which
is calculated based on the distance between the playing strength and the ideal
reference point (the lower bound of the optimization problem).
Assuming that team i competes with team j at week t, the chance of each team to
defeat another team can be expressed as (Kashan 2009) follows:

f xti  ^f ptj
3:1
f xt  ^f pti
j

where xti and xtj formation of teams i and j at week t; f xti and f xtj playing
strength of teams i and j at week t; ^f ideal reference point; pti chance of team
i to defeat team j at week t; and ptj chance of team j to defeat team i at week t.
Because the chances of both teams to win the match are evaluated based on the
specic point, the ratio of distance is identied as the teams winning portion.
According to idealized rule 3, the relationship of the chances of teams i and j at
week t can be expressed as follows:

pti ptj 1 3:2

Based on Eqs. (3.1) and (3.2), the chance of team i to defeat team j at week t is
given by (Kashan 2014):
 
f xtj  ^f
pti   3:3
f xti f xj  2^f
t

In LCA, in order to specify the winner of the match, a random number between
[0,1] is generated randomly. If the generated number is equal to or less than pti , team
i defeats team j at week t. Otherwise, team j defeats team i at week t.

3.6 Setting Up a New Team Formation

Before applying any strategy to team i, in order to change the formation of team i at
next week, coaches should identify strengths and weaknesses of the team and
players (individuals). Based on these strengths and weaknesses, coaches determine
the formation of the team in next week to enhance the performance of the team. An
articial match analysis can be performed to specify the opportunities and threats.
Strengths and weaknesses are internal factors while opportunities and threats are
external factors. In LCA, the internal factors are evaluated based on the teams
26 H. Rezaei et al.

Fig. 3.3 Procedure of the articial match analysis in LCA

performance at the last week (week t), while evaluating the external factors is based
on the opponents performance at week t. The articial match analysis helps prepare
team i for next week (week t + 1). In the modeling process, if team i wins (loses)
the match at week t, it is assumed that the success (failure) is directly related to the
strengths (weaknesses) of team i or weaknesses (strengths) of its opponent team
j (idealized rule 5). The procedure of modeling and evaluating the articial match
analysis for team i at week t is displayed in Fig. 3.3. The left side in Fig. 3.3 shows
the evaluation of hypothetical internal factors and the right side shows the way of
evaluating the external factors.
According to the results of the articial match analysis applied to team i in order
to determine its performance, the coach should take some possible actions to
improve the teams performance. The possible actions (SWOT analysis) are shown
in Table 3.2. The SWOT analysis is adjusted based on idealized rule 6. Table 3.2
shows different strategies (S/T, S/O, W/T, and W/O) that can be adopted for
team i in different situations. For instances, if team i has won the last match and
team l has lost its match at the last week, it is reasonable for team i to focus on
strengths which give it more chance to win team l at next match. Therefore,
adopting the S/O strategy for team i is efcient. Table 3.2 also displays,
3 League Championship Algorithm (LCA) 27

Table 3.2 Hypothetical SWOT analysis derived from the articial match analysis
Adopt S/T strategy Adopt S/O strategy Adopt W/T strategy Adopt W/O strategy
Team i has won Team i has won Team i has lost Team i has lost
Team l has won Team l has lost Team l has won Team l has lost
Focusing on Focusing on Focusing on focusing on
S Own strengths or Own strengths or
weaknesses of weaknesses of
team j team j
W Own weaknesses or Own weaknesses or
strengths of team j strengths of team j
O Weaknesses of team Weaknesses of team
l or strengths of l or strengths of
team k team k
T Strengths of team Strengths of team
l or weaknesses of l or weaknesses of
team k team k

in a metaphorical way, the SWOT analysis matrix which is used for planning in the
future matches.
The aforementioned analysis must be performed by all participants at week t to
plan for next match and choose a suitable formation for upcoming match. After
adopting a suitable strategy for team i based on the SWOT matrix, all teams should
ll their gaps. For instance, assume that in a soccer match team i has lost the match
at week t to team j and the results of the match analysis process have specied that
the type of defensive state (man-to-man defensive state) is the reason of loss
Therefore, a gap exists between the current sensitive defensive state and the state
which ensures a man-to-man pressure defense at week t 1.
According to the league schedule, team l is the competitor of team i (i = 1, 2, , L)
at week t 1; team j is the competitor of team i at week t; and team k is the competitor
of team l at week t. As aforementioned, Xit ; Xjt and Xkt , respectively, denote the
formations of teams i, j, and k at week t. Xkt  Xit denes the gap between playing
styles of teams i and k, which highlights the strengths of team k. This case applies
when team i wants to play with team l at week t + 1 and team k wins team l at week t
by Xkt s formation. Therefore, if team i uses the playing style of team k at week t (Xkt ) to
compete with team l at week t 1, it is highly possible for team i to win team l at
week t 1. Similarly, Xit  Xkt is used if we want to focus on the weaknesses
of team k. In this case, team i should not use the playing style of team k at week t
against team l. Xit  Xjt and Xjt  Xit also can be dened. Due to the principle that
each team should play with the best formation that is selected from playing experi-
ence up to now and by considering the results of the articial match analysis in last


week, new formation of team i at week t 1 Xit 1 xti1 1 ; xi2t1
; . . .; xtin 1 can be
set up by one of the following equations:
If teams i and l have won the match at week t, the new formation of team i will
be generated based on the S/T strategy:
28 H. Rezaei et al.

S=T strategy : xtim 1 btim ytim x1 r1im xtim  xtkm x1 r2im xtim  xtjm
3:4
8m 1; 2; . . .; n

If team i has won and team l has lost at week t, the new formation of team i will
be generated based on the S/O strategy:

S=O strategy : xtim 1 btim ytim x2 r1im xtKm  xtim x1 r2im xtim  xtjm
3:5
8m 1; 2; . . .; n

If team i has lost and team l has won at week t, the new formation of team i will
be generated based on the W/T strategy:

W=T strategy : xtim 1 btim ytim x1 r1im xtim  xtkm x2 r2im xtjm  xtim
3:6
8m 1; 2; . . .; n

If teams i and l have lost the match at week t, the new formation of team i will be
generated based on the W/O strategy:
t1
W=O strategy : xim btim ytim x2 r1im xtkm  xtim x2 r2im xtjm  xtim
3:7
8m 1; 2; . . .; n

where m number of team members; r1im and r2im random numbers between
[0,1]; x1 and x2 coefcients used to scale the contribution of approach or retreat
components; and ytim binary variable that species whether or not the mth player
must change in the new formation (only ytim 1 allows to change). Note that
different signs in the parentheses are the consequence of acceleration towards the
winner or recess from the loser. Yit yti1 ; yti2 ; . . .; ytin denotes a binary change
variable. The summation of the changes needed for next match yti 1 is equal to
qti . Changes in all aspects of the team (players and styles) by coaches are not
common (just a few changes in the team can be required). In order to calculate the
number of changes in the teams formation for next match, a truncated geometric
distribution is applied in LCA. The truncated geometric distribution lets LCA to
control the number of changes with emphasis on the smaller rates of changes in Bti .
The truncated geometric distribution can be expressed as follows:
" #
ln1  1  pc nq0 1 r
qi
t
q0  1 : qti 2 fq0 ; q0 1; . . .; q0 ng 3:8
ln1  pc

where r random number between [0,1]; pC control parameter pc \1; pc 6 0;

and q0 the least number of changes. If pc \0, the situation is reversed so that by a
more negative value of pc , The emphasis in LCA is placed on a greater rate of change
in the teams formation. q0 is determined during the match analysis (note that the
minimum value of q0 is equal to zero). pc in the truncated geometric distribution is
the probability of success. In LCA, after calculating the value of qti by using Eq. (3.8),
the players of Bti are randomly selected and changed based on Eqs. (3.4)(3.7).
3 League Championship Algorithm (LCA) 29

3.7 Pseudo Code of LCA

Begin
Generate initial teams formation randomly [ ( X it ( xit1 , xit2 ,..., xint ) ], i 1, 2, ,L]

Generate league schedule for L teams

For m=1: L ( S 1)

Evaluate the strengths for all teams

Calculate the chance of each team to defeat its competitors in next match ( p it )

Generate a random number between [0,1] (Rn)

If Rn p it

Team i wins the match

Else

Team j wins the match

End if

Generate a random number between [0,1] (r)

Calculate the number of changes in teams best formation ( Bit ) for next match

based on the truncated geometric distribution ( qit )

qit players are selected randomly from Bit and changed by the SWOT matrix

If team i and team l have won

Select the S/T strategy

Else if team i has won and team l has lost

Select the S/O strategy

Else if team i has lost and team l has won

Select the W/T strategy

Else if team i has lost and team l has lost

Select the W/O strategy

End if

End for m

End
30 H. Rezaei et al.

3.8 Conclusions

This chapter described the league championship algorithm (LCA), which stemmed
from the concept of league championship in sport. This chapter also presented a
literature review of LCA, and its algorithmic fundamental and pseudo code.

References

Abdulhamid, S. M., & Latiff, S. A. (2014). League championship algorithm (LCA) based job
scheduling scheme for infrastructure as a service cloud. 5th International Graduate Conference
on Engineering, Science and Humanities, UTM Postgraduate Student Societies, Johor,
Malaysia, 1921 August.
Abdulhamid, S. M., Latiff, M. S., & Abdullahi, M. (2015a). Job scheduling technique for
infrastructure as a service cloud using an enhanced championship algorithm. 2nd International
Conference on Advanced Data and Information Engineering, Lecture Notes in Electrical
Engineering, Bali, Indonesia, 2526 April.
Abdulhamid, S. M., Latiff, M. S., & Idris, I. (2015b). Tasks scheduling technique using league
championship algorithm for makespan minimization in IaaS cloud. ARPN Journal of
Engineering and Applied Sciences., 9(12), 25282533.
Abdulhamid, S. M., Lattif, M. S. A., Madni, S. H. H., & Oluwafemi, O. (2015c). A survey of
league championship algorithm: prospects and challenges. Indian Journal of Science and
Technology.
Jalili, S., Kashan, A. H., & Husseinzadeh, Y. (2016). League championship algorithms for
optimization design of pin-jointed structures. Journal of Computing in Civil Engineering.
doi:10.1061/(ASCE)CP.1943-5487.0000617
Kashan, A. H. (2009). League championship algorithm: A new algorithm for numerical function
optimization. International Conference on Soft Computing and Pattern Recognition, IEEE
Computer Society, Malacca, Malaysia, 47 December.
Kashan, A. H. (2011). An efcient algorithm for constrained global optimization and application to
mechanical engineering design: League championship algorithm (LCA). Computer-Aided
Design, 43(2011), 17691792.
Kashan, A. H. (2014). League championship algorithm (LCA): an algorithm for global
optimization inspired by sport championships. Applied Soft Computing, 16(2014), 171200.
Kashn, A. H., Karimiyan, S., Karimiyan, M., & Kashan, M. H. (2012). A modied League
Championship Algorithm for numerical function optimization via articial modeling of the
between two halves analysis. The 6th International Conference on Soft Computing and
Intelligent Systems, and the 13th International Symposium on Advanced Intelligence Systems,
University of Aizu, Kobe, Japan, 2024 November.
Lenin, K., Reddy, B. R., & Kalavati, M. S. (2013). League championship algorithm (LCA) for
solving optimal reactive power dispatch problem. International Journal of Computer and
Information Technologies., 1(3), 254272.
Sajadi, S. M., Kashan, A. H., & Khaledan, S. (2014). A new approach for permutation flow-shop
scheduling problem using league championship algorithm. In Proceedings of CIE44 and
IMSS14, 2014.
Xu, W., Wang, R., & Yang, J. (2015). An improved league championship algorithm with free
search and its application on production scheduling. Journal of Intelligent Manufacturing.
doi:10.1007/s10845-015-1099-4
Chapter 4
Anarchic Society Optimization
(ASO) Algorithm

Atiyeh Bozorgi, Omid Bozorg-Haddad and Xuefeng Chu

Abstract Due to limited resources and equipment in most engineering projects, it

is necessary to use optimization techniques. Older optimization techniques,
including derivative and other mathematical methods may not be practical to new
complex problems. Therefore new optimization algorithms are needed. In the past
decades many algorithms were developed and used for different optimization
problems, which can be divided into three categories including classic, evolutionary
and heuristic algorithms. The evolutionary and heuristic algorithms which are used
widely in recent years are based on animals life. In this chapter, one of the heuristic
algorithms named Anarchic Society Optimization (ASO) algorithm based on
human societies, is introduced. After a brief literature review of the ASO algorithm,
more technical details on this method and its performance are described.

4.1 Introduction

Anarchic is derived from the Greek word anarkos meaning no boss and Anarchia
means lack of government. The term Anarchism refers to a political opinion and
movement believing that any political power and authority are obscene and
unnecessary and that any government should be overthrown and replaced with free
associations and volunteer groups. Because the Anarchism believes that the gov-
ernment causes a nations social miseries. Overall, Anarchists are opposed to any

A. Bozorgi
O. Bozorg-Haddad (&)

Department of Irrigation and Reclamation Engineering, Faculty of Agricultural Engineering
and Technology, College of Agriculture and Natural Resources, University of Tehran,
31587-77871 Karaj, Tehran, Iran
e-mail: [email protected]
A. Bozorgi
e-mail: [email protected]
X. Chu
Department of Civil and Environmental Engineering, North Dakota State University,
Dept 2470, Fargo, ND 58108-6050, USA
e-mail: [email protected]

Springer Nature Singapore Pte Ltd. 2018 31

O. Bozorg-Haddad (ed.), Advanced Optimization by Nature-Inspired Algorithms,
Studies in Computational Intelligence 720, DOI 10.1007/978-981-10-5221-7_4
32 A. Bozorgi et al.

government authority and consider the democracy as the Tyranny of the majority.
They emphasize individual freedom. This emphasis results in opposition to any
external authority, especially government, which is construed as a barrier for free
individual growth and excellence.
The Anarchist thought is based on a variety of principles including individual-
ism, humanism, libertarians, lawlessness, anarchic, and absolute freedom.
According to these principles, Anarchism opposes any religious or non-religious
social institutions and considers the human as an absolute free creature. At the heart
of Anarchism, there is a reckless utopia orientation, believing in natural wellness, or
at least mankind potential wellness (Nettlau 2000).
Anarchism contains a variety of branches and anarchism theorists follow one of
them. According to the view of the Communist Anarchism, human is inherently
social, and a society and its individuals benet each other. Human and society
conformity is possible when negating the powerful social institutions especially
government. The Syndicate-oriented Anarchism looks for the salvation in economic
strife not in the political strife of proletariat. Followers of this faction organize labor
unions and syndicates to quarrel with the power structure. According to their point
of view, the current government will be eventually annihilated as a result of a
revolution and the new economic order will be formed based on syndicates.
Nowadays, such thoughts have become a mass movement in some South American
and European countries. The followers of Individualist Anarchism believe that the
human has the right to do whatever he/she will and whatever deprive him/her from
freedom must be destroyed (Woodcook 2004).
In general, there are three major insufciencies that Anarchism suffers from.
First, Anarchism has the unrealistic goal that is related to collapse of government
and all other forms of political authority, while economic and social development
has been always accompanied with government roles. Second, the Anarchism
objects to powerful institutions like parties that can play an effective and efcient
role in development of a society. Third, Anarchism lacks a series of distinct and
coherent political beliefs, which causes many disputes.
During the recent centuries, many countries have undergone certain types of
anarchy, including France (The revolution period), Jamaica (1720), Russia (during
Civil Wars), Spain (1936), Albania (1997), and Somali (19912006). According to
the view of Anarchists, a society can be managed without the need of the central
government and only based on individuals or volunteer groups. In this case, indi-
viduals or groups will be able to determine the right direction without being ordered
by a ruling power and only based on their or others previous experiences.
Although this viewpoint has not been prosperous in stable management of a society
so far, it can be used as a basis for developing optimization methods in engineering
sciences.
In this method, each individual selects his/her next position according to the
personal experiences, group or syndicate experiences, and historical experiences.
Finally, after a number of moves, at least one of the group members would reach a
near-optimal answer. Employing this algorithm causes the total decision space to be
fully searched and prevents being stuck at local optima.
4 Anarchic Society Optimization (ASO) Algorithm 33

The ASO algorithm was rst introduced by Ahmadi-Javid (2011). He compared

the answers obtained from the ASO algorithm with those from the Genetic
Algorithm (GA) and the particle swarm optimization (PSO) algorithm and proved
that the ASO algorithm resulted in better answers. The author also claimed that the
anarchy community algorithm converged to the global optimum with the proba-
bility of 100%. Shayeghi and Dadashpour (2012) compared ASO with PSO, vector
evaluated particle swarm optimization (VEPSO) and craziness based particle swarm
optimization (CRPSO) algorithm for voltage oscillation damping problem and the
best results were obtained from the ASO algorithm. Ahmadi-Javid and
Hooshangi-Tabrizi (2012) expanded the ASO algorithm to two objective functions
and the comparison between ASO and PSO algorithms showed that ASO provided
much better results. Ahmadi Javid and Hooshangi-Tabrizi (2015) expanded and
applied the ASO algorithm for permutation flow-shop scheduling problem with
integer and linear objective functions namely ASO(I) and ASO(II). The numerical
results obtained showed that the ASO algorithm had higher efciency for that
problem.

4.2 Formulation

The detailed procedures of the ASO algorithm are shown in Fig. 4.1.
For a solution space S, f : S ! R is a function that should be minimized in
S. Consider that a community, consisting of N member(s), is being searched in an

Start

End
Generating the initial
population Yes

Evaluating the objective Satisfy stopping

function criteria
No

Sorting of Solutions
Combining movement policies

Determining the values of iK*,

Pbest and Gbest

Updating position of each Updating position of each

Updating position of each member based on other position member based on past position
member based on current

Computing external Computing internal

Computing fickleness index irregularity index irregularity index

Determining MpCurrent Determining MpSociety Determining MpPast

Fig. 4.1 Flowchart of the ASO algorithm

34 A. Bozorgi et al.

unknown territory (the solution space) for discovering the best place to live (i.e., the
overall minimum of f on S). Xi(k) presents the position of the ith member in the kth
iteration; X*(k) denotes the best position experienced by all members in the kth
iteration; and Gbest is the best position experienced by the ith member during the
rst k iterations.

4.3 Algorithm Procedure

As shown in Fig. 4.1, rst, a number of community members are selected randomly
within the solution space. Then, the tness of every member is determined.
According to the calculated tness value and comparison with X*(k), Pbest
i and Gbest ,
the movement policy and a new position of the member will be determined. After
an adequate number of iterations, at least one of the members will reach the optimal
position. Table 4.1 lists the characteristics of the ASO.

4.4 Movement Policy Based on Current Positions

The rst movement policy for the ith member in the kth iteration [MPicurrent(k)] is
adopted based on the current position. The Fickleness Index FIi(k) for member i in
iteration k is used to select this movement policy (Ahmadi-Javid 2011). This index
measures the satisfaction of the current position of the ith member compared with
other members positions. If the objective function is positive in S, the Fickleness
Index can be expressed as one of the following forms (Ahmadi-Javid 2011):

f X  k f Pi k
FIi k 1  ai  1  ai 4:1
f X i k f Xi k

Table 4.1 Characteristics of the ASO algorithm

General algorithm Anarchic society optimization
Decision variable Society members positions in each dimension
Solution Society members positions
Old solution Old positions of society members
New solution New positions of society members
Best solution Best position
Fitness function Desirability of members positions
Initial solution Random positions
Selection
Process of generating new solution Combination of movement policy
4 Anarchic Society Optimization (ASO) Algorithm 35

f G k f Pi k
FIi k 1  ai  1  ai 4:2
f X i k f Xi k

where ai is a non-negative number in [0,1]. Thus, the Fickleness Index is a number

in the range of [0,1]. According to the values of Fickleness Index, the ith member
would select his/her next position. If FIi k is smallest, the ith member has the best
position among all members. So it is better to select the movement policy based on
X  k . Otherwise, the ith member has a totally unpredictable movement. Therefore
the movement policy for the ith member according to the value of FIi k can be
expressed as:

moving towards X k 0  FIi k  ai
MPcurrent k 4:3
i moving towards a random Xi k ai  FIi k  1

4.5 Movement Policy Based on Positions of Other

Members

The second movement policy for the ith member in the kth iteration [MPisociety(k)] is
adopted based on the positions of other members. Although each member should
move in the direction of Gbest logically, the movement of the member is not pre-
dictable due to the anarchist nature of the member and may move toward another
community member. Therefore, the external irregularity index EIi(k) for the ith
member in the kth iteration can be calculated by (Ahmadi-Javid 2011):

EIi k 1  ehi f Xi kf Gk 4:4

EIi k 1  edi Dk 4:5

in which hi and di are positive numbers and Dk is an appropriate dispersion

measure like coefcient of variation CV k . Equation (4.4) denes the distance of
community member i from Gbest . If the community member is close to Gbest , it will
have a more logic behavior. Otherwise, it shows an anarchic behavior based on
anarchy. Equation (4.5) denes a diversity index in community which has a direct
relationship with the diversity of the community members. In the case that this
index is selected, the community members are supposed to behave more logically
and they are less diversied. Thus, with consideration of a threshold for EIi k, it is
possible to dene the movement policy based on the positions of other members as
follows:
36 A. Bozorgi et al.


moving towards Gbest 0  EIi (k)  threshold
MPsociety (k) 4:6
i
moving towards a random Xi (k) thereshold  EIi (k)  1

The closer the threshold is to zero, the more illogical the member movements
would be. As the threshold converges to one, the members would behave more
logically.

4.6 Movement Policy Based on Previous Positions

The third movement policy for the ith member in the kth iteration [MPipast(k)] is
adopted based on the previous positions of the individual member. In order to select
this movement policy, the position of the ith member in the kth iteration is com-
pared to Pbest best
i . If the position of the member is close to Pi , the member behaves
more logically. Otherwise, the member shows illogical behavior. To determine the
movement policy based on previous positions, the internal irregularity index
IIi(k) for the ith member in the kth iteration is dened as follows:

IIi k 1  ebi f Xi kf Pi k 4:7

where bi is a positive number. Like the previous policy, with selection of a

threshold for IIi k , the movement policy can be dened based on previous posi-
tions as follows:

moving towards Pbest 0  IIi (k)  threshold
MPpast
i (k) i 4:8
moving towards a random Xi (k) thereshold  IIi (k)  1

The closer the threshold is to zero, the more illogical the member movements
would be. As the threshold converges to one, the members would behave more
logically.

4.7 Combination of Movement Policies

In order to select the nal movement policy, the three policies discussed above are
combined with each other. After the movement policies are calculated, each
member should combine these policies with a method and move toward a new
position. One of the simplest methods is to select the policy with the best answer.
The next alternative is to combine the movement policies with each other
sequentially which is referred to as the sequential combination rule. The crossover
method can either be used for continuous problems coded as chromosomes, or used
in a sequential way to combine the movement policies. Ahmadi Javid (2011)
demonstrated that the ASO algorithm is a more general state of the PSO algorithm.
4 Anarchic Society Optimization (ASO) Algorithm 37

4.8 Pseudo Code of the ASO

38 A. Bozorgi et al.

4.9 Conclusion

The evolutionary and heuristic algorithms are widely used for solving optimization
problems in the engineering and science elds. In the last few decades, various
algorithms have been introduced mostly based on insects, animal lives, and bio-
logical concepts. In this chapter, the Anarchic Society Optimization algorithm, rst
introduced by Ahmadi-Javid (2011), was described.
This algorithm has been investigated for solving electrical and industrial engi-
neering problems (e.g., power) and optimizing water networks and reservoir
operation. The ASO algorithm has some advantages such as its relatively simple
structure and its potential to achieve better performance. In denition of the ASO
algorithm three indices are used. It seems that changing the use of these indices to
reach a new position for each member or even dening a new index can lead to
superior convergence.
The ASO algorithm was adopted from the life of human communities. Since
human societies are more complicated than animal or insect groupings, it is
expected that ASO as the rst algorithm based on human societies is a turning point
of the performance and capabilities of optimization algorithms.

References

Ahmadi-Javid, A. (2011). Anarchic Society Optimization: A human-inspired method. In IEEE

Congress on Evolutionary Computation (CEC) (pp. 25862592), New Orleans, LA.
Ahmadi-Javid, A., & Hooshangi-Tabrizi, P. (2012). An Anarchic Society Optimization Algorithm
for a flow-shop scheduling problem with multiple transporters between successive machines.
International Conference on Industrial Engineering and Operations Management (ICIEOM),
Istanbul, Turkey, 36 July.
Ahmadi-Javid, A., & Hooshangi-Tabrizi, P. (2015). A mathematical formulation and anarchic
society optimisation algorithms for integrated scheduling of processing and transportation
operations in a flow-shop environment. International Journal of Production Research, 53(19),
59886006.
Nettlau, M. (2000). A short history of anarchism (1st ed.). Freedom Press. ISBN-13: 978-
0900384899.
Shayeghi, H., & Dadashpour, J. (2012). Anarchic society optimization based PID control of an
Automatic Voltage Regulator (AVR) system. Electrical and Electronic Engineering, 2(4),
199207.
Woodcook, G. (2004). Anarchism. Toronto: Higher Education Division, University of Toronto
Press. ISBN 13: 978-1551116297.
Chapter 5
Cuckoo Optimization Algorithm (COA)

Saba Jafari, Omid Bozorg-Haddad and Xuefeng Chu

Abstract The cuckoo optimization algorithm (COA) is used for continuous

non-linear optimization. COA is inspired by the life style of a family of birds called
cuckoo. These birds life style, egg laying features, and breeding are the basis of the
development of this optimization algorithm. Like other evolutionary approaches,
COA is started by an initial population. There are two types of the population of
cuckoos in different societies: mature cuckoos and eggs. The basis of the algorithm
is made by the attempt to survive. While competing for being survived, some of
them are demised. The survived cuckoos immigrate to better areas and start
reproducing and laying eggs. Finally, the survived cuckoos are converged in a way
that there is a cuckoo society with the same prot rate.

5.1 Introduction

Rajabioun (2011) proposed a new evolutionary algorithm called cuckoo opti-

mization algorithm (COA), which was inspired by the life of cuckoos. Kahramanli
(2012) developed a modied cuckoo optimization algorithm (MCOA) and used it to
solve two constrained continuous engineering optimization problems. The results
showed that MCOA was a powerful optimization method that yielded better
solutions to engineering problems. Rabiee and Sajedi (2013) used COA for solving

S. Jafari
O. Bozorg-Haddad (&)

Department of Irrigation and Reclamation Engineering, Faculty of Agricultural Engineering
and Technology, College of Agriculture and Natural Resources, University of Tehran, Karaj,
Tehran 31587-77871, Iran
e-mail: [email protected]
S. Jafari
e-mail: [email protected]
X. Chu
Department of Civil and Environmental Engineering, North Dakota State University,
Dept 2470, Fargo, ND 58108-6050, USA
e-mail: [email protected]

Springer Nature Singapore Pte Ltd. 2018 39

O. Bozorg-Haddad (ed.), Advanced Optimization by Nature-Inspired Algorithms,
Studies in Computational Intelligence 720, DOI 10.1007/978-981-10-5221-7_5
40 S. Jafari et al.

job scheduling in a grid computational design. The disadvantages of the evolu-

tionary techniques such as genetic algorithm (GA), simulated annealing (SA),
particle swarm optimization (PSO), and ant colony optimization (ACO) to solve the
problems in the grid schedule, are early convergence and trapping in local optima in
large-scale problems. Their results showed that the proposed method was more
efcient and had better performance than GA and PSO. Balochian and Ebrahimi
(2013) optimized parameters for Sugeno-type fuzzy logic controllers (S-FLCs) that
were applied for liquid level control. A programmable logic controller (PLC) was
used with fuzzy controller. The results indicated that the optimized FLC by COA
had better performance than the one with manual setting of the system parameters
for different datasets. Khajeh and Jahanbin (2014) developed a solid phase
extraction method using a new sorbent (zinc oxide nanoparticles-chitosan) for
pre-concentration and determination of uranium from water samples. A coupled
cuckoo optimization algorithmarticial neural network (COAANN) model was
developed for simulation and optimization. The optimum limit of detection and the
enrichment factor of uranium were 0.5 lg L1 and 125, respectively. Mellal and
Williams (2015a) studied the multipass turning process, one of the widely used
machining methods in manufacturing industry. They considered minimization of
the unit production cost as a key objective of the operation. In their work, cutting
parameters were optimized by using COA. The results showed that COA was better
than a wide range of other optimization algorithms. Singh and Rattan (2014)
employed the COA for the optimization of linear and non-uniform circular antenna
arrays. COA was used to ascertain a set of parameters of antenna elements that
provided a required radiation pattern. The effectiveness of COA for design of
antenna arrays was conrmed by their numerical results. The results showed the
superior performance of COA compared to other methods for designing linear and
circular antenna arrays. Shadkam and Bijari (2014) evaluated the performance of
COA with the Rastrigin function, a continuous non-linear function that was used for
evaluating optimization algorithms. The aforementioned function was solved with
articial bee colony (ABC) and the refly algorithm (FA). Comparison of the
results showed that COA had better performance than other algorithms.
Shokri-Ghaleh and Al (2014) designed an optimal controller for synchronization
of bilateral teleoperation systems with the aim of reducing factors such as time
delay in communication channels and modeling uncertainties. A novel
meta-heuristic algorithm named COA was used. Comparative simulations were
performed to determine the feasibility of the proposed control method. The results
showed that CO yielder better solutions to the problem than other algorithms
including biogeography-based optimization, imperialist competition, and articial
bee colony. Khajeh and Golzary (2014) developed the zinc oxide
nanoparticles-chitosan based on extraction of the solid phase for separation and
pre-concentration of a trace amount of methyl orange from water samples. The
COA-ANN model was used for optimization and simulation. The optimum limit of
detections factor of methyl orange was 0.7 lg L1. Mellal and Williams (2015b)
studied a complex engineering optimization problem called CHPED (combined
heat and power economic dispatch). The aim was to minimize the system
5 Cuckoo Optimization Algorithm (COA) 41

production costs by taking different constraints into consideration. In their study,

COA was employed by a penalty function (PFCOA) to solve the CHPED problem.
Two case studies of CHPED were presented and the results were compared with
those obtained by several other optimization methods, which proved the superior
performance of PFCOA. Moezi et al. (2015) determined the location and depth of a
crack by measuring its natural frequency changes and used the MCOA numerical
method for open edge-crack detection in an EulerBernoulli cantilever beam. The
crack was modeled by a torsional spring and the coefcient was calculated by using
the crack dimensions. The objective function was the weighted squared difference
between the calculated and measured natural frequencies. The results of the
numerical simulations and experimental tests showed the high accuracy in nding
the location and depth of the crack. Mellal and Williams (2016) considered
parameter optimization of advanced machining processes to produce complex
proles and high quality products. COA and the Hoopoe heuristic algorithm were
used for optimization of the parameters for two conventional machining processes
(drilling process and grinding process) and four advanced machining processes
(abrasive jet machining, abrasive water jet machining, ultrasonic machining, and
water jet machining). Finally, the results were compared with those from other
optimization algorithms.

5.2 Cuckoo Life Style

All the 9000 existing birds in the world have the same reproduction way: egg
laying. None of the birds give birth. They lay eggs and raise the baby birds outside
their bodies. The larger the eggs are, the less the probability is for a female bird to
have more than one egg in her body simultaneously, because on one hand, bigger
eggs make flying tough and require more energy to fly. On the other hand, eggs are
a rich source of protein for the predators, so it is necessary for birds to nd a secure
place for egg laying and hatching their eggs. Finding a secure place for egg laying,
hatching, and raising the birds until being independent of their parents is of vital
importance, which is intellectually solved by birds. They use an artistry and a
complicated engineering to do this. The variety of nest-making and the architecture
of the nests are absolutely stunning. Most birds make their nests segregated and
hidden to prevent being detected by the predators. Some of them hide their nests so
skillfully that human beings are not able to recognize and see them.
There are some birds that detached themselves from the challenge of
nest-making and use a cunning way to raise their families. These birds are called
Brood Parasites that never build a nest. They lay their eggs in other species nests
and wait for them to take care of their young. Cuckoo is the most famous Brood
Parasite that is an expert in deceiving cruelly. The strategy of cuckoos includes
their speed, being stealthy and surprising. A mother cuckoo destroys the hosts eggs
and lay her own eggs among others in the nest and flies away from the location fast
and lays caring on the host bird. This process is hardly more than 10 s. Cuckoos
42 S. Jafari et al.

make other nests parasitized by their eggs and mimic the color and the patterns of
existing eggs carefully so that new eggs in the nest look like the previous eggs.
Each female cuckoo is specialized on specic species of birds. This is one of the
main secrets of nature about how female cuckoos imitate a special kind of other
birds eggs accurately. Some of the birds recognize cuckoos eggs and sometimes
they even throw the eggs out of the nest. Some of them completely leave the nest
and build a new one. In fact, cuckoos continuously improve their mimicry from the
eggs in the target nests and host birds learn new ways to recognize the strange eggs
as well. This struggle for survival among different birds and cuckoos is a constant
and continuous process.
A suitable habitat for cuckoos should provide food sources (specially insects)
and locations for laying eggs, so the main necessity of brood parasites will be the
habitats of the host species. Cuckoos are found in a variety of places. Most of the
species are found in forests, especially evergreen rain forests. Some of the cuckoo
species select a wider range of places to live, which can even include dry areas and
deserts. Immigrant species select vast environments to make maximum misuse of
the host birds. Most of the cuckoo species are non-immigrant but there are several
ones that have seasonal immigration as well. Some of the species have partial
immigrations in their habitat range. Some species (e.g., channel-billed cuckoos)
have diurnal immigration, while others (e.g., yellow-billed cuckoo) have nocturnal
immigration. For those cuckoos that live in mountainous areas, availability of the
foods necessitates to immigrate to tropical areas. Long-tailed Koel cuckoos which
live and lay eggs in New Zealand, immigrate to Micrones, Melanesia, and
Polynesia in winters. Yellow-billed species and black-billed species that breed in
North America, pass the Caribbean Sea in a non-stop 4000-km flight. Other
long-distance immigrations include lesser cuckoos that fly over Indian Ocean from
India to Kenya (about 3000 km). Ten types of cuckoos perform polarized
intra-continental migration in a way that they spend non-breeding seasons in
tropical areas of the continent and then immigrate to dry and desert areas for egg
laying.
About 52 old species and 3 new species are brood parasite. They lay their eggs
in other birds nests. These species are obligate brood parasites since this is the only
way to their reproduction. Cuckoo eggs hatch earlier than their hosts eggs. In most
cases, a cuckoo chick throws the hosts eggs or the hosts chicks out of the nest.
This is completely instinctive and the cuckoo chick has no time to learn it.
A cuckoo chick makes the host provide a food suitable to its growth and beg for
food again and again. The cuckoo chick announces its need for food by an open
mouth because an open mouth to the mother is an indication for hunger.
Female cuckoos are skillful and expert in producing eggs similar to their hosts
eggs due to natural selection. Some birds recognize the eggs and throw them out
though. Parasite cuckoos are divided into different categories and each category is
expert in a special host. It is proved that the cuckoos in one category are genetically
different from those in another category. Specialization in imitating the hosts eggs
gradually improves and evolves.
5 Cuckoo Optimization Algorithm (COA) 43

5.3 Details of COA

Figure 5.1 shows the flowchart of COA. Like other evolutionary algorithms, COA
starts with an initial population (population of cuckoos). These cuckoos have got
some eggs that will be laid in other species nests. Some of these eggs that look like
the hosts eggs are more probable to be raised and turned into cuckoos. Other eggs
are detected by the host and are demised. The rate of the raised eggs shows the
suitability of the area. If there are more eggs to be survived in an area, there is more
prots to that area. Thus the situation in which more eggs are survived will be a
parameter for the cuckoos to be optimized.
Cuckoos search for the best area to maximize their eggs life lengths. After
hatching and turning into mature cuckoos, they form societies and communities.
Each community has its habitat to live. The best habitat of all communities will be
the next destination for cuckoos in other groups. All groups immigrate to the best
current existing area. Each group will be the resident in an area near the best current
existing area. An egg laying radius (ELR) will be calculated regarding the number
of eggs each cuckoo lays and its distance from the current optimized area.

Start

Find the radius of laying egg for each

Determine parameters and inputs
cuckoo

Put eggs in various nests

Move all cuckoos toward best local

Recognize some eggs and destroy them

Determine cuckoos societies

Destroy cuckoos in
No Whether the population is lower than max value?
unsuitable areas
Find the nests with best rate of life

yes

Calculate benefit (check probability of eggs life)

Whether stop condition is confirmed? No Grow eggs

yes

End

Fig. 5.1 Flowchart of the COA

44 S. Jafari et al.

Table 5.1 Characteristics of the COA

General algorithm Cuckoo optimization algorithm
Decision variable Cuckoo habitat
Solution Habitat
Old solution Old habitat
New solution New habitat
Best solution Habitat with best rate of life
Fitness function Distance between best habitat and recent habitat
Initial solution Random eggs for all cuckoos
Selection
Process of generating new solution Emigration cuckoos toward best area

Afterwards, cuckoos start laying eggs randomly in the nests within their egg
laying radii. This process continues until reaching the best place for egg laying
(a zone with the most prot). This optimized zone is the place in which the maximum
number of cuckoos gathers together. Table 5.1 lists the characteristics of the COA.

5.4 Cuckoos Initial Residence Locations

It is necessary to form variables as an array so that an optimization problem can be

solved. In GA and Particle Swarm Optimization, these arrays are identied by
chromosome and particles positions, but in COA, this array is called habitat.
In a one-dimensional Nvar optimization problem, habitat is a 1  Nvar array
that shows the current position of cuckoos life. It is dened as:

Habitat x1 ; x2 ; . . .; xNvar  5:1

The amount of prot or suitability rate for the current habitat can be obtained by
prot function evaluation. Thus

Porofit fp habitat fp x1 ; x2 ; . . .; xNvar 5:2

COA is an algorithm that maximizes the prot function. To use COA, the cost
function should be multiplied by a minus sign so that the problem could be solved.
To start optimization, a habitat matrix sized Npop  Nvar is generated.
Afterwards, a number of random eggs are specied for each habitat matrix. Each
cuckoo lays 520 eggs in nature. These numbers are used as the maximum and
minimum limits in the egg specication of each cuckoo in different iterations. Each
real cuckoo lays eggs in a specic range. Thus, the maximum range of egg laying is
5 Cuckoo Optimization Algorithm (COA) 45

the ELR. In an optimization problem, with the upper and lower limits of varhi and
varlow, each cuckoo has an ELR which is proportionate to the total number of eggs,
current number of eggs, and the upper/lower limits of variables of the problem.
So ELR is dened as (Rajabioun 2011):

Number of current cuckoos eggs

ELR a   varhi  varlow 5:3
Total number of eggs

in which a is a variable, by which the maximum ELR is set.

5.5 Cuckoos Egg Laying Approach

Each cuckoo randomly lays eggs in its host birds nest within its ELR. Figure 5.2
shows the egg laying radius or the maximum range of egg laying.
After all of the cuckoos lay their eggs, some of the eggs which are less similar to
the hosts eggs are recognized and thrown out. Thus after each egg laying process, p
% of all eggs (usually 10%) whose prot function value is less will be destroyed.
The rest of chicks in the hosts nest are fed and raised.
Another interesting point about cuckoo chicks is that only one egg has the
opportunity to be raised in each nest. Because when cuckoo chicks hatch, they
throw out the hosts eggs. If the hosts chicks hatch earlier, the cuckoo chick has
eaten the largest amount of food (because its body is three times larger and it knocks
other chicks over) and after several days the hosts own chicks will die from hunger
and only cuckoo chick will survive.

Fig. 5.2 Random egg laying

in ELR (the black circle is the
cuckoos initial habitat with
three eggs; and the white
circles are the eggs at new
positions)
46 S. Jafari et al.

5.6 Cuckoos Immigration

When cuckoo chicks grow up and become mature, they live in the surrounding
environment and in their communities for a while. But when the egg laying time is
close, they immigrate to better habitats in which the chances for the survival of their
eggs are higher. After forming cuckoo groups in various environments (search
space of the problem), the group with the best position will be selected as the target
group for other cuckoos for immigration.
It is difcult to recognize which group each cuckoo belongs when mature
cuckoos live in several environment zones. To solve this problem, classication of
cuckoos is done by K-means clustering (a number of K between 3 and 5 sufces).
The average prot of a group is calculated after all groups are formed to obtain
the relative optimality of the living area of each group. Afterwards, the group with
the highest value of average optimization will be selected as the target group and all
others will immigrate toward this group. While immigrating toward the target point,
cuckoos will not fly the whole way to the target place. They just pass a portion of
the distance and they may even digress from the target too. This movement is
shown in Fig. 5.3.
As shown in Fig. 5.3, each cuckoo only flies k% of the entire distance toward
the current ideal target and has a deviation of u too. These two parameters help
cuckoos search more space. k is a random number between 0 and 1 and u is a
number from p/6 to p/6. When all cuckoos immigrate to the target point and their
habitat points are determined, each cuckoo has a number of eggs. An ELR is
determined for each cuckoo and then egg laying is started.

Fig. 5.3 Immigration of a sample cuckoo to the target habitat

5 Cuckoo Optimization Algorithm (COA) 47

5.7 Demising Cuckoos Laid in Inappropriate Positions

According to the fact that there is always a balance among the populations of birds
in nature, a number Nmax is used to control the maximum number of cuckoos that
can live in a place. This balance is due to competing for limited foods, being hunted
by predators, and nding improper nests for eggs.

5.8 Pseudo Code for COA

After several repetitions, all cuckoos will attain a point of optimization with the
maximum similarity of their eggs to the hosts eggs and with the maximum food
sources. This location has the most total prot and the least chance for the eggs to
be ruined. Convergence of more than 95% of all cuckoos toward one point will
nalize the optimization process. The main steps of COA are shown in the fol-
lowing pseudo code:
48 S. Jafari et al.

5.9 Capabilities of COA

The capabilities of COA in different elds can be summarized as follows:

1. Solving complicated non-linear optimization problems accurately;
2. Teaching articial neuron networks, considering the fact that other approaches
of optimization cannot provide an assured optimized solution due to the high
number of parameters;
3. Using for easy and assured design of PID controllers for MIMO systems;
4. Finding the balance point in games quickly;
5. Optimizing the antenna design;
6. Optimizing the segmentation of pictures; and
7. Using for optimization problems that can be formulated as a target function.

5.10 Conclusion

The cuckoo optimization algorithm inspired by the life style of cuckoos was
explained. The cuckoos specic and unique feature in egg laying and raising is the
basis of this algorithm. In COA, each cuckoo has a habitat in which eggs are laid. If
the eggs survive, they are raised and become mature. Afterwards, they immigrate to
the best habitat found for reproduction. The variety associated with cuckoos
movement toward the target habitat, provides more space for search. This algorithm
is considered as a successful imitation of nature and is suitable for optimization
problems in different elds.

References

Balochian, S., & Ebrahimi, E. (2013). Parameter optimization via cuckoo optimization algorithm
of fuzzy controller for liquid level control. Journal of Engineering, 2013.
Kahramanli, H. (2012). A modied cuckoo optimization algorithm for engineering optimization.
International Journal of Future Computer and Communication, 1(2), 199.
Khajeh, M., & Golzary, A. R. (2014). Synthesis of zinc oxide nanoparticleschitosan for
extraction of methyl orange from water samples: Cuckoo optimization algorithmarticial
neural network. Spectrochimica Acta Part A, 131, 189194.
Khajeh, M., & Jahanbin, E. (2014). Application of cuckoo optimization algorithmarticial neural
network method of zinc oxide nanoparticleschitosan for extraction of uranium from water
samples. Chemometrics and Intelligent Laboratory Systems, 135, 7075.
Mellal, M. A., & Williams, E. J. (2015a). Cuckoo optimization algorithm for unit production cost
in multi-pass turning operations. The International Journal of Advanced Manufacturing
Technology, 76(14), 647656.
Mellal, M. A., & Williams, E. J. (2015b). Cuckoo optimization algorithm with penalty function for
combined heat and power economic dispatch problem. Energy, 93, 17111718.
5 Cuckoo Optimization Algorithm (COA) 49

Mellal, M. A., & Williams, E. J. (2016). Parameter optimization of advanced machining processes
using cuckoo optimization algorithm and hoopoe heuristic. Journal of Intelligent
Manufacturing, 27(5), 927942.
Moezi, S. A., Zakeri, E., Zare, A., & Nedaei, M. (2015). On the application of modied cuckoo
optimization algorithm to the crack detection problem of cantilever Euler-Bernoulli beam.
Computers & Structures, 157, 4250.
Rabiee, M. and Sajedi, H. (2013). Job scheduling in grid computing with cuckoo optimization
algorithm. International Journal of Computer Applications, 62(16).
Rajabioun, R. (2011). Cuckoo optimization algorithm. Elsevier, 11(8), 55085518.
Singh, U., & Rattan, M. (2014). Design of linear and circular antenna arrays using cuckoo
optimization algorithm. Progress in Electromagnetics Research C, 46, 111.
Shadkam, E., & Bijari, M. (2014). Evaluation the efciency of cuckoo optimization algorithm.
International Journal on Computational Sciences and Applications (IJCSA), 4, 3947.
Shokri-Ghaleh, H., & Al, A. (2014). Optimal synchronization of teleoperation systems via
cuckoo optimization algorithm. Nonlinear Dynamics, 78(4), 23592376.
Chapter 6
Teaching-Learning-Based Optimization
(TLBO) Algorithm

Parisa Sarzaeim, Omid Bozorg-Haddad and Xuefeng Chu

Abstract This chapter is prepared to describe the Teaching-Learning-Based

Optimization (TLBO) algorithm, a novel metaheuristic optimization method
inspired by an educational classroom environment. It has an interesting exclusivity
which may facilitate the solution process of optimization problems. In this chapter,
a brief literature review of the TLBO algorithm is rst presented. Then, the working
process and two phases of TLBO (teacher phase and learner phase) are depicted.
Eventually, a pseudocode of TLBO is presented.

6.1 Introduction

The Teaching-Learning-Based Optimization (TLBO) algorithm was rst proposed

by Rao et al. (2011). The elitist version of the TLBO algorithm, in which the worst
individuals were replaced by elite individuals for the next generation, was devel-
oped by Rao and Patel (2012). The modied version of TLBO for multi-objective
optimization problems was also developed by Rao and Patel (2013a, b). Despite the
fact that the TLBO algorithm is a new metaheuristic optimization method, it has
been applied to various engineering and science elds such as mechanical, civil,

P. Sarzaeim
O. Bozorg-Haddad (&)

Department of Irrigation and Reclamation Engineering, Faculty of Agricultural Engineering
and Technology, College of Agriculture and Natural Resources, University of Tehran, Karaj,
Tehran 31587-77871, Iran
e-mail: [email protected]
P. Sarzaeim
e-mail: [email protected]
X. Chu
Department of Civil and Environmental Engineering, North Dakota State University,
Dept 2470, Fargo, ND 58108-6050, USA
e-mail: [email protected]

Springer Nature Singapore Pte Ltd. 2018 51

O. Bozorg-Haddad (ed.), Advanced Optimization by Nature-Inspired Algorithms,
Studies in Computational Intelligence 720, DOI 10.1007/978-981-10-5221-7_6
52 P. Sarzaeim et al.

electrical, and environmental engineering. For example, Rao and Kalyankar (2012)
used the TLBO algorithm for mechanical design optimization problems. Toan
(2012) optimized the design of planar steel frames by using the TLBO algorithm.
Baghlani and Makiabadi (2013) used the TLBO algorithm to optimize the design of
truss structures. Garca and Mena (2013) presented an optimum design of dis-
tributed generation by using a modied version of the TLBO algorithm. Roy (2013)
and Roy et al. (2013) obtained an optimum solution to a hydrothermal scheduling
problem for hydropower plants by the TLBO algorithm. Sultana and Roy (2014)
applied the TLBO algorithm to minimize power loss and energy cost in power
distribution systems. Bouchekara et al. (2014) applied the TLBO algorithm to solve
the power flow problem. Ji et al. (2014) applied a modied TLBO algorithm to
improve the forecast accuracy of water supply system operation. Bayram et al.
(2015) used the TLBO algorithm to predict the concentrations of dissolved oxygen
in surface water. Thus, the TLBO algorithm has a variety of applications because it
is easy to use and convenient to be adapted for different problems.

6.2 Mapping a Classroom

into the Teaching-Learning-Based Optimization
Algorithm

All evolutionary and metaheuristic algorithms need some controlling parameters

which vary in different problems. These controlling parameters are divided into two
general groups: (1) general parameters such as population size, and number of
generations or number of iterations, and (2) specic parameters which depend on
the type of the algorithm. For instance, the Genetic Algorithm (GA) requires
crossover rate and mutation rate. In the GA algorithms, proper selection of specic
parameters has a signicant effect on their performance, computation time, and
modeling outputs. On the other hand, if the specic controlling parameters are
selected improperly, the solution may be stuck in a local optimum or the improper
parameters may lead to reduction of the solution quality. In such a situation, sen-
sitivity analysis should be performed by user to identify the best measures of the
specic parameters though it needs more time. Compared with other optimization
algorithms, the TLBO algorithm does not require any of such specic parameters. It
only requires general parameters such as population size and number of genera-
tions. It seems to be an interesting property which simplies the application of the
algorithm. Thus, TLBO algorithm is self-regulating.
The principle of the TLBO algorithm is inspired by the teacherstudents relation
in an educational classroom environment, the influence of teacher on learners or
students, and the interactions of learners and their effects on each other. Teacher and
learners are two main sectors of the algorithm which are named teacher phase and
learner phase, respectively.
6 Teaching-Learning-Based Optimization (TLBO) Algorithm 53

6.2.1 Teacher Phase

Imagine a classroom where there are two major groups: a teacher who teaches the
class and some students who are learning. The responsibility of the teacher is to
improve the knowledge level of the whole class to lead to better performance of
students in exams. In the teacher phase, the teacher tries to transfer the knowledge
to the learners. Thus, the teacher has a high knowledge level in the classroom and
endeavors to raise the level of class. Suppose that there are n students (j = 1, 2, ,
n) in a classroom whose average grade in exam i is Mi and the best learner who
achieves the best grade XT,i is supposed to be the teacher. The difference between
the classroom average grade (Mi) and the best grade (XT,i) can be computed by

Diff i ri XT;i  TF Mi 6:1

where Diff i = difference between the average grade and the best grade; ri = random
number in 0; 1 in iteration i; XT;i = grade of the best learner (teacher) in iteration i;
TF = teacher factor which depends on teaching quality and is either 1 or 2; and
Mi = average of learners grades in iteration i.

TF is also a random number which is given

TF round1 rand0; 1f2  1g 6:2

Then, by using Diff i , the new grade of student j in iteration i can be expressed as
0
Xj;i Xj;i Diffi 6:3

0
where Xj;i = new grade of student j in iteration i and Xj;i = old grade of student j in
0 0
iteration i. If Xj;i is better than Xj;i , Xj;i will go through to the learner phase.
Otherwise, Xj;i will go through to the learner phase.

6.2.2 Learner Phase

In a classroom, successful students try to help other students in order to increase

their level of knowledge. In other words, the students help each other in some ways,
for instance doing assignments in a group, to learn course materials better than the
teacher teaches them only. Suppose that two students, students A and B, are selected
randomly among the students in a classroom. The way that they help each other can
be expressed as
54 P. Sarzaeim et al.

start

Initializing randomly for population size (Xj)

Computing and evaluating the objective function

Teacher phase

Teacher selection and computation of Diff

Computing of X'j

Which one is better, Xj or X'j

Learner phase

Random selection of two students and computation

of X''j

Comparison X''j with before amount

Are stop criteria acceptable? No

Yes

Report the final population

Fig. 6.1 Flowchart of the TLBO algorithm

6 Teaching-Learning-Based Optimization (TLBO) Algorithm 55

 0 0 0 0 0
00 XA;i ri XA;i  XB;i if XA;i [ XB;i
XA;i 0 0 0 0 0 6:4
XA;i ri XB;i  XA;i if XB;i [ XA;i

00 0 00 0
If XA;i is better than XA;i , XA;i will go through the next iteration. Otherwise, XA;i
will go through the next iteration.
The steps of the working method of TLBO are summarized as follows:
(1) Initialize the grades (Xj;i ) of n students (population size) randomly for iteration i.
(2) Evaluate the objective function for n students.
(3) Select the best objective function (teacher) and compute Diff i by using
Eq. (6.1).
0
(4) Compute Xj;i for n students and iteration i by using Eq. (6.3).
0
(5) Compare Xj;i with Xj;i . The better one goes to next step and the other one is
removed.
00
(6) Select pairs of the students randomly and compare with each other. Then Xj;i is
computed for every student by using Eq. (6.4).
00 0
(7) Compare Xj;i with Xj;i . The better one goes to next step and the other one is
removed.
(8) Evaluate the objective function for n students. Check if the stop criteria are
satised. If yes, the best solution is achieved; otherwise, go back to Step (3).
The flowchart of the TLBO algorithm is shown in Fig. 6.1 to comprehend the
working process of TLBO better and the algorithms characteristics are shown in
Table 6.1.

Table 6.1 Characteristics of the TLBO algorithm

General algorithm Teaching-learning-based optimization
Decision variable Grades of learners
Solution Evaluating the grades of learners
Old solution Old grades of learners
New solution New grades of learners
Best solution Best grade of learners
Fitness function The grade is better than the one before
Initial solution Randomly selected grades
Selection
Process of generating new solution Comparison of the old and new grades of learners
56 P. Sarzaeim et al.

6.3 Pseudo Code of the TLBO Algorithm

6 Teaching-Learning-Based Optimization (TLBO) Algorithm 57

6.4 Conclusion

In this chapter, the teaching-learning-based optimization algorithm, a novel meta-

heuristic optimization method, was described. First, a brief literature review of
development and applications of the TLBO algorithm was presented. Then, fun-
damental details on the TLBO algorithm, including the two main phases (teacher
phase and learner phase), all computational steps, and the optimization process
were described. Finally, a pseudocode of TLBO was presented to demonstrate the
implementation of this optimization approach. As aforementioned, the TLBO
algorithm is a new user-friendly optimization algorithm that can be applied in
different elds. As aforementioned, the TLBO algorithm does not require any
specic parameters, and thus it is a valuable method that can achieve the nal
optimization solution, without the need to specify any parameters.

References

Baghlani, A., & Makiabadi, M. H. (2013). Teaching-learning-based optimization algorithm for

shape and size optimization of truss structures with dynamic frequency constraints. IJST,
Transactions of Civil Engineering, 37, 409421.
Bayram, A., Uzlu, E., Kankal, M., & Dede, T. (2015). Modeling stream dissolved oxygen
concentration using teaching-learning based optimization algorithm. Environmental Earth
Sciences, 73(10), 65656576.
Bouchekara, H. R. E. H., Abido, M. A., & Boucherma, M. (2014). Optimal power flow using
teaching-learning-based optimization technique. Electric Power Systems Research, 114, 4959.
Garca, J. A. M., & Mena, A. J. G. (2013). Optimal distributed generation location and size using a
modied teaching-learning based optimization algorithm. Electrical Power and Energy
Systems, 50, 6575.
Ji, G., Wang, J., Ge, Y., & Liu, H. (2014). Urban water demand forecasting by LS-SVM with
tuning based on elitist teaching-learning-based optimization. In Proceeding of 26th Chinese
Control and Decision Conference, Changsha, China, May 31June 2.
Rao, R. V., Savsani, V. J., & Vakharia, D. P. (2011). Teaching-learning-based optimization: A
novel method for constrained mechanical design optimization problems. Computer-Aided
Design, 43(3), 303315.
Rao, R. V., & Patel, V. (2012). An elitist teaching-learners-based optimization algorithm for
solving complex constrained optimization problems. International Journal of Industrial
Computations, 3(4), 535560.
Rao, R. V., & Patel, V. (2013a). Multi-objective optimization of heat exchangers using a modied
teaching-learning-based optimization algorithm. Applied Mathematical Modelling, 37(3),
11471162.
Rao, R. V., & Patel, V. (2013b). Multi-objective optimization of two stage thermoelectric cooler
using a modied teaching-learning-based optimization algorithm. Engineering Applications of
Articial Intelligence, 26(1), 430445.
Rao, R. V., & Kalyankar, V. D. (2012). Parameters optimization of machining process using a new
optimization algorithm. Materials and Manufacturing Processes, 27(9), 978985.
Roy, P. K. (2013). Teaching learning based optimization for short-term hydrothermal scheduling
problem considering valve point effect. Electrical Power and Energy systems, 53, 1019.
58 P. Sarzaeim et al.

Roy, P. K., Sur, A., & Pradhan, D. K. (2013). Optimal short-term hydro-thermal scheduling using
quasi-oppositional teaching learning based optimization. Engineering Applications of Articial
Intelligence, 26(10), 25162524.
Sultana, S., & Roy, P. K. (2014). Optimal capacitor placement in radial distribution systems using
teaching learning based optimization. Electrical Power and energy systems, 54, 387398.
Toan, V. (2012). Design of planar steel frames using teaching-learning based optimization.
Engineering Structures, 34, 225232.
Chapter 7
Flower Pollination Algorithm (FPA)

Marzie Azad, Omid Bozorg-Haddad and Xuefeng Chu

Abstract This chapter is designed to describe the flower pollination algorithm

(FPA) which is a new metaheuristic algorithm. First, the FPA applications in dif-
ferent problems are summarized. Then, the natural pollination process and the
flower pollination algorithm are described. Finally, a pseudocode of the FPA is
presented.

7.1 Introduction

The flower pollination algorithm (FPA) was proposed by Yang (2012) for global
optimization. This new metaheuristic algorithm is inspired by the pollination
phenomenon of flowing plants in nature. Yang et al. (2013) used the eagle strategy
with FPA to balance exploration and exploitation. Sharawi et al. (2014) employed
FPA in a wireless sensor network for efcient selection cluster heads and compared
with the Low-Energy Adaptive Clustering Hierarchy (LEACH). The results indi-
cated that FPA outperformed the LEACH. Sakib et al. (2014) used FPA and the bat

M. Azad
O. Bozorg-Haddad (&)

Department of Irrigation and Reclamation Engineering, Faculty of Agricultural Engineering
and Technology, College of Agriculture and Natural Resources, University of Tehran,
31587-77871 Karaj, Tehran, Iran
e-mail: [email protected]
M. Azad
e-mail: [email protected]
X. Chu
Department of Civil and Environmental Engineering, North Dakota State University,
Dept 2470, Fargo, ND 58108-6050, USA
e-mail: [email protected]

Springer Nature Singapore Pte Ltd. 2018 59

O. Bozorg-Haddad (ed.), Advanced Optimization by Nature-Inspired Algorithms,
Studies in Computational Intelligence 720, DOI 10.1007/978-981-10-5221-7_7
60 M. Azad et al.

algorithm (BA) to solve continuous optimization problems. They tested and com-
pared the two algorithms on the benchmark functions. Emary et al. (2014) applied
FPA to a retinal vessel segmentation optimization problem. Platt (2014) used FPA
in the calculation of dew point pressure in a system that exhibited double retrograde
vaporization. El-henawy and Ismail (2014) combined FPA with the particle swarm
optimization (PSO) algorithm to solve large integer programming problems and
demonstrated that FPA was useful for betterment searching accuracy. Abdel-Raouf
et al. (2014) formulated Sudoku puzzles as an optimization problem, and then
employed a hybrid optimization method, flower pollination algorithm with the
Chaotic Harmony Search (FPCHS) to obtain the optimal solutions. Yang et al.
(2014) used a novel version of FPA to solve several multi-objective test functions.
Trivedi et al. (2015) used FPA for optimization of relay coordination in a wide
electrical network with the aim of increasing the selectivity and at the same time
reducing the fault clearing time to improve reliability of the system. Bensouyad and
Saidouni (2015) applied the discrete flower pollination algorithm (DFPA) for
solving a graph coloring problem. Lukasik and Kowalski (2015) tested FPA for a
number of continuous benchmark problems. Dubey et al. (2015) applied a modied
flower pollination algorithm in the modern power systems to nd out the solutions
to economic dispatch problems solutions. They added a scaling factor to control the
local pollination and compression of the exploitation stage to achieve the best
solution. Bibiks et al. (2015) used DFPA in order to solve combinatorial opti-
mization problems. Alam et al. (2015) applied the FPA technique for determining
optimal parameters of a single diode and two diodes that were used to describe
photovoltaic systems. In the design of a structural system, the optimal values of
design variables cannot be obtained analytically and a structural engineering
problem has different design constraints, so optimization is an important part of the
structural design process. For this purpose Nigdeli et al. (2016) used FPA to solve
structural engineering problems related to pin-jointed plane frames, truss systems,
deflection minimization of I-beams, tubular columns, and cantilever beams. Nabil
(2016) developed a Modied Flower Pollination Algorithm (MFPA) from the
hybridization FPA with the Clonal Selection Algorithm (CSA) and performed tests
on 23 optimization benchmark problems to investigate the efciency of the new
algorithm. Then, the results of MFPA were compared with those of Simulated
Annealing (SA), Genetic Algorithm (GA), FPA, Bat Algorithm (BA), and Firefly
Algorithm (FA). The results showed that the proposed MFPA was able to nd more
accurate solutions than FPA and the four other algorithms. Abdelaziz et al. (2016)
applied FPA to drive the optimal sizing and allocations of the capacitors in different
water distribution systems.
7 Flower Pollination Algorithm (FPA) 61

7.2 Flower Pollination Process

Pollination is a natural mechanism for the reproduction of flowering plants and is

dened as transfer of pollen from one flower flag to the pistil stigma of the same
flower or another flower of the same plant species. Pollen has both vegetative and
reproductive cells. After sitting pollen on the pistil stigma, the vegetative cell
multiplies and forms a pollen tube. A reproductive cell can be divided into two cells
along its patch, reaching the ovary by a pollen tube. One of the reproductive cells is
fertilized by an egg cell, forming a zygote. Thus, a new plant forms whit growth
zygote. There are two types of pollination according to pollen transfer methods:
(1) biotic pollination, and (2) abiotic pollination. For most flowering plants, biotic
pollination is done by pollinators such as insects or animals. But abiotic pollination
does not require the transfer of pollen by living organisms. Instead, it is done by
water, wind, or gravity as pollinators. When pollens are delivered from one plant to
another of the same type, such pollination is called cross-pollination and
self-pollination occurs when pollen is delivered to the same flower or flowers of the
same plant.
Almost 90% of flowering plants have biotic pollination in which the pollens are
transferred by pollinators such as insects or animals. Pollination by insects is more
relevant among plants. Flowers that pollinated with aim of insects attract insects by
their colors, odors, and nectars. Generally, the size of flowers is consistent with the
insects bodies so that the insects can enter the flowers and their bodies are in
contact with the pollen and pistil. Approximately 10% of flowering plants perform
abiotic pollination that does not need pollinators and the pollens are transferred by
wind or water. Most of the bush plants and trees do not need insects for pollination.
Pollens of these plants are spread in air and transferred by wind. Although most of
the pollens are being lost, some of them can be trapped by ripen stigma of female
flowers. Wind pollination takes place in the plants that have female and male
flowers and these flowers exist in separate trees. In such flowers, their pollens are
released by shaking the stamens by wind. Numerous ne pollens of these flowers
can travel long distances with wind. The stigma of a flower has feathery rami-
cations and is outside of the flower, which increases the chance of the stigma to trap
pollen transferred by wind. For the flowers that are pollinated by wind, their
petioles are absent or very small and they have no nectar. Pollinator insects are
often associated with a specic flower type, which is dened as flower constancy.
That is, the pollinators tend to sit on certain flower species. Therefore, flower
constancy helps quantify the cost of searching for each of pollinators. For biotic
pollination, the pollinators such as flies, birds, and bats can fly long distances. Thus,
they can be considered as global pollination. Likewise, step-jump or flying of birds
or bees can be described as levy flight.
62 M. Azad et al.

7.3 Flower Pollination Algorithm

The biotic pollination, cross-pollination, abiotic pollination and self-pollination

strategies are dened in domain optimization and embedded in the flower polli-
nation algorithm. The pollination process includes a series of complex mechanisms
in plant production strategies. A flower and its pollen gametes form a solution of the
optimization problem. Flower constancy as a tted solution is perceptible. In global
pollination, the pollinators transfer pollen in long distances towards high tting. On
the other hand, local pollination within a limited area of a unique flower takes place
under shading by wind or water. Global pollination occurs with a probability that is
called switch probability. If this step is removed, local pollination replaces it. In the
FPA algorithm four rules are followed: (1) live pollination and cross-pollination are
considered as global pollination and the carriers or pollen pollinators move in a way
that follows levy ght; (2) abiotic and self-pollination are considered as local
pollination; (3) pollinators including insects can develop flower constancy. Flower
constancy is production probability that is proportional to the similarity of two
involved flowers; and (4) the interaction of global and local pollination can be
controlled by switch probability. The rst and third rules can be expressed as (Yang
2012):

xit 1 xti c  Lk  g  xti 7:1

where xti = pollen or solution vector at iteration t; g = the current best solution
among all current generation solutions; c = a scale factor for controlling step size;
and L = strength of pollination, which is a step size related to the levy distribution.
Levy flight is a bunch of random processes where the length of each jump follows
the levy probability distribution function and has innite variance. Following Yang
(2012), L for a levy distribution is given by:

k  Ck  sin pk 1
L 2
 1 k S  S0 0; 7:2
p S

where Ck = standard gamma function.

For local pollination, the second and third rules are given by:
 
xit 1 xti e xtj  xtk 7:3
7 Flower Pollination Algorithm (FPA) 63

Table 7.1 Characteristics of the FPA

General algorithm Flower pollination algorithm
Decision variable Flowers or pollen gametes in each dimension
Solution Flowers or pollen gametes
Old solution Old flower or old pollen gamete
New solution New flower or new pollen gamete
Best solution Current best solution
Fitness function
Initial solution Selection Random selection
Process of generating new solution Flying and local random walk

where xtj and xtk = two pollens from different flowers of the same plant.
Mathematically, if xtj and xtk come from the same species or are selected from the
same population, this becomes a local random walk if has a uniform distribution
in [0,1].
Table 7.1 lists the characteristics of the FPA and Fig. 7.1 shows the flowchart of
the FPA.

7.4 User-Dened Parameters of the FPA

The size of the population of solutions (n), the scale factor for controlling step size
c, the levy distribution parameter Lk, and the switch probability (P) are
user-dened parameters in the FPA. Determining the optimal parameters of the
FPA is a time-consuming work and needs to run the algorithm many times. It
should be noted that the optimal parameters of the algorithm for one problem are
different from those of other problems. Considering a mixture of parameters is an
appropriate method for nding the suitable values of the algorithm parameters. The
algorithm can be done for several times for one mixture of parameters, and the
similar process can be repeated for other mixtures of parameters. Finally, the results
for different sets of parameters can be compared and the best value can be deter-
mined. Yang (2012) suggested to start the modeling with P = 0.5 and k = 1.5.
64 M. Azad et al.

start

Determination parameters FPA: the size initial population and the

maximum number of iteration and the amount of p

Construction initial population with random solution

Calculation the objective function corresponding to random

solutions and selection the best solution in current population

For each iteration t

No Yes
Rand > p

Do global pollination using the

Do local pollination Levy distribution and create a
new population

Calculate the objective function corresponding to new solutions

Update the best current solution

No
t > maximum number of
t=t+1
iteration

Yes

Display the best solution

End

Fig. 7.1 Flowchart of the FPA

7 Flower Pollination Algorithm (FPA) 65

7.5 Pseudo Code of FPA

66 M. Azad et al.

7.6 Conclusion

This chapter described the flower pollination algorithm (FPA), which is based on
the pollination phenomenon of flowing plants in nature. The chapter presented a
summary of the FPA applications in different problems. Then, the natural process of
pollination, and the flower pollination algorithm and a pseudocode of the FPA are
presented.

References

Abdelaziz, A. Y., Ali, E. S., & Abd Elazim, S. M. (2016). Optimal sizing and locations of
capacitors in radial distribution systems via flower pollination optimization algorithm and
power loss index. Engineering Science and Technology, 19(1), 610618.
Abdel-Raouf, O., & Abdel-Baset, M. (2014). A new hybrid flower pollination algorithm for
solving constrained global optimization problems. International Journal of Applied
Operational Research-An Open Access Journal, 4(2), 113.
Abdel-Raouf, O., El-Henawy, I., & Abdel-Baset, M. (2014). A novel hybrid flower pollination
algorithm with chaotic harmony search for solving sudoku puzzles. International Journal of
Modern Education and Computer Science, 6(3), 38.
Alam, D. F., Yousri, D. A., & Eteiba, M. B. (2015). Flower pollination algorithm based solar PV
parameter estimation. Energy Conversion and Management, 101, 410422.
Bekda, G., Nigdeli, S. M., & Yang, X. S. (2015). Sizing optimization of truss structures using
flower pollination algorithm. Applied Soft Computing, 37, 322331.
Bibiks, K., Li, J. P., & Hu, F. (2015). Discrete flower pollination algorithm for resource
constrained project scheduling problem. International Journal of Computer Science and
Information Security, 13(7), 8.
Dubey, H. M., Pandit, M., & Panigrahi, B. K. (2015). A biologically inspired modied flower
pollination algorithm for solving economic dispatch problems in modern power systems.
Cognitive Computation, 7(5), 594608.
El-henawy, I., & Ismail, M. (2014). An improved chaotic flower pollination algorithm for solving
large integer programming problems. International Journal of Digital Content Technology and
its Applications, 8(3).
Emary, E., Zawbaa, H. M., Hassanien, A. E., Tolba, M. F., & Snel, V. (2014). Retinal vessel
segmentation based on flower pollination search algorithm. In Proceedings of the Fifth
International Conference on Innovations in Bio-Inspired Computing and Applications IBICA,
2014 (pp. 93100). Springer International Publishing.
ukasik, S., & Kowalski, P. A. (2015). Study of flower pollination algorithm for continuous
optimization. In Intelligent Systems, 2014 (pp. 451459). Springer International Publishing.
Nabil, E. (2016). A modied flower pollination algorithm for global optimization. Expert Systems
with Applications, 57, 192203.
Nigdeli, S. M., Bekda, G., & Yang, X. S. (2016). Application of the flower pollination algorithm
in structural engineering. In Metaheuristics and optimization in civil engineering (pp. 2542).
Springer International Publishing.
Platt, G. M. (2014). Computational experiments with flower pollination algorithm in the
calculation of double retrograde dew points. International Review of Chemical Engineering, 6
(2), 9599.
Sakib, N., Kabir, M. W. U., Subbir, M., & Alam, S. (2014). A comparative study of flower
pollination algorithm and bat algorithm on continuous optimization problems. International
Journal of Soft Computing and Engineering, 4(2014), 1319.
7 Flower Pollination Algorithm (FPA) 67

Sharawi, M., Emary, E., Saroit, I. A., & El-Mahdy, H. (2014). Flower pollination optimization
algorithm for wireless sensor network lifetime global optimization. International Journal of
Soft Computing and Engineering, 4(3), 5459.
Trivedi, I. N., Purani, S. V., & Jangir, P. K. (2015). Optimized over-current relay coordination
using Flower Pollination Algorithm. In Advance Computing Conference (IACC), 2015 IEEE
International (pp. 7277). IEEE.
Yang, X. S. (2012). Flower pollination algorithm for global optimization. In International
Conference on Unconventional Computing and Natural Computation (pp. 240249). Berlin:
Springer.
Yang, X. S., Karamanoglu, M., & He, X. (2014). Flower pollination algorithm: A novel approach
for multiobjective optimization. Engineering Optimization, 46(9), 12221237.
Yang, X. S., Deb, S., & He, X. (2013). Eagle strategy with flower algorithm. In 2013 International
Conference on Advances in Computing, Communications and Informatics (ICACCI)
(pp. 12131217). IEEE.
Chapter 8
Krill Herd Algorithm (KHA)

Babak Zolghadr-Asli , Omid Bozorg-Haddad and Xuefeng Chu

Abstract The krill herd algorithm (KHA) is a new metaheuristic search algorithm
based on simulating the herding behavior of krill individuals using a Lagrangian
model. This algorithm was developed by Gandomi and Alavi (2012) and the pre-
liminary studies illustrated its potential in solving numerous complex engineering
optimization problems. In this chapter, the natural process behind a standard KHA
is described.

8.1 Introduction

In the past decades, metaheuristic optimization techniques have been widely

employed in many elds to solve complex optimization problems, due to their
advantages over the conventional optimization techniques. Generally, these algo-
rithms have two main features: (1) intensication and (2) diversication (Gandomi
et al. 2013c). The former denotes searching through the current candidates for the
optimal solution, while the latter indicates expanding the searching horizon to
ensure that the nal result is a global optimum, instead of a local one. Each newly
proposed algorithm attempts to improve these two main features, either by
decreasing the distance of the reported solutions and the actual global optima or by
reducing the solution searching time.

B. Zolghadr-Asli
O. Bozorg-Haddad (&)

Department of Irrigation and Reclamation Engineering, Faculty of Agricultural Engineering
and Technology, College of Agriculture and Natural Resources, University of Tehran,
3158777871 Karaj, Iran
e-mail: [email protected]
B. Zolghadr-Asli
e-mail: [email protected]
X. Chu
Department of Civil and Environmental Engineering, North Dakota State University,
Dept 2470, Fargo, ND 58108-6050, USA
e-mail: [email protected]

Springer Nature Singapore Pte Ltd. 2018 69

O. Bozorg-Haddad (ed.), Advanced Optimization by Nature-Inspired Algorithms,
Studies in Computational Intelligence 720, DOI 10.1007/978-981-10-5221-7_8
70 B. Zolghadr-Asli et al.

Although the basic principles of these algorithms are similar and contain an
iterative mechanism, the iteration process differs in each algorithm. The main
objective of such a process is to search through the decision space for arrays of
decision variables that produce an optimum result. This process is usually inspired
by the natural phenomena, and is intended to imitate a natural feature that has been
evolved over millions of years (Gandomi and Alavi 2012). Consequently, there is
no limitation to the source of inspiration for these bio-inspired algorithms, and they
can imitate a vast domain of features, from the genetic evolution process of a
species to the foraging mechanism of bacteria. Swarm intelligence, which is an
imitation of an animal groups behavior, could serve as an inspiration source to
develop such algorithms.
Many studies have focused on capturing the underlying mechanism that governs
the development of formation grouping of various species of marine animals,
including the Antarctic krill (Flierl et al. 1999). The krill herds are aggregations
with no parallel orientation in both temporal and spatial scales (Brierley and Cox
2010). These creatures that can form large swamps are the source of inspiration for
the krill herd algorithm (KHA). The herding of the krill individuals is a
multi-objective process, including two main goals: (1) increasing krill density, and
(2) reaching the food. Density-dependent attraction of krill (increasing density) and
nding food (areas of high food concentration) are used as objectives, which nally
cause krill to herd around the global optima. In this process, an individual krill
moves toward the best solution when it searches for the highest density and food.
The imaginary distance of krill individuals serve as objective functions, and min-
imizing them is the priority of the optimization process. Hence the closer the
distance to the high density and food, the less the objective function (better)
(Gandomi and Alavi 2012).
The engineering optimization problems, which mostly have a nonlinear decision
space, are complicated, due to their numerous decision variables and complex
constraints. Such conditions can be regarded as an advantage for the metaheuristic
algorithms over the conventional optimization techniques. KHA is a new and novel
metaheuristic search algorithm based on the herding behavior of krill individuals,
using a Lagrangian model at its core. This algorithm was rst introduced by
Gandomi and Alavi (2012), and the preliminary studies have demonstrated its
potential to outperform the existing algorithms for solving the complicated engi-
neering problems (Gandomi et al. 2013a, b). Additionally, KHA was further vali-
dated for various engineering problems, including optimal design of civil structures
(Gandomi et al. 2013a, b; Gandomi and Alavi 2016), power flow optimization
(Mukherjee and Mukherjee 2015), and optimum operation of power plants (Mandal
et al. 2014). Similarly, studies have illustrated that the power of the classical KHA
tends towards global exploration (Bolaji et al. 2016). Some modications have been
made to the standard KHA, and the modied algorithms include: chaotic-particle
swarm krill herd (CPKH) (Wang et al. 2013), fuzzy krill herd algorithm
(FKH) (Fattahi et al. 2016), and discrete-based krill herd algorithm (DKH) (Bolaji
et al. 2016). As a compatible and efcient algorithm, KHA can be a promising
alternative for solving engineering optimization problems.
8 Krill Herd Algorithm (KHA) 71

8.2 Krill Swarms Herding Pattern

The basic core of a standard KHA is its krill herding simulator. The krill herd
defuses after a hypothetical attack from a predator. This is the initial step in the
standard KHA. Each krill after such an event has two priorities, which are
decreasing its distance from both the food source and the highest density of the krill
swarm. These imaginary distances are acting as the objective function, and mini-
mizing these distances is considered as the goal of each krill individual.
Consequently, the time-dependent position of an individual krill is governed by the
motion induced by other krill individuals (Ni), foraging motion (Fi), and physical
diffusion (Di). As any efcient optimization algorithm should be compatible with
arbitrary dimensions, since each arbitrary dimension is to represent a decision
variable, the following Lagrangian model is generalized for an n-dimensional
decision space (Gandomi and Alavi 2012):

dXi
Ni Fi Di 8:1
dt

in which Xi = location of the ith krill individual in the decision space.

Equation (8.1), which simulates the movement of each individual krill, implies
that the movement of each krill is affected by three factors: (1) the behavior of the
group, (2) the location of food, and (3) a random factor.

8.3 Motion Induced by the Krill Herd

Theoretically speaking, the krill herd has a tendency to move in a group. In other
words, forming a high-density swarm is considered as an advantage for the krill
community. Thus, krill individuals try to maintain a high density and move due to
their mutual effects. The motion induced by the krill herd can be expressed as
(Gandomi and Alavi 2012):

Ninew N max  ai xn  Niold 8:2

in which Ninew = motion of the ith krill individual induced by the krill herd at the
current iteration; Nmax = maximum induced speed (according to the measured
values, it is around 0.01 m/s) (Hofmann et al. 2004); ai = direction of motion of the
ith krill individual induced by the krill swamp; wn = inertia weight of the motion
induced in the range [0,1]; and Niold = motion of the ith krill individual induced by
the krill herd in the previous iteration. wn is one of the model parameters and it acts
as a weight for the previously calculated motion induced by the krill herd. A lower
value of wn decreases the influence of the Niold .
72 B. Zolghadr-Asli et al.

The direction of motion of the ith krill individual induced by the krill swamp
(ai), however, is influenced by both the nearby krill individuals (local effect) and the
target swarm density (target effect), and it is given by Gandomi and Alavi (2012):

ai alocal
i atarget
i 8:3

X
NN
alocal
i b i;j  X
K b i;j 8:4
j1

b i;j  Xj  X
X i 8:5
Xj  Xi  e

b i;j Ki  Kj
K 8:6
K worst  K best

in which X b i;j = local effect induced by the jth neighboring krill individual for the
ith krill individual; K b i;j = target direction effect provided by the best krill indi-
vidual; Ki and Kj = tness values of the ith and jth krill individuals, respectively;
Kbest and Kworst = best and worst tness values for krill individuals, respectively;
and NN = number of neighboring krill individuals for the ith krill.
Equation (8.5) represents the unit vector that connects the ith krill to the jth krill,
while Eq. (8.6) calculates the normalized tness value, which plays the role of a
weight for the unit vector in Eq. (8.5). In fact, each calculated X b i;j  K
b i;j char-
acterizes the effect
 of the jth neighboring krill. This influence can be (1) an
b
attractive one K i;j [ 0 , which indicates that both krill individuals are moving
 
toward one another; (2) a repulsive one K b i;j \0 , which refers to a situation
where bothkrill individuals are moving away from each other; and (3) a futile one

b i;j 0 , which suggests that both krill individuals are incurious toward one
K
another. The summation of these weighted vectors shows the influence of the
neighboring krill individuals on the motion induced by the ith krill.
To choose the number of neighboring krill individuals for any given krill, dif-
ferent strategies can be implemented. For instance, a neighboring ratio can be
simply dened to nd the number of the closest krill individuals. Reportedly, using
the actual behavior of the krill individuals suggests that a sensing distance (ds) is a
proper value to determine the neighboring krill individuals (Fig. 8.1). The sensing
distance for each krill individual in each iteration can be determined by Gandomi
and Alavi (2012):
1 XN  
Xi  Xj 
ds;i 8:7
5N j1
8 Krill Herd Algorithm (KHA) 73

Fig. 8.1 Schematic representation of the sensing ambit around a krill individual

in which N = number of krill individuals. It should be noted that this is merely a

suggestion and other techniques could be used, instead. In fact, the number of
neighboring krill individuals (NN) is also one of the algorithms parameters.
Equation (8.7) suggests that if the distance between two krill individuals is less
than the calculated ds, they should be considered as neighbors, and thus, they could
influence the movement of one another in a direct manner. An analytical review of
Eq. (8.7) also reveals that this formula has a tendency to calculate higher ratios of ds
for those krill individuals that are separated from the herd (placed in a low-density
position). In this way there is a better chance that the neighbors of such a krill could
make a greater impact on its motion so that it could decrease its distance from the
herd.
The known target vector of each krill swarm has the lowest tness of an indi-
vidual krill in that herd. The effect of the individual krill with the best tness on the
ith individual krill is taken into account using Eq. (8.8). Such an effect can increase
the chance of nding the global optima (Gandomi and Alavi 2012).

atarget
i
b i;best  X
C best  K b i;best 8:8

in which atarget
i = effective coefcient of the krill individuals with the best tness to
the ith krill individual. atarget
i leads the solution to the probable location of global
optima and hence it should be more effective than other krill individuals such as
neighboring krill individuals. Herein, Cbest in Eq. (8.8) is dened as (Gandomi and
Alavi 2012):
74 B. Zolghadr-Asli et al.

 
I
C best
2 rand 8:9
Imax

where rand = a random value in the range of [0,1] and it has a uniform distribution;
I = current iteration number; and Imax = maximum number of iterations.
Equation (8.9) suggests that the effect of the target krill is enhanced in each
iteration.

8.4 Foraging Motion

The foraging motion, which is centered around the krill herds tendency to nd
nutrition, has two terms in its structure: (1) the location of food, and (2) the previous
experiences about the food location encountered by each individual krill. The above
mechanism can be formulated for each individual krill as follows (Gandomi and
Alavi 2012):

Fi Vf  bi xf  Fiold 8:10

where

bi bfood
i bbest
i 8:11

in which Vf = foraging speed (according to the measured values, it is 0.02 m/s)

(Price 1989); wf = inertia for the foraging motion in the range of [0,1]; bfood
i = food
attraction parameter; and bbest
i = effect value of the best tness of the ith krill.
Equations (8.10) and (8.11) encourage each individual krill to decrease its dis-
tance from the location of food, which is the probable location of the global optima
in the decision space. However, the location of food and its attraction are, in fact,
the result of the optimization process. In each iteration, the location of food and its
attraction for the krill herd can only be estimated. While the estimation process can
also influence the KHA efcacy, the following formula can be adapted to
approximate the virtual center of the food concentration, using the tness distri-
bution of krill individuals. This approach, inspired by the center of the mass
notion, can be expressed as (Gandomi and Alavi 2012):

P
N
1
Ki  Xi
i1
X food
8:12
P
N
1
Ki
i1

in which Xfood = estimated location of the food.

8 Krill Herd Algorithm (KHA) 75

Thus, the attraction of the food is given by

bfood
i
b i;food  X
Cfood  K b i;food 8:13

in which Cfood = food concentration. To ensure that the food attraction is

decreasing for the krill herd during the iteration procedure, this food concentration
term is added to the food attraction calculation process, and it is given by Gandomi
and Alavi (2012):
 
I
Cfood 2 1  8:14
Imax

The main reason behind the food attraction is to ensure that the krill swarm nds
the global optima. As a result, when the krill herd is randomly spread through the
decision space, this motion can help the herd to gather around the plausible location
of the food (global optima). However, as the searching process advances in each
iteration, the herd must be able to spread in a limited space, to locate the best
solution. Thus, as shown in the formula, this motion decreases with time. This can
be considered as an efcient global optimization strategy that helps improve the
efciency of KHA.
Each individual krill is also moving due to its visited memory of the previously
spotted locations of food. The effect of the best tness of the ith krill individual can
also be expressed as (Gandomi and Alavi 2012):

bbest
i
b i;best  X
K b i;best 8:15

in which K(i,best) = best previously encountered position of the ith krill individual.

8.5 Physical Diffusion

The two mechanisms behind inducing motion to each individual krill (motion
induced by the krill herd and foraging motion) are to ensure that after the initial
separation of the krill herd throughout the decision space, the herd gathers around
what is considered to be the global optima. Yet, to ensure that the decision space is
inspected thoroughly by the krill herd, a random process is required to spread an
enough number of krill individuals in the decision space, in a random-based manner.
If the random process is too strong, the herd will not gather around a center location;
yet, lack of such a mechanism could interrupt a proper search throughout the
decision space. The physical diffusion term is introduced in the KHA as a random
process. This motion can be expressed in terms of maximum diffusion speed, a
random direction vector, and a mathematical mechanism to ensure the decreasing
effects of this term as searching for the global optimal solution continues. Thus, the
physical diffusion term can be formulated as follows (Gandomi and Alavi 2012):
76 B. Zolghadr-Asli et al.

 
I
Di D max
 1 d 8:16
Imax

in which Dmax = maximum diffusion speed, which has the range of [0.002, 0.010]
(m/s) (Gandomi and Alavi 2012); and d = random directional vector and its arrays
are random values between 1 and 1. A random selection can also be employed to
determine the value of Dmax. The physical diffusion motion introduced in Eq. (8.16)
works on the basis of a geometrical annealing schedule, and the random speed
linearly decreases with time.

8.6 Motion Process of the KHA

The above three mechanisms allow one to calculate the direction and speed of
relocation for each individual krill at any given iteration. In other words, the dened
motions frequently change the position of a krill individual toward the position that
is expected to be the best one. The motion induced by other krill individuals and the
foraging motion are working in parallel, which resultantly makes the KHA a
potentially powerful algorithm for solving complex optimization problems.
The KHA formulation suggests that if any of Kj, Kbest, Kfood, and Kifood can illustrate
a better performance than the ith krill individual, they can have an attractive effect,
which can inspire this krill to move toward any of these locations, in the hope that
such an action would improve its tness value. Such a mechanism can also have a
negative effect. The Kj, Kbest, Kfood, and Kifood can repulse the ith krill individual,
causing it to move away from the aforementioned locations. Additionally, the
physical diffusion can spread the krill herd throughout the decision space for a
comprehensive search of the plausible arrays of decision variables. After calculating
the motion for every krill in the herd, the position vector of the ith krill individual
during the time interval from t to t + Dt is given by

dXi
Xi t Dt Xi t Dt  8:17
dt

It should be noted that t is one of the most important model parameters since it
works as a scale factor for the speed vector. Thus, it should be carefully set for the
optimization problem. Suggestively, t, which completely depends on the search
space, can be estimated by Gandomi and Alavi (2012):

X
NV  
Dt Ct  UBj  LBj 8:18
j1

in which NV = number of variables; and LBj and UBj = lower and upper bounds of
the jth variable, respectively. It is empirically found that Ct is a constant within
8 Krill Herd Algorithm (KHA) 77

(0,2]. It is also obvious that the low values of Ct let the krill individual search the
space in a slower, yet more careful pace. One should bear in mind that, this
parameter is the most important parameter of the model, and thus, needs to be
carefully adapted to each optimization problem.
Finally, it should be pointed out that although the above-mentioned mechanisms
are the core concept of a standard KHA, this algorithm is compatible to implement
a few external searching operators, including but not limited to genetic operators

Fig. 8.2 Simplied flowchart of the KHA

Table 8.1 The characteristics of the KHA

General algorithm Krill herd algorithm
Decision variable Krill individuals position in each dimension
Solution Krill individuals position
Old solution Old position of the krill individual
New solution New position of the krill individual
Best solution Any krill with the best tness
Fitness function Distance between krill individual and food and the densest
location in the herd
Initial solution Randomly
Selection
Process of generating new Motion induced by krill herd, foraging activity, and physical
solution diffusion
78 B. Zolghadr-Asli et al.

such as crossover and mutation. While this capability surely enhances the perfor-
mance of the standard KHA, their presence is not obligatory (Gandomi and Alavi
2012). A basic representation of the KHA is shown in Fig. 8.2. Additionally,
Table 8.1 summarizes the characteristics of the standard KHA.

8.7 Pseudo Code of KHA

Begin

Define population size (N) and maximum iteration number (Imax)

Set the iteration counter I=1

Initialize the population by generating Xi for i = 1, 2, 3, N

Set the inertia weight of the motion induced (n) and foraging motion (f)

Define t

Evaluate each krill individual according to its position

While (the stop criterion is not satisfied or I < Imax )

For i = 1: N

Perform the following motion calculation:

Movement induced by other krill individuals

Foraging activity

Physical diffusion

Update the krill individual position in the search space

Evaluate each krill individual according to its position

End for i

Sort the population/krill from best to worst and find the current best

End while

Post-processing the results and visualization.

End
8 Krill Herd Algorithm (KHA) 79

8.8 Conclusion

This chapter described the krill herd algorithm (KHA), which is a novel, yet rel-
atively newly introduced metaheuristic optimization algorithm. After a brief review
of the vast applications of KHA, including complex engineering optimization
problems, the standard KHA and its mechanism were described. In the nal section,
a pseudo code of the standard KHA was also presented.

References

Bolaji, A. L. A., Al-Betar, M. A., Awadallah, M. A., Khader, A. T., & Abualigah, L. M. (2016).
A comprehensive review: Krill Herd algorithm (KH) and its applications. Applied Soft
Computing, 49, 437446.
Brierley, A. S., & Cox, M. J. (2010). Shapes of krill swarms and sh schools emerge as aggregation
members avoid predators and access oxygen. Current Biology, 20(19), 17581762.
Fattahi, E., Bidar, M., & Kanan, H. R. (2016). Fuzzy krill herd (FKH): An improved optimization
algorithm. Intelligent Data Analysis, 20(1), 153165.
Flierl, G., Grnbaum, D., Levins, S., & Olson, D. (1999). From individuals to aggregations: The
interplay between behavior and physics. Journal of Theoretical Biology, 196(4), 397454.
Gandomi, A. H., & Alavi, A. H. (2012). Krill herd: A new bio-inspired optimization algorithm.
Communications in Nonlinear Science and Numerical Simulation, 17(12), 48314845.
Gandomi, A. H., & Alavi, A. H. (2016). An introduction of krill herd algorithm for engineering
optimization. Journal of Civil Engineering and Management, 22(3), 302310.
Gandomi, A. H., Alavi, A. H., & Talatahari, S. (2013a). Structural optimization using krill herd
algorithm. Chapter 15 in swarm intelligence and bio-inspired computation: Theory and
applications. London, UK: Elsevier Publication.
Gandomi, A. H., Talatahari, S., Tadbiri, F., & Alavi, A. H. (2013b). Krill herd algorithm for
optimum design of truss structures. International Journal of Bio-Inspired Computation, 5(5),
281288.
Gandomi, A. H., Yang, X. S., & Alavi, A. H. (2013c). Cuckoo search algorithm: A metaheuristic
approach to solve structural optimization problems. Engineering with Computers, 29(1), 1735.
Hofmann, E. E., Haskell, A. E., Klinck, J. M., & Lascara, C. M. (2004). Lagrangian modelling
studies of Antarctic krill (Euphausia superba) swarm formation. ICES Journal of Marine
Science, 61(4), 617631.
Mandal, B., Roy, P. K., & Mandal, S. (2014). Economic load dispatch using krill herd algorithm.
International Journal of Electrical Power & Energy Systems, 57, 110.
Mukherjee, A., & Mukherjee, V. (2015). Solution of optimal power flow using chaotic krill herd
algorithm. Chaos, Solitons & Fractals, 78, 1021.
Price, H. J. (1989). Swimming behavior of krill in response to algal patches: A mesocosm study.
Limnology and Oceanography, 34(4), 649659.
Wang, G. G., Gandomi, A. H., & Alavi, A. H. (2013). A chaotic particle-swarm krill herd
algorithm for global numerical optimization. Kybernetes, 42(6), 962978.
Chapter 9
Grey Wolf Optimization
(GWO) Algorithm

Hossein Rezaei, Omid Bozorg-Haddad and Xuefeng Chu

Abstract This chapter describes the grey wolf optimization (GWO) algorithm as
one of the new meta-heuristic algorithms. First, a brief literature review is presented
and then the natural process of the GWO algorithm is described. Also, the opti-
mization process and a pseudo code of the GWO algorithm are presented in this
chapter.

9.1 Introduction

Grey wolf optimization (GWO) is one of the new meta-heuristic optimization

algorithms, which was introduced by Mirjalili et al. (2014). Gholizadeh (2015)
developed the GWO algorithm to solve an optimization problem of double-layer
grids considering the nonlinear behavior. The results illustrated that GWO had a
better performance than other algorithms in nding the optimal design of nonlinear
double-layer grids. Mirjalili (2015) used the GWO algorithm to learn multi-layer
perceptron (MLP) for the rst time. In the study, the results of GWO were com-
pared with those from particle swarm optimization (PSO), genetic algorithm (GA),
ant colony optimization (ACO), and evolution strategy (EA), and indicated the
higher performance of GWO. Saremi et al. (2015) coupled GWO with the evolu-
tionary population dynamic (EPD) to improve the performance of the basic GWO

H. Rezaei
O. Bozorg-Haddad (&)

Department of Irrigation and Reclamation Engineering, Faculty of Agricultural Engineering
and Technology, College of Agriculture and Natural Resources, University of Tehran,
31587-77871 Karaj, Tehran, Iran
e-mail: [email protected]
H. Rezaei
e-mail: [email protected]
X. Chu
Department of Civil and Environmental Engineering, North Dakota State University,
Dept 2470, Fargo, ND 58108-6050, USA
e-mail: [email protected]

Springer Nature Singapore Pte Ltd. 2018 81

O. Bozorg-Haddad (ed.), Advanced Optimization by Nature-Inspired Algorithms,
Studies in Computational Intelligence 720, DOI 10.1007/978-981-10-5221-7_9
82 H. Rezaei et al.

algorithm by removing weak individuals from the society. Comparison with the
basic GWO illustrated that the proposed algorithm had a better performance in
conversion rate and exploration, and also avoided trapping into local optima.
Sulaiman et al. (2015) used GWO to solve an optimal reactive power dispatch
(ORPD) problem and compared with swarm intelligence (SI), evolutionary com-
putation (EC), PSO, harmony search algorithm (HAS), gravity search algorithm
(GSA), invasive weed optimization, and modied imperialist competitive algorithm
with invasive weed optimization (MICA-IWO). The results demonstrated that
GWO had more desirable optimal solution than others.

9.2 Natural Process of the GWO Algorithm

GWO is inspired by social hierarchy and the intelligent hunting method of grey
wolves. Usually, grey wolves are at the top of the food chain in their life areas. Grey
wolves mostly live in a pack of 512 individuals. In particular, in grey wolves life
there is a strict social hierarchy. As shown in Fig. 9.1, the leaders of a pack of grey
wolves (alpha) are a male and female wolves that often are responsible for making
decisions for their pack such as sleep place, hunting, and wake-up time. Mostly,
other individuals of the pack must obey the decision made by alpha. However,
some democratic behaviors in the social hierarchy of grey wolves can be observed
(alpha may follow other individuals of the pack). In gatherings, individuals conrm
the alphas decision by holding their tails down. It is also interesting to know that it

Fig. 9.1 Social hierarchy of

grey wolves

9 Grey Wolf Optimization (GWO) Algorithm 83

is not necessary for the alpha to be the strongest ones in the pack. Managing the
pack is the main role of the alpha. In a pack of grey wolves, discipline and orga-
nization are the most important. The level next to alpha in the social hierarchy of
grey wolves is beta and the role of beta is to help alpha in making decisions. Beta
can be either male or female wolves and beta can be the best candidate of substi-
tution for alpha when one of them becomes old or dies. The beta must respect alpha,
but he/she can command other individuals. Beta is the consultant of alpha and
responsible for disciplining the pack. The beta reinforces the orders of alpha and
gives alpha the feedbacks. The weakest level in a pack of grey wolves is omega that
plays a role of scapegoat. The wolves at the level of omega have to obey other
individuals orders and they are the last wolves that are allowed to eat food. Omega
seems to be the least important individuals in the pack, but without omega, internal
ght and other problems can be observed. This can be attributed to the omegas
venting role of violence and frustration of other wolves, which helps satisfy other
individuals and maintain the dominant structure of grey wolves. Sometimes, omega
plays the role of babysitter in the pack. The remaining wolves, other than alpha,
beta, and omega, are called subordinate (delta). The wolves at the level of delta
obey the alpha and beta wolves and dominate the omega wolves. They act as
scouts, sentinels, elders, hunters, and caretakers in the pack. Scouts are responsible
for looking after boundaries and territory and also they should alarm the pack in
facing to danger. Sentinels are in charge of security establishment. Elders are the
experienced wolves that are candidates for alpha and beta. Hunters help alpha and
beta in hunting and preparing food for the pack, while caretakers should look after
the weak, ill, and wounded wolves.
In addition to the social hierarchy in a pack of grey wolves, group hunting is one
of the interesting social behaviors of grey wolves too. According to Muro et al.
(2011) grey wolves hunting includes the following three main parts:
(1) Tracking, chasing, and approaching the prey.
(2) Pursuing, encircling, and harassing the prey till it stops moving.
(3) Attacking the prey.
These two social behaviors of grey wolves pack (social hierarchy and hunting
technique) are modeled in the GWO algorithm.

9.3 Mathematical Model of the GWO Algorithm

In this section, mathematical modeling of the social hierarchy of grey wolves, and
their hunting technique (tracking, encircling, and attacking prey) in the GWO
algorithm is detailed.
84 H. Rezaei et al.

9.3.1 Social Hierarchy

In order to mathematically model the social hierarchy of grey wolves in the GWO
algorithm, the best solution is considered as alpha a. Therefore, the second and
third best solutions are respectively considered as beta b and delta d, and other
solution is assumed to be omega x. In the GWO algorithm, hunting (optimiza-
tion) is guided by a, b, and d, and x wolves follow them.

9.3.2 Encircling the Prey

As aforementioned, grey wolves in the process of hunting, encircle the prey. The
grey wolves encircling behavior to hunt for a prey can be expressed as (Mirjalili
et al. 2014):
 
! ! ! 
D  C : X p t  Xt 9:1

! ! !!
X t 1 X p t  A : D 9:2

! ! !
where t iteration number; A and C coefcient vectors; X P vector of the
! !
preys positions; X vector of the grey wolfs positions; and D calculated
! !
vector used to specify a new position of the grey wolf. A and C can be calculated
by Mirjalili et al. (2014):
!
A 2!
a :r! !
1  a 9:3
!
C 2:r!
2 9:4

where ! a vector set to decrease linearly from 2 to 0 over the iterations; and
!
r1 and ! r 2 random vectors in [0,1]. As shown in Fig. 9.2, a grey wolf at x; y
can change its position based on the position of prey at x0 ; y0 . Different places to
the best agent can be achieved with respect to the current position by regulating the
! ! ! !
A and C . For instance, by setting A 1; 0 and C 1; 1, the position of the
grey wolf is updated to x0  x; y0 .
Note that the random ! r1 and ! r2 vectors let the grey wolf select any
positions/nodes in Fig. 9.2. Therefore, a grey wolf can be placed in each random
position around the prey that is calculated by using Eqs. (9.1) and (9.2). Following
the same way, in an n-dimensional decision space grey wolves can move to any
nodes of a hypercube around the best solution (position of the prey). They can
distinguish the position of the prey from others and encircle it. Usually, hunting
operation is guided by a, and b and d provide support for a. In a decision space of
an optimization problem we do not have any idea about the optimum solution.
9 Grey Wolf Optimization (GWO) Algorithm 85

Fig. 9.2 Attacking toward

prey versus searching for prey

Thus, in order to simulate the hunting behavior of grey wolves, we assume that a
(best candidate for the solution), b, and d have more knowledge about the potential
position of the prey. Therefore, the algorithm saves three best solutions achieved so
far and forces others (i.e., omega wolves) to update their positions to achieve the
best place in the decision space. In the optimization algorithm, such a hunting
behavior can be modeled by Mirjalili et al. (2014):
     
! ! ! ! ! ! ! ! ! ! ! !
D a  C 1 : X a  X ; D b  C 2 : X b  X ; D d  C 3 : X d  X  9:5

! ! ! ! ! ! ! ! !
X 1 X a  A1 : D a ; X 2 X b  A2 : D b ; X 3 X d  A2 : D d 9:6
! ! !
X 1 X a  A1 : D a 9:7

Figure 9.3 shows how the search agent updates the positions of a, b, and d in a
2D search space. As shown in Fig. 9.3, the nal position (solution) is inside a circle
that is specied based on the positions of a, b, and d in the decision space. In other
words, a, b, and d estimate the positions of prey and other wolves and then update
their new positions, randomly around the prey.

9.3.3 Attacking the Prey

As aforementioned, grey wolves nish the hunting process by attacking the prey
until it stops moving. In order to model the attacking process, the value of !
a can be
86 H. Rezaei et al.

Fig. 9.3 Updating of positions in the GWO algorithm

Fig. 9.4 Attacking toward prey and searching for prey

!
decreased in different iterations. Note that as !
a decreases the fluctuation rate of A
!
decreases too. In other words, A is a random value in the range of 2a; 2a where
!
a decreases from 2 to 0 over iterations. When the random value of A is being in the
range of 1; 1. The next position of a wolf can be between the current position
and the prey position. As illustrated in Fig. 9.4, when j Aj\1 grey wolves will
attack the prey.
9 Grey Wolf Optimization (GWO) Algorithm 87

By using the operators provided so far, the GWO algorithm lets the search agent
to update its position based on the positions of a, b, and d(move toward the prey). It
is true that the encircling process provided as an operator in the GWO algorithm
limits the solutions around local optima, but GWO also has many other operators to
discover new solutions.

9.3.4 Searching for the Prey (Exploration)

Grey wolves often search for the prey according to the positions of a, b, and d.
They diverge from each other to explore the position of prey and then converge to
!
attack the prey. In order to mathematically model the divergence of grey wolves, A
!
can be utilized. A is a random vector that is greater than 1 or less than 1 to force
the search agent to diverge from the prey,
 which emphasizes the global search in
!
GWO. Figure 9.4 illustrates that when  A  [ 1, the grey wolf is forced to move
away from the prey (local optimum) to search for better solutions in the decision
space. !
The GWO algorithm has another component C that assists the algorithm to
!
discover new solutions. As shown in Eq. (9.4), the elements of vector C are within
the range of 0; 2. This component provides random weights for the prey to ran-
domly emphasize C [ 1 or deemphasize C\1 the impact of the prey in
dening the distance in Eq. (9.1). This component helps the GWO algorithm to
behave more randomly and in favor of exploration, and keep the search agent away
from local optima during the optimization process. Note that unlike A, C decreases
nonlinearly. C is required in the GWO algorithm because not only in the initial
iteration but also in the nal iteration, it provides a global search in the decision
space. This component is very useful in avoidance of local optima, especially in the

Table 9.1 Characteristics of the GWO algorithm

General algorithm Grey wolf optimization algorithm
Decision variable Grey wolf
Solution Position of grey wolf
Old solution Old position of grey wolf
New solution New position of grey wolf
Best solution Position of prey
Fitness function Distance between grey wolf and prey
Initial solution Initial random position of grey wolf
Selection
Process of generating new solution Hunting operators (encircling and attacking prey)
88 H. Rezaei et al.

nal iteration. The C vector can be used as a hedge of approaching the prey in
nature. Generally, the hedge can be seen in a nature hunting process of grey wolves.
This hunting technique prevents grey wolves from quickly approaching the prey
(this is truly what C does in the optimization process of the GWO algorithm).
Table 9.1 presents the characteristics of the GWO algorithm.

9.4 Optimization Process in GWO Algorithm

The optimization process of GWO starts with creating random population of grey
wolves (candidate solutions). Over the iterations, a, b, and d wolves estimate the
probable position of the prey (optimum solution). Grey wolves update their posi-
tions based on their distances from the prey. In order to emphasize exploration and
exploitation
  during the search process, parameter a should decrease from 2 to 0. If
!
 A  [ 1, the candidate solutions diverge from the prey; and if j Aj\1, the candidate
solutions converge to the prey. This process continues and the GWO algorithm is
terminated if the stopping criteria are satised. To understand how the GWO
algorithm solves optimization problems theatrically, some notes can be summarized
as follows:
The concept of social hierarchy in the GWO algorithm helps grade the solutions
and save the best solutions up to the current iteration.
The encircling mechanism denes a 2D circle-shaped neighbor and the solution
(in higher dimensions, the 2D circle can be extended to a 3D hyper-sphere).
The random parameters (A and C) help grey wolves (candidate solutions) to
dene different hyper-spheres with random radii.
The hunting approach implemented in the GWO algorithm allows grey wolves
(candidate solutions) to locate the probable position of the prey (optimum
solution).
The adaptive values of parameters A and a guarantee exploration and
exploitation in the GWO algorithm and also allow it to easily transfer between
exploration and exploitation.
By
!decreasing
  the values of A, a half of iterations are assigned to exploration
 
 A  [ 1 and the other half of iterations are assigned to exploitation j Aj\1.
a and C are two main parameters of the GWO algorithm.
Figure 9.5 shows the flowchart of the GWO algorithm with details on the
optimization process.
9 Grey Wolf Optimization (GWO) Algorithm 89

Fig. 9.5 Flowchart of the

GWO algorithm
90 H. Rezaei et al.

9.5 Pseudocode of GWO

Begin

Initialize the population of grey wolves Xi (i = 1, 2, ... , n)

Initialize a, A, and C

Calculate the fitness values of search agents and grade them. (X= the best solution in

the search agent, X= the second best solution in the search agent, and X= the third

best solution in the search agent)

t= 0

While (t < Max number of iterations)

For each search agent

Update the position of the current search agent by Equation (9.7)

End for

Update a, A, and C

Calculate the fitness values of all search agents and grade them

Update the positions of X, X, and X

t= t+1

End while

End

9.6 Conclusions

This chapter described the grey wolf optimization (GWO) algorithm as one of the
new meta-heuristic algorithms. The GWO algorithm was inspired by the lift style of
the pack of grey wolves (social hierarchy and hunting mechanism). Also, this
chapter presented a brief literature review of GWO, described the natural process of
grey wolves life style and the mathematical equations of GWO, and nally pre-
sented a pseudocode of GWO.
9 Grey Wolf Optimization (GWO) Algorithm 91

References

Gholizadeh, S. (2015). Optimal design of double layer grids considering nonlinear behaviour by
sequential grey wolf algorithm. Journal of Optimization in Civil Engineering, 5(4), 511523.
Mech, L. D. (1999). Alpha status, dominance, and division of labor in wolf packs. Canadian
Journal of Zoology, 77(8), 11961203.
Mirjalili, S., Mirjalili, S. M., & Lewis, A. (2014). Grey wolf optimizer. Advances in Engineering
Software, 69(2014), 4661.
Mirjalili, S. (2015). How effective is the grey wolf optimizer in training multi-layer perceptron.
Applied Intelligence, 43(1), 150161.
Mirjalili, S. M., & Mirjalili, S. Z. (2015). Full optimizer for designing photonic crystal
waveguides: IMoMIR framework. IEEE Photonics Technology Letters, 27(16), 17761779.
Mirjalili, S. M., Mirjalili, S., & Mirjalili, S. Z. (2015). How to design photonic crystal LEDs with
articial intelligence techniques. Electronics Letters, 51(18), 14371439.
Muro, C., Escobedo, R., Spector, L., & Coppinger, R. (2011). Wolf-pack (Canis Lupus) hunting
strategies emerge from simple rules in computational simulations. Behavioral Processes, 88(3),
192197.
Naderizadeh, M., & Baygi, S. J. M. (2015). Statcom with grey wolf optimizer algorithm based pi
controller for a grid Connected wind energy system. International Research Journal of Applied
and Basic Sciences, 9(8), 1421.
Noshadi, A., Shi, J., Lee, W. S., Shi, P., & Kalam, A. (2015). Optimal PID-type fuzzy logic
controller for a multi-input multi-output active magnetic bearing system. Neural Computing
and Applications, 27(7), 116.
Saremi, S., Mirjalili, S. Z., & Mirjalili, S. M. (2015). Evolutionary population dynamics and grey
wolf optimizer. Neural Computing and Applications, 26(5), 12571263.
Sulaiman, M. H., Mustaffa, Z., Mohamed, M. R., & Aliman, O. (2015). Using the grey wolf
optimizer for solving optimal reactive power dispatch problem. Applied Soft Computing, 32
(2015), 286292.
Wong, L. I., Sulaiman, M. H., & Mohamed, M. R. (2015). Solving economic dispatch problems
with practical constraints utilizing grey wolf optimizer. Applied Mechanics and Materials, 785
(2015), 511515. Trans Tech Publications.
Yusof, Y., & Mustaffa, Z. (2015). Time series forecasting of energy commodity using grey wolf
optimizer. In Proceedings of the international multiconference of engineers and computer
scientists (IMECS 2015), Hong Kong, 1820 March.
Chapter 10
Shark Smell Optimization (SSO)
Algorithm

Sahar Mohammad-Azari, Omid Bozorg-Haddad and Xuefeng Chu

Abstract In this chapter, the shark smell optimization (SSO) algorithm is pre-
sented, which is inspired by the sharks ability to hunt based on its strong smell
sense. In Sect. 10.1, an overview of the implementations of SSO is presented. The
underlying idea of the algorithm is discussed in Sect. 10.2. The mathematical
formulation and a pseudo-code are presented in Sects. 10.3 and 10.4, respectively.
Section 10.5 is devoted to conclusion.

10.1 Introduction

Generally, all animals have abilities that ensure their survival in the nature. Some
species have special abilities which distinguish them from others (Costa and
Sinervo 2004). Finding the prey and the movement of hunter toward the prey are
two important factors in the hunting process. Animals that are able to nd the prey
in a short time with a correct movement, are a successful hunter. Shark is one of the
most well-known and superior hunter in the nature. The reason of this superiority is
the sharks ability to nd the prey in a short time based on its strong smell sense in a
large search space.
Based on this sharks ability, Abedinia et al. (2014) developed a meta-heuristic
algorithm named shark smell optimization (SSO), and evaluated its efciency based

S. Mohammad-Azari
O. Bozorg-Haddad (&)

Department of Irrigation and Reclamation Engineering, Faculty of Agricultural Engineering
and Technology, College of Agriculture and Natural Resources, University of Tehran,
3158777871 Karaj, Iran
e-mail: [email protected]
S. Mohammad-Azari
e-mail: [email protected]
X. Chu
Department of Civil and Environmental Engineering, North Dakota State University,
Dept 2470, Fargo, ND 58108-6050, USA
e-mail: [email protected]

Springer Nature Singapore Pte Ltd. 2018 93

O. Bozorg-Haddad (ed.), Advanced Optimization by Nature-Inspired Algorithms,
Studies in Computational Intelligence 720, DOI 10.1007/978-981-10-5221-7_10
94 S. Mohammad-Azari et al.

on some standard benchmark functions and compared with other meta-heuristic

algorithms. Then, they investigated the performance of SSO in a real frequency
control engineering problem in the electric power systems.
Abedinia and Amjadi (2015) applied a hybrid prediction model based on the
neural network and chaotic SSO for wind power forecasting. The number of hidden
nodes in a neural network was optimized by using the chaotic SSO model. In order
to evaluate the efciency of their proposed model, two case studies were considered
and the results were compared with 14 other prediction methods. The results
demonstrated the capability of their proposed model to cope with the variability and
intermittency of wind power time series for providing accurate predictions.
Gnanaskaran et al. (2016) applied the SSO algorithm as an efcient method for
nding the optimal size and location of shunt capacitors to minimize the cost due to
energy loss and reactive power compensation of a distribution system. The results
indicated the superiority of SSO compared with other classical algorithms. The
reason of this superiority was that acquiring optimal solutions through simple
formulation satised the problem constraints.
Ghaffari et al. (2016) investigated optimal economic load dispatch using the SSO
algorithm. In their problem, risk constraints for unpredictable and uncertain
behavior of wind were considered. In fact, they investigated the balance between
cost and risk over a 30-bus power system with SSO and compared the results with
those from other conventional methods.
Ehteram et al. (2016) evaluated the capability of SSO as a meta-heuristic opti-
mization algorithm for operation of the single-reservoir (Bazoft) and multi-reservoir
(Larson) systems. Then, the obtained results were compared with those from GA
and the PSO algorithm based on certain performance criteria. The results indicated
the superiority of SSO over two other algorithms due to the higher reliability and
lower vulnerability. In addition, the application of the SSO algorithm was suggested
for complex multi-objective reservoir systems with several operators.

10.2 Underlying Idea of SSO

The olfactory system in each animal is the primary sensory system which responds
to the chemical signal from a remote source. In shes, the smell receptors are
located in the olfactory pits which are positioned on the sides of their heads. Each
pit has two outside openings through which water flows in and out. The mechanism
of water movement inside the pit is formed through the wave motion of tiny hair on
the cells lining the pit and the force caused by the movement of sh in water.
Dissolved chemicals connect to a pleated surface in the olfactory nerve endings
(Abedinia et al. 2014). In vertebrates, unlike other sensory nerves, the olfactory
receptors are directly connected to their brains without any nerve intermediaries.
10 Shark Smell Optimization (SSO) Algorithm 95

Fig. 10.1 Schematic of sharks movement toward the source of the smell

The smell impulses are received by the portion in the front of the brain called
olfactory bulb. Fishes have two olfactory bulbs, each of which is located in an
olfactory pit. The allocation of the larger surface of olfactory pits to smell nerves
and larger smell information centers in the brain make the shes smell sense
stronger (Magnuson 1979). Eels and sharks have the largest olfactory bulbs for
smell information processing. About 400 million years ago, the rst sharks
appeared in the oceans as the superior hunters in the nature. One of the reasons of
shark survival in the nature is its ability to capture the prey with its strong smell
sense.
The sharks smell sense is one of its most effective senses. When a shark swims
in water, the water flows through its nostrils which are located along the sides of its
snout. After the water enters the olfactory pits, it flows through the folds of skin
which is covered with sensory cells. Some sharks have this ability to detect the
slightest trace of blood due to the sensory cells (Sfakiotakis et al. 1999). For
example, a shark can detect one drop of blood in a large swimming pool.
Accordingly, the shark can smell an injured sh from up to one kilometer away
(Abedinia et al. 2014).
The sharks smell sense can consider as a guide for it. The smell that comes from
the left side of the shark passes the left pit before entering the right pit. This process
helps shark to nd the source of the smell (Wu and Yao-Tsu 1971). The schematic
of sharks movement toward the source of the smell is shown in Fig. 10.1.
In this movement, concentration plays an important role to guide the shark to its
prey. In other words, a higher concentration results in a true movement of the shark.
This characteristic is the base of development of an optimization algorithm to nd
the optimal solution of a problem.
96 S. Mohammad-Azari et al.

10.3 Formulation of the SSO Algorithm

Some assumptions are considered to construct the mathematical formulation. They

include:
(1) Fish is injured and this occurrence leads to blood injection into the sea (the
search space). As a result, the velocity of the injured sh can be neglected
compared against the shark movement velocity. In other words, the source
(prey) is assumed to be xed.
(2) Blood injection to sea water occurs regularly. The effect of water flow on
distorting particles of odor is neglected. It is obvious that the odor particles are
stronger near the injured sh. Consequently, following the odor particles helps
shark to approach the prey.
(3) One injured sh results in one odor source in the search environment of the
shark (Abedinia et al. 2014).

10.3.1 Initialization of SSO: Finding Initial Odor Particles

The search process begins when the shark smells odor. In fact, the particles of odor
have a weak diffusion from an injured sh (prey). In order to model this process, a
population of initial solutions are randomly generated for an optimization problem
in the feasible search space. Each of these solutions represents a particle of odor
which shows a possible position of the shark at the beginning of the search process.

1 1 
x1 ; x2 ; . . .; x1NP ; 10:1

where x1i = ith initial position of the population vector or ith initial solution; and
NP = population size. The related optimization problem can be expressed as:
h i
x1i x1i; 1 ; x1i; 2 ; . . .; x1i; ND i 1; 2; . . .; NP, 10:2

where x1i; j = jth dimension of the sharks ith position or jth decision variable of ith
 
position of the shark x1i ; and ND = number of decision variables in the opti-
mization problem.
The odor intensity at each position reflects its closeness to the prey. This process
is modeled in the SSO algorithm through an objective function. Assuming a
maximization problem and considering the general principle, a higher value of the
10 Shark Smell Optimization (SSO) Algorithm 97

objective function represents stronger odor (or more odor particles). Consequently,
this process represents a closer position of the shark to its prey. The SSO algorithm
initiates according to this view (Abedinia et al. 2014).

10.3.2 Shark Movement Toward the Prey

The shark at each position moves with a velocity to become closer to the prey.
Based on the position vectors, the initial velocity vector can be expressed as:


V11 ; V21 ; . . .; VNP
1
10:3

In Eq. (10.3), the velocity vectors have components in each dimension.

h i
Vi1 Vi;1
1
; Vi;2
1
; . . .; Vi;ND
1
i 1; . . .; ND 10:4

The shark follows the odor and the direction of its movement is determined
based on the intensity of odor. The velocity of the shark is increased due to the
increased concentration of odor. From the optimization point of view, this move-
ment is modeled mathematically by the gradient of the objective function. The
gradient indicates the direction in which the function increases with the highest rate.
Equation (10.5) shows this process (Abedinia et al. 2014):

Vik gk :R1:rOFjxk i 1; . . .; NP k 1; . . .; kmax ; 10:5

i

where Vik = velocity of the shark which is approximately constant; OF = objective

function; r = gradient of the objective function; kmax = maximum number of
stages for forward movement of the shark; k = number of stages; gk = a value in
the interval [0,1]; and R1 = a random value which is uniformly distributed in the
interval [0,1].
gk is in the interval [0,1] as it may be impossible for a shark to reach the velocity
determined by the gradient function. The parameter R1 gives more random search
inherent to the SSO algorithm. The idea of considering R1 has been taken from the
gravitational search algorithm (GSA). The velocity in each dimension can be cal-
culated by Eq. (10.6) (Abedinia et al. 2014):

@ OF
Vi;j gk :R1:
k
i 1; . . .; NP j 1; . . .; ND k 1; . . .; kmax 10:6
@xj xk
i;j
98 S. Mohammad-Azari et al.

Due to the existence of inertia, acceleration of the shark is limited and its
velocity depends on its previous velocity. This process is modeled by a modied
Eq. (10.6) as follows:

@ OF
Vi;jk gk :R1: ak :R2:Vi;jk1
@xj xk 10:7
i;j

i 1; . . .; NP j 1; . . .; ND k 1; . . .; kmax ;

where ak = rate of momentum or inertia coefcient that has a value in the interval of
[0,1] and becomes a constant for stage k; and R2 = random number generator with a
uniform distribution on the interval [0,1], which is intended for the momentum
term. A larger value of ak indicates higher inertia and hence the current velocity is
more dependent on the previous velocity. From the mathematical point of view, the
application of momentum leads to smoother search paths in the solution space. R2
increases
 the diversity of the search in the algorithm. For the velocity in the rst
stage Vi;j1 , it is possible to neglect the initial velocity of the shark before starting
 
the search process Vi;j0 or allocate a very small value to it.
The velocity of the shark can be increased up to a specied limit. Unlike most
shes, sharks do not have swim bladders to help them stay afloat. So, they cannot
be static and must swim upward in a direction even with a low velocity. This
process occurs using the strong tail n which acts as a propulsion (Wu and Yao-Tsu
1971). The normal velocity of a shark is about 20 km/h which is increased up to
80 km/h when the shark tends to attack. The ratio of the highest to lowest velocities
of the sharks is limited (For example, 8020 4). The velocity limiter used for each
stage of the SSO algorithm can be expressed as (Abedinia et al. 2014):
"   #
   @ OF   
 k  k1   k1 
Vi;j  min gk :R1: ak :R2:Vi;j ; bk :Vi;j  ;
 @xj xk  10:8
i;j

i 1; . . .; NP; j 1; . . .; ND; k 1; . . .; kmax

where bk = velocity limiter ratio for stage k. The value of Vi;jk is calculated by
Eq. (10.8) and it has the same sign as the term selected by the minimum operator in
Eq. (10.8). Due to forward movement of the shark, its new position Yik 1 is
determined based on its previous position and velocity:

Yik 1 Xik Vik :Dtk i 1; . . .; NP k 1; . . .; kmax ; 10:9

where Dtk = time interval of stage k. Dtk is assumed one for all stages for the
purpose of simplicity. Each component of Vi;jk j 1; . . .; ND of vector Vik is
obtained by Eq. (10.8).
10 Shark Smell Optimization (SSO) Algorithm 99

Fig. 10.2 Rotational movement of a shark

In addition to forward movement, the shark also has a rotational movement in its
direction to nd stronger odor particles. In fact, this improves its progress (Yao-Tsu
1971). The simulation of this movement is shown in Fig. 10.2.
As shown in Fig. 10.2, the rotational movement of the shark takes place along a
closed contour which is not necessarily a circle. From the optimization view, in
order to nd better solutions the shark does a local search in each stage. This local
search in the SSO algorithm is modeled by Abedinia et al. (2014):

Zik 1;m Yik 1 R3:Yik 1 m 1; . . .; M i 1; . . .; NP k 1; . . .; kmax ;

10:10

where Zik 1; m position of point m in the local search; R3 = a random number

with a uniform distribution in the interval [1, 1]; and M = number of points in the
local search of each stage.
Since this operator does a local search around Yik 1 , the limit of R3 can be
considered in the interval [1, 1]. M points of the local search Zik 1; m are in the
vicinity of Yik 1 (If the random number generator generates zero, Yik 1 is obtained).
A closed contour is obtained by the connections of M points, which is similar to the
rotational movement of the shark. During the rotational movement, if the shark
nds a point with a stronger odor it follows the search from that point. This process
is shown in Fig. 10.2. This characteristic in the SSO algorithm can be expressed as
(Abedinia et al. 2014):
     
xki 1 arg maxfOF Yik 1 ; OF Zik 1; i ; . . .; OF Zik 1; M i 1; 2; . . .; NP
10:11
100 S. Mohammad-Azari et al.

Start

Determine SSO algorithms parameters

Generate initial population for SSO

Initialize the stage counter

k =1

Calculate each component of velocity vector

Determine the new position of shark based on

the forward movement.
.

Determine the new position of shark based on k= k+1

the raotational movement.

Select the next position of shark

No
k = kmax ?

Yes

Select the best position of shark in last stage

End

Fig. 10.3 Flowchart of the SSO algorithm

10 Shark Smell Optimization (SSO) Algorithm 101

Table 10.1 Characteristics of the SSO algorithm

General algorithm Gradient evolution algorithm
Decision variable Odor particle which shows possible position of shark
Solution Odor intensity
Old solution Old position of shark
New solution New position of shark
Best solution Position of shark with the best tness function
Fitness function Odor particle intensity which represents the closer position of
shark to prey
Initial solution Randomly generated position of shark
Selection Select sharks position based on the forward and rotational
movements
Process of generating new Forward movement and rotational movement
solution

In Eq. (10.11), the objective function (OF) must be maximized. In other words,
among Yik 1 obtained from forward movement and Zik 1; m (m = 1, 2, ,
M) obtained from the rotational movement, a solution with the highest objective
function is selected as the next position of the shark (xki 1 ). The cycle of forward
and rotational movement will continue until k is equal to kmax .
Like other meta-heuristic optimization methods, the SSO algorithm also has a
number of parameters that must be dened by users, including NP, kmax , g, a, and b
in each stage. Changing these parameters during the SSO evolution based on an
adaptive mechanism is an effective method in the applications. For example, such a
mechanism may start adaptively from larger values of g and b and a smaller value
of a, and then the values of g and b will be decreased and the value of a will be
increased. So, in the initial stage of the evolution process, the algorithm will
continue with large steps in order to enhance the search ability and for the last stage
(when the algorithm approaches the optimal solution) the steps will be smaller to
increase the resolution of search around the optimal solution. After setting the
parameters, the population and the stage counter of the SSO are initialized.
Population will be evolved by the operators of forward and rotational move-
ments. Finally, the best solution in the nal stage is selected for the optimization
problem. The search operators in the SSO algorithm including the gradient-based
forward movement and local search based rotational movement are not used in any
other meta-heuristic algorithms (Abedinia et al. 2014).
The flowchart of the SSO algorithm is shown in Fig. 10.3, and the parameters
and variables used in the SSO algorithm are listed in Table 10.1.
102 S. Mohammad-Azari et al.

10.4 Pseudo-Code of SSO

Begin
Step 1. Initialization

Set parameters NP, k max , k , k , and k k 1, 2 , ..., k max

Generate initial population with all individuals

Generate each decision randomly within the allowable range

Initialize the stage counter k = 1

For k = 1 : k max

Step 2. Forward movement

Calculate each component of the velocity

vector, v i , j i 1, ..., NP , j 1,..., ND

Obtain new position of shark based on forward movement, Yik 1 i 1, ..., NP

Step 3. Rotational movement

Obtain position of shark based on rotational movement, Z ik 1,m m 1, .., M

Select next position of shark based on the two movements X ik 1 i 1, ..., NP

End for k

Set k = k+1

Select the best position of shark in the last stage which has the highest OF value
End

10.5 Conclusion

In this chapter, the shark smell optimization (SSO) algorithm was introduced as one
of the new meta-heuristic optimization methods. This algorithm was developed
based on the hunting ability of sharks to use their smell sense. This is a stochastic
search optimization algorithm which initiates with a sets of random solutions and
10 Shark Smell Optimization (SSO) Algorithm 103

continues the search to nd the optimal solution. In fact, this algorithm applies a
gradient-based forward movement and a local search based rotational movement
during the optimization process. In this chapter, the SSO algorithm was introduced
and its mathematical formulation was presented.

References

Abedinia, O., & Amjadi, N. (2015). Short-Term wind power prediction based on hybrid neural
network and chaotic shark smell optimization. International Journal of Precision Engineering
and Manufacturing-Green Technology, 2(3), 245254.
Abedinia, O., Amjady, N., & Ghasemi, A. (2014). A new metaheuristic algorithm based on shark
smell optimization. Complexity. doi:10.1002/cplx.21634
Costa, D. P., & Sinervo, B. (2004). Field physiology: physiological insights from animals in
nature. Annual Review of Physiology, 66, 209238.
Ehteram, M., Karimi, H., Musavi, S. F., & EL-Shae, A. (2017). Optimizing dam and reservoirs
operation based model utilizing shark algorithm approach. Knowledge-Based Systems
(In Press). doi:10.1016/j.knosys.2017.01.026
Ghaffari, S., Aghajani, Gh, Noruzi, A., & Hedayati-Mehr, H. (2016). Optimal economic load
dispatch based on wind energy and risk constrains through an intelligent algorithm.
Complexity, 21(S2), 494506.
Gnanasekaran, N., Chandramohan, S., Sathish Kumar, P., & Mohamed Imran, A. (2016). Optimal
placement of capacitors in radial distribution system using shark smell optimization algorithm.
Ain Shams Engineering Journal, 7, 907916.
Magnuson, J. J. (1979). 4 Locomotion by Scombrid shes: hydromechanics, morphology and
behavior. Fish Physiology, 7, 239313.
Sfakiotakis, M., Lane, D. M., & Davies, J. B. C. (1999). Review of sh swimming modes for
aquatic locomotion. IEEE Journal of Oceanic Engineering, 24, 237252.
Wu, T. Yao-Tsu. (1971). Hydromechanics of swimming propulsion. Part 1. Swimming of a
two-dimensional flexible plate at variable forward speeds in an inviscid fluid. Journal of Fluid
Mechanics, 46(2), 337355.
Chapter 11
Ant Lion Optimizer (ALO) Algorithm

Melika Mani, Omid Bozorg-Haddad and Xuefeng Chu

Abstract This chapter introduces the ant lion optimizer (ALO), which mimics the
hunting behavior of antlions in the larvae stage. Specically, this chapter includes
literature review, details of the ALO algorithm, and a pseudo-code for its
implementation.

11.1 Introduction

Mirjalili (2015) introduced the ant lion optimizer (ALO) algorithm and proved its
capability by solving 19 different mathematical benchmark problems and three
classical engineering problems including three-bar truss design, cantilever beam
design, and gear train design. In addition, ALO was used for optimizing the shape
of two ship propeller as a challenging constrained problem with a diverse search
space, which showed the ability of the ALO algorithm for solving real complex
problems. Yamany et al. (2015) used ALO for determining weights and biases in a
training process of multilayer perceptron (MLP) for having a minimum error and an
appropriate classication rate. In the research, the performance of ALO was
compared with those of genetic algorithm (GA), particle swarm optimization
(PSO) algorithm, and ant colony optimization (ACO) algorithm to show its capa-
bility. Zawbaa et al. (2015) applied ALO to an optimization problem of feature

M. Mani
O. Bozorg-Haddad (&)

Department of Irrigation and Reclamation Engineering, Faculty of Agricultural Engineering
and Technology, College of Agriculture and Natural Resources, University of Tehran,
31587-77871 Karaj, Tehran, Iran
e-mail: [email protected]
M. Mani
e-mail: [email protected]
X. Chu
Department of Civil and Environmental Engineering, North Dakota State University,
Dept 2470, 58108-6050 Fargo, ND, USA
e-mail: [email protected]

Springer Nature Singapore Pte Ltd. 2018 105

O. Bozorg-Haddad (ed.), Advanced Optimization by Nature-Inspired Algorithms,
Studies in Computational Intelligence 720, DOI 10.1007/978-981-10-5221-7_11
106 M. Mani et al.

selection to maximize the classication performance. They compared ALO with

PSO and GA, and demonstrated the better performance of ALO. Ali et al. (2016)
employed ALO to optimize allocation and sizing of distributed generation (DG),
which contained photovoltaic and wind turbines. The results showed the superiority
of ALO over several other optimization algorithms such as GA and PSO. Dubey
et al. (2016) applied ALO for solving a hydrothermal power generation scheduling
(HTPGS) problem with wind integration, which was a nonlinear, non-convex, and
highly complex optimization problem. The results showed high ability of ALO for
nding powerful solutions in a complex decision space.
Talatahari (2016) implemented ALO for an optimal design of skeletal structures.
The efciency of ALO was illustrated by comparing with various classical opti-
mization algorithms. Zawbaa et al. (2016) enhanced the performance of ALO for a
feature selection problem by developing a chaotic version of ALO. Kamboj et al.
(2016) used ALO for a non-convex optimization problem of economic load dis-
patch of electric power systems. Petrovi et al. (2016) employed ALO for solving a
combinatorial optimization problem of integrated planning and scheduling. Raju
et al. (2016) presented an application of ALO for simultaneous optimization of
different controllers. Saxena and Kothari (2016) used ALO for solving antenna
array synthesis and other electromagnetic optimization problems. Yao and Wang
(2016) developed the dynamic adaptive ant lion optimizer (DAALO) by replacing
random walk of ants with levy flight and adjusting size of traps. The results from
the unmanned aerial vehicle (UAV) route planning problem showed high efciency,
robustness, and feasibility of DAALO. Mirjalili et al. (2017) developed
multi-objective ALO (MOALO) and used standard unconstrained and constraint
test functions to show the efciency of the algorithm. Also, MOALO was applied to
solve several engineering design problems including cantilever beam design,
brushless dc wheel motor design, disk brake design, 4-bar truss design, safety
isolating transformer design, speed reduced design, and welded beam design. The
results showed the capability of the algorithm in solving challenging real-world
engineering problems. Kaur and Mahajan (2016) applied ALO for optimization of
the communities in large networks. Rajan et al. (2017) applied ALO to determine
optimal reactive power dispatch of systems, which was a highly nonlinear,
non-convex challenging optimization problem. They modied ALO by introducing
a noble weighted elitism concept in the elitism phase of the original ALO. Kaushal
and Singh (2017) used ALO to optimize the allocation of stocks in portfolio and
compared with GA, showing that ALO outperformed GA for portfolio designing.

11.2 Mapping Antlions Hunting Mechanism into the ALO

Antlions are in the family of Myrmeleontidae and belong to the order of

Neuroptera. The life cycle of antlions includes two main phases, larva and adult,
which takes 2 to 3 years. The antlions life cycle mostly occurs in larvae and
adulthood has only 3 to 5 weeks. The larvae of antlions are also known as
11 Ant Lion Optimizer (ALO) Algorithm 107

Fig. 11.1 Antlion hunting behavior

doodlebugs, which have a predatory habit. Adult antlions, which are less well
known, can fly and maybe are mistakenly identied as dragonflies or damselflies.
The name of antlions best describes their unique hunting behavior and their
favorite prey which is ants. The larvae of some antlions species dig cone-shaped
pits with different sizes and wait at the bottom of the pits for ants or other insects to
slip on the loose sands and fall in, as shown in Fig. 11.1a.
When an insect is in a trap, the antlion will try to catch it while the trapped insect
will try to escape. The antlion intelligently tries to slide the prey into the bottom of
the pit by throwing sands toward the edge of the pit. After catching the prey, the
antlion pulls it under the soil and consumes it (Fig. 11.1b). After feeding is com-
pleted, the antlion flicks the leftovers of the prey out of the pit and prepares the pit
for next hunting.
It should be noted that the size of the antlions trap depends on the level of antlion
hunger and the shape of the moon. Antlions dig larger pits when they become hungry
and also when the moon is full (Goodenough 2009). For larger pits, the chance of
successful hunting increases. The ALO algorithm is inspired by this intelligent
hunting behavior of antlions and the interaction with their favorite prey, ants. So the
main steps of antlions hunting are mathematically modeled in the ALO algorithm.
In the ALO algorithm, ants are search agents and move over the decision space,
and antlions are allowed to hunt them and become tter. In each iteration, the
position of each ant is updated with respect to the selected antlion based on roulette
wheel and elite (best antlion obtained so far). By the roulette wheel selection
operator, solutions with the better tness function have more chance to be selected
as the antlion with a larger trap has more chance to hunt more ants. Table 11.1 lists
the characteristics of the ALO algorithm.
In the ALO algorithm, the rst positions of antlions and ants are initialized ran-
domly and their tness functions are calculated. Then, the elite antlion is determined.
In each iteration for each ant, one antlion is selected by the roulette wheel operator and
its position is updated with the aid of two random walk around the roulette selected
antlion and elite. The new positions of ants are evaluated by calculating their tness
functions and comparing with those of antlions. If an ant becomes tter than its
corresponding antlion, its position is considered as a new position for the antlion in the
108 M. Mani et al.

Table 11.1 Characteristics of the ALO algorithm

General algorithm Antlion optimizer
Decision variable Antlions and Ants positions in each dimension
Solution Antlions position
Old solution Old position of antlion
New solution New position of antlion
Best solution Elite antlion
Fitness function Desirability of elite
Initial solution Random antlion
Selection Roulette wheel
Process of generating new solution Random walk over the decision space

Fig. 11.2 Flowchart of the ALO algorithm (It = iteration counter; and IT = number of iterations)

next iteration. Also, the elite will be updated if the best antlion achieved in the current
iteration becomes tter than the elite. These steps are repeated until the end of itera-
tions. The flowchart of the ALO algorithm is illustrated in Fig. 11.2.
11 Ant Lion Optimizer (ALO) Algorithm 109

11.2.1 Initialization of Positions of Ants and Antlions

and Evaluation of Their Fitness Functions

In the ALO algorithm, there are two populations, ants and antlions. As afore-
mentioned, ants are the search agents in the decision space and antlions which hide
somewhere in the decision space, can hunt them, and catch their positions to
become tter. In an optimization problem with N decision variables (N-dimensional
optimization problem), the ants/antlions positions in an N-dimensional space are
the decision variables. So each dimension of the ant/antlions position belongs to
one of the decision variables which can be expressed as:

Ant's location A1 ; A2 ; . . .; An ; . . .; AN  11:1

Antlion's location Al1 ; Al2 ; . . .; ALn ; . . .; AlN ; 11:2

where An/ALn = nth decision variable; and N = number of decision variables.

The ALO algorithm starts with randomly generated matrices of the positions of
ants and antlions within their specied ranges as follow:
8 9
>
> A1;1 A1;2 . . . A1;n . . . A1;N > >
>
> A2;1 A2;2 . . . A2;n . . . A1;N > >
>
> >
>
>
>. . . . . . >
>
>
< .. .. .. .. .. .. >
=
Pant .. .. ; 11:3
> Am;1 Am;2 .
> Am;n . >
Am;N >
>
> >
>
>
> .. .. .. .. .. .. >
>
>
> . . . . . . >
>
>
: >
;
AM;1 AM;2 . . . AM;n . . . AM;N

where Pant = matrix of ants positions; Am,n = nth decision variable of the mth ant;
and M = number of ants.
8 9
>
> Al1;1 Al1;2 . . . Al1;n ... Al1;N >
>
>
> Al2;1 Al2;2 . . . ... >
>
>
> Al2;n Al1;N >
>
>
> >
>
< ...
> ..
.
..
.
..
.
..
.
..
.
>
=
Pantlion . .. ; 11:4
>
>
> Alm;1 Alm;2 .. Alm;n . Alm;N >
>
>
>
> .. .. .. .. .. .. >
>
>
> >
>
>
> . . . . . . >
>
: ;
AlM;1 AlM;2 . . . AlM;n . . . AlM;N

where Pantlion = matrix of antlions positions; and Alm,n = nth decision variable of
the mth antlion.
For evaluating the ants and antlions, a tness function is utilized and their tness
values are calculated during optimization and saved in the following matrices. In
this process, the best antlion (antlion with the best tness) is selected as elite.
110 M. Mani et al.

8 9
>
> f A1;1 A1;2 ... A1;n ... A1;N  >
>
>
> f A2;1 ... ... A1;N  >
>
>
> A2;2 A2;n >
>
>
> >
>
< ...
> ..
.
..
.
..
.
..
.
..
.
>
=
Fant .. .. ; 11:5
>
> f Am;1 Am;2 . Am;n . Am;N  >
>
>
> >
>
>
> .. .. .. .. .. .. >
>
>
> . . . . . . >
>
>
: >
;
f AM;1 AM;2 . . . AM;n . . . AM;N 

where Fant = matrix of the ants tness functions; and f = tness function.
8 9
>
> f Al1;1 Al1;2 ... Al1;n ... Al1;N >
>
> f Al
> ... ... Al1;N >
>
>
> 2;1 Al2;2 Al2;n >
>
>
> . .. .. .. .. .. >
>
>
< .. >
=
. . . . .
Fantlion .. .. ; 11:6
>
> f Alm;1 Alm;2 . Alm;n . Alm;N >
>
>
> >
>
>
> .. .. .. .. .. .. >
>
>
> . . . . . . >
>
>
: >
;
f AlM;1 AlM;2 . . . AlM;n . . . AlM;N

where Fantlion = matrix of the antlions tness functions.

11.2.2 Digging Trap

In this step, by using the roulette wheel operator for each ant, an antlion is selected.
It should be noted that in the ALO algorithm each ant can fall into only one antlion
trap in each iteration. The roulette wheel selected antlion for each ant is the one that
has trapped the ant. By using the roulette wheel operator, the solution with a better
tness function has more chance to be selected, as an antlion with a larger trap can
hunt more ants.

11.2.3 Sliding Ants Toward Antlion

When an ant falls into the trap, the antlion starts shooting sand outward the center of
the pit for sliding down the ant which is trying to escape. This behavior is math-
ematically modeled by shrinking the radius of the ants random walk. So the range
of boundary for all decision variables is decreased and updated, as expressed in
Eqs. (11.7) and (11.8) (Mirjalili 2015).
11 Ant Lion Optimizer (ALO) Algorithm 111

cIt
cIt 11:7
R
dIt
dIt ; 11:8
R

where c(It) = modied vector including the minimum of all decision variables at
the Itth iteration; c(It) = vector including the minimum of all decision variables at
the Itth iteration; R = a ratio given by Eq. (11.9); d(It) = modied vector including
the maximum of all decision variables at the Itth iteration; and d(It) = vector
including the maximum of all decision variables at the Itth iteration (Mirjalili 2015).

It
R 10w ; 11:9
IT

where w = a constant dened based on the iteration number given by:

8
>
> 2 If It [ 0:1IT
>
>
<3 If It [ 0:5IT
w 4 If It [ 0:75IT 11:10
>
>
>
> 5 If It [ 0:9IT
:
6 If It [ 0:95IT

When the iteration number in Eq. (11.10) increases, the radius of random walk
decreases, which guarantees convergence of the ALO algorithm.

11.2.4 Entrapping Ants Inside Pits

Antlion traps affect the random walk of ants. In order to mathematically model this
behavior, the boundary of ant random walk is adjusted in each iteration so that the
ant moves in a hyper-sphere around the selected antlion trap. The lower and upper
bounds of the ant random walk for each dimension in each iteration can be cal-
culated by the following equations (Mirjalili 2015):

cm It Antlionl It cIt 11:11

dm It Antlionl It dIt; 11:12

where cm It = vector including the minimum of all decision variables for the mth
ant in the Itth iteration; Antlionl It = position of the selected lth antlion in the Itth
iteration; and dm It = vector including the minimum of all decision variables for
the mth ant. Equations (11.11) and (11.12) show that ants random walk is in the
hyper-sphere, which is dened by vectors c and d around the roulette wheel selected
antlion.
112 M. Mani et al.

11.2.5 Random Walk of Ants

As ants move randomly in the nature for food, random walk is used for modeling
their movement, which can be expressed as the following equation (Mirjalili 2015):

X 0; cumsum2r1  1; cumsum2r2  1; . . .; cumsum2rIt

 1; . . .; cumsum2rIT  1; 11:13

where X = vector of random walk positions; cumsum = cumulative sum; and

r (It) = stochastic function which is calculated by:

1 if rand [ 0:5
rIt ; 11:14
0 otherwise

where rand = random value generated with a uniform distribution between 0 and 1.
Figure 11.3 shows three random walks in 100 iterations. From Fig. 11.3,
completely different behaviors of random walk can be observed.
To keep the random walks within the decision space, they are normalized by the
min-max normalization method (Mirjalili 2015):

Xn It  an dn It  cn It
Zn It ; 11:15
bn  an

where Zn(It) = normalized random walk position of the nth decision variable at the
Itth iteration; Xn(It) = random walk position of the nth decision variable at the Itth
iteration before normalization; an = minimum of random walks for the nth decision
variable; bn = maximum of random walks for the nth decision variable;
cn(It) = minimum of the nth decision variable at the Itth iteration; and
dn(It) = maximum of the nth decision variable at the Itth iteration.

Fig. 11.3 Three random walk curves in one dimension started at zero
11 Ant Lion Optimizer (ALO) Algorithm 113

11.2.6 Elitism

Elitism is an important aspect of the evolutionary algorithms which allows them to

keep the best solution during the optimization process. In the ALO algorithm, the
best antlion in each iteration is saved as the elite antlion and the average of both
random walks around the roulette wheel selected antlion and the elite is considered
for generating new positions of ants, as expressed in Eq. (11.16).

RItl RIte
AntItm ; 11:16
2

where AntItm = position of the selected mth antlion in the Itth iteration; RItl = random
walk around the lth roulette wheel selected antlion at the Itth iteration; and
RIte = random walk around the elite antlion at the Itth iteration.

11.2.7 Catching Prey and Reconstruct the Trap

At the nal step of antlion hunting, an ant falls into the bottom of the trap and is
caught by the antlions jaw. The antlion pulls the ant into sand and consumes it. In
the ALO algorithm, catching the prey occurs when an ants tness function
becomes better than its corresponding antlion. In this situation, the antlion changes
its position to the position of the hunted ant. This process can be mathematically
expressed as:
AntlionItl AntItm If f AntItm [ f AntlionItl : 11:17

11.3 Termination Criteria

In the evolutionary algorithms, at the end of each iteration the termination criterion
is applied to decide if the algorithm is to be stopped or continues the next iteration.
A good termination criterion should guarantee convergence of the algorithm. To
meet this purpose in the ALO algorithm by increasing the iteration number, the
radius of random walk decreases and the maximum number of iterations is con-
sidered as a termination criterion.

11.4 User-Dened Parameters of the ALO Algorithm

In the ALO algorithm, only the number of search agents and the number of iter-
ations are user-dened parameters. So, one of the main advantages of the ALO
algorithm is that it has very few parameters to be adjusted. Generally, the
114 M. Mani et al.

evolutionary algorithms require a large number of user-dened parameters and the

optimal settings of these parameters (to nd an appropriate solution in a reasonable
time) are different for each problem. Determining a good parameter setting often
requires numerous careful, time-consuming experiments. For the ALO algorithm,
however, only two user-dened parameters are required.

11.5 Pseudo-Code of the ALO Algorithm

Begin

Input parameters of the algorithm and initial data

Generate the initial population of ants and antlions randomly

Calculate the fitness function values for ants and antlions

Select the fittest antlion as the elite

While the iteration number is smaller than the maximum iteration number

For each ant

Select an antlion by utilizing the roulette wheel operator

Update the lower and upper bounds by using Equations (11.7) and (11.8)

Update and using Equations (11.11) and (11.12)

Generate two random walks around the roulette selected antlion and the elite
and normalize them by Equation (11.15)

Update the ant positions by Equation (11.16)

End for

Calculate the fitness function values for all ants and replace them in the
corresponding matrix

Replace an antlion with its corresponding ant if it becomes fitter [Equation (11.17)]

Update the elite value if an antlions fitness function value becomes better than the
elite

End while

Report the best solution (elite)

End
11 Ant Lion Optimizer (ALO) Algorithm 115

11.6 Conclusion

This chapter introduced the antlion optimizer (ALO), which is inspired by the
hunting behavior of antlions. In this chapter, after a brief literature review of the
ALO algorithm and an explanation of the hunting behavior of antlions, the algo-
rithmic fundamentals of the ALO algorithm are detailed and a pseudo-code is also
presented.

References

Ali, E. S., Elazim, S. A., & Abdelaziz, A. Y. (2016). Optimal allocation and sizing of renewable
distributed generation using Antlion optimization algorithm. Electrical Engineering. doi:10.
1007/s00202-016-0477-z
Dubey, H. M., Pandit, M., & Panigrahi, B. K. (2016). Antlion optimization for short-term wind
integrated hydrothermal power generation scheduling. International Journal of Electrical
Power & Energy Systems, 83(1), 158174.
Goodenough, J., McGuire, B., & Jakob, E. (2009). Perspectives on animal behavior (3rd ed.).
New York, USA: John Wiley and Sons.
Kamboj, V. K., Bhadoria, A., & Bath, S. K. (2016). Solution of non-convex economic load
dispatch problem for small-scale power systems using Antlion optimizer. Neural Computing
and Applications, 25(5), 112.
Kaur, M. & Mahajan, A. (2016). Community Detection in Complex Networks: A Novel Approach
Based on Antlion Optimizer. Sixth International Conference on Soft Computing for Problem
Solving, Punjab, India. December 2324.
Kaushal, K., & Singh, S. (2017). Allocation of stocks in a portfolio using Antlion algorithm:
Investors perspective. IUP Journal of Applied Economics, 6(1), 3449.
Mirjalili, S. (2015). The Ant lion optimizer. Advances in Engineering Software, 83(1), 8098.
Mirjalili, S., Jangir, P., & Saremi, S. (2017). Multi-objective Antlion optimizer: a multi-objective
optimization algorithm for solving engineering problems. Applied Intelligence, 46(1), 7995.
Petrovi, M., Petronijevi, J., Miti, M., Vukovi, N., Miljkovi, Z., & Babi, B. (2016). The
Antlion optimization algorithm for integrated process planning and scheduling. Applied
Mechanics and Materials, 834(1), 187192.
Rajan, A., Jeevan, K., & Malakar, T. (2017). Weighted elitism based Antlion optimizer to solve
optimum VAr planning problem. Applied Soft Computing, 55(1), 352370.
Raju, M., Saikia, L. C., & Sinha, N. (2016). Automatic generation control of a multi-area system
using Antlion optimizer algorithm based PID plus second order derivative controller.
International Journal of Electrical Power & Energy Systems, 80(1), 5263.
Saxena, P., & Kothari, A. (2016). Antlion optimization algorithm to control side lobe level and
null depths in linear antenna arrays. AEU-International Journal of Electronics and
Communications, 70(9), 13391349.
Talatahari, S. (2016). Optimum design of skeletal structures using Antlion optimizer. International
Journal of Optimization in Civil Engineering, 6(1), 1325.
Yamany, W. et al. (2015, September 2022). A new multi-layer perceptrons trainer based on
Antlion optimization algorithm. Fourth International Conference on Information Science and
Industrial Applications, Beijing, China.
116 M. Mani et al.

Yao, P., & Wang, H. (2016). Dynamic adaptive Antlion optimizer applied to route planning for
unmanned aerial vehicle. Soft Computing. doi:10.1007/s00500-016-2138-6
Zawbaa, H. M., Emary, E., & Grosan, C. (2016). Feature selection via chaotic antlion
optimization. PLoS ONE, 11(3), e0150652.
Zawbaa, H. M., Emary, E., & Parv, B. (2015, November 2325). Feature selection based on
antlion optimization algorithm. Third World Conference on Complex Systems, Marrakech,
Morocco.
Chapter 12
Gradient Evolution (GE) Algorithm

Mehri Abdi-Dehkordi, Omid Bozorg-Haddad and Xuefeng Chu

Abstract In this chapter, a meta-heuristic optimization algorithm named gradient

evolution (GE) is discussed, which is based on a gradient search method. First, the
GE algorithm and the underlying idea are introduced and its applications in some
studies are reviewed. Then, the mathematical formulation and a pseudo-code of GE
are discussed. Finally, the conclusion is presented.

12.1 Introduction

The gradient evolution (GE) algorithm is a meta-heuristic optimization algorithm

which is derived from a gradient-based optimization method. A gradient denes the
curve of a function at each point. Negative, zero, and positive gradients indicate the
decreasing, flat, and increasing functions, respectively. Accordingly, the optimal
solution is located at certain extreme point. This concept is the basis of the
gradient-based search methods including Newton, quasi-Newton, and conjugate
direction methods. Many researchers have applied the gradient-based methods or
their combinations with other methods in solving optimization problems. The
non-differentiable functions in complex problems often limit the application of the
gradient-based methods. Kuo and Zulvia (2015) developed the GE algorithm, a novel
meta-heuristic optimization algorithm which applied a modied gradient-based

M. Abdi-Dehkordi
O. Bozorg-Haddad (&)

Department of Irrigation and Reclamation Engineering, Faculty of Agricultural Engineering
and Technology, College of Agriculture and Natural Resources, University of Tehran,
3158777871 Karaj, Iran
e-mail: [email protected]
M. Abdi-Dehkordi
e-mail: [email protected]
X. Chu
Department of Civil and Environmental Engineering, North Dakota State University,
Dept 2470, Fargo, ND 58108-6050, USA
e-mail: [email protected]

Springer Nature Singapore Pte Ltd. 2018 117

O. Bozorg-Haddad (ed.), Advanced Optimization by Nature-Inspired Algorithms,
Studies in Computational Intelligence 720, DOI 10.1007/978-981-10-5221-7_12
118 M. Abdi-Dehkordi et al.

method as the main updating rule. The GE algorithm explored the search space of an
optimization problem using a set of vectors. Kuo and Zulvia also considered a set of
operators in order to enhance the ability of their model in nding the optimal solution.
They further evaluated the performance of the GE algorithm by using 15 benchmark
test functions in three stages. In the rst stage, the effects of changing parameters on
the obtained results of the GE algorithm were investigated and the best parameter
setting was determined. Then, the results of the GE algorithm were compared with
those from other meta-heuristic algorithms including particle swarm optimization
(PSO), differential evolution (DE), continuous genetic algorithm (GA), and articial
bee colony (ABC). The results indicated the better performance of the GE algorithm
than those of the other meta-heuristic algorithms.
Kuo and Zulvia (2016) also proposed a K-means clustering algorithm based on
the GE algorithm in order to derive the hidden stored information in data sets. The
reason of development of this new algorithm was the dependency of the K-means
algorithm on the initial centroids of clusters. In the proposed algorithm, the GE
algorithm was utilized to nd a good center for the K-means algorithm. The
algorithm was validated by a number of benchmark datasets and the obtained
results were compared with those from other meta-heuristic based K-means algo-
rithms, indicating the superiority of the GE based K-means algorithm over the other
meta-heuristic algorithms. Kuo and Zulvia (2016) nally proposed to consider the
similarities between clusters as well as the similarities within clusters in a
multi-objective structure.

12.2 Underlying Idea of the GE Algorithm

The GE algorithm is a population-based optimization method and its main rule is

updated by the gradient search methods (including the NewtonRaphson method
and Taylor series expansion), so it can be considered as a center differencing
approach. Consequently, this algorithm can be applied to different optimization
problems with differentiable and non-differentiable functions (Kuo and Zulvia
2015).

12.2.1 Gradient

If f(x) is a differentiable function at point x, changing x causes a change in f

(x) (Larson et al. 2007). In a single-variable function, the derivative is dened as the
slope of the tangent or zero gradient at point x. The gradient of function f(x) at a
point like x0 is determined by a secant line between f x0 and f x0 Dx
(Fig. 12.1).
12 Gradient Evolution (GE) Algorithm 119

Fig. 12.1 Gradient

determination

Thus, the gradient is determined by:

Dy f x0 Dx  f x0
m 12:1
Dx Dx

where m = gradient; and Dy = change of f(x) between points x0 and x0 Dx.

The values of f x0 Dx and f x0 become closer for a smaller value of Dx,
which increases the accuracy of calculation. So if Dx approaches zero, the secant
line is tangent of f(x) at point x0 . The derivative of f(x), which is the gradient at the
point x0 , is dened by:

f x0 Dx  f x0
m Lim
Dx 12:2
x!0

The information about the function contour or gradient can be obtained based on
the correlation between the derivative and the slope of the tangent (Miller 2011). An
optimization is dened as nding the optimal solution which maximizes or mini-
mizes an objective function. If f(x) is not strongly convex or concave, there will be
several stationary points x.
In addition, there is a maximum or minimum point in a flat part, which indicates
that the optimal solution is located at a stationary point. In fact, the optimal solution
of an optimization problem is located at a point with a zero gradient. So, the optimal
solution can be obtained by determination of the function contour. Accordingly, the
derivative, which is based on the limit [Eq. (12.2)], is an important concept in
optimization problems. Although many real-world problems have discrete vari-
ables, the limit concept just applies in continuous functions. Furthermore, for some
continuous functions, the calculation of limit is so difcult because of the com-
plexity of functions. In this case, derivative is calculated using numerical methods
such as Newton interpolation, Lagrange and cubic spine (Patil and Verma 2006).
The GE algorithm was developed using the following Taylor-series-based rst- and
second-order derivatives:
120 M. Abdi-Dehkordi et al.

f x Dx  f x  Dx
f 0 x 12:3
2 : Dx
f x Dx  2 f x f x  Dx
f 00 x 12:4
Dx2

where f 0 x and f 00 x = rst and second order derivatives of f(x), respectively (Kuo
and Zulvia 2015).

12.2.2 Gradient-Based Algorithm

Optimization methods can be divided into two main groups: direct search and
gradient-based search methods. Both groups start the search from a point and
evaluate the other points in order to nd an optimal solution till the stopping
criterion is satised. In the direct search methods (region eliminating), the contour
of a function is determined based on two or more points. Then, the search direction
is limited to the search space which has a better initial point. The search space is
decreased iteratively in different successful iterations in order to achieve the optimal
solution. Golden search and Fibonacci search are two direct search methods
(Bazaraa et al. 2013). The rst- and second-order derivatives are applied in the
function analysis using the gradient-based methods such as NewtonRaphson
(Ypma 1995). The concept of this method applied in the GE algorithm is presented
as follows.
In the NewtonRaphson method, the search process starts from an initial point
and continues its movement to the point with a zero gradient. If the search is located
at point xt in iteration t, in next iteration it will be at point xt 1 which is located Dxt
from xt . Since nding an extreme point is considered, the rst derivative must be
equal to zero. The Taylor series expansion is applied in order to estimate the rst-
and second-order derivatives [Eqs. (12.3) and (12.4)]. Furthermore, xt 1 is deter-
mined by:

Dxt f xt Dxt  f xt  Dxt

xt 1 xt  : 12:5
2 f x Dxt  f xt f xt  Dxt
t

The gradient-based methods apply numerical methods, instead of direct

derivation, due to non-differentiability of the functions of optimization problems.
The global gradient of point ! xt can be determined based on the sets of exper-
!
imental candidates ti i 1; 2; . . .; k in iteration t (Salomon 1998). In order to
!
achieve ti , a random mutation, !zi , is applied at current point !
xt . !
et includes all the
information related to the experimental candidates and points in the gradient esti-
mation direction.
12 Gradient Evolution (GE) Algorithm 121

!
!
et
gt
12:6
kgt k

k   
X  ! !
! !
gt f ti  f !
xt ti  xt ; 12:7
i1

where gt = gradient direction in each iteration.

This method is improved in the evolutionary gradient search method (EGS). In
this regard, the concept of central differencing is applied, instead of forward dif-
ferencing (Salomon 1998). In the pseudo-code of GE, a simpler method is applied
for gradient estimation (Wen et al. 2003). In the method, modication in direction
from ! xi to !
xj in Rn ; diri; j is dened by:
8
<1 if xik  xjk
diri; j 0 if xik xjk 8 k 1; 2; . . .; n 12:8
:
1 if xik  xjk

Although the GE algorithm applied a central differencing approach-based gra-

dient estimation method as well as EGS, the formulations of these two methods are
completely different (Kuo and Zulvia 2015).

12.3 Mathematical Formulation of the GE Algorithm

As aforementioned, the GE algorithm applies a gradient-based method to determine

the search direction. It is also based on population. This population includes a
number of vectors which indicate the possible solutions. All vectors are updated by
three operators: updating, jumping, and refreshing. Description of the application of
the GE algorithm is presented for a minimization optimization problem in RD as
follows.

12.3.1 Solution Representation and Algorithm Initialization

In an optimization problem with D variables, the GE algorithm encodes the initial

solution as a vector with D dimensions. Population Pt in iteration
 t is represented
by
 t   i 1; . . .; N
P Xi i 1; . . .; N which includes n vectors. Xi xij 
t t t
is the
j 1; . . .; D
ith vector in iteration t which corresponds a feasible solution for the problem. In the
initial step, it is necessary to determine six parameters including number of
122 M. Abdi-Dehkordi et al.

iterations T, number of vectors N, size of initial step k, rate of jump Jr, rate of
refreshing Sr, and rate of reduction e.
Determining the number of iterations and vectors depends on the complexity of
the problem. In complex problems, more iterations and more vectors are consid-
ered. Jr, Sr, and e are in the interval [0, 1]. Jr is used when there is considerable
modication in the vector direction. Sr and e also manage vector regeneration and
refreshing. Moreover, the value of e is effective for acceleration of vector
refreshing. It is necessary to determine the initial points for all vectors in the GE
algorithm. In this regard, a simplest method is employed to generate random
numbers.

12.3.2 Vector Updating

The updating rule, which controls the vector movement in order to reach a better
position, includes two gradient-based and acceleration factor parts. The rst part is
the core of updating rule and is derived from the gradient-based methods, which are
started from an arbitrary initial point and move gradually to the next point in a
certain direction determined by the gradient.
Movement to the points with better values of objective function in the search
direction is illustrated in Fig. 12.2. The GE algorithm also determines the search
space which has better solutions. Due to the complexity of the problem, the gradient
is not considered as the rst-order derivative of the objective function and the GE
algorithm applies central differencing instead. The search process of the GE algo-
rithm is shown in Fig. 12.3.
As shown in Fig. 12.3a, the GE algorithm explores the search space to achieve a
better area. If all the vectors in population move simultaneously to the same area,
the search space becomes narrow. If there is a distraction direction for vector
movement, the search process of the algorithm is performed in a wider range
(Fig. 12.3b). The gradient-based updating rule applies the NewtonRaphson
equation [Eq. (12.5)]. Since the main updating rule is individual-based, in order to

Fig. 12.2 Search direction

for the original gradient-based
method
Search direction

12 Gradient Evolution (GE) Algorithm 123

(a) (b)

tion tion
direc direc
rch rch
Sea Sea

Fig. 12.3 Search direction for the GE algorithm

(a) (b)

Fig. 12.4 Gradient approximation method modied from individual-based search to

population-based search: a individual-based, b population-based

apply this method in the GE algorithm, which is population-based, some modi-

cations are required. The updating rule in Eq. (12.5) includes the neighboring
points of X t which are X t  DX and X t DX. It is necessary to evaluate the
neighboring points in order to reach the next point X t 1 . This evaluation is time
consuming in population-based search for each vector. So, the GE algorithm
replaces points X t  DX and X t DX with the positions of two other vectors in
population.
The transformation from the individual-based updating rule to the
population-based updating rule is shown in Fig. 12.4. In individual-based
  search,
Xit has two neighboring points Xit DXit and Xit  DXit . If f Xit is a minimization
     
problem, f Xit  DXit  f Xit  f Xit DXit , which implies that point Xit  DXit is
better than point Xit and point Xit DXit is worse than point Xit . Since the GE
124 M. Abdi-Dehkordi et al.

algorithm is based on the population-based search method, there are many possible
solutions in population. In fact, in addition to the worst and best vectors, there is a
vector Xit which may have a neighboring with a worse or better position.
To update Xit in the GE algorithm, point Xit  DXit is replaced with vector
   
XiB [XiB 2 Pt ; f XiB  f Xit ] and Xit DXit is replaced with XiW
   
[XiW 2 Pt ; f XiW  f Xit ]. So, points Xit  DXit and Xit DXit are substituted by
vectors XiB and XiW , respectively. The GE algorithm also applies position Xit instead
 
of the tness of position f Xit because applying a tness value is time consuming.
In order to expand the search, a random number rg  N 0; 1 is added to the
updating rule. Considering rg ensures the distribution of vector movement in the GE
algorithm. The updating rule in Eq. (12.5) can be transformed to GradientMove by:
  !
Dxtij ij  xij
xW B
Gradient Move rg : : W ; 8j 1; . . .; D 12:9
2 xij  xtij xBij
   
 t   t
xij  xBij  xW
ij  x ij 
Dxtij ; 8j 1; . . .; D 12:10
2

The acceleration factor Acc is used to accelerate the convergence of each vector.
This process uses the best vector of a direction which is expressed by Eq. (12.11).
In the GE algorithm, it is assumed that the best vector is the vector closest to the
optimal solution. So, all other vectors will move in a better direction by considering
the position of the best vector. Similar to the gradient updating rule, the acceleration
factor in Eq. (12.11) is also multiplied by a random number ra  N 0; 1, which
ensures different sizes of steps for each vector (Kuo and Zulvia 2015).
 
Acc ra : yi  xtij ; 8j 1; . . .; D; 12:11

where Y fyi jj 1; 2; . . .; Dg, the best vector. Finally, the vector updating is
conducted by (Kuo and Zulvia 2015):

utij xtij  GradientMove Acc

  !
Dxtij   12:12
ij  xij
xW B
uij xij  rg :
t t
: W ra y i  x t
8j 1; . . .; D
2 xij  xtij xBij ij

where utij 2 Uit = transition vector which is obtained by updating Xit . Since vector
Xit includes worse and better vectors, an additional process is necessary in order to
determine the worst and best vectors in Pt because the main updating rule includes
the neighboring vectors which have a worse or better tness value. The worst and
best vectors do not have any neighboring vector which is worse or better than
themselves. In this regard, an additional process is required to determine the gra-
dient of the objective function.
12 Gradient Evolution (GE) Algorithm 125

If W t and Bt are the worst and best vectors of Xit , respectively, the values of xW
ij
and xBij can be replaced with wj and bj , respectively by Eqs. (12.13)(12.16) (Kuo
and Zulvia 2015):

bj xtij  Dxtij 8j 1; . . .; D 12:13

 
 t
r  xW
ij  x ij 
Dxtij 8j 1; . . .; D 12:14
2
wj xtij Dxtij 8j 1; . . .; D 12:15
 
 t 
xij  xBij  c
Dxtij 8j 1; . . .; D 12:16
2

where c = size of the initial step which is predened. c can be a static or dynamic
number. It can be decreased by increasing the number of iterations if it is a dynamic
number. There are two ways for solving a maximization problem: (1) switching
worse and better neighbors and (2) transforming into a minimization problem.

12.3.3 Vector Jumping

An appropriate search method must be able to explore the search space widely and
deeply. The vector updating and vector jumping operators focus on deep and wide
search, respectively. In the GE algorithm, the vector jumping operator is applied to
avoid local optima. This operator just performs on a selective vector and modies
the movement direction. In the GE algorithm,  Jr is considered
 for determining
whether or not the vector must jump. If rj  Jr rj  N 0; 1 , vector jumping to a
transition vector Uit is given by (Kuo and Zulvia 2015):
 
utij utij rm : utij  xtkj 8j 1; . . .; D; 12:17

where xtkj 2 Xkjt in each random neighboring vector Pt ; 8i 6 k, and rm = random

number in the interval (0, 1). The process of vector jumping is schematically shown
in Fig. 12.5.
126 M. Abdi-Dehkordi et al.

Fig. 12.5 Vector jumping

operator
t
Xk

t
Xi
t
t rm X t X k
i Xt
X i k
t X t
t Xk
t rm X i
i X t
k
Xi

12.3.4 Vector Refreshing

The GE algorithm uses an elitist strategy. In iteration t, the transition vector Uit
records the updated results and the jumped vector Xit . The next position of vector i,
 
Xit 1 , is replaced by Uit when the tness value f Uit is better than tness value
 
f xti . Otherwise, it remains at Xit in next iteration (i.e., Xit 1 Xit ). Using the elitist
strategy, the GE algorithm ensures that each vector always moves to a better
position. If the determination of a better position is difcult, problem arises. This
situation takes place in the complex problems and the problems with many local
optima. In this case, vector refreshing is performed for such a problematic vector.
The GE algorithm records the position of vector Xit and the history of vector
updating. The history of vector i which is si 2 0; 1, provides information about the
number of iterations that the vector cannot move to a better position. The newly
generated vector si is set to one. When vector i is stuck in the same position, si is
reduced by:

si si  e:si ; 12:18

where e = rate of reduction. The predened parameters Sr and e are continuous

numbers in the interval [0, 1]. They will be set when vector regeneration is required.
If si is less than Sr, vector i must be regenerated (Kuo and Zulvia 2015). The
flowchart of the GE algorithm is shown in Fig. 12.6. The parameters and variables
used in the GE algorithm are presented in Table 12.1.
12 Gradient Evolution (GE) Algorithm 127

Start

Set GE algorithm parameters

For each vector X it

Vector updating U it

Yes
rj Jr ? U it = vector jumping

No

Calculate fitness U it

No X it = X it 1
f ( U it ) f ( X it 1 ) ?
s i = s i .s i
Yes

X it = U it

Yes Vector refreshing

si = S r ?

No
i

No
Is stopping criterion
satisfied?
Yes
Select best vector

End

Fig. 12.6 Flowchart of the GE algorithm

128 M. Abdi-Dehkordi et al.

Table 12.1 Characteristics of the GE algorithm

General algorithm Gradient Evolution Algorithm
Decision variable Vector which indicates the possible solutions
Solution Movement to the point with a zero gradient
Old solution Old vector of possible solutions
New solution New vector of possible solutions
Best solution Vector with the best tness function
Fitness function Gradient obtained by a gradient-based method
Initial solution Randomly generated vector
Selection
Process of generating new solution Vector updating, vector jumping, and vector refreshing

12.4 Pseudo-Code of GE
12 Gradient Evolution (GE) Algorithm 129
130 M. Abdi-Dehkordi et al.

12.5 Conclusion

This chapter introduced the gradient evolution (GE) algorithm, a population-based

meta-heuristic algorithm based on the concept of gradient. The gradient-based
optimization methods evaluate the search space based on an individual-based
approach. In order to apply these methods in the population-based algorithms,
considerable modications must be performed. The GE algorithm explores the
search space based on a set of vectors in optimization problems and applies three
main operators including vector updating, vector jumping, and vector refreshing.
The vector updating operator based on the Tylor series expansion theorem trans-
forms the updating rule for population-based search. The vector jumping operator
avoids local optima and the vector refreshing operator is applied when a vector
cannot move to another position in several iterations. This chapter detailed the
underlying idea and the mathematical formulation of the GE algorithm.

References

Bazaraa, M. S., Sherali, H. D., & Shetty, C. M. (2013). Nonlinear programming: Theory and
algorithms (3rd ed.). New Jersey, USA: Wiley.
Kuo, R. J., & Zulvia, F. E. (2015). The gradient evolution algorithm: A new metaheuristic.
Information Sciences, 316, 246265.
Kuo, R. J., & Zulvia, F. E. (2016, July 2529). Cluster analysis using a gradient evolution-based
k-means algorithm. In IEEE congress on evolutionary computation (CEC). Beijing, China:
Peking University.
Larson, R., Hostetler, R. P., & Edwards, B. H. (2007). Essential Calculus: Early Transcendental
Functions (1st ed.). New York, USA: Houghton Mifflin Company.
Miller, H. R. (2011). Optimization: Foundations and applications (1st ed.) New York, USA:
Wiley.
Patil, P. B., & Verma, U. P. (2006). Numerical computational methods (1st ed.). Oxford, UK:
Alpha Science International.
Salomon, R. (1998). Evolutionary algorithms and gradient search: Similarities and differences.
IEEE Transactions on Evolutionary Computation, 2(2), 4555.
Wen, J. Y., Wu, Q. H., Jiang, L., & Cheng, S. J. (2003). Pseudo-gradient based evolutionary
programming. Electronics Letters, 39(7), 631632.
Ypma, T. J. (1995). Historical development of the Newton-Raphson method. SIAM Review, 37(4),
531551.
Chapter 13
Moth-Flame Optimization
(MFO) Algorithm

Mahdi Bahrami, Omid Bozorg-Haddad and Xuefeng Chu

Abstract This chapter introduces the Moth-Flame Optimization (MFO) algorithm,

along with its applications and variations. The basic steps of the algorithm are
explained in detail and a flowchart is represented. In order to better understand the
algorithm, a pseudocode of the MFO is also included.

13.1 Introduction

The MFO is a novel nature-inspired optimization algorithm based on the navigation

method of moths through night by maintaining a xed angle with respect to the
Moon. Introduced by Mirjalili (2015) and tested on 29 benchmark functions and
seven real engineering problems, MFO showed promising results in comparison to
other nature-inspired algorithms. Since being introduced, MFO has been applied to
real engineering problems in different elds of research.
Yamany et al. (2015) trained a multi-layer perceptron (MLP) using MFO.
MFO-MLP was used to search for the weights and biases of the MLP to achieve the
minimum error and the highest classication rate. Li et al. (2016a) proposed a new
hybrid annual power load forecasting model based on the least squares support
vector machine (LSSVM) and the MFO algorithm to forecast the annual power load
essential for the planning, operation, and maintenance of an electric power system.

M. Bahrami
O. Bozorg-Haddad (&)

Department of Irrigation and Reclamation Engineering, Faculty of Agricultural Engineering
and Technology, College of Agriculture and Natural Resources, University of Tehran,
31587-77871 Karaj, Tehran, Iran
e-mail: [email protected]
M. Bahrami
e-mail: [email protected]
X. Chu
Department of Civil and Environmental Engineering, North Dakota State University,
Dept 2470, Fargo, ND 58108-6050, USA
e-mail: [email protected]

Springer Nature Singapore Pte Ltd. 2018 131

O. Bozorg-Haddad (ed.), Advanced Optimization by Nature-Inspired Algorithms,
Studies in Computational Intelligence 720, DOI 10.1007/978-981-10-5221-7_13
132 M. Bahrami et al.

The parameters for LSSVM were optimally determined using the MFO algorithm.
Raju et al. (2016) used MFO for simultaneous optimization of secondary controller
gains in a cascade controller proposed for automatic generation control of a
two-area hydro-thermal system under a deregulated scenario. Zawbaa et al. (2016)
proposed a feature selection algorithm based on MFO and applied it to machine
learning for feature selection to nd the optimal feature combination using the
wrapper-based feature selection mode. MFO was exploited as a searching method
to nd the optimal feature set, maximizing classication performance. Ceylan
(2016) used MFO to solve the harmonic elimination problem and minimize the total
harmonic distortion. The simulation results showed that the MFO model solved the
harmonic elimination problem and the total harmonic distortion minimization
problem efciently. Lal and Barisal (2016) applied MFO to evaluate the optimal
gains of the fuzzy-based proportional, integral and derivative (PID) controllers in a
microgird power generation system interconnected with a single area reheat thermal
power system. Gope et al. (2016) used MFO to obtain an optimal bidding strategy
of supplier considering double sided bidding under a congested power system.
Jangir et al. (2016) solved ve constrained benchmark functions of engineering
problems using MFO and compared the results with other recognized optimization
algorithms. MFO provided better results in various design problems in comparison
to other optimization algorithms. Parmar et al. (2016) solved the optimal power
flow (OPF) problem using MFO involving fuel cost reduction, active power loss
minimization, and reactive power loss minimization. Comparing with other tech-
niques such as flower pollination algorithm (FPA) and particle swarm optimization
(PSO), MFO showed a better performance. Bentouati et al. (2016) applied MFO to
solve the problem of OPF in the interconnected power system for the Algerian
power system network with different objective functions. The results were com-
pared with those obtained by articial bee colony (ABC) and other metaheuristics.
Allam et al. (2016) utilized MFO for the parameter extraction process of the
three-diode model for the multi-crystalline solar cell\module, with the results being
compared with those obtained by FPA and hybrid evolutionary (DEIM) algorithms.
The results showed that MFO achieved the least root mean square error (RMSE),
mean bias error (MBE), absolute error at the maximum power point (AEMPP), and
best coefcient of determination. Buch et al. (2017) applied MFO to various
nonconvex, nonlinear optimum power flow objective functions with ve single
objective functions. Comparing MFO with other stochastic methods showed that
MFO obtained the optimum value with rapid and smooth convergence. Garg and
Gupta (2017) used MFO to optimize the performance of open shortest path rst
(OSPF) algorithm which is the widely used, efcient algorithm to select the shortest
path between the source and destination. The results for different scenarios showed
the reduction in delay and energy consumption in the optimized OSPF compared to
the traditional OSPF. Khalilpourazari and Pasandideh (2017) solved a
multi-constrained Economic Order Quantity (EOQ) model using MFO and the
interior-point method. To compare the results of the two methods, three measures,
including objective function value, computation time, and the number of function
evaluations were used. The results indicated that there was no signicant difference
13 Moth-Flame Optimization (MFO) Algorithm 133

between the average objective function values of MFO and the interior-point
method, but MFO required signicantly less computation time and fewer function
evaluations.
Different versions of the MFO algorithm have been developed by other
researchers in order to improve the performance of the algorithm in different elds
of research. Bhesdadiya et al. (2016) developed a hybrid optimization algorithm
based on PSO and MFO. PSO was used for the exploitation and MFO was utilized
for the exploration phase. The proposed algorithm was tested on some unconstraint
benchmark test functions along with some constrained/complex design problems
and the obtained results demonstrated its effectiveness comparing to the standard
PSO and MFO algorithms. Nanda (2016) modied the original MFO to handle
multi-objective optimization problems. The proposed MOMFO used concepts such
as the archive grid, coordinate based distance for sorting and non-dominance of
solutions. In the tests of six benchmark mathematical functions, MOMFO achieved
better accuracy and shorter computational time than non-dominated Sorting Genetic
Algorithm-II (NSGA-II) and Multi-objective PSO (MOPSO). Muangkote et al.
(2016) proposed an improved version of MFO for image segmentation to enhance
the optimal multilevel thresholding of satellite images. The proposed multilevel
thresholding moth-flame optimization algorithm (MTMFO) was tested for various
satellite images. MTMFO provided more effective results with better accuracy than
MFO.
Soliman et al. (2016) proposed two modied versions of MFO and used as a
prediction tool for terrorist groups, and compared with the original MFO as well as
ant lion optimizer (ALO), grey wolf optimization (GWO), PSO, and genetic
algorithm (GA). The results proved that the modied versions of MFO achieved an
advance over the original MFO algorithm. Li et al. (2016b) proposed an improved
version of MFO based on the Lvy-flight strategy (LMFO) to improve the con-
vergence and precision of MFO. The new LMFO algorithm increased the diversity
of the population against premature convergence and made the algorithm jump out
of local optimum more effectively. Compared with MFO and other heuristic
methods, LMFO demonstrated its superior performance. Trivedi et al. (2016)
applied MFO to economic load dispatch problems. They integrated MFO with Lvy
flights to achieve the competitive results in case of both discrete and continuous
control parameters.

13.2 Mapping the Navigation Method of Moths

into Moth-Flame Optimization

MFO is a nature-inspired optimization algorithm based on the moths navigation

mechanism in the night, known as the transverse orientation. Moths can travel long
distances in a straight line by maintaining a xed angle with the Moon (Frank 2006).
134 M. Bahrami et al.

Fig. 13.1 Moths spiral

flying path around a light
source

When encountered articial lights, moths try to maintain a similar angle to the light
source and because of the close distance they get caught in a spiral path (Fig. 13.1).
The MFO assigns moths to different solutions in the solution space of the
optimization problem with each moth having its own tness function value. Each
moth also has a flame which stores the best solution found by that moth. In each
iteration, the moths search the solution space by flying through a spiral path toward
their flames and update their positions.
MFO starts with the positions of moths randomly initialized within the solution
space. The tness values of the moths are calculated which are the best individual
tness values so far. The flame tags the best individual position for each moth. In
the next iteration, the moths positions are updated based on a spiral movement
function toward their best individual positions tagged by a flame and the positions
of the flames are updated with new best individual positions. The MFO algorithm
continues updating the positions of the moths and flames and generating new
positions until the termination criteria are met. Table 13.1 lists the characteristics of
the MFO. Figure 13.2 shows the flowchart of the MFO.
13 Moth-Flame Optimization (MFO) Algorithm 135

Table 13.1 Characteristics of the MFO algorithm

General algorithm Moth-flame optimization
Decision variable Moths positions in each dimension
Solution Moths positions
Old solution Old positions of moths
New solution New positions of moths
Best solution Positions of flames
Fitness function Distance between moth and flame
Initial solution Random positions of moths
Selection _
Process of generating new solution Flying in a spiral path toward a flame

Start

Define the parameters of the algorithm

Generate initial moths randomly

Calculate the fitness functions and tag the best positions by flames

Update flame number, t, and r

Calculate D for the corresponding moth

Update M(i, j) for the corresponding moth

No Yes
Are termination criteria satisfied?

Report the best position among the moths

End

Fig. 13.2 Flowchart of the MFO algorithm

136 M. Bahrami et al.

13.3 Creating the Initial Population of Moths

In MFO, each moth is assumed to have a position in a D-dimensional solution

space. The set of moths can be expressed as (Mirjalili 2015):
2 3
m1;1 m1;2 ... . . . m1;d
6 m2;1 m2;2 ... . . . m2;d 7
6 7
M6
6 : : : : : 7 7 13:1
4 : : : : : 5
mn;1 mn;2 ... . . . mn;d

in which n = number of moths; and d = number of dimensions in the solution

space. The corresponding tness function values for the moths are sorted in an array
as follows (Mirjalili 2015):
23
OM1
6 OM2 7
6 7
OM 6
6 : 7
7 13:2
4 : 5
OMn

Two other components of the MFO are the flame matrix representing the flames
in the D-dimensional space and their corresponding tness function vector, which
can be respectively expressed as (Mirjalili 2015):
2 3
F1;1 F1;2 ... ... F1;d
6 F2;1 F2;2 ... ... F2;d 7
6 7
F6
6 : : : : : 7 7 13:3
4 : : : : : 5
Fn;1 Fn;2 ... ... Fn;d
2 3
OF1
6 OF2 7
6 7
OF 6
6 : 7
7 13:4
4 : 5
OFn

In MFO, moths and flames represent solutions, with moths searching the solu-
tion space in each iteration to nd the optimal solution and the flames representing
the best solution found by each moth. In other words, each moth searches the space
around its flame and each time it nds a better solution. The position of the flame is
then updated.
13 Moth-Flame Optimization (MFO) Algorithm 137

13.4 Updating the Moths Positions

MFO uses three functions to initialize the random positions of the moths (I), move
the moths in the solution space (P), and terminate the search operation (T):

MFO I; P; T 13:5

Any random distribution can be used to initialize the moths positions in the
solution space. The implementation of the I function can be written as (Mirjalili
2015):

Mi; j ubi  lbj  rand lbi 13:6

in which ub and lb = arrays that respectively dene the upper and lower bounds of
variables.
The movement of moths in the solution space is based on the transverse ori-
entation and modeled by using a logarithmic spiral subjected to the following
conditions (Mirjalili 2015):
Spirals initial point should start from the moth
Spirals nal point should be the position of the flame
Fluctuation of the range of spiral should not exceed the search space.
Hence, the P function for the movement is dened as:

SMi ; Fj Di :ebt : cos2p t Fj 13:7

in which b = a constant to dene the shape of the logarithmic spiral; t = a random

number between [1, 1]; and Di = distance between the i-th moth and the j-th
flame, which is dened as (Mirjalili 2015):
 
D i F j  M i  13:8

The spiral movement of the moth around the flame guarantees the exploration
and exploitation of the solution space. In order to prevent the moths getting trapped
in local optima, the best solutions (flames) are sorted in each iteration and each
moth flies around its corresponding flame based on the OF and OM matrices. In
other words, the rst moth flies around the best obtained solution, while the last
moth flies around the worst obtained solution.
138 M. Bahrami et al.

13.5 Updating the Number of Flames

In order to improve the exploitation of the MFO algorithm, Eq. (13.9) is used to
decrease the number of the flames, and hence the moths only fly around the best
solution in the nal steps of the algorithm (Mirjalili 2015):
 
Nl
Flame no round N  l  13:9
T

in which l = current number of iterations; N = maximum number of flames; and

T = maximum number of iterations. The decrease in the number of the flames
balances the exploration and exploitation of the solution space.

13.6 Termination Criteria

The termination criterion determines when the algorithm is terminated. Selecting a

good termination criterion has an important role to ensure a correct convergence of
the algorithm. The number of iterations, the amount of improvement, and the
running time are common termination criteria for the MFO.

13.7 Performance of the MFO

Mirjalili (2015) tested MFO on 29 benchmark functions and seven real engineering
problems, and compared the results with those obtained by other well-known
nature-inspired algorithms such as PSO, gravitational search algorithm (GSA), bat
algorithm (BA), FPA, states of matter search (SMS), refly algorithm (FA), and
GA. MFO showed promising and competitive results for the benchmark test
functions and the results of the real problems demonstrated the MFOs ability in
dealing with challenging problems with constrained and unknown search spaces.
13 Moth-Flame Optimization (MFO) Algorithm 139

13.8 Pseudocode of the MFO

Begin

Input parameters of the algorithm and the initial data

Initialize the positions of moths and evaluate their fitness values

While (the stop criterion is not satisfied or I < Imax)

Update flame no.

= Fitness Function( )

If iteration = 1

= sort( )

= sort( )

Else

= sort( 1, )

= sort( 1, )

End if

For = 1: N

For = 1: D

Update and

Calculate with respect to the corresponding moth

Update ( , ) with respect to the corresponding moth

End for j
End for i

End While

Post-processing the results and visualization.

End
140 M. Bahrami et al.

13.9 Conclusion

This chapter described moth-flame optimization (MFO), a new nature-inspired

algorithm based on the transverse orientation method used by moths. MFO searches
the decision space using a set of moths which report the tness functions at each
time step and tag the best solution by a flame. The movement of moths is based on
the flames in a spiral path around their flames. In this chapter, a literature review of
MFO was presented, showing the success of the algorithm for different optimization
problems, along with different variations of the MFO algorithm developed by
researchers. The flowchart and the pseudocode of MFO were also presented to help
understand the detailed computational procedures of the algorithm.

References

Allam, D., Yousri, D. A., & Eteiba, M. B. (2016). Parameters extraction of the three diode model
for the multi-crystalline solar cell/module using Moth-Flame Optimization Algorithm. Energy
Conversion and Management, 123, 535548.
Bentouati, B., Chaib, L., & Chettih, S. (2016). Optimal Power Flow using the Moth Flam
Optimizer: A case study of the Algerian power system. Indonesian Journal of Electrical
Engineering and Computer Science, 1(3), 431445.
Bhesdadiya, R. H., Trivedi, I. N., Jangir, P., Kumar, A., Jangir, N., & Totlani, R. (2016, August
1213). A novel hybrid approach particle swarm optimizer with moth flame optimizer
algorithm. In International Conference on Computer, Communication and Computational
Sciences (ICCCCS), Advances in Intelligent Systems and Computing. Ajmer, India.
Buch, H., Trivedi, I. N., & Jangir, P. (2017). Moth flame optimization to solve optimal power flow
with non-parametric statistical evaluation validation. Cogent Engineering, 4(1).
Ceylan, O. (2016, November 35). Harmonic elimination of multilevel inverters by moth-flame
optimization algorithm. In International Symposium on Industrial Electronics (INDEL).
Republic of Srpska, Bosnia and Herzegovina: IEEE.
Frank, K. D. (2006). Effects of articial night lighting on moths. In C. Rich & T. Longcore (Eds.),
Ecological consequences of articial night lighting (pp. 305344). Washington, DC: Island
Press.
Garg, P., & Gupta, A. (2017). Optimized open shortest path rst algorithm based on moth flame
optimization. Indian Journal of Science and Technology, 9(48).
Gope, S., Dawn, S., Goswami, A. K., & Tiwari, P. K. (2016, November 2225). Moth Flame
Optimization based optimal bidding strategy under transmission congestion in deregulated
power market. In Region 10 Conference (TENCON). Marina Bay Sands, Singapore: IEEE.
Jangir, N., Pandya, M. H., Trivedi, I. N., Bhesdadiya, R. H., Jangir, P., & Kumar, A. (2016, March
56). Moth-Flame Optimization algorithm for solving real challenging constrained engineering
optimization problems. In Students Conference on Electrical, Electronics and Computer
Science (SCEECS). Bhopal, India: IEEE.
Khalilpourazari, S., & Pasandideh, S. H. R. (2017). Multi-item EOQ model with nonlinear unit
holding cost and partial backordering: Moth-flame optimization algorithm. Journal of
Industrial and Production Engineering, 34(1), 4251.
Lal, D. K., & Barisal, A. K. (2016, December 2728). Load frequency control of AC microgrid
interconnected thermal power system. In International Conference on Advanced Material
Technologies (ICAMT). Andhra Pradesh, India.
13 Moth-Flame Optimization (MFO) Algorithm 141

Li, C., Li, S., & Liu, Y. (2016a). A least squares support vector machine model optimized by
moth-flame optimization algorithm for annual power load forecasting. Applied Intelligence, 45
(4), 11661178.
Li, Z., Zhou, Y., Zhang, S., & Song, J. (2016b). Lvy-flight moth-flame algorithm for function
optimization and engineering design problems. Mathematical Problems in Engineering.
doi:10.1155/2016/1423930.
Mirjalili, S. (2015). Moth-flame optimization algorithm: A novel nature-inspired heuristic
paradigm. Knowledge-Based Systems, 89, 228249.
Muangkote, N., Sunat, K., & Chiewchanwattana, S. (2016, July 1315). Multilevel thresholding
for satellite image segmentation with moth-flame based optimization. In The 13th International
Joint Conference on Computer Science and Software Engineering. Khon Kaen, Thailand.
Nanda, S. J. (2016, September 2124). Multi-objective Moth Flame Optimization. In Advances in
Computing, Communications and Informatics (ICACCI). Jaipur, India: IEEE.
Parmar, S. A., Pandya, M. H., Bhoye, M., Trivedi, I. N., Jangir, P., & Ladumor, D. (2016, April 7
8). Optimal active and Reactive Power dispatch problem solution using Moth-Flame Optimizer
algorithm. In International Conference on Energy Efcient Technologies for Sustainability
(ICEETS). Nagercoil, India: IEEE.
Raju, M., Saikia, L. C., & Saha, D. (2016, November 2225). Automatic generation control in
competitive market conditions with moth-flame optimization based cascade controller. In
Region 10 Conference (TENCON). Marina Bay Sands, Singapore: IEEE.
Soliman, G. M. A., Khorshid, M. M. H., & Abou-El-Enien, T. H. M. (2016, July). Modied
moth-flame optimization algorithms for terrorism prediction. International Journal of
Application or Innovation in Engineering and Management, 5, 4758.
Trivedi, I. N., Kumar, A., Ranpariya, A. H., & Jangir, P. (2016, April 78). Economic Load
Dispatch problem with ramp rate limits and prohibited operating zones solve using Levy Flight
Moth-Flame optimizer. In International Conference on Energy Efcient Technologies for
Sustainability (ICEETS). Nagercoil, India.
Yamany, W., Fawzy, M., Tharwat, A., & Hassanien, A. E. (2015, December 2930). Moth-flame
optimization for training multi-layer perceptrons. In 11th International Computer Engineering
Conference (ICENCO). Giza, Egypt: IEEE.
Zawbaa, H. M., Emary, E., Parv, B., & Sharawi, M. (2016, July 2429). Feature selection
approach based on moth-flame optimization algorithm. In Evolutionary Computation (CEC).
IEEE.
Chapter 14
Crow Search Algorithm (CSA)

Babak Zolghadr-Asli , Omid Bozorg-Haddad and Xuefeng Chu

Abstract The crow search algorithm (CSA) is novel metaheuristic optimization

algorithm, which is based on simulating the intelligent behavior of crow flocks.
This algorithm was introduced by Askarzadeh (2016) and the preliminary results
illustrated its potential to solve numerous complex engineering-related optimization
problems. In this chapter, the natural process behind a standard CSA is described at
length.

14.1 Introduction

In the last several decades, optimization played a crucial role in many aspects of
various problems, including but not limited to engineering problems. Often, such
problems include complicated objective functions, numerous decision variables,
and a considerable number of constraints, which adds complexity to an already
complicated optimization problem. The aforementioned characteristics limit the
efciency of traditional optimization techniques. Consequently, the search for an
alternative method leads to a new eld of studyswarm intelligence (SI), which
was introduced by Beni and Wang in the late 1980s (Bei and Wang 1993). SI,
ultimately, aims to imitate the social intelligence of the natures group living
creatures (Bonabeau et al. 1999). Each newly proposed algorithm attempts to

B. Zolghadr-Asli
O. Bozorg-Haddad (&)

Department of Irrigation and Reclamation Engineering, Faculty of Agricultural Engineering
and Technology, College of Agriculture and Natural Resources, University of Tehran,
3158777871 Karaj, Iran
e-mail: [email protected]
B. Zolghadr-Asli
e-mail: [email protected]
X. Chu
Department of Civil and Environmental Engineering, North Dakota State University,
Dept 2470, Fargo, ND 58108-6050, USA
e-mail: [email protected]

Springer Nature Singapore Pte Ltd. 2018 143

O. Bozorg-Haddad (ed.), Advanced Optimization by Nature-Inspired Algorithms,
Studies in Computational Intelligence 720, DOI 10.1007/978-981-10-5221-7_14
144 B. Zolghadr-Asli et al.

improve two main features: (1) decreasing the distance between the reported
solutions and the actual global optima; and/or (2) reducing the solution searching
time. Although each proposed optimization algorithm has its unique characteristics,
with both merits and drawbacks, it has been proven that there is no single algorithm
that could outperform all its rivals (Wolpert and Macready 1997). Subsequently, a
wide range of alternative novel optimization algorithms have been proposed, each
of which has its exclusive advantages.
One of these newly proposed algorithms is the crow search algorithm (CSA),
which was initially introduced by Askarzadeh (2016). CSA attempts to imitate the
social intelligence of crow flock and their food gathering process. The primary
results illustrated the improved efciency of CSA over many conventional opti-
mization algorithms, such as genetic algorithm (GA), particle swarm optimization
(PSO), and harmony search (HS), in both convergence time and the accuracy of the
results (Askarzadeh 2016). Ultimately, it can be concluded that CSA is a proper
alternative method for solving complex engineering optimization problems.

14.2 Crow Flocks Food Gathering Imitation

Crows are a widely distributed genus of birds, which have been credited with
intelligence throughout folklore. Recent experiments investigating the cognitive
abilities of crows have begun to reveal the intelligence capability of these species
(Emery and Clayton 2004, 2005; Prior et al. 2008). The studies have demonstrated
that some species of crows are not only superior in intelligence to other birds but
also rival many nonhuman primates. Observations of the crows tool use in the wild
are an example of their complex cognition (Emery and Clayton 2004). Further
studies have also revealed their self-awareness, face recognition capabilities,
sophisticated communication techniques, and food storing and retrieving skills
across seasons (Emery and Clayton 2005; Prior et al. 2008).
Interestingly, a crow individual has a tendency to tap into the food resources of
other species, including the other crow members of the flock. In fact, each crow
attempts to hide its excess food in a hideout spot and retrieve the stored food in the
time of need. However, the other members of the flock, which have their own food
reservation spots as well, try to tail one another to nd these hiding spots and
plunder the stored food. Nevertheless, if a crow senses that it has been pursuited by
other members of the flock, in order to lose the tail and deceive the plunderer, it
maneuvers its path into a fallacious hideout spot (Clayton and Emery 2005).
Plainly, the aforementioned is the core principles of the CSA, in which each crow
individual searches the decision space for hideouts with the best food resources
(i.e., the global optima from the point of view of optimization). Thus, each crow
individuals motion is induced by two main features: (1) nding the hideout spots
of the other members of the flock; and (2) protecting its own hideout spots.
In the standard CSA, the flock of crows spread and search throughout the
decision space for the perfect hideout spots (global optima). Since any efcient
14 Crow Search Algorithm (CSA) 145

optimization algorithm should be compatible with arbitrary dimensions and each

arbitrary dimension is to represent a decision variable, a d-dimensional environment
is assumed for the search space. Initially, it is assumed that N crow individuals (the
flock size) occupy a position in the d-dimensional space, randomly. The position of
the ith crow individual at the tth iteration in the search space is represented by x(i,t),
which is, in fact, a feasible array of the decision variables. Additionally, each crow
individual can memorize the location of the best encountered hideout spot. At the
tth iteration, the position of the hideout spot of the ith crow individual is represented
by m(i,t), which is the best position that the ith crow individual has spotted, so far.
Subsequently, each crow individual shall make a motion based on the two basic
principles of the CSA: (1) protecting its own hideout spot; and (2) detecting the
other members hideout spots. Assume that at the tth iteration, the jth crow indi-
vidual attempts to retrieve food from its hideout spot [m(j,t)], while the ith crow
decides to tail the jth crow individual, in order to plunder its stored food. In such
circumstances, two situations may occur: (1) the jth crow individual could not
detect that it has been tailed leading to the reveal of the hideout spot to the ith crow
individual; or (2) the jth crow individual senses the presence of a plunderer, which
leads to a deceiving maneuver by the jth crow.
In the rst case, the lack of attention of the jth crow individual would enable the
ith crow to spot and plunder the jth crows hideout spot. In such a case, the
repositioning of the ith crow can be obtained as follows (Askarzadeh 2016):
 
xi;t 1 xi;t ri  fli;t  mj;t  xi;t 14:1

in which ri = a random number with the uniform distribution and the range of [0,1];
and fl(i,t) = flight length of the ith crow individual at the tth iteration.
It is worth to be mentioned that fl(i,t) is one of the algorithms parameters and it
can affect the searching capability of the algorithm. Assume that smaller values of fl
lead to the local search at the vicinity of x(i,t), while larger values of fl would widen
the searching space. In terms of optimization, smaller values of fl would help
intensify the results, while larger values of fl would diversify the results. Both
well-intensication and -diversication are the characteristics of an efcient opti-
mization algorithm (Gandomi et al. 2013).
There could also be the case, in which the jth crow individual would sense that it
had been tailed by one of the members of the flock (say the ith crow). As a result, in
order to protect its food supply from the plunderer, the jth crow would deceitfully
fly over a non-hideout spot. To imitate such an action in the CSA, a random place in
the d-dimensional decision space would be assumed for the ith crow.
In summary, the tailing motion of crow individuals for the aforementioned two
cases can be expressed as (Askarzadeh 2016)
  
xi;t ri  fli;t  mj;t  xi;t rj  APj;t
xi;t 1 14:2
a random position otherwise
146 B. Zolghadr-Asli et al.

in which rj = a random number with the uniform distribution and the range of [0,1];
and AP(j,t) = the awareness probability of the jth crow at the tth iteration.
As mentioned previously, an efcient metaheuristic algorithm should provide a
good balance between diversication and intensication (Yang 2010). In the CSA,
intensication and diversication are mainly controlled by two parameters: the
flight length (fl) and the awareness probability (AP). By decreasing the awareness
probability, the chance of detecting the hideout spots by the members of the crow
flock would increase. As a result, CSA tends to focus the search on the vicinity of
the hideout spots. Thus, it can be assumed that smaller values of AP would amplify
the intensication aspect of the algorithm. On the other hand, by increasing the
awareness probability, the flock of crows is more likely to search the decision space
in a random manner for, in fact, such an action would decrease the chance of
discovering the real hideout spots by the plunderers. As a result, larger values of AP
would amplify the diversication aspect of the algorithm.

14.3 CSA Implementation for Optimization

For an efcient implementation of a metaheuristic algorithm, one needs to tune the

parameters of the algorithm. Parameter setting, however, is a time-consuming
process. Thus, the algorithms with a limited number of parameters are easier to be
implemented in different optimization problems. The aforementioned addresses one
of the major advantages of the CSA over many conventional metaheuristic algo-
rithms, since it has only two major parameters that require tuning (i.e., fl and AP).
After the parameter adjustment, the flock size (N) and the maximum number of
iterations (T) are assumed, as well.
The rst step is to locate N crows, randomly, in a d-dimensional decision space.
Since the crows have no experiences at the initial iteration, it is assumed that they
have hidden their foods at their initial positions. After the initial step, the CSA
relocates each crow individual, say the ith crow, as follows: the ith crow individual
would assume the role of a plunderer for a randomly selected member of the flock,
say the jth crow. Using Eq. (14.2), the new position of the ith crow is calculated.
To avoid unfeasible answers, it is suggested in the standard CSA to check the
feasibility of the new location in the decision space. If an unfeasible location is
generated in the latter process, the crow must stay still. An alternative for such a
procedure is the implementation of a penalty function for unfeasible answers. In any
case, the crow updates its memory as follows (Askarzadeh 2016):
  
xi;t 1 if f xi;t 1 is better than f mi;t 
mi;t 14:3
mi;t otherwise
14 Crow Search Algorithm (CSA) 147

Fig. 14.1 Flowchart of the standard CSA

Table 14.1 Characteristics of the CSA

General algorithm Crow search algorithm
Decision variable Crows position in each dimension
Solution Crows position
Old solution Old position of a crow
New solution New position of a crow
Best solution Any crow with the best tness
Fitness function The value of discovered hideout spots
Initial solution Randomly selected position of a crow
Selection
Process of generating new solution Locating the flock members hideout spots

in which f[] = objective function. These steps are repeated until the termination
criterion is satised. At that point, the best position that is memorized by the
members of the crow flock is reported as the optimum solution. Figure 14.1
illustrates the flowchart of the standard CSA. Additionally, Table 14.1 summarizes
the characteristics of the CSA.
148 B. Zolghadr-Asli et al.

14.4 Pseudo Code of the CSA

14 Crow Search Algorithm (CSA) 149

14.5 Conclusion

This chapter described the crow search algorithm (CSA), which is a novel, yet
relatively new metaheuristic optimization algorithm, based on the intelligent
behavior of crows. CSA is a population-based optimization algorithm, with mainly
two adjustable parameters (flight length and awareness probability). Such charac-
teristics make CSA a viable option for complex engineering optimization problems.
In the nal section, a pseudocode of the standard CSA was also presented.

References

Askarzadeh, A. (2016). A novel metaheuristic method for solving constrained engineering

optimization problems: Crow search algorithm. Computers & Structures, 169, 112.
Beni, G., & Wang, J. (1993). Swarm intelligence in cellular robotic systems. In P. Dario,
G. Sandini, & P. Aebischer (Eds.), Robots and Biological Systems: Towards a New Bionics?
Berlin, New York, NY: Springer.
Bonabeau, E., Dorigo, M., & Theraulaz, G. (1999). Swarm intelligence: From natural to articial
systems. New York, NY: Oxford University Press.
Clayton, N., & Emery, N. (2005). Corvid cognition. Current Biology, 15(3), R80R81.
Emery, N. J., & Clayton, N. S. (2004). The mentality of crows: Convergent evolution of
intelligence in corvids and apes. Science, 306(5703), 19031907.
Emery, N. J., & Clayton, N. S. (2005). Evolution of the avian brain and intelligence. Current
Biology, 15(23), R946R950.
Gandomi, A. H., Yang, X. S., & Alavi, A. H. (2013). Cuckoo search algorithm: A metaheuristic
approach to solve structural optimization problems. Engineering with Computers, 29(1), 1735.
Prior, H., Schwarz, A., & Gntrkn, O. (2008). Mirror-induced behavior in the magpie (Pica
pica): Evidence of self-recognition. PLoS Biology, 6(8), e202.
Wolpert, D. H., & Macready, W. G. (1997). No free lunch theorems for optimization. IEEE
Transactions on Evolutionary Computation, 1(1), 6782.
Yang, X. S. (2010). Nature-inspired metaheuristic algorithms. Frome, UK: Luniver press.
Chapter 15
Dragonfly Algorithm (DA)

Babak Zolghadr-Asli , Omid Bozorg-Haddad and Xuefeng Chu

Abstract The dragonfly algorithm (DA) is a new metaheuristic optimization

algorithm, which is based on simulating the swarming behavior of dragonfly
individuals. This algorithm was developed by Mirjalili (2016) and the preliminary
studies illustrated its potential in solving numerous benchmark optimization
problems and complex computational fluid dynamics (CFD) optimization problems.
In this chapter, the natural process behind a standard DA is described at length.

15.1 Introduction

In the past decades, the natural swarming behavior of species has been the source of
inspiration for a wide range of metaheuristic optimization algorithms. In fact, the
aforementioned is the fundamental basis of swarm intelligence (SI), which was rst
proposed by Beni and Wang in the late 1980s (Beni and Wang 1993). SI, ulti-
mately, aims to simulate the collective and social intelligence of natures group
living creatures (Bonabeau et al. 1999). Although both SI and traditional evolu-
tionary algorithms (EAs), such as genetic algorithm (GA), have undeniable

B. Zolghadr-Asli
O. Bozorg-Haddad (&)

Department of Irrigation and Reclamation Engineering, Faculty of Agricultural Engineering
and Technology, College of Agriculture and Natural Resources, University of Tehran,
3158777871 Karaj, Iran
e-mail: [email protected]
B. Zolghadr-Asli
e-mail: [email protected]
X. Chu
Department of Civil and Environmental Engineering, North Dakota State University,
Dept 2470, Fargo, ND 58108-6050, USA
e-mail: [email protected]

Springer Nature Singapore Pte Ltd. 2018 151

O. Bozorg-Haddad (ed.), Advanced Optimization by Nature-Inspired Algorithms,
Studies in Computational Intelligence 720, DOI 10.1007/978-981-10-5221-7_15
152 B. Zolghadr-Asli et al.

advantages over traditional optimization methods in dealing with complex opti-

mization problems, some may prefer the SI-based algorithms over EAs. First, there
are fewer controlling parameters in the SI-based algorithms. Second, the SI-based
algorithms are equipped with fewer operators than most traditional EAs (Mirjalili
2016).
The basic principles of the SI-based algorithms are centered around an iterative
process, in which the SI searches through the decision space for arrays of decision
variables, resulting in an optimum solution. This nature-inspired process is intended
to imitate a natural feature that has been evolved over millions of years (Gandomi
and Alavi 2012). Since the introduction of the SI-based algorithms, many promising
metaheuristic algorithms have been proposed. These algorithms intend to increase
the pace and accuracy of the searching process of the decision space, making them
more suitable to solve complicated optimization problems.
Dragonfly algorithm (DA) is a newly proposed SI-based optimization algorithm.
Dragonflies are majestic creatures. According to the discovered fossils, they may
have evolved for more than 300 million years. Accordingly, there are up to 3000
different species of this insect around the world (Thorp and Rogers 2014).
Dragonflies are considered as small carnivorous predators, eating a wide variety of
insects ranging from small midges and mosquitoes to butterflies, moths, and
damselflies. Although dragonflies are swift and agile fliers, some predators, such as
swallows, are fast enough to hunt them as preys. A fascinating fact about dra-
gonflies is perhaps their unique swarming behavior. They may swarm only for
hunting or migration purposes (Mirjalili 2016). The former is referred to as the
static (feeding) swarm (SS), and the latter is known as the dynamic (migratory)
swarm (DS).
In the static swarm, dragonflies form small groups and maneuver over small
areas. In this swarming behavior, which can be characterized by the local move-
ments and swift changes in the flying orientation, each dragonfly individual tends to
hunt flying preys. In a dynamic swarm, however, a massive number of dragonfly
individuals form a migrating swarm, which would travel in one direction, and over
long distances (Russell et al. 1998; Wikelski et al. 2006).
The aforementioned swarming behaviors of dragonflies are the main source of
inspiration for the DA, for in fact, these behaviors are in line with the two main
characteristics of a metaheuristic optimization algorithm: intensication (also
known as exploitation) and diversication (also known as exploration) (Gandomi
et al. 2013). The static swarming behavior enables dragonflies to create sub-swarms
and investigate the presence of promising optima in numerous, yet small areas of
the decision space (diversication). In a dynamic swarm, however, dragonflies will
migrate in massive swarms toward what are the most promising locations for the
global optimum, which in other words is a description of the intensication phase of
an optimization algorithm. These are, in fact, the basic mechanism behind the DA.
15 Dragonfly Algorithm (DA) 153

The DA was initially proposed by Mirjalili (2016), and the preliminary studies
have demonstrated its potential to outperform existing algorithms in solving both
benchmark test problems and complicated engineering problems of computational
fluid dynamics (CFD). The DA was also modied to better deal with binary [binary
dragonfly algorithm (BDA)], and multi-objective optimization problems
[multi-objective dragonfly algorithm (MODA)] (Mirjalili 2016). As a compatible
and efcient algorithm, the DA can be an already promising alternative for solving
complex engineering optimization problems. The following sections will focus on
the basic principles of a standard DA.

15.2 Dragonflies Swarming Patterns

The behavior of swarms follows three primitive principals (Reynolds 1987):

Separation: The static collision avoidance of the dragonfly individuals in a
neighborhood.
Alignment: The velocity matching of dragonfly individuals in a neighborhood.
Cohesion: The tendency of dragonfly individuals toward a neighborhoods
center of the mass.
Additionally, any swarms of living creatures would also follow their survival
instincts. Hence, all of the dragonfly individuals should be attracted toward food
sources (food attraction) and distracted outward predators (predator distraction). In
result, the swarming behavior of the dragonfly community can be explained by
these ve main factors (Fig. 15.1).

Fig. 15.1 Primitive corrective patterns of dragonfly individuals in a swarm: a Separation;

b Alignment; c Cohesion; d Food Attraction; and e Predator distraction
154 B. Zolghadr-Asli et al.

In order to simulate the swarming behavior of dragonflies, the aforementioned

characteristics must be mathematically modeled as follows:
The separation motion can be expressed as Reynolds (1987):

X
N
Si;t  Xi;t  Xj;t 15:1
j1

in which X(i,t) = position of the ith dragonfly individual in the tth iteration; X(j,
t) = position of the jth neighboring dragonfly individual in the tth iteration;
N = number of neighboring dragonfly individuals; and S(i,t) = separation motion for
the ith dragonfly individual in the tth iteration.
The alignment motion is calculated by Mirjalili (2016):
PN
j1 Vj;t
Ai;t 15:2
N

in which V(j,t) = velocity of the jth neighboring dragonfly individual in the tth
iteration; and A(i,t) = alignment motion for the ith dragonfly individual in the tth
iteration.
The cohesion motion can be measured by Mirjalili (2016):
PN
j1 Xj;t
Ci;t  Xi;t 15:3
N

in which C(i,t) = cohesion motion for the ith dragonfly individual in the tth iteration.
The food attraction motion is calculated by Mirjalili (2016):

Fi;t Xfood;t  Xi;t 15:4

in which X(food,t) = position of the food source in the tth iteration; and F(i,t) = food
attraction motion for the ith dragonfly individual in the tth iteration. The food is
considered as the dragonfly individual with the best objective function observed so
far.
The predator distraction is quantied by Mirjalili (2016):

Ei;t Xenemy;t Xi;t 15:5

in which X(enemy,t) = position of the predator in the tth iteration; and E(i,t) = predator
distraction motion for the ith dragonfly individual in the tth iteration. The predator
is considered as the dragonfly individual with the worst objective function observed
so far.
The combination of the aforementioned motions can predict the corrective
pattern of dragonfly individuals in each iteration. The positions of dragonflies
15 Dragonfly Algorithm (DA) 155

individuals are updated in each iteration using the current position of the dragonfly
individual [X(i,t)] and a step vector [DX(i,t)]. In fact, the introduced step vector is
analogous to the velocity vector in the particle swamps optimization
(PSO) algorithm, and the procedure for updating the positions of dragonfly indi-
viduals in the DA is based on the framework of the PSO algorithm. The step vector,
which demonstrates the motion orientation for each dragonfly individual, is dened
as Mirjalili (2016):

DXi;t 1 s  Si;t a  Ai;t c  Ci;t f  Fi;t e  Ei;t w  DXi;t

15:6

in which s = separation weight; a = alignment weight; c = cohesion weight;

f = food attraction weight; e = predator distraction weight; and w = inertia weight.
After calculating the step vector, the updated position vectors are calculated by:

Xi;t 1 Xi;t DXi;t 15:7

By tampering the separation, alignment, cohesion, food attraction, and predator

weights (s, a, c, f, e, and w), different diversication and intensication behaviors
can be achieved by the optimization. The aforementioned weights are, in fact, the
DAs parameters and should be modied for each set of optimization problems in
order to achieve preferable results. Additionally, neighboring of dragonfly indi-
viduals is crucial to the performance of the DA. In fact, an improper neighboring
detecting mechanism could interfere with the convergence of the DAs results. This
is due to the fact that in each algorithm, to detect the potential location of global
optima, the searching mechanism should initially investigate the entire decision
space thoroughly. This phase, which is better known as the desperation, requires the
dragonfly individuals to spread through the search space. Yet, to locate the global
optima, these individual dragonflies are required to converge and move toward the
plausible locations of the global optima. This phase is known as the intensication.
As discussed earlier, the swarming behavior of dragonflies can be categorized into
two general motions: static swarming and dynamic swarming. Dragonflies tend to
align their flying while maintaining proper separation and cohesion in a dynamic
swarming (intensication). In a static swarm, however, alignments are very low
while cohesion is high to attract prey. Therefore, we assign dragonflies with high
alignment and low cohesion weights when exploring the search space. For tran-
scending between diversication to intensication, the radii of neighborhoods are to
increase proportionally to the number of iterations. This way, as the optimization
proceeds, more dragonfly individuals can induce the motion of one another, causing
the swarm to converge toward the possible location of the global optimum. Another
way to balance diversication and intensication is to adaptively tune the DAs
parameters (i.e., s, a, c, f, e, and w) during the optimization.
156 B. Zolghadr-Asli et al.

However, in order to increase the odds of investigating the entire decision space
by any optimization algorithm, a random motion needs to be introduced to the
searching mechanism. As a result, to improve the randomness, stochastic behavior,
and exploration of the dragonfly individuals, they are required to fly around the
search space using random walk (Lvy flight) when no neighboring solutions in the
vicinity are detected. In this case, the positions of dragonflies are updated by
Mirjalili (2016):

Xi;t 1 Xi;t Levyd  Xi;t 15:8

in which d = number of decision variables; and Lvy(d) = Lvy flight function that
is given by Yang (2010):
r1  r
Levyd 0:01  1 15:9
j r2 j b

in which r1 and r2 = two random numbers in the range of [0,1]; b = constant,

which is equal to 1.5 according to the measured values of the dragonfly individuals
movement (Mirjalili 2016); and r is calculated by Yang (2010):

0   1b1
C1 b  sin pb
r@  A
2
 15:10
C 1 2 b  b  2 2
b1

in which C(x) = (x 1)!.

15.3 Optimization Procedure of the DA

Like most SI-based optimization algorithms, the DA starts the optimization process
by creating a set of random solutions for a given optimization problem. Naturally,
the number of initial dragonfly individuals (M) can influence the performance of the
DA. The bigger populations increase the chance of nding the global optima, while
increasing the calculation time for each iteration, and in turn the entire optimization
problem. After determining the positions of dragonflies within the lower and upper
boundaries of any given variable, the position of each dragonfly individual is
updated in each iteration by calculating the step position vector for each individual
dragonfly, using the motions induced by separation, alignment, cohesion, food
attraction, and predator distraction. The position updating process is continued
iteratively until the termination criterion is satised. Table 15.1 lists the charac-
teristics of the DA. Additionally, the flowchart of the DA is shown in Fig. 15.2.
15 Dragonfly Algorithm (DA) 157

Table 15.1 Characteristics of the DA

General algorithm Dragonfly algorithm
Decision variable Dragonflys position in each dimension
Solution Dragonflys position
Old solution Old position of a dragonfly
New solution New position of a dragonfly
Best solution Any dragonfly with the best tness
Fitness function Distance from the food source, the predator, center of the swarm,
velocity matching, and collision avoidance
Initial solution Randomly selected position of a dragonfly
Selection
Process of generating Flying with a specic velocity and direction
new solution

Fig. 15.2 Flowchart of the DA

158 B. Zolghadr-Asli et al.

15.4 Pseudo-Code of the DA

Begin
Define population size (M)

Set the iteration counter t=1

Initialize the population by generating Xi for i = 1, 2, 3, M

Calculate the objective function values of all dragonflies

Update the food and the predators location

While (the stop criterion is not satisfied)

For i=1:M

Update neighborhood radii (or update w, s, a, c, f, and e)

If a dragonfly has at least one neighborhood dragonfly

Calculate
Separation motion

Alignment motion

Cohesion motion

Food attraction motion

Predator distraction motion

Update position vector

Else
Update position vector using the Lvy flight function

End if
End for i

Sort the population/dragonflies from best to worst and find the current best

End while
Post-process and visualize the results
End
15 Dragonfly Algorithm (DA) 159

15.5 Conclusion

This chapter described the dragonfly algorithm (DA) which is a novel, yet newly
introduced metaheuristic optimization algorithm. After a brief review of the pre-
vious applications of the DA, the standard DA and its mechanism were described.
In the nal section, a pseudo-code of the standard DA was also presented.

References

Beni, G., & Wang, J. (1993). Swarm intelligence in cellular robotic systems. In: P. Dario, G.
Sandini, & P. Aebischer (Eds.), Robots and biological systems: Towards a new bionics?
Berlin, Heidelberg, New York, NY: Springer.
Bonabeau, E., Dorigo, M., & Theraulaz, G. (1999). Swarm intelligence: From natural to articial
systems. New York, NY: Oxford University Press.
Gandomi, A. H., & Alavi, A. H. (2012). Krill herd: A new bio-inspired optimization algorithm.
Communications in Nonlinear Science and Numerical Simulation, 17(12), 48314845.
Gandomi, A. H., Yang, X. S., & Alavi, A. H. (2013). Cuckoo search algorithm: A metaheuristic
approach to solve structural optimization problems. Engineering with Computers, 29(1),
1735.
Mirjalili, S. (2016). Dragonfly algorithm: A new meta-heuristic optimization technique for solving
single-objective, discrete, and multi-objective problems. Neural Computing and Applications,
27(4), 10531073.
Reynolds, C. W. (1987). Flocks, herds and schools: A distributed behavioral model. In
Proceedings of the 14th annual conference on computer graphics and interactive techniques,
New York, NY, July 2731.
Russell, R. W., May, M. L., Soltesz, K. L., & Fitzpatrick, J. W. (1998). Massive swarm migrations
of dragonflies (Odonata) in eastern North America. The American Midland Naturalist, 140(2),
325342.
Thorp, J. H., & Rogers, D. C. (Eds.). (2014). Thorp and Covishs freshwater invertebrates:
Ecology and general biology (Vol. 1). Amsterdam, Netherland: Elsevier.
Wikelski, M., Moskowitz, D., Adelman, J. S., Cochran, J., Wilcove, D. S., & May, M. L. (2006).
Simple rules guide dragonfly migration. Biology Letters, 2(3), 325329.
Yang, X. S. (2010). Nature-inspired metaheuristic algorithms. Frome, UK: Luniver Press.