You are on page 1of 22

Applied Soft Computing Journal 75 (2019) 84–105

Contents lists available at ScienceDirect

Applied Soft Computing Journal


journal homepage: www.elsevier.com/locate/asoc

β -Chaotic map enabled Grey Wolf Optimizer



Akash Saxena a , , Rajesh Kumar b , Swagatam Das c
a
Department of Electrical Engineering, Swami Keshvanand Institute of Technology, Management & Gramothan, Jaipur, India
b
Department of Electrical Engineering, Malaviya National Institute of Technology, Jaipur, India
c
Electronics and Communication Sciences Unit, Indian Statistical Institute, Kolkata, India

highlights

• A bridging mechanism inspired from β function is proposed for Grey Wolf Optimizer.
• Evolved β -GWO is benchmarked on shifted and biased and CEC-2017 functions.
• Comparative analyses with variants of GWO and other algorithms are presented.
• Applications of β -GWO are reported on real problems.

article info a b s t r a c t

Article history: The diversification (exploration) and intensification (exploitation) are two main attributes of any
Received 11 June 2018 population-based metaheuristic algorithm. It is essential for any algorithm that in exploration phase
Received in revised form 20 September 2018 the search space is utilized and explored properly through random behavior, on the other hand, the
Accepted 24 October 2018
progression of the search in a viable direction to obtain global minima, should be performed through
Available online 7 November 2018
strategic behavior in exploitation phase. A proper balance between these two can be achieved by an
Keywords: adaptive mechanism in every algorithm. Robustness of an algorithm is judged by the efficacy of these
β -Chaotic map two attributes along with the efficiency of the bridging mechanism. In literature, the positive impact
Grey Wolf Optimizer of inculcation of chaotic sequences on the efficacy of these attributes has been reported. With this
Congress on Evolutionary Computation motivation, the paper presents an adaptive bridging mechanism based on β -chaotic sequence for the
(CEC)
improvement of Grey Wolf Optimizer (GWO). The control vector of classical GWO is integrated with
Chaos theory
the β -chaotic sequence for better exploration and exploitation virtues. The new variant β -GWO is
benchmarked on two benchmark suites 1 and 2 that include 12 shifted and biased functions and 29
Congress on Evolutionary Computation-2017 (CEC-2017) functions. Sensitivity Dependence of Initial
Conditions (SDIC) is performed for tuning the initial parameters. The comparison of the proposed variant
with other contemporary algorithms is carried out and different statistical tests are performed to judge
the efficacy of the proposed variant. Further, the applicability of the proposed variant is checked with two
real engineering problems namely frequency modulated sound waves parameter estimation problem and
strategic bidding in the energy market. Results reveal that the proposed chaotic variant exhibits better
exploration and exploitation qualities.
© 2018 Elsevier B.V. All rights reserved.

1. Introduction problems, the metaheuristic algorithms have been developed by


the researchers for over last four decades or more. Metaheuristics,
Non-convex and black-box optimization problems involving founded on the basis of simulation and mimicry of nature, provide
multi-dimensional real search spaces, often pose a severe chal- the advantages in terms of requiring no derivative information of
lenge to the traditional mathematical programming based solvers. the function, being insensitive to initialization and being adaptive
In addition, some instances of such problems may not clearly de- as well as simple. In contrast to the traditional problem-specific
fine the bounds on the decision variables while involving stochastic heuristic methods, these metaheuristic optimization algorithms
parameters (often due to noise). These problems can be continu- can be applied in a black-box manner and without presuming
ous, discontinuous, constrained, or unconstrained in nature. While domain knowledge about the problem at hand [1].
searching the powerful paradigms to solve such hard optimization A metaheuristic method is a problem independent higher level
heuristic method that can be employed to solve many hard op-
∗ Corresponding author. timization problems. The use of metaheuristic algorithms in real
E-mail address: akash@skit.ac.in (A. Saxena). applications has increased from last few years. The metaheuristic

https://doi.org/10.1016/j.asoc.2018.10.044
1568-4946/© 2018 Elsevier B.V. All rights reserved.
A. Saxena, R. Kumar and S. Das / Applied Soft Computing Journal 75 (2019) 84–105 85

intelligence. In this category, Particle Swarm Optimization (PSO)


Acronym is the most popular technique and experimented algorithm [9].
GWO Grey Wolf Optimizer This algorithm is based on ‘‘follow the leader’’ philosophy. Parti-
mGWO Modified Grey Wolf Optimizer cles (search agents) search for the solution in multidimensional
search space and update the position and velocity with the help
OGWO Oppositional Grey Wolf Optimization
of cognitive and social intelligence. Many experiments have been
InGWO Intelligent Grey Wolf Optimizer
done in order to improve the performance of this optimizer. Other
IGWO Improved Grey Wolf Optimizer algorithms, in this category are: Bat Algorithm [10], Spider Monkey
GGWO Grouped Grey Wolf Optimizer Optimization Algorithm [11], Cuckoo Search Optimization Algo-
CEC-2017 Congress on Evolutionary computation rithm [12], Fruit Fly Optimization [13], Firefly Optimization [14]
SDIC Sensitive Dependence on Initial Condi- and many more.
tions In category 3, concepts of physics are employed to derive
CSA Crow Search Algorithm the optimization process. Gravitational Search Algorithm [15],
GSA Gravitational Search Algorithm Big Bang–Big Crunch Algorithm [16], Black Hole Algorithm [17]
FPA Flower Pollination Algorithm and many more algorithms fall in this category. These algorithms
MCP Market Clearing Price are based on the laws of physics. According to No Free Lunch
Theorem [18] no metaheuristic is suitable for all the applications
SBP Strategic Bidding Problem
this means that a particular metaheuristic is suitable for one ap-
EMP Exercise of Market Power
plication and may not be suitable for other application. Hence, it is
always necessary that some modifications should be incorporated
to make the metaheuristic suitable for a particular application.
algorithms reported in the literature are mainly based on mimicry With this motivation, many improvements have been sug-
of social behavior of various plants, animals, and biological sys- gested in the existing algorithms to make them compatible with
tems [2]. In short, all the metaheuristic algorithms possess the particular application areas. An adaptive harmony search with
following common characteristics: best-based search strategy has been proposed in work [19]. A
variant of Social Emotional Optimization Algorithm (SEOA) based
• These algorithms are nature inspired algorithms and mimic on a local search strategy has been proposed in the work [20].
the laws of nature in the inner structure to solve complex Along with the inculcation of different search strategies, the ap-
engineering design problems. plication of chaos theory in the metaheuristics for improvement is
• These algorithms possess stochastic nature, this nature is considered as a beneficial tool [21]. Chaotic sequences are symbolic
simulated by the incorporation of the random component in of noisy behavior and randomness. Chaos theory is a branch of
the search process and in other decision-making points. mathematics, which focuses on the behavior of the dynamical
• These algorithms are derivative-free algorithms, this fact system which is highly sensitive to initial conditions. The chaos
makes these algorithms less time-consuming and simple. sequences are very vulnerable to initial conditions, Hence the
• These algorithms possess several parameters, those have to term Sensitive Dependence on Initial Conditions (SDIC) is used to
be fitted according to the nature of the problem. identify chaotic complex systems [22]. These number sequences
In recent years, metaheuristic algorithms are in the spotlight possess constant feedback loops, underlying patterns, random,
due to various reasons. The derivative-free mechanism enables self-repetitive, fractals. Even a small change in one state of a
the metaheuristic algorithms to apply to the real-world problems deterministic nonlinear system results in a large difference in a
where the shape and characteristics of the objective function are later state. This phenomenon is known as butterfly effect and
unknown. The applicability of these algorithms is not questionable this term was coined by Sir Edward Lorenz [23]. Application of
because they do not compute the derivative of functions. These chaotic sequences is reported in secure communication and spread
algorithms possess very simple structure and based on the con- systems [24], optimization algorithms [25,21,26–28] and [29], im-
cepts of life, phenomena, and behavior. The simplicity of these age encryption [30], cryptographic applications [31], underwater
algorithms is proven as a boon to the researchers for implemen- acoustic communication [32] and many more.
tation of the algorithms to real problems and that too without Many optimization algorithms have employed chaotic
changing the structure, hybridize these algorithms with different sequences over the random walk (random numbers generation)
concepts or with each other and simulate them on computers. due to the fact that the random walk not always implement global
These algorithms are flexible; the flexible nature of the algorithms search well. Thus in some cases, the algorithm development is
indicates the compatibility towards the application part. Some based on chaotic variables instead of random variables, so these
applications of the metaheuristic can easily be seen in [3,4]. Fur- algorithms are called chaotic algorithms [26]. A chaotic Firefly
ther, the metaheuristic algorithms are classified into three main algorithm was proposed in [26]. In this work attractive move-
categories namely: ment of fireflies was simulated with ten chaotic maps. Further,
Chaotic Crow search algorithm was employed to carry out fea-
1. Evolutionary computing based algorithms ture selection for different standard classification datasets [25].
2. Swarm Intelligence (SI) based algorithms −
→ −

Chaotic sequences are used for parameter a and A in Chaotic
3. Physics postulates based algorithms
GWO approach [27]. Chaos enhanced Accelerated Particle Swarm
In category 1, the algorithm which is widely used is Genetic Optimization (CAPSO) was proposed by Gandomi et al. [21]. An
Algorithm (GA) and is based on survival of the fittest Darwinian attraction parameter was tuned with normalized chaotic maps in
Theory [5]. Other examples of these algorithms are Evolution Strat- that work. Chaotic bat algorithm was proposed in [28], the tuning
egy (ES) [6], Evolutionary Programming (EP) [7] and Biogeography of the crucial parameter of this algorithm was done with the help
Based Optimization (BBO) [8]. The flow of these algorithms starts of chaotic maps.
with the initial set of solution and successive improvements in runs On the basis of the literature review, this paper is a proposal of
with the incorporation of different operators. a new variant of Grey Wolf Optimizer (GWO), β -GWO. A Chaotic
In category 2, the algorithms which mimic the behavior of sequence inspired from β functions is embedded with the control
swarms, school of fish, birds and based on cognitive and social vector through a normalization process [33]. Effect of this mapping
86 A. Saxena, R. Kumar and S. Das / Applied Soft Computing Journal 75 (2019) 84–105

on the performance of the algorithm is evaluated through the 2.2. Hunting the prey
benchmarking of the algorithm on 12 standard shifted and biased
benchmark functions from benchmark suite 1 and 29 unimodal, In the search space, the exact identification of the prey is not
multimodal, hybrid and complex functions proposed for the com- possible , Hence, the best three solutions are kept to oblige the
petition on bound-constrained real parameter optimization under social hierarchy (α, β & δ ). The mathematical representation of this
the IEEE Congress on Evolutionary Computation-2017 (CEC-2017) fact can be visualized in Eqs. (3)–(5).
named as benchmark suite 2. Further, to test the applicability of the −
→ ⏐−
⏐→ − → − →⏐⏐ −
→ ⏐−
⏐→ −
→ −
→⏐⏐
proposed variant on real engineering problems, two real problems Dα = ⏐ C1 . Xα − X ⏐ , Dβ = ⏐ C2 . Xβ − X ⏐ ,
named as Frequency Modulated Sound Waves Parameter estima- −
→ ⏐⏐− →− → − →⏐⏐ (3)
Dδ = ⏐ C3 . Xδ − X ⏐
tion problem and strategic bidding problem in energy market have
been solved with the application of the proposed β -GWO. In short,

→ − → − → −−→ − → − → − → −−→
the following research objectives are framed for this study: X1 = Xα − A1 .(Dα ), X2 = Xβ − A2 .(Dβ ),

→ − → − → −→ (4)
1. To investigate the application of β -chaotic sequence for X3 = Xδ − A3 .(Dδ )
improving the existing linearly decreasing mechanism of
bridging phase of explorative and exploitative phases of −
→ −
→ −


→ ( X1 + X2 + X3 )
GWO by introducing β -chaotic sequence through normal- X (t + 1) = (5)
3
ization process.
2. To investigate the performance of this new variant with re- The equations represent the position updation according to alpha,

→ − → −

cently developed GWO variants and some of the contempo- where Dα , Dβ and Dδ are the distance of the prey from α, β and δ

→ −
→ −

rary algorithms on variety of benchmark functions including wolves respectively and X1 , X2 and X3 are positions of α, β and δ
the latest ones (CEC-2017). wolves respectively.
3. To perform different statistical tests including box plot anal-
ysis and Wilcoxon rank-sum for validation of the results, 2.3. Attacking prey
obtained from the independent runs.
4. To compare the performance of proposed β -GWO with some This phase is responsible for exploitation and is handled by
of the leading rank holders algorithm from CEC competition. −

linear decrement in a . The Fig. 1 shows this phenomenon. The
5. To test the applicability of the proposed variant over the linear decrement in this parameter enables grey wolves to attack
real world problems and present a comparative analysis


the prey while it stops moving. Fluctuations in A is also controlled
with other developed variants of GWO and contemporary −
→ −

by a i.e. if the value of a is high then there are more fluctuations
metaheuristic techniques. −

in A .
The rest of the paper is organized in the following way: In GWO has attracted the interest of the researchers due to follow-
Section 2, details of chaotic sequence and related literature review ing reasons:
is given, in Section 3, development of β -GWO is given, in Section • GWO is based on the leadership hierarchy of the grey wolves,
4, details of benchmark suites are given, in Section 5, results and and segregate the population according to the characteris-
discussions are presented, in Section 6, real applications of the tics of the wolf. GWO possesses very simple structure, this
proposed variant are presented. Last but not the least, the major structure allows researchers to hybridize and to perform
findings of the work are reported in conclusion section. experiments to develop new variants [35,36].
• GWO is a derivative-free algorithm and control parameters
2. Grey wolf optimizer: An overview are very less, these features make GWO more flexible, adapt-
able and user-friendly paradigm.
In the year 2014, Mirjalili et al. [34] introduced a meta-heuristic
• This algorithm has been applied to real-world applications
algorithm named as Grey Wolf Optimizer (GWO). This algorithm
because it can handle discontinuous, non-differentiable and
was based on the leadership hierarchy of grey wolves and hunting
stochastic functions [37,38].
strategy of the wolves. Grey wolves live in a pack and present the
fine example of leadership hierarchy. The pack is divided into four Many approaches have been reported regarding the application
categories: Alpha (dominant wolves) the decision makers, Beta the of new variants of GWO in engineering problems. On the basis of
subordinate wolves, Omega the space goat or babysitters and Delta. researches on GWO, the developed variants can be subdivided into
The major steps involved in the algorithm are the following: four main categories:
• Updating in existing mechanism [36,39]
2.1. Encircling the prey
• Addition of new operators [40]
The grey wolf encircles the prey and the mathematical equation • Encoding scheme of individuals [35,41,42]
for the same can be given as: • Change in population structure and hierarchy [36,38] and
[43]

→ ⏐−
⏐→ −
→ −
→ ⏐
D = ⏐ C . Xp (t) − X (t)⏐ (1)

Updation of the mechanism between the exploration and ex-
ploitation phase is a potential area of research. A nonlinear mod-

→ −
→ −
→−→ ulation index has been proposed recently in [44] and named as
X (t + 1) = Xp (t) − A . D (2) Improved Grey Wolf Optimizer (IGWO). Similarly, a monotonic

→ exponential function has been employed instead of monotoni-
where t indicates the current iteration, Xp (t) is position vector of −


→ cally linearly decreasing a in [39]. This variant was named as
the prey and X (t + 1) updation in the position of grey wolf in next Modified Grey Wolf Optimizer (mGWO). On the basis of encoding
−→ −

iteration. A and C are the coefficient vectors. These coefficients scheme, two different schemes, namely random scout and hunting

→ −→− → − → −
→ −

vectors are defined as under A = 2 a . r1 − a and C = 2. r2 group have been presented in [42] , for proper utilization of the


where a decreases linearly from 2 to 0 and r1 , r2 are the random search space and to enhance the exploitation virtue of GWO. The
numbers. algorithm was named as Grouped Grey Wolf Optimizer (GGWO)
A. Saxena, R. Kumar and S. Das / Applied Soft Computing Journal 75 (2019) 84–105 87

and applied to design maximum power point tracking of doubly-


fed induction generator. In [43] authors proposed a new position
update mechanism on the basis of the fact that the solution can
be explored by omega wolves. Hence, a newly updated encoding
scheme for an individual was proposed which includes the new
position update vector based on delta and omega wolves’ positions.
The proposed variant (MGWO) has been applied to the transmis-
sion line expansion problem. A binary variant of GWO has been
employed to carry out large-scale unit commitment and feature
selection [35,41]. An optimal reactive power dispatch problem
was solved with two-archive multi-objective Grey Wolf Optimizer
(2arch MGWO) [45]. Ambient air quality classification has been Fig. 1. Variation of control vector in GWO.
carried out for different parts of India by using GWO [46]. Appli-
cation of GWO in Automatic Generation Control has been explored
in [37]. Economic load dispatch problem has been carried out in
two approaches by using hybrid version [40] and oppositional con-
cept [38]. Tuning of fuzzy control systems with reduced parametric
sensitivity has been addressed in [47]. Strategic bidding problem
has been addressed by Intelligent Grey Wolf Optimizer in [36].
These applications are strong evidence of global acceptability of
GWO as a strong metaheuristic algorithm for solving various com-
plex optimization problems.
Further, in following section critical review of literature along
with the chronological development of β -GWO is presented.

3. Development of β-GWO

The β functions are most common distributions for parametric Fig. 2. β -chaotic sequence.
studies, image compression, image object detection and recogni-
tion and biomedical signal composition [33,48]. In [30] new chaotic
maps inspired from β function were introduced to carry out image
The Fig. 2 shows the chaotic sequence generated for an
encryption. The studies reported that new maps can be formed
iterative process. The values of these constants are taken as
with a small change in parameters [30]. Successful application
from [30].
results on image encryption and other applications reported in ref-
erences [30,48,33] motivated authors to employ this sequence for (b) The normalization function is employed to distribute this se-
improving the bridging mechanism of existing GWO. Hence, in this quence between maximum and minimum bias before it can


be biased with the control vector a . This process is shown
paper, a β function inspired chaotic sequence is embedded with
the mapping through a normalize function to the control vector in Fig. 3. The mathematical expression for this function at

→a . In classical GWO, this vector is responsible for the exploration any iteration t can be given as:
and exploitation phase over the whole course of iterations. Nm max − Nm min
In the initial phase, the search agents take large steps to explore Nm (t) = Nm max − ( )t (8)
the search space in an effective manner and in later phase these T
steps are reduced with the help of linear decrement in the control where T is maximum no. of iteration, Nm max and Nm min
−→
vector a . This work focuses on this linear variation and employs are the maximum and minimum values of normalization
β -chaotic sequence through a normalized function. The major function respectively. Normalize β -chaotic sequence can be
motivation to perform this experiment is to seek the possibility of given as per following equation:
better exploration and exploitation by introducing the β -chaotic
sequence in each iteration. In the classical GWO, this parameter C (t) = Nm (t)xt (9)
decreases linearly, which means that an algorithm either performs (c) The instantaneous value of β -chaotic sequence embedded
diversification (exploration) or intensification (exploitation). In −→
in control vector a for β -GWO will be given in following


this work, authors have changed the control vector a chaotically
equation:
so that the exploration virtue can be kept alive in the final steps of
iterations. aβ−GWO (t) = aGWO (t) + C (t) (10)
The following steps are followed to develop β -GWO.
The Fig. 4 shows this mathematical process in a lucid man-
(a) Generation of β -chaotic sequence through the initialization ner. Further, a discussion is presented in the next subsection
of the parameters (µ, ν, x1 , x2 ) is carried out. A generalized in the support of the proposed developed philosophy.
equation for β distribution as given in [33] is as follows:
⎧( )µ ( )ν
⎨ x − x1 x2 − x 3.1. Comparison between β -GWO and other GWO variants
if x ∈ [x1 , x2 ]
β (x; µ, ν, x1 , x2 ) =
⎩ xc − x1 x2 − xc
0 other w ise Authors are motivated to perform the comparative analysis of
different virtues of GWO variants with the proposed β -GWO. After
(6)
the careful inspection of the various virtues of these variants, here
where (µ, ν, x1 , x2 ) ∈ R and x1 < x2 . β -Chaotic sequence at we underline some distinct feature of the β -GWO:
any iteration t will be given as:
• β -GWO is inspired from the β function and its distribution.
xt +1 = kβ (xt ; µ, ν, x1 , x2 ) (7)


By inspecting the variation of a , it is empirical to judge that
88 A. Saxena, R. Kumar and S. Das / Applied Soft Computing Journal 75 (2019) 84–105

cannot be ensured every time in the search process. In [27]


the variables associated with the GWO have been varied
according to chaotic sequences.
• In [49] the division operators are replaced with chaotic vari-
ables, which again follow some specific patterns. In β -GWO,
we modify the mechanism between two phases. The pro-
posed modification does not change the search pattern and
original structure of the algorithm.
From these points, it is evident that the implications of chaos
theory have been significantly studied in metaheuristic al-
gorithms, in order to achieve better exploitation and explo-
ration virtues. However, less work is reported on the im-
plications of chaos on the bridging mechanism between ex-
ploration and exploitation phases. This fact also motivated
authors to propose a chaotic bridging mechanism for GWO.

Fig. 3. Normalized β -chaotic sequence.


3.3. Motivation

It is a well-known fact that the metaheuristic algorithms pos-


sess stochastic nature and due to this nature, the robustness of a
new variant of the algorithm has paramount importance. Although
the GWO is a very well structured, hierarchy-based algorithm,
which mimics the discipline of grey wolves in hunting, approach-
ing and encircling to the prey. Yet, the studies reported by Muro
et al. [50] motivated us to embed the random behavior in diver-
sification and intensification process of all wolves simultaneously.
The authors of [50] proposed a hypothesis that wolf-pack hunting
strategies can be explained as emergent phenomena which do not
necessarily rely on the presence of a hierarchical social structure,
intentional behavior, or communication skills. This research moti-
vates us to see the wolf as an individual in pack, where the behavior
of the wolf can be chaotic while hunting and encircling. Following
Fig. 4. Variation of control vector in β -GWO. key points are worth mentioning:

1. In this work, the philosophy of GWO is obeyed by decreas-




ing the control vector a linearly with due course of the
although variation is not linearly monotonically decreasing iteration. However, with the inculcation of the chaotic se-
(chaotic variation) yet it is bound in a closed monotonically quence in control vector enhance the diversification virtue
decreasing envelope. This virtue of β -GWO enables the explo- of the GWO till the last iteration. The proposed approach
ration and exploitation processes alive till the last iteration. which embeds chaotic sequence can be a beneficial tool for
However, in [39] the exponential monotonically decreasing −

researchers to tune the control vector a dynamically.


trend in a is analyzed, which shows that both exploration
2. The chaotic sequence does not follow any special mono-
and exploitation process cannot be alive in an ongoing simu-
tonically increasing or decreasing trends. Hence, it provides
lation. −

random values for control vector a over the course of the
• From the behavior of wolves, it is empirical to say that they
optimization process.
live in a pack, hence the population structure that places them −

3. This sequence changes the value of control vector a ran-
in different directions is not a matter of interest [38]. Hence,
domly hence assist the search agents to release themselves
in β -GWO, we did not modify the population structure.
from the local minima trap.
• Since, the positions of the non-dominant wolves are decided
4. The transition from diversification phase to intensification
by the dominant wolves, the proposed variant also advocates
phase can be achieved slowly with the employment of
the hierarchy in grey wolves. Hence, the position updating
chaotic sequence enabled adaptive approach. This change
mechanism has not been changed. However, in [43] the posi-
tion update was partially governed by non-dominant wolves. makes control vector an adaptive and random concurrently.
In following section, the details of benchmark functions which
3.2. Comparison between β -GWO and other chaotic algorithms are employed to test the efficacy of proposed β -GWO are illus-
trated.
The algorithm, that uses chaotic variables instead of random
numbers to induce stochasticity is known as a chaotic algorithm. 4. Details of benchmark suites for evaluation of β-GWO
These algorithms employ chaotic maps in two ways, the first way
is to change the decision operator according to chaotic sequence
For characterization, evaluation and measurement of the per-
instead of a machine generated (pseudo) random number and the
formance of optimization algorithms, set of functions with known
second way is swarming through chaotic sequences instead of
characteristics (location of minima, maxima, global and local min-
random numbers. To underline the features of β -GWO, following
ima and maxima along with their numerical values) are employed.
points are derived from the literature:
These functions are known as benchmark functions.
• Chaotic numbers are generated with reference to a specific In this paper, We test efficacy of the proposed variants on two
initial starting point, thus the avoidance of local minima benchmark suites:
A. Saxena, R. Kumar and S. Das / Applied Soft Computing Journal 75 (2019) 84–105 89

Table 1
Benchmark suite 1.
Function Dimension Range Minimum value
Unimodal benchmark function
∑n
F1 (x) = i=1 (xi + 40)2 − 80 30 [−100,100] −80
∑n ∏n
F2 (x) = i=1 |xi + 7| + i=1|xi + 7| − 80 30 [−10,10] −80
∑n (∑ )2
i
F3 (x) = i=1 j−1 (xj + 60) − 80 30 [−100,100] −80
F4 (x) = maxi {|(xi + 60)| , 1 ≤ i ≤ n} − 80 30 [−100,100] −80
∑n−1 )2
[100 (xi+1 + 60) − (xi + 60)2 + (xi + 60 − 1)2 ] − 80
(
F5 (x) = i=1 30 [−30,30] −80
∑n−1
F6 (x) = i=1 ([(xi + 60) + 0.5]) − 80
2
30 [−100,100] −80
Multimodal benchmark function
∑n (√ )
F7 (x) = i=1 −(xi + 300) sin |(xi + 300)| 30 [−500,500] −418.9829 × 32
∑n [
i=1 (xi + 2) − 10 cos (2π (xi + 2)) + 10 − 80
2
]
F8 (x) = 30 [−5.12,5.12] −80
( √ ∑ )
n
F9 (x) = −20 exp −0.2 1n 2
− exp 1n ni=1 cos(2π (xi + 20)) + 20 + e − 80
( ∑ )
i=1 (xi + 20) 30 [−32,32] −80
∑n ∏n ( )
1 2 (xi +400)
F10 (x) = 4000 i=1 (xi + 400) − i=1 cos
√ + 1 − 80 30 [−600,600] −80
i
{ }
π
∑n−1
F11 (x) = n 10 sin(π y1 ) + i=1 (yi − 1) [1 + 10sin (π yi+1 )] + (yn − 1)2 +
2 2

∑n 30 [−50,50] −80
i=1 u((xi + 30), 10, 100, 4) − 80
(xi +30)+1
yi = 1 + 4

⎨ k(xi − a)m xi > a
u(xi , a, k, m) = 0 −a < xi < a
xi < −a
⎩k(−x − a)m
i
∑n
F12 (x) = 0.1{sin2 (3π (x1 + 30)) + i=1 ((xi + 30) − 1)2 [1 + sin2 (3π (xi + 30) + 1)] + 30 [−50,50] −80
∑n
((xn + 30) − 1)2 [1 + sin2 (2π (xn + 30))]} + i=1 u((xi + 30), 5, 100, 4) − 80

Fig. 5. Benchmark suite 1.

• Benchmark Suite 1: 12 shifted and biased standard bench- benchmarking the proposed variant. In this suite, 29 func-
mark functions having unimodal and multimodal character- tions of different nature are incorporated. The details of these
istics are used here. The details of benchmark suite 1 are taken functions such as the search range, composition and proper-
from [29,51,52] and [53]. The mathematical definitions and ties are described in Table 2. Functions of this suite can be
other relevant details of these functions such as boundaries characterized as unimodal, multimodal, hybrid and compos-
of variable and dimension of the problem are incorporated ite functions. Mathematical definitions of these functions are
in Table 1. Two dimensional version of the benchmark func- given in [54].
tions are shown in Fig. 5. In this suite, six unimodal and six To benchmark the proposed variant, total 41 functions are
multimodal shifted and biased functions are considered for employed from suite 1 and suite 2. The following section
benchmarking. reports the simulation results and numerical experiments
• Benchmark Suite 2: Real parameter single objective opti- performed on these two suites by various variants of GWO
mization problems at CEC 2017 [54] are also used here for and some other contemporary algorithms.
90 A. Saxena, R. Kumar and S. Das / Applied Soft Computing Journal 75 (2019) 84–105

Table 2
Details of CEC-2017 (benchmark suite 2).
Function nature Function name Optima
Shifted and Rotated Bent Cigar function (F-1) 100
Unimodal
Shifted and Rotated Zakharov function (F-3) 300
Shifted and Rotated Rosenbrock’s function (F-4) 400
Shifted and Rotated Rastrigin’s function (F-5) 500
Shifted and Rotated Expanded Scaffer’s function (F-6) 600
Simple multimodal functions Shifted and Rotated Lunacek Bi_Rastrigin Function (F-7) 700
Shifted and Rotated Non-continuous Rastrigin’s Function (F-8) 800
Shifted and Rotated Levy Function (F-9) 900
Shifted and Rotated Schwefel’s Function (F-10) 1000
Hybrid Function 1 (N = 3) (F-11) 1100
Hybrid Function 2 (N = 3) (F-12) 1200
Hybrid Function 3 (N = 3) (F-13) 1300
Hybrid Function 4 (N = 4) (F-14) 1400
Hybrid Function 5 (N = 4) (F-15) 1500
Hybrid functions
Hybrid Function 6 (N = 4) (F-16) 1600
Hybrid Function 7 (N = 5) (F-17) 1700
Hybrid Function 8 (N = 5) (F-18) 1800
Hybrid Function 9 (N = 5) (F-19) 1900
Hybrid Function 10 (N = 6) (F-20) 2000
Composition Function 1 (N = 3) (F-21) 2100
Composition Function 2 (N = 3) (F-22) 2200
Composition Function 3 (N = 4) (F-23) 2300
Composition Function 4 (N = 4) (F-24) 2400
Composition Function 5 (N = 5) (F-25) 2500
Composition function
Composition Function 6 (N = 5) (F-26) 2600
Composition Function 7 (N = 6) (F-27) 2700
Composition Function 8 (N = 6) (F-28) 2800
Composition Function 9 (N = 3) (F-29) 2900
Composition Function 10 (N = 3) (F-30) 3000

5. Numerical experiments and simulation results Table 3


Details of control parameters.

In this section, numerical experiments and results on both Algorithm Parameter Value

benchmark suite 1 and 2 are reported. In subsections, authors Awareness probability 0.1
CSA
Flight length 2
present following analysis to establish the efficacy of proposed
chaotic bridging mechanism: Probability switch 0.8
FPA
Gama 1.5
1. The impact of Sensitivity Dependence on Initial Conditions α 20
(SDIC) on β -GWO. GSA
G0 100
Rnorm 2
2. The comparison of the performance of the β -GWO with
Rpower 1
other contemporary algorithms and significance analysis by
Nm max 0.2
Wilcoxon rank-sum test [55] on benchmark suite 1. β -GWO
Nm min 1e−10
3. The comparison of the performance of the β -GWO with
other different variants of GWO and with other contempo-
rary algorithms on benchmark suite 2 along with statistical
comparison by Wilcoxon rank-sum test [55] . β -GWO (x1 = 0.8) and β -GWO (x1 = 0.9)) is carried out. From the
results reported in Table 4, it is empirical to judge that for each
5.1. Sensitivity dependence on initial conditions starting chaos point (x1 ), algorithm behaves differently. However,
due to stochastic nature of the algorithms the best results obtained
It is empirical to state that the chaotic behavior is highly sen- from 20 independent runs are reported in this set.
sitive to its initial condition, surprisingly the effect is addressed From this experiment, following conclusions can be drawn:
very less in literature with reference to optimization performance
of chaotic algorithms. In this section, investigation on the effect 1. It is observed that the mean values are optimal for 5 func-
of initial starting point of chaotic sequence on the performance tions (F1, F2, F5, F8 and F10) out of 12 functions for initial
of proposed β -GWO is performed. The parameters of β -GWO is point 0.8, (β -GWO (0.8)). It is observed that standard devia-
defined in Table 3. tion values for functions (F5, F6, F8 and F10) are optimal for
The experiment results for this study are presented in Table 4. this variant too. The low values of standard deviation (SD)
First 6 functions are unimodal benchmark functions and these are is an indicator of better solution quality. However, for initial
used for test the exploitation capability of the proposed algorithm.
point 0.7, β -GWO (0.7), also shows promising results in a
The hypothesis is that initial conditions effect the performance
few cases.
of chaotic algorithms. To judge that, different initial points for
2. The effect of SDIC can be observed from the results of Ta-
generation of the chaos is chosen and on the basis of initial gen-
eration points (Values of initial points x1 varies from 0.4 to 0.9 ) six ble 4. β -GWO (0.8) gives favorable results as the mean and
different algorithms are framed with six different chaos vector and standard deviation obtained for the functions are optimal
−→
embedded the same in control vector a . In this study, comparison and convergence of the algorithm is faster. Hence, it can be
of the performance of these six different algorithms i.e. (β -GWO (x1 concluded that for these shifted benchmark functions the
= 0.4), β -GWO (x1 = 0.5), β -GWO (x1 = 0.6), β -GWO (x1 = 0.7), (β -GWO (0.8)) algorithm is best suited.
A. Saxena, R. Kumar and S. Das / Applied Soft Computing Journal 75 (2019) 84–105 91

Table 4
Sensitivity analysis of β -GWO with different initial starting point of chaos.
Function x1 = 0.4 x1 = 0.5 x1 = 0.6 x1 = 0.7 x1 = 0.8 x1 = 0.9
Mean SD Mean SD Mean SD Mean SD Mean SD Mean SD
F1 1.82E+03 1.59E+03 1.83E+03 1.08E+03 2.33E+03 1.52E+03 2.21E+03 1.39E+03 1.58E+03 1.64E+03 1.76E+03 1.74E+03
F2 −2.12E+01 1.67E+01 −1.77E+01 2.47E+01 −1.46E+01 2.34E+01 −2.51E+01 1.67E+01 −2.64E+01 2.01E+01 −2.43E+01 2.09E+01
F3 7.57E+03 3.13E+03 8.12E+03 2.70E+03 9.15E+03 3.76E+03 6.52E+03 2.28E+03 7.12E+03 2.60E+03 6.74E+03 3.09E+03
F4 −7.89E+01 3.01E−01 −7.91E+01 2.20E−01 −7.89E+01 2.85E−01 −7.90E+01 3.12E−01 −7.90E+01 2.96E−01 −7.89E+01 2.76E−01
F5 2.68E+09 8.73E+08 2.88E+09 8.13E+08 2.45E+09 4.84E+08 2.43E+09 4.62E+08 2.43E+09 4.50E+08 2.56E+09 7.61E+08
F6 2.11E+03 2.48E+03 1.94E+03 2.22E+03 1.94E+03 2.77E+03 1.01E+03 2.67E+03 1.39E+03 2.19E+03 1.33E+03 2.65E+03
F7 −1.10E+04 1.52E+03 −1.16E+04 9.92E+02 −1.17E+04 1.14E+03 −1.15E+04 1.14E+03 −1.15E+04 1.10E+03 −1.17E+04 1.22E+03
F8 −4.33E+00 1.20E+01 3.02E+00 2.14E+01 2.02E+00 2.74E+01 −3.37E+00 1.46E+01 −4.98E+00 1.14E+01 5.21E+00 1.98E+01
F9 −7.18E+01 7.95E+00 −6.99E+01 7.65E+00 −7.20E+01 6.32E+00 −7.22E+01 8.20E+00 −6.87E+01 6.96E+00 −7.23E+01 7.02E+00
F10 −6.94E+01 1.73E+01 −6.11E+01 2.04E+01 −7.31E+01 1.45E+01 −6.79E+01 2.21E+01 −7.52E+01 1.18E+01 −6.69E+01 1.90E+01
F11 8.00E+05 3.58E+06 2.40E+06 5.86E+06 −7.99E+01 6.39E−02 8.00E+05 3.58E+06 8.00E+05 3.58E+06 −7.99E+01 8.52E−02
F12 3.91E+06 1.20E+07 −7.99E+01 3.20E−02 −7.99E+01 3.23E−02 3.91E+06 1.20E+07 1.95E+06 8.73E+06 5.21E+06 1.30E+07

Fig. 6. Convergence curves of benchmark suite-1.


92 A. Saxena, R. Kumar and S. Das / Applied Soft Computing Journal 75 (2019) 84–105

Table 5
Results of benchmark suite-I.
Function Parameter GWO [34] OGWO [38] β -GWO GGWO [42] mGWO [39] IGWO [44] MGWO [43] GSA [15] CSA [56] FPA [57]
p-values 6.20E−01 6.80E−08 – 6.80E−08 4.25E−01 9.17E−08 6.80E−08 1.48E−07 6.80E−08 1.48E−07
F1 SD 1.63E+03 3.18E+03 1.40E+03 1.33E+04 1.97E+03 2.40E+03 2.11E+03 1.91E+03 3.61E+03 3.29E+03
Mean 2.14E+03 3.44E+04 1.99E+03 8.85E+04 2.38E+03 9.93E+03 1.66E+04 7.91E+03 2.01E+04 1.19E+04
p-values 6.50E−01 6.80E−08 – 6.80E−08 2.62E−01 6.80E−08 6.80E−08 6.75E−01 6.80E−08 6.80E−08
F2 SD 2.40E+01 3.87E+16 1.90E+01 7.47E+15 2.37E+01 5.11E+07 9.99E+10 2.38E+01 1.08E+27 1.32E+05
Mean −1.80E+01 9.12E+15 −2.22E+01 3.45E+15 −6.45E+00 1.15E+07 2.77E+10 −2.20E+01 2.63E+26 4.10E+04
p-values 7.64E−02 6.80E−08 – 6.80E−08 1.35E−03 6.80E−08 6.80E−08 6.80E−08 6.80E−08 6.80E−08
F3 SD 3.97E+03 2.25E+06 3.61E+03 2.59E+06 2.80E+03 2.79E+04 1.55E+05 1.47E+06 9.69E+05 2.17E+04
Mean 9.79E+03 1.00E+07 7.29E+03 1.42E+07 8.93E+03 3.65E+04 3.98E+05 8.06E+06 1.73E+06 4.37E+04
p-values 4.09E−01 8.01E−09 – 6.80E−08 1.23E−07 6.71E−08 6.80E−08 6.80E−08 3.37E−08 6.80E−08
F4 SD 3.18E−01 0.00E+00 1.58E−01 5.46E+00 5.21E−01 8.43E−11 1.63E−01 2.59E+00 6.18E+00 1.86E+01
Mean −7.90E+01 −2.00E+01 −7.91E+01 5.47E+01 −7.79E+01 −2.00E+01 −2.01E+01 −1.87E+01 −2.36E+01 −1.97E+01
p-values 1.20E−06 6.80E−08 – 6.80E−08 1.79E−04 3.94E−07 6.80E−08 1.80E−06 7.41E−05 2.22E−04
F5 SD 6.94E+08 1.68E+09 6.88E+08 3.51E+09 5.88E+08 2.48E+09 1.08E+09 6.31E+08 3.28E+09 2.63E+09
Mean 2.60E+09 2.79E+10 2.53E+09 3.80E+10 2.93E+09 6.30E+09 1.48E+10 4.40E+09 8.31E+09 7.97E+09
p-values 1.64E−01 6.80E−08 – 6.80E−08 8.59E−02 6.80E−08 6.80E−08 6.80E−08 6.80E−08 2.56E−07
F6 SD 2.98E+03 8.40E+03 2.59E+03 1.64E+04 1.84E+03 6.02E+03 4.13E+03 3.09E+03 7.97E+03 8.52E+03
Mean 1.87E+03 7.85E+04 1.36E+03 1.34E+05 1.09E+03 2.36E+04 3.61E+04 3.83E+04 3.64E+04 1.78E+04
p-values 3.23E−01 6.80E−08 – 6.80E−08 7.56E−01 4.68E−05 6.80E−08 6.66E−08 6.04E−08 3.35E−04
F7 SD 1.25E+03 8.41E+02 1.32E+03 6.87E+02 1.00E+03 8.29E+02 6.04E+02 7.82E+02 1.52E+03 5.06E+02
Mean −1.13E+04 −4.88E+03 −1.22E+04 −5.05E+03 −1.15E+04 −1.01E+04 −5.75E+03 −5.12E+03 −2.31E+05 −1.09E+04
p-values 3.37E−01 6.80E−08 – 6.80E−08 9.03E−01 7.90E−08 6.80E−08 3.79E−01 5.22E−07 6.79E−08
F8 SD 6.44E+08 2.20E+00 7.11E+08 4.16E+01 2.34E+01 3.54E+01 2.18E+00 1.01E+09 2.52E+26 1.89E+09
Mean 1.83E+08 3.79E+01 1.39E+00 4.27E+02 1.60E+00 5.80E+01 3.12E+01 2.52E+08 6.72E+25 5.30E+08
p-values 1.35E−05 1.41E−05 – 6.80E−08 3.94E−01 1.77E−06 6.80E−08 8.29E−08 5.72E−09 1.10E−03
F9 SD 6.17E+08 1.86E−01 6.54E+08 9.14E−02 7.29E+00 3.03E−01 2.89E−01 1.02E+09 2.55E+26 1.91E+09
Mean 2.06E+08 −6.03E+01 −6.84E+01 −5.87E+01 −6.80E+01 −6.01E+01 −6.01E+01 3.15E+08 8.31E+25 6.51E+08
p-values 7.35E−01 6.80E−08 – 6.80E−08 1.55E−02 1.92E−07 6.80E−08 6.80E−08 6.68E−08 1.43E−07
F10 SD 6.37E+08 6.82E+01 6.20E+08 1.98E+02 3.27E+01 9.40E+01 5.04E+01 1.03E+09 8.01E+25 1.92E+09
Mean 2.81E+08 7.66E+02 −7.11E+01 1.16E+03 −5.71E+01 1.66E+02 2.90E+02 3.82E+08 3.28E+25 7.79E+08
p-values 5.40E−01 6.80E−08 – 6.80E−08 5.63E−04 6.80E−08 6.80E−08 1.60E−05 6.80E−08 2.69E−06
F11 SD 6.82E+08 3.48E+07 6.84E+08 1.26E+09 4.92E+06 1.67E+07 3.14E+07 1.03E+09 8.08E+25 1.91E+09
Mean 3.26E+08 3.38E+08 −7.99E+01 4.47E+09 1.60E+06 2.31E+07 1.06E+08 4.52E+08 3.85E+25 9.14E+08
p-values 1.20E−03 6.80E−08 – 6.80E−08 5.87E−06 1.92E−07 6.80E−08 1.16E−04 6.80E−08 1.50E−04
F12 SD 6.16E+08 6.79E+07 6.38E+08 1.23E+09 8.73E+06 5.05E+07 5.12E+07 1.02E+09 8.02E+25 1.90E+09
Mean 3.44E+08 8.53E+08 3.10E+08 6.01E+09 1.95E+06 5.79E+07 2.64E+08 5.26E+08 4.45E+25 1.05E+09

Hence, the results of the performance of (β -GWO (0.8)) with other The blank (–) entry refers that the proposed algorithm is com-
contemporary algorithms are exhibited in following study. pared with the other opponents to ensure the significant statistical
difference. For unimodal functions the β -GWO results are com-
5.2. Performance of β -GWO with different variants of GWO and other petitive as the p-values are less than 0.05 for most of the cases.
contemporary algorithms (benchmark suite 1) Similarly, for multimodal functions F8, F9, F10 and F11 the β -GWO
shows promising results. The convergence curves of the algorithms
In this study, the optimization performance of (β -GWO (0.8)) are shown in Fig. 6. In disparity to unimodal functions, multimodal
is compared (which will be referred as β -GWO) with GWO [34] functions possess several optima and hence considered as difficult
and variants of GWO namely OGWO [38], GGWO [42], mGWO [39] functions. The algorithms which possess loose mechanism have
,IGWO [44] and MGWO [43]. Some of the well-known algorithms the high probability to be trapped in local minima. The results of
are also chosen for the comparison of the proposed variant. These β -GWO are competitive for these functions too.
algorithms are Flower Pollination Algorithm (FPA) [57], Gravita- For these functions obtained p-values indicate that the results
tional Search Algorithm (GSA) [15] and Crow Search Algorithm of β -GWO are competitive and β -GWO and GWO are the most sub-
(CSA) [56]. To maintain a fair competitive environment, the stop-
stantial algorithms. The results of β -GWO are optimal for 10 func-
ping criterion for all the algorithms is maximum iteration. The
tions out of 12. Fig. 7 shows the mesh plots of solution obtained for
variables like search agents, no. of maximum iteration are kept
different run for functions F2, F3, F4, F7, F9 and F11. It is observed
same for all the algorithms. The other parameters of the respective
that the population diversity does exist in the obtained positions
algorithms are given in Table 3 and these are taken from the
of wolves. This fact advocates positive implication of embedded
references [57,15,56]. The Table 5 shows the optimization results
of the algorithms. chaos sequence. Also, for these 5 functions the proposed algorithm
From the results of unimodal functions it can be concluded that performs better than other opponents. However, for function F7
the mean values of β -GWO is optimal for all six functions. To obtain population diversity is not observed prominently, which reflects
fair comparison Wilcoxon rank-sum test [55] is performed and p- in the performance of the algorithm for this function. In short, for
values have been obtained for the comparison with the confidence the most of the cases positive implication of chaotic sequence can
interval of 5 percent. It is worth mentioning here that the values of be observed with the convergence curves and values of different
standard deviation for unimodal functions are also competitive for statistical parameters. Further, it can be concluded that the adap-
β -GWO. These low values are indicator of better solution quality tive chaotic mechanism substantially improved the convergence
obtained from independent runs. (exploitation) and exploration capability of classical GWO.
A. Saxena, R. Kumar and S. Das / Applied Soft Computing Journal 75 (2019) 84–105 93

Fig. 7. Effect of β -chaotic sequence on the solutions.

5.3. Performance of β -GWO with different variants of GWO and other variant and other algorithms as the obtained p-values for
contemporary algorithms (Benchmark Suite 2) other algorithms are less than 0.05. The blank entry (–) is
shown in the column of β -GWO as it is compared with others.
The optimization results of proposed variant are presented in Hence it can be concluded that for 23 functions it outperforms
this section on benchmark suite 2. Each function of the suite others. However, in some cases other algorithms outperform
is iterated over 51 independent runs. The initial population is β -GWO. The results obtained from this analysis advocates
generated by random sequence generator. Two main important supremacy of the proposed variant.
parameters mean and standard deviation (SD) of these runs are • Convergence curves of this experiment is shown in Fig. 8 for
reported. The search range and other relevant details of this suite is function 4, 5, 6 and 7, in Fig. 9 for function 8, 9, 10 and 11, in
given in Table 2. The performance of β -GWO is compared with the Fig. 10 for function 15, 16, 19 and 29 and in Fig. 11 for function
GWO [34] and variants of GWO namely, OGWO [38], GGWO [42], 17, 18, 22 and 24. By inspecting the convergence curves of this
mGWO [39], IGWO [44], MGWO [43]. Further the comparison of experiment, it is empirical to state that a significant improve-
the optimization performance of β -GWO with GSA [15], CSA [56] ment has been observed in convergence characteristics by the
and FPA [57] is also added in this analysis to establish the efficacy proposed chaotic mechanism.
of the proposed variant. The results reported in this section have
been divided into two folds i.e. results on 10 dimension (10-D) and 5.3.2. Optimization results of 30-D problems
30 dimension (30-D) problems. The analysis on these two multi- The analysis is aggregated towards high dimensions by consid-
dimension problems are carried out for testing the scalability of ering the dimension ‘30’. The results of this analysis have been
the proposed variant. reported in Table 8 for first 14 functions and rest are reported in
9.
5.3.1. Optimization Results of 10-D problems
The results on 10 dimension problems are reported into two • Inspecting the mean values of 51 runs, it is observed that
folds: In Table 6 results of first 14 functions are reported and mean values of 20 functions are optimal for β -GWO as com-
remaining results are reported in Table 7. Wilcoxon rank-sum test pared with the other opponents. The similar performance
results are also incorporated in this analysis. Following points can of proposed β -GWO for high dimension problem, makes it
be worth mentioning from the results: a suitable choice for complex real applications. It is worth
mentioning here that in some cases, especially for function
• It has been observed that out of 29 functions, the mean 13, 14 and 18 the performance of CSA is much better as a huge
values observed for 23 functions are optimal for the proposed numerical difference exist between the mean values of other
variant. The performance of β -GWO is competitive. However, participants.
for function 1, 3, 12, 13, 18 and 30 the performance of CSA is • For these 20 functions the p-values of Wilcoxon rank-sum
better as indicated by low mean values. It is empirical to state test is lower than 0.05 for most of the cases. This fact shows
that the performance of the proposed variant is better than that the proposed β -GWO is statistically different from the
other variants of GWO. other opponents and hence, it can be concluded that it out
• Wilcoxon rank-sum test is also conducted and the perfor- performs others in these cases. For remaining cases the per-
mance of β -GWO is statistically compared with other oppo- formance of algorithm is competitive.
nents by considering 5% level of significance. By inspecting p- • As per the standard deviation values obtained from indepen-
values of the Wilcoxon rank-sum test, it can be concluded that dent runs, it is empirical to state that the performance of β -
there is a significant difference exists between the proposed GWO is competitive. The β -GWO is able to provide a fair trade
94 A. Saxena, R. Kumar and S. Das / Applied Soft Computing Journal 75 (2019) 84–105

Table 6
Results of CEC 2017 for 10-D problems [54].
Functions Statistical parameters Algorithms
GWO [34] OGWO [38] β -GWO GGWO [42] mGWO [39] IGWO [44] MGWO [43] GSA [15] CSA [56] FPA [57]
Mean 2.09E+06 5.98E+09 1.46E+07 1.64E+10 1.69E+07 4.68E+07 1.44E+08 2.24E+02 1.00E+02 1.57E+06
F-1 SD 7.99E+06 1.61E+09 6.71E+07 4.59E+09 8.01E+07 1.23E+08 1.31E+08 1.59E+02 3.28E−10 9.08E+05
p-values 3.30E−18 5.97E−01 – 3.30E−18 1.63E−02 7.43E−04 5.00E−16 3.50E−18 3.30E−18 1.82E−12
Mean 6.76E+02 8.48E+03 6.06E+02 1.87E+05 4.72E+02 3.99E+03 1.42E+03 3.24E+02 3.00E+02 1.11E+03
F-3 SD 8.02E+02 2.35E+03 7.57E+02 7.28E+05 3.11E+02 3.42E+03 1.53E+03 3.98E+01 2.54E−14 5.50E+02
p-values 3.30E−18 2.28E−01 – 3.30E−18 9.79E−01 2.19E−11 5.19E−12 9.46E−10 2.22E−19 2.63E−11
Mean 4.10E+02 6.78E+02 4.00E+02 2.09E+03 4.12E+02 4.28E+02 4.26E+02 4.02E+02 4.08E+02 4.17E+02
F-4 SD 1.26E+01 1.15E+02 4.71E+00 6.51E+02 1.62E+01 3.06E+01 2.07E+01 2.40E−01 5.08E−14 2.60E+01
p-values 3.30E−18 4.42E−01 – 3.30E−18 2.77E−02 3.99E−04 9.14E−16 3.30E−18 1.07E−18 5.91E−02
Mean 5.11E+02 5.82E+02 5.10E+02 6.45E+02 5.11E+02 5.21E+02 5.32E+02 5.35E+02 5.19E+02 5.36E+02
F-5 SD 4.86E+00 1.44E+01 5.70E+00 2.32E+01 5.54E+00 1.08E+01 5.16E+00 5.84E+00 6.69E+00 1.04E+01
p-values 3.30E−18 1.90E−01 – 3.30E−18 7.23E−01 8.60E−08 2.26E−17 9.95E−18 1.96E−08 3.57E−17
Mean 6.00E+02 6.48E+02 6.00E+02 6.84E+02 6.00E+02 6.05E+02 6.09E+02 6.02E+02 6.03E+02 6.25E+02
F-6 SD 4.88E−01 8.39E+00 4.56E−01 1.32E+01 5.57E−01 5.72E+00 3.37E+00 2.83E+00 2.54E+00 1.01E+01
p-values 3.30E−18 1.22E−01 – 3.30E−18 8.54E−05 6.95E−16 3.30E−18 6.87E−02 5.28E−16 3.30E−18
Mean 7.23E+02 8.06E+02 7.11E+02 1.13E+03 7.24E+02 7.35E+02 7.50E+02 7.23E+02 7.24E+02 7.93E+02
F-7 SD 7.79E+00 1.13E+01 7.98E+00 6.83E+01 7.04E+00 1.06E+01 6.20E+00 6.48E−01 5.30E+00 2.95E+01
p-values 3.30E−18 9.25E−01 – 3.30E−18 4.45E−01 1.56E−08 1.60E−17 3.71E−18 3.56E−01 3.30E−18
Mean 8.11E+02 8.52E+02 8.10E+02 9.30E+02 8.10E+02 8.14E+02 8.23E+02 8.18E+02 8.12E+02 8.40E+02
F-8 SD 4.38E+00 9.07E+00 4.45E+00 1.51E+01 4.37E+00 5.70E+00 4.51E+00 3.44E+00 4.00E+00 1.34E+01
p-values 3.30E−18 2.61E−01 – 3.30E−18 8.30E−01 2.71E−05 1.32E−16 2.60E−13 6.56E−02 3.57E−17
Mean 9.03E+02 1.32E+03 9.00E+02 4.62E+03 9.01E+02 9.39E+02 9.20E+02 9.03E+02 9.01E+02 1.42E+03
F-9 SD 8.97E+00 1.56E+02 8.59E+00 1.08E+03 2.79E+00 3.94E+01 2.02E+01 0.00E+00 7.12E−01 2.91E+02
p-values 3.30E−18 7.63E−01 – 3.30E−18 5.12E−01 7.56E−12 4.15E−14 1.39E−20 8.37E−02 3.30E−18
Mean 1.41E+03 2.82E+03 1.37E+03 3.56E+03 1.41E+03 1.78E+03 1.90E+03 2.63E+03 1.73E+03 1.87E+03
F-10 SD 2.40E+02 2.13E+02 2.27E+02 2.37E+02 2.07E+02 2.82E+02 2.67E+02 2.61E+02 2.70E+02 2.23E+02
p-values 3.30E−18 5.34E−01 – 3.30E−18 2.58E−01 5.42E−11 8.94E−14 3.30E−18 6.53E−09 5.10E−14
Mean 1.12E+03 1.50E+03 1.12E+03 1.21E+04 1.12E+03 1.15E+03 1.15E+03 1.13E+03 1.13E+03 1.19E+03
F-11 SD 2.19E+01 6.21E+02 1.77E+01 1.49E+04 9.89E+00 6.37E+01 1.75E+01 7.60E+00 1.47E+01 7.93E+01
p-values 3.50E−18 6.49E−01 – 3.30E−18 3.28E−01 1.85E−07 1.34E−15 1.60E−07 1.20E−03 8.50E−14
Mean 3.10E+05 7.17E+07 3.16E+05 1.64E+09 2.80E+05 5.18E+05 1.19E+06 8.65E+03 1.98E+03 4.06E+06
F-12 SD 5.78E+05 9.79E+07 5.97E+05 8.46E+08 4.53E+05 6.73E+05 1.16E+06 1.83E+03 6.19E+02 4.81E+06
p-values 3.30E−18 2.87E−01 – 3.30E−18 7.38E−01 5.91E−02 2.89E−10 6.58E−16 3.30E−18 6.59E−14
Mean 8.16E+03 3.37E+04 8.01E+03 9.99E+07 6.98E+03 7.56E+03 8.73E+03 1.03E+04 1.58E+03 1.77E+04
F-13 SD 2.91E+03 2.47E+04 3.46E+03 1.07E+08 3.14E+03 3.73E+03 1.85E+03 1.82E+03 1.97E+02 1.25E+04
p-values 1.56E−16 9.31E−01 – 3.30E−18 1.07E−01 4.26E−01 5.47E−01 1.07E−03 3.30E−18 5.14E−05
Mean 1.75E+03 4.70E+03 1.44E+03 2.02E+06 1.53E+03 1.68E+03 1.92E+03 5.63E+03 1.53E+03 1.57E+03
F-14 SD 9.90E+02 9.75E+02 5.13E+02 3.27E+06 5.14E+02 8.53E+02 1.14E+03 1.36E+03 1.35E+01 6.67E+01
p-values 1.07E−17 1.39E−01 – 3.50E−18 9.73E−01 3.76E−03 5.05E−10 7.96E−18 1.17E−09 3.79E−16

Fig. 8. Convergence curves of other algorithms with β -GWO for functions (4–7).
A. Saxena, R. Kumar and S. Das / Applied Soft Computing Journal 75 (2019) 84–105 95

Table 7
Results of CEC 2017 for 10-D problems [54].
Functions Statistical parameters Algorithms
GWO [34] OGWO [38] β -GWO GGWO [42] mGWO [39] IGWO [44] MGWO [43] GSA [15] CSA [56] FPA [57]
Mean 2.12E+03 1.48E+04 1.54E+03 7.90E+04 2.08E+03 3.20E+03 2.91E+03 1.43E+04 1.79E+03 2.53E+03
F-15 SD 1.15E+03 7.15E+03 7.08E+02 9.25E+04 1.12E+03 1.87E+03 1.42E+03 3.56E+03 3.85E+01 6.00E+02
p-values 6.30E−18 2.86E−02 – 4.99E−18 1.75E−02 6.38E−08 4.51E−12 3.30E−18 2.43E−03 2.21E−12
Mean 1.70E+03 2.00E+03 1.65E+03 2.57E+03 1.65E+03 1.72E+03 1.74E+03 2.08E+03 1.69E+03 1.69E+03
F-16 SD 9.89E+01 1.46E+02 5.52E+01 2.16E+02 6.39E+01 9.95E+01 1.20E+02 7.97E+01 9.18E+01 7.63E+01
p-values 5.29E−18 4.33E−02 – 3.30E−18 1.56E−01 1.74E−04 1.22E−06 3.30E−18 8.20E−01 1.25E−02
Mean 1.74E+03 1.81E+03 1.73E+03 2.26E+03 1.73E+03 1.76E+03 1.76E+03 1.80E+03 1.74E+03 1.77E+03
F-17 SD 2.68E+01 1.35E+01 1.14E+01 1.54E+02 1.22E+01 2.34E+01 1.33E+01 6.34E+01 1.13E+01 2.12E+01
p-values 3.30E−18 3.15E−01 – 3.30E−18 4.14E−01 1.68E−08 7.75E−16 1.47E−16 2.14E−07 2.18E−15
Mean 2.54E+04 3.93E+05 2.64E+04 4.37E+08 2.68E+04 1.34E+04 2.82E+04 6.68E+03 1.89E+03 2.14E+04
F-18 SD 1.39E+04 4.79E+05 1.36E+04 5.47E+08 1.40E+04 1.09E+04 8.57E+03 1.92E+03 5.73E+01 1.29E+04
p-values 2.43E−16 7.99E−01 – 3.30E−18 8.51E−01 3.75E−06 6.54E−01 1.81E−13 3.30E−18 1.34E−01
Mean 3.12E+03 2.27E+05 1.91E+03 2.99E+07 2.14E+03 4.37E+03 4.26E+03 8.19E+03 1.91E+03 2.79E+03
F-19 SD 3.42E+03 4.07E+05 3.06E+03 1.01E+08 1.58E+03 4.71E+03 4.50E+03 1.83E+03 8.99E+00 8.26E+02
p-values 1.42E−17 8.72E−01 – 3.30E−18 2.58E−01 2.14E−05 2.63E−05 2.21E−13 1.11E−07 2.99E−13
Mean 2.05E+03 2.21E+03 2.05E+03 2.48E+03 2.05E+03 2.07E+03 2.08E+03 2.20E+03 2.05E+03 2.08E+03
F-20 SD 4.07E+01 6.29E+01 4.21E+01 1.29E+02 4.45E+01 5.02E+01 4.96E+01 3.52E+01 2.56E+01 2.79E+01
p-values 2.57E−16 2.11E−01 – 3.30E−18 1.50E−01 1.49E−05 1.64E−09 5.95E−18 2.43E−03 7.14E−08
Mean 2.30E+03 2.34E+03 2.20E+03 2.42E+03 2.30E+03 2.21E+03 2.29E+03 2.33E+03 2.20E+03 2.21E+03
F-21 SD 3.01E+01 5.58E+01 3.89E+01 3.51E+01 3.61E+01 1.63E+01 5.43E+01 2.18E+01 8.68E−01 1.95E+00
p-values 1.01E−04 8.62E−01 – 4.49E−17 7.89E−01 2.65E−10 1.95E−05 2.10E−12 3.76E−17 2.54E−10
Mean 2.30E+03 2.72E+03 2.29E+03 3.78E+03 2.30E+03 2.32E+03 2.33E+03 2.30E+03 2.30E+03 2.29E+03
F-22 SD 5.38E+00 1.17E+02 9.24E+01 4.21E+02 1.39E+01 3.05E+01 7.94E+00 1.80E−11 1.91E+01 2.92E+01
p-values 5.97E−17 1.48E−01 – 3.94E−18 2.18E−01 2.22E−03 5.70E−15 3.30E−18 1.82E−06 4.54E−02
Mean 2.61E+03 2.75E+03 2.60E+03 2.82E+03 2.61E+03 2.63E+03 2.64E+03 2.67E+03 2.60E+03 2.64E+03
F-23 SD 6.06E+00 2.84E+01 6.54E+00 3.97E+01 6.79E+00 1.21E+01 7.34E+00 1.60E+01 9.90E+01 1.33E+01
p-values 3.30E−18 5.69E−01 – 3.30E−18 1.58E−01 8.00E−13 4.99E−18 3.30E−18 5.56E−09 1.49E−15
Mean 2.74E+03 2.81E+03 2.54E+03 2.96E+03 2.74E+03 2.74E+03 2.75E+03 2.54E+03 2.51E+03 2.55E+03
F-24 SD 7.14E+00 1.02E+02 6.89E+00 5.65E+01 1.39E+01 5.26E+01 5.63E+01 1.04E+02 5.50E+01 1.01E+02
p-values 2.86E−02 7.99E−01 – 3.30E−18 1.50E−01 1.48E−04 2.75E−14 3.74E−10 7.34E−16 8.60E−08
Mean 2.93E+03 3.17E+03 2.91E+03 4.13E+03 2.93E+03 2.93E+03 2.93E+03 2.94E+03 2.92E+03 2.95E+03
F-25 SD 2.21E+01 7.48E+01 1.88E+01 4.76E+02 2.04E+01 1.96E+01 1.34E+01 1.20E+01 2.20E+01 3.95E+01
p-values 3.30E−18 3.32E−01 – 3.30E−18 6.06E−01 9.73E−01 2.13E−01 7.11E−01 1.19E−05 2.58E−04
Mean 2.92E+03 3.78E+03 2.86E+03 4.70E+03 2.92E+03 3.06E+03 3.02E+03 2.86E+03 2.89E+03 3.03E+03
F-26 SD 1.44E+02 2.36E+02 2.97E+01 4.41E+02 1.27E+02 1.75E+02 5.62E+01 1.92E+02 6.11E+01 9.75E+01
p-values 3.30E−18 6.44E−01 – 3.30E−18 6.25E−10 1.20E−15 7.75E−16 3.79E−16 6.88E−12 8.19E−16
Mean 3.09E+03 3.23E+03 3.09E+03 3.36E+03 3.10E+03 3.09E+03 3.10E+03 3.17E+03 3.10E+03 3.10E+03
F-27 SD 2.28E+00 3.43E+01 1.20E+01 1.02E+02 1.78E+01 4.39E+00 1.00E+01 1.94E+01 7.97E+00 3.27E+00
p-values 3.50E−18 8.53E−03 – 3.30E−18 6.83E−01 3.12E−02 3.38E−14 1.00E−17 2.33E−13 1.18E−12
Mean 3.33E+03 3.71E+03 3.12E+03 3.99E+03 3.34E+03 3.22E+03 3.37E+03 3.41E+03 3.12E+03 3.22E+03
F-28 SD 1.10E+02 1.04E+02 1.03E+02 1.87E+02 1.01E+02 8.38E+01 1.03E+02 1.66E+01 7.13E+01 6.26E+01
p-values 3.30E−18 6.06E−01 – 1.42E−17 4.82E−01 1.68E−05 2.72E−04 1.94E−03 1.34E−15 4.00E−06
Mean 3.17E+03 3.41E+03 3.16E+03 3.83E+03 3.16E+03 3.19E+03 3.21E+03 3.25E+03 3.18E+03 3.20E+03
F-29 SD 3.50E+01 1.00E+02 3.17E+01 1.83E+02 2.00E+01 4.26E+01 2.91E+01 7.31E+01 1.96E+01 3.22E+01
p-values 7.96E−18 3.84E−01 – 3.30E−18 9.89E−01 2.63E−05 1.32E−11 3.56E−14 1.48E−04 6.03E−09
Mean 5.23E+05 1.34E+07 4.80E+05 9.59E+07 5.07E+05 1.01E+05 7.37E+05 1.03E+05 3.91E+03 1.84E+05
F-30 SD 6.66E+05 1.22E+07 6.30E+05 5.29E+07 5.64E+05 1.94E+05 9.20E+05 1.56E+04 4.79E+02 2.72E+05
p-values 1.34E−17 9.52E−01 – 3.30E−18 4.58E−01 7.79E−01 2.79E−04 6.09E−02 4.49E−17 8.20E−01

off between convergence properties and obtained solution • For 10-D problems, it is observed that box plots are super
quality. narrow for F-5, F-10, F-20 and F-23 for β -GWO. Narrow box
plots suggest that the obtained solutions possess high level
5.4. Box plot analysis of agreement with each other. We observed that for above
mentioned functions solution quality obtained by proposed
variant is superior. Ranges for function 5, 10, 11, 16, 17 and
To study the distributional characteristics of the experiments
22 are optimal. These results advocate the supremacy of the
performed in previous subsections, a box plot analysis is conducted
β -GWO over GWO.
and a few of boxplots for 10-D and 30-D problems are shown in
• Similarly, we observed that for 30-D problems where the IQR
Figs. 12 and 13. Box plot is a standardized way of demonstrating is narrow for F-1, F-8, F-18, F-23, F-26 and F-27 in the case of
the distribution of the data. For this analysis the objective function β -GWO . The range is optimal for all other functions. It can
values obtained from 51 runs are sorted. The comparison of the be clearly observed that the interquartile range and median
proposed variant and GWO is carried out. We analyze the box of the proposed β -GWO is optimal as compared to GWO. It is
plot on the basis of Inter Quartile Range (IQR) that is the width of worth mentioning here that upper whisker represents scores
box and range that is the vertical distance between optimal and outside the middle 50%. These upper whiskers are observed
maximum values. Following observations are worth mentioning prominently in case of GWO, that represents the variation in
from this analysis the obtained solutions.
96 A. Saxena, R. Kumar and S. Das / Applied Soft Computing Journal 75 (2019) 84–105

Fig. 9. Convergence curves of other algorithms with β -GWO for functions (8–11).

Fig. 10. Convergence curves of other algorithms with β -GWO for functions (15, 16, 19 and 29).

5.5. Comparison of β -GWO with the participants of CEC and front • Inspecting the p-values and mean values of β -GWO it can
ranker algorithms be concluded that the proposed variant can be a suitable
choice to handle multimodal, complex and hybrid functions.
Further, the comparison of the β -GWO has been executed with
some of the participants and front rankers of the CEC. The partici- However, the results of β -GWO are not competitive in a few
pants are Differential Evolution (DE) [58] variants namely Success- cases.
History based Adaptive DE (SHADE) [59] and LSHADE [60]. In this • For 30-D problems, mean values obtained from the runs are
experiment we compare the variants on the basis of the mean
optimal for 18 functions out of 29. This fact advocates the
values obtained from independent runs and p-values obtained
from Wilcoxon rank-sum test for both 10-D and 30-D problems. competency of the β -GWO to solve the hard optimization
From Table 10, the comparative analysis of the performance of problems. It is interesting to judge here that for almost all
proposed variant can be studied. Following are the major findings
cases the p-values obtained are less than 0.05. This suggests
of this experiment:
that the variant is statistically different from the opponents.
• It is empirical to state that for 10-D problems mean values
obtained for 51 independent runs are optimal for 16 functions From these analyses, it can be concluded that the performance
out of 29 functions. To judge the significance of these results of proposed β -GWO is competitive as compared to GWO variants,
p-values are also calculated. From the results of Wilcoxon
contemporary algorithms and CEC front ranker algorithms. A brief
rank-sum test [55] it can be concluded that for these 16
functions p-values are less than 0.05. This suggests that β - discussion on the results is presented in next subsection, which
GWO is statistically different from the LSHADE and SHADE. encapsulates different virtues of this chaotic algorithm.
A. Saxena, R. Kumar and S. Das / Applied Soft Computing Journal 75 (2019) 84–105 97

Fig. 11. Convergence curves of other algorithms with β -GWO for functions (17–18, 22 and 24).

Table 8
Results of CEC-2017 for 30-D problems [54].
Functions Statistical parameters Algorithms
GWO [34] OGWO [38] β -GWO GGWO [42] mGWO [39] IGWO [44] MGWO [43] GSA [15] CSA [56] FPA [57]
Mean 7.33E+08 4.16E+10 9.64E+08 1.10E+11 9.02E+08 6.58E+09 8.83E+09 2.11E+03 3.47E+03 1.07E+08
F-1 SD 6.09E+08 3.95E+09 6.60E+08 1.64E+10 8.12E+08 3.67E+09 1.89E+09 9.55E+02 3.76E+03 3.78E+07
p-values 3.30E−18 7.84E−02 – 3.30E−18 4.07E−01 2.42E−15 3.30E−18 3.30E−18 3.30E−18 6.78E−11
Mean 2.72E+04 8.19E+04 2.18E+04 2.00E+08 2.42E+04 4.36E+04 5.48E+04 7.87E+04 3.00E+02 1.10E+05
F-3 SD 8.32E+03 5.64E+03 8.36E+03 8.86E+08 1.07E+04 1.38E+04 8.17E+03 6.37E+03 3.43E−01 2.29E+04
p-values 3.30E−18 1.32E−03 – 3.30E−18 3.22E−01 1.27E−13 7.51E−18 3.30E−18 3.30E−18 3.30E−18
Mean 5.41E+02 1.13E+04 4.47E+02 4.19E+04 5.47E+02 8.59E+02 9.93E+02 5.31E+02 5.00E+02 5.79E+02
F-4 SD 3.34E+01 1.62E+03 3.75E+01 1.15E+04 3.89E+01 4.63E+02 1.88E+02 1.44E+01 2.49E+01 4.59E+01
p-values 3.30E−18 3.49E−01 – 3.30E−18 9.15E−01 1.30E−12 3.30E−18 4.83E−02 4.26E−10 5.27E−04
Mean 5.81E+02 9.00E+02 5.80E+02 1.15E+03 5.95E+02 6.45E+02 7.42E+02 7.20E+02 6.38E+02 8.06E+02
F-5 SD 3.28E+01 2.16E+01 2.28E+01 5.43E+01 3.27E+01 5.29E+01 2.03E+01 1.54E+01 2.39E+01 5.37E+01
p-values 3.30E−18 5.25E−01 – 3.30E−18 1.14E−02 1.43E−12 3.30E−18 3.30E−18 6.01E−15 3.50E−18
Mean 6.04E+02 6.91E+02 6.03E+02 7.32E+02 6.04E+02 6.22E+02 6.44E+02 6.46E+02 6.30E+02 6.69E+02
F-6 SD 2.93E+00 4.82E+00 1.84E+00 9.34E+00 1.76E+00 7.99E+00 6.05E+00 3.24E+00 6.17E+00 9.05E+00
p-values 3.30E−18 2.02E−02 – 3.30E−18 1.78E−03 3.94E−18 3.30E−18 3.30E−18 3.30E−18 3.30E−18
Mean 8.46E+02 1.36E+03 7.54E+02 3.23E+03 8.34E+02 9.54E+02 1.03E+03 7.54E+02 8.79E+02 1.33E+03
F-7 SD 3.87E+01 3.63E+01 3.01E+01 2.47E+02 3.99E+01 7.95E+01 3.68E+01 7.43E+00 3.62E+01 1.25E+02
p-values 3.30E−18 3.62E−02 – 3.30E−18 6.16E−01 4.86E−15 3.30E−18 3.30E−18 3.01E−11 3.30E−18
Mean 8.74E+02 1.14E+03 8.72E+02 1.38E+03 8.81E+02 9.06E+02 9.94E+02 9.46E+02 9.10E+02 1.07E+03
F-8 SD 1.80E+01 2.22E+01 1.30E+01 4.34E+01 2.42E+01 3.21E+01 1.79E+01 1.25E+01 1.96E+01 4.42E+01
p-values 3.30E−18 6.98E−01 – 3.30E−18 5.23E−02 1.96E−10 3.30E−18 3.30E−18 1.07E−14 3.30E−18
Mean 1.30E+03 1.06E+04 1.25E+03 3.36E+04 1.31E+03 3.88E+03 4.13E+03 1.91E+03 1.86E+03 1.31E+04
F-9 SD 2.58E+02 1.18E+03 2.57E+02 5.02E+03 3.26E+02 1.68E+03 6.32E+02 2.85E+02 4.23E+02 3.38E+03
p-values 3.30E−18 3.70E−01 – 3.30E−18 4.78E−01 2.39E−17 3.30E−18 2.56E−15 1.21E−13 3.30E−18
Mean 4.07E+03 8.48E+03 3.86E+03 1.06E+04 3.86E+03 4.48E+03 6.45E+03 4.84E+03 4.46E+03 6.20E+03
F-10 SD 9.80E+02 4.04E+02 6.38E+02 4.55E+02 8.00E+02 1.12E+03 4.44E+02 3.78E+02 6.49E+02 5.48E+02
p-values 3.30E−18 2.31E−01 – 3.30E−18 7.03E−01 6.19E−03 3.30E−18 2.63E−11 9.11E−06 4.43E−18
Mean 1.42E+03 7.21E+03 1.20E+03 4.51E+04 1.43E+03 3.62E+03 2.22E+03 1.20E+03 1.24E+03 1.54E+03
F-11 SD 3.78E+02 1.16E+03 3.57E+02 2.09E+04 3.44E+02 1.31E+03 5.55E+02 2.60E+01 4.14E+01 1.10E+02
p-values 3.30E−18 4.03E−01 – 3.30E−18 7.08E−01 5.64E−17 2.57E−16 1.84E−16 1.82E−11 1.64E−13
Mean 3.70E+07 1.16E+10 2.89E+07 2.34E+10 3.46E+07 1.67E+08 5.21E+08 1.22E+06 4.64E+05 4.84E+07
F-12 SD 5.37E+07 2.17E+09 4.61E+07 5.38E+09 5.61E+07 2.93E+08 2.54E+08 3.78E+06 4.02E+05 3.61E+07
p-values 3.30E−18 4.58E−01 – 3.30E−18 2.18E−01 4.50E−07 3.72E−18 1.34E−15 3.57E−17 2.08E−06
Mean 5.93E+06 8.24E+09 1.44E+07 2.01E+10 1.04E+07 2.92E+07 1.44E+08 2.83E+04 2.53E+04 2.61E+06
F-13 SD 2.46E+07 3.21E+09 4.96E+07 6.77E+09 4.11E+07 1.09E+08 1.51E+08 5.11E+03 1.55E+04 2.39E+06
p-values 3.30E−18 2.08E−01 – 3.30E−18 3.66E−01 4.34E−01 2.12E−14 2.09E−11 2.01E−12 3.30E−11
Mean 9.29E+04 2.32E+06 1.02E+05 5.47E+07 6.27E+04 3.43E+05 4.06E+05 1.71E+05 1.62E+03 1.66E+05
F-14 SD 1.69E+05 1.22E+06 1.66E+05 4.46E+07 9.74E+04 5.35E+05 4.13E+05 5.61E+04 5.49E+01 1.26E+05
p-values 4.99E−18 9.04E−01 – 3.30E−18 8.83E−01 8.18E−02 1.23E−08 1.64E−09 3.50E−18 8.54E−05
98 A. Saxena, R. Kumar and S. Das / Applied Soft Computing Journal 75 (2019) 84–105

Table 9
Results of CEC-2017 for 30-D problems [54].
Functions Statistical parameters Algorithms
GWO [34] OGWO [38] β -GWO GGWO [42] mGWO [39] IGWO [44] MGWO [43] GSA [15] CSA [56] FPA [57]
Mean 2.14E+05 7.69E+07 7.57E+04 5.37E+09 2.19E+05 4.49E+05 1.28E+06 9.21E+03 6.58E+03 4.22E+05
F-15 SD 5.54E+05 1.05E+08 2.40E+05 2.69E+09 4.98E+05 9.77E+05 1.40E+06 1.61E+03 2.68E+03 3.76E+05
p-values 3.30E−18 3.19E−01 – 3.30E−18 4.54E−01 7.08E−02 3.57E−17 3.01E−17 7.51E−18 6.31E−17
Mean 2.35E+03 5.23E+03 2.30E+03 7.88E+03 2.30E+03 2.59E+03 3.03E+03 3.03E+03 2.53E+03 3.46E+03
F-16 SD 2.54E+02 4.62E+02 2.84E+02 1.58E+03 2.92E+02 2.38E+02 2.93E+02 2.72E+02 2.35E+02 3.25E+02
p-values 3.30E−18 1.19E−01 – 3.30E−18 9.09E−01 4.06E−08 5.70E−15 1.49E−15 2.80E−06 1.27E−17
Mean 1.93E+03 3.57E+03 1.91E+03 1.28E+04 1.92E+03 2.16E+03 2.21E+03 2.75E+03 2.10E+03 2.71E+03
F-17 SD 1.21E+02 3.76E+02 1.10E+02 1.20E+04 1.13E+02 2.16E+02 1.65E+02 2.12E+02 1.54E+02 2.91E+02
p-values 3.30E−18 3.66E−01 – 3.30E−18 8.94E−01 6.03E−09 2.48E−14 3.30E−18 6.79E−09 6.68E−18
Mean 1.17E+06 1.91E+07 5.42E+05 5.83E+08 5.70E+05 2.66E+06 1.06E+06 2.08E+05 1.57E+04 3.17E+06
F-18 SD 3.76E+06 1.54E+07 7.00E+05 3.84E+08 6.10E+05 5.22E+06 1.15E+06 7.67E+04 9.26E+03 2.75E+06
p-values 4.43E−18 3.09E−01 – 3.30E−18 3.42E−01 5.73E−02 4.20E−07 2.37E−02 3.30E−18 1.27E−13
Mean 5.69E+05 2.38E+08 3.72E+05 5.88E+09 6.58E+05 9.10E+05 2.72E+06 6.46E+03 5.71E+03 2.30E+06
F-19 SD 1.26E+06 1.32E+08 1.21E+06 2.20E+09 1.52E+06 1.50E+06 1.20E+06 1.30E+03 6.26E+03 1.55E+06
p-values 3.30E−18 6.47E−02 – 3.30E−18 6.42E−04 7.07E−07 1.39E−16 1.91E−12 2.90E−14 6.68E−15
Mean 2.37E+03 2.99E+03 2.33E+03 3.67E+03 2.29E+03 2.51E+03 2.49E+03 3.01E+03 2.37E+03 2.64E+03
F-20 SD 1.40E+02 1.54E+02 1.14E+02 1.84E+02 1.21E+02 1.84E+02 1.13E+02 1.70E+02 1.01E+02 1.29E+02
p-values 3.30E−18 1.19E−01 – 3.30E−18 8.91E−02 5.55E−07 2.33E−10 3.50E−18 6.00E−02 1.57E−15
Mean 2.37E+03 2.72E+03 2.37E+03 2.90E+03 2.38E+03 2.41E+03 2.51E+03 2.53E+03 2.41E+03 2.57E+03
F-21 SD 1.96E+01 3.19E+01 2.48E+01 4.72E+01 3.51E+01 4.34E+01 1.98E+01 2.15E+01 2.49E+01 4.44E+01
p-values 3.30E−18 8.15E−01 – 3.30E−18 2.67E−01 1.17E−09 4.18E−18 3.50E−18 4.10E−12 3.50E−18
Mean 3.80E+03 8.79E+03 3.99E+03 1.18E+04 4.17E+03 5.17E+03 4.83E+03 4.71E+03 2.30E+03 6.11E+03
F-22 SD 1.39E+03 8.22E+02 1.39E+03 6.43E+02 1.72E+03 1.34E+03 2.05E+03 2.42E+03 1.60E+00 2.20E+03
p-values 3.30E−18 2.81E−01 – 3.30E−18 8.51E−01 3.97E−05 3.68E−02 9.25E−01 3.30E−18 4.27E−06
Mean 2.74E+03 3.66E+03 2.72E+03 3.86E+03 2.73E+03 2.88E+03 2.96E+03 3.48E+03 2.91E+03 3.00E+03
F-23 SD 4.21E+01 1.36E+02 2.31E+01 1.67E+02 2.60E+01 7.32E+01 3.14E+01 1.32E+02 8.19E+01 6.86E+01
p-values 3.30E−18 5.65E−02 – 3.30E−18 2.26E−01 5.29E−18 3.30E−18 3.30E−18 4.99E−18 3.30E−18
Mean 2.90E+03 3.93E+03 2.90E+03 4.18E+03 2.92E+03 3.09E+03 3.14E+03 3.18E+03 3.05E+03 3.13E+03
F-24 SD 3.65E+01 1.54E+02 4.70E+01 2.42E+02 5.67E+01 6.54E+01 2.78E+01 5.36E+01 6.71E+01 7.67E+01
p-values 3.30E−18 3.25E−01 – 3.30E−18 1.29E−01 5.33E−17 3.30E−18 3.30E−18 1.57E−15 1.13E−17
Mean 2.95E+03 4.42E+03 2.92E+03 1.62E+04 2.95E+03 3.09E+03 3.13E+03 2.93E+03 2.92E+03 2.97E+03
F-25 SD 2.41E+01 3.02E+02 2.24E+01 2.95E+03 2.85E+01 1.10E+02 4.94E+01 1.12E+01 2.09E+01 4.11E+01
p-values 3.30E−18 6.88E−01 – 3.30E−18 1.62E−01 6.68E−17 3.30E−18 1.07E−06 1.29E−04 2.49E−03
Mean 4.40E+03 1.02E+04 3.36E+03 1.53E+04 4.45E+03 5.73E+03 6.52E+03 4.75E+03 3.65E+03 7.63E+03
F-26 SD 2.52E+02 4.76E+02 1.64E+02 1.59E+03 2.68E+02 7.70E+02 3.71E+02 1.86E+03 1.47E+03 8.75E+02
p-values 3.30E−18 1.41E−01 – 3.30E−18 7.95E−02 2.13E−17 3.30E−18 9.89E−01 4.13E−06 3.30E−18
Mean 3.23E+03 4.67E+03 3.23E+03 5.13E+03 3.23E+03 3.32E+03 3.39E+03 4.10E+03 3.41E+03 3.34E+03
F-27 SD 1.19E+01 2.98E+02 1.46E+01 5.31E+02 1.15E+01 5.08E+01 6.06E+01 2.60E+02 7.71E+01 5.49E+01
p-values 3.30E−18 7.94E−01 – 3.30E−18 2.55E−01 4.24E−16 3.30E−18 3.30E−18 3.30E−18 5.29E−18
Mean 3.35E+03 6.11E+03 3.15E+03 1.25E+04 3.35E+03 3.74E+03 3.68E+03 3.24E+03 3.20E+03 3.35E+03
F-28 SD 5.48E+01 3.96E+02 7.08E+01 1.48E+03 6.22E+01 2.76E+02 1.07E+02 4.32E+01 4.30E+01 5.01E+01
p-values 3.30E−18 3.32E−01 – 3.30E−18 6.64E−01 1.84E−16 1.51E−17 1.27E−13 5.61E−18 3.29E−01
Mean 3.67E+03 6.83E+03 3.66E+03 2.39E+04 3.66E+03 4.00E+03 4.49E+03 4.55E+03 4.00E+03 4.64E+03
F-29 SD 1.59E+02 6.43E+02 1.30E+02 1.61E+04 1.52E+02 2.00E+02 2.00E+02 1.62E+02 2.15E+02 3.48E+02
p-values 3.30E−18 6.93E−01 – 3.30E−18 3.84E−01 8.94E−14 3.30E−18 3.30E−18 9.55E−12 3.30E−18
Mean 4.26E+06 5.79E+08 3.77E+06 3.07E+09 4.39E+06 8.63E+06 3.78E+07 5.01E+04 7.26E+04 6.79E+06
F-30 SD 3.89E+06 3.93E+08 2.67E+06 1.27E+09 3.01E+06 7.43E+06 2.31E+07 8.82E+03 4.53E+04 4.64E+06
p-values 3.30E−18 8.57E−01 – 3.30E−18 2.81E−01 3.01E−04 4.70E−18 3.30E−18 3.30E−18 1.78E−04

5.6. Discussion 6. Real applications of β-GWO

Based on the observations reported in this section, it can be 6.1. Parameter estimation for frequency modulated sound waves [61]
concluded that embedding chaos in bridging mechanism between
exploration and exploitation phases improves this mechanism. To judge the efficacy of the proposed β -GWO the problem of
It also empowers search agents with better exploration and ex- frequency parameter estimation is addressed. In year 2011, Das
ploitation capability till the last iteration. It also increases search et al. [61] presented a set of real world engineering problems to test
speed after the exploration phase which results in better conver- the efficacy of the evolutionary algorithms. Parameter estimation
gence characteristics. In classical GWO, the control vector changes is the one from that set. Parameter estimation of Frequency Mod-
linearly which is the indication of a linear decrement shift from ulated (FM) synthesizer is a six dimensional optimization problem
exploration to exploitation phase, by embedding chaos in this and a part of FM sound wave synthesis. The problem is formulated
vector, the exploration phase is accelerated and the exploitation as the parameter estimation for generation of the sound as per the
phase can be speeded up. Both of these virtues can be kept alive target sound. The problem is complex and multimodal in nature.
with inculcation of the chaos. In following section, the applications The minima for the objective function is at zero. The parameter
of proposed variant on real engineering problems are investigated. vector is to be estimated through optimization process. The vector
A. Saxena, R. Kumar and S. Das / Applied Soft Computing Journal 75 (2019) 84–105 99

Fig. 12. Box plot analysis of β -GWO for 10-D problems.

Fig. 13. Box plot analysis of β -GWO for 30-D problems.

has six parameters as per following equation: To solve this optimization problem, results are averaged over 30
independent runs. Fig. 14 along with Table 11 show the results
Z = (a1 , ω1 , a2 , ω2 , a3 , ω3 ) (11) of proposed variant with some contemporary optimizers. It is
The expressions for estimated and the target sound waves are as observed from the results that the mean and minimum values are
follows: optimal for β -GWO. These values are shown in bold face. After
careful observations on the box plots and convergence curve in
S (t ) = a1 . sin (ω1 .t .θ + a2 . sin (ω2 .t .θ + a3 . sin (ω3 .t .θ))) (12) Fig. 15, it is observed that the performance of β -GWO is competi-
tive with other variants of GWO. The mean values as shown in box
S0 (t ) = (1.0) . sin ((5.0) .t .θ − (1.5) . sin ((4.8) .t .θ plot is optimal for proposed variant and IQR is also optimal for β -
GWO. It is worth mentioning here that the performance of IGWO
+ (2.0) . sin ((4.9) t .θ))) (13)
is very competitive with the proposed variant as IQR is low. From
this experiment it can be easily concluded that the performance of
100 β -GWO is competitive and aligned with modern optimizers. This

→ ∑
analysis also advocates the applicability of β -GWO on real world
Min f ( Z ) = (S(t) − S0 (t))2 (14)
t =0
engineering problem.
100 A. Saxena, R. Kumar and S. Das / Applied Soft Computing Journal 75 (2019) 84–105

Table 10
Comparison of the β -GWO with some of the CEC participants algorithms.
Function 10-D Problems 30-D Problems
LSHADE [60] SHADE [59] β -GWO LSHADE [60] SHADE [59] β -GWO
Mean p-values Mean p-values Mean p-values Mean P-values Mean p-values Mean p-values
F-1 1.00E+02 1.39E−20 1.00E+02 2.68E−19 7.61E+08 – 1.00E+02 1.39E−20 1.00E+02 1.39E−20 4.42E+05 –
F-3 3.00E+02 8.97E−20 3.00E+02 1.25E−18 2.31E+04 – 3.00E+02 1.39E−20 3.00E+02 1.39E−20 4.51E+02 –
F-4 4.00E+02 9.02E−20 4.53E+02 1.59E−18 5.94E+02 – 4.29E+02 2.30E−06 4.00E+02 1.39E−20 4.13E+02 –
F-5 5.07E+02 3.30E−18 5.15E+02 3.30E−18 5.15E+02 – 5.03E+02 1.20E−17 5.03E+02 8.38E−17 5.01E+02 –
F-6 6.00E+02 5.19E−20 6.00E+02 3.30E−18 6.00E+02 – 6.00E+02 2.79E−20 6.00E+02 1.39E−20 6.00E+02 –
F-7 7.37E+02 3.30E−18 7.45E+02 3.30E−18 7.37E+02 – 7.12E+02 1.60E−17 7.13E+02 3.21E−16 7.07E+00 –
F-8 8.07E+02 3.30E−18 8.15E+02 3.30E−18 8.01E+02 – 8.02E+02 3.72E−18 8.03E+02 1.51E−17 8.02E+02 –
F-9 9.00E+02 1.39E−20 9.00E+02 1.27E−18 1.32E+03 – 9.00E+02 1.39E−20 9.00E+02 1.39E−20 9.00E+02 –
F-10 2.11E+03 3.30E−18 2.68E+03 1.02E−15 3.72E+03 – 1.07E+03 4.01E−17 1.06E+03 5.13E−15 1.02E+03 –
F-11 1.11E+03 3.30E−18 1.13E+03 3.30E−18 1.11E+03 – 1.10E+03 2.31E−19 1.10E+03 1.17E−19 1.10E+03 –
F-12 2.24E+03 3.30E−18 2.40E+03 3.30E−18 2.80E+07 – 1.22E+03 2.44E−18 1.28E+03 3.17E−18 2.22E+05 –
F-13 1.32E+03 3.30E−18 1.34E+03 3.30E−18 4.04E+06 – 1.30E+03 2.16E−18 1.30E+03 3.24E−18 7.22E+03 –
F-14 1.42E+03 3.30E−18 1.43E+03 3.30E−18 9.93E+04 – 1.40E+03 7.08E−19 1.40E+03 3.22E−18 1.37E+03 –
F-15 1.50E+03 3.30E−18 1.52E+03 3.30E−18 8.17E+05 – 1.50E+03 3.30E−18 1.50E+03 3.30E−18 1.50E+03 –
F-16 1.65E+03 3.30E−18 1.88E+03 5.71E−12 1.65E+03 – 1.60E+03 3.50E−18 1.60E+03 3.30E−18 1.60E+03 –
F-17 1.76E+03 8.87E−17 1.75E+03 1.05E−16 1.75E+03 – 1.72E+03 1.02E−15 1.70E+03 3.30E−18 1.70E+03 –
F-18 1.82E+03 3.30E−18 1.87E+03 3.30E−18 5.11E+05 – 1.80E+03 3.30E−18 1.80E+03 3.30E−18 2.73E+04 –
F-19 1.91E+03 3.30E−18 1.91E+03 3.30E−18 6.88E+05 – 1.90E+03 3.18E−18 1.90E+03 3.30E−18 1.90E+03 –
F-20 2.07E+03 3.30E−18 2.07E+03 2.01E−17 2.02E+03 – 2.00E+03 3.19E−17 2.00E+03 9.48E−19 2.00E+03 –
F-21 2.25E+03 1.39E−20 2.32E+03 3.30E−18 2.22E+03 – 2.20E+03 1.39E−20 2.26E+03 1.82E−12 2.20E+03 –
F-22 2.35E+03 4.03E−19 2.30E+03 6.87E−20 4.17E+03 – 2.34E+03 5.03E−16 2.30E+03 2.83E−19 2.30E+03 –
F-23 2.80E+03 2.21E−13 2.67E+03 3.30E−18 2.66E+03 – 2.64E+03 3.30E−18 2.60E+03 9.14E−16 2.60E+03 –
F-24 3.32E+03 1.08E−15 2.84E+03 3.30E−18 2.84E+03 – 2.79E+03 1.78E−13 2.71E+03 6.80E−10 2.71E+03 –
F-25 2.91E+03 2.30E−16 2.89E+03 3.30E−18 2.86E+03 – 2.90E+03 1.81E−13 2.92E+03 4.62E−05 2.90E+03 –
F-26 4.49E+03 6.02E−04 3.68E+03 3.30E−18 3.44E+03 – 2.90E+03 8.78E−16 2.90E+03 4.03E−19 2.90E+03 –
F-27 3.43E+03 3.30E−18 3.21E+03 8.50E−14 3.21E+03 – 3.12E+03 2.61E−18 3.09E+03 7.49E−14 3.09E+03 –
F-28 3.19E+03 3.16E−18 3.14E+03 1.06E−18 3.12E+03 – 3.10E+03 3.63E−20 3.20E+03 6.28E−10 3.04E+03 –
F-29 3.22E+03 3.30E−18 3.36E+03 3.30E−18 3.22E+03 – 3.14E+03 4.44E−13 3.14E+03 5.93E−11 3.14E+03 –
F-30 3.92E+03 3.30E−18 5.09E+03 3.30E−18 4.47E+06 – 3.28E+03 3.06E−18 9.20E+04 1.12E−13 3.76E+05 –

Table 11
Statistical parameters of error in objective function values for FM Sound Waves Synthesis.
Algorithms Min Max Mean SD
GWO [34] 2.53E−01 2.56E+01 1.72E+01 5.95E+00
β -GWO 6.60E−02 2.52E+01 1.44E+01 7.16E+00
OGWO [38] 8.14E−02 2.51E+01 1.49E+01 7.14E+00
IGWO [44] 8.75E+00 2.51E+01 1.79E+01 4.81E+00
MGWO [43] 1.06E+01 2.57E+01 2.04E+01 6.01E+00
mGWO [39] 8.56E−01 2.51E+01 1.59E+01 5.99E+00
GGWO [42] 2.84E+01 4.56E+01 3.28E+01 3.66E+00
ECGOA [62] 1.34E+01 2.75E+01 2.13E+01 3.89E+00
GOA [63] 8.42E+00 2.67E+01 2.01E+01 4.67E+00
CPSOH [64,65] 3.45E+00 4.25E+01 2.71E+01 6.06E+01
TRIBES-D [66,67] 2.22E+00 2.22E+01 1.47E+01 4.57E+00
CGSA [29] 8.42E+00 2.47E+01 1.74E+01 4.16E+00
G-CMA-ES [64,66,68] 3.33E+00 5.51E+01 3.88E+01 1.68E+01

6.2. Strategic bidding problem in a competitive energy market a generating company, the rival bid price information plays an im-
portant role. In literature, there are many mathematical paradigms
In this section, we present a case study of a power market, suggested by researchers to model this uncertain behavior. Re-
where participating generating company exercises the market cently, a Monte Carlo simulation based approach has been em-
power and earns profit by bidding strategically through the appli- ployed in [36]. In these approaches, rival bids are considered as
cation of β -GWO. Hence, the strategic biding problem is a profit normal probability distribution functions.
maximization problem with several operating constraints. Strategic bidding problem (SBP) is considered as a hard opti-
mization problem due to the involvement of the uncertain rivals
6.2.1. Preface bidding behavior. This is overcome by generating random samples
Power markets are the place where the system operator exe- with different rival bids. In approaches [36,69] the probability dis-
cutes the trading of the power as per the submitted bid and the tribution of the block bid prices are modeled with normal distribu-
consumer demand. In recent years, many countries have adopted tion.However, there can be many mechanisms exist for modeling
this process. The market model considered in this work is a pool rival generator’s behavior [70]. In the following section, the steps
market model. Every bid includes the bid price and amount of for execution of the strategic bidding problem is illustrated.
power. Submitted bids are sorted in increasing order and the
intersection between total demand and supply is obtained. This 6.2.2. Problem formulation
point is known as market equilibrium. The bids lower than this In this work, we have simulated the case of five thermal gen-
are accepted bid. In this illustration, we consider a uniform price erators which are participating in an auction and wish to sell the
rule where the uniform Market Clearing Price (MCP) calculates the energy to consumers in an energy market. Generator-1 wishes to
revenue earned by the company. While framing the strategy for maximize the profit by bidding strategically, for that the engineers
A. Saxena, R. Kumar and S. Das / Applied Soft Computing Journal 75 (2019) 84–105 101

Fig. 14. Results of FM Sound Waves Synthesis parameter estimation.

Fig. 15. Box plot and convergence curve for frequency synthesis parameter estimation problem.

develop an optimization routine considering normally distributed where P1m (t) is the power output of Generator-1 in mth
rival bids for different capacity blocks under different price caps. In block at time t. The above expression is used for handling
this work, we consider three capacity blocks with three different the optimization process in reference [69]. This optimization
upper price caps. The data for this simulation study has been taken
process is subjected to some operating constraints which are
from reference [36]. Following stepwise process is followed to
develop this stochastic optimization problem. given as:

1. Generate the samples of rival bids by using the probability • Generation limit
density function as per Eq. (15). 100 samples of rival bids are
shown in Fig. 16. P1m (t) ≤ P1m (t) ≤ P1m (t) ∀t ∈ H (17)
k 2
( )
1 (bidm k − µm ) where P1m (t) is the lower and P1m (t) is the upper limit
pdf (bidm k ) = √ exp − (15)
2π σm
2
k
2(σm k ) on generation at time t. H is an indicator of hours in a
trading day.
µm k and σm k are the mean and standard deviation of bid
• Limitation on bid price
price of kth rival (k ∈ (1, 2...4)) for mth block respectively.
2. Develop the optimization routine by defining the objective Cm (t) ≤ bidm (t) ≤ bidmax ∀t ∈ H (18)
function as a maximization preposition with different oper-
ating constraints. The objective function for this study can In this illustration, we are considering the case of a single
be given as:
trading hour and it is assumed that Generator-1 along with
H
∑ M
∑ its four rivals participate in an auction and all of them are
Maximize F (MCP(t), P1m (t)) = MCP(t) × P1m (t) working at their full capacity. The generator cost character-
t =1 m=1 istics and other relevant data are taken from the work [36].
(16)
102 A. Saxena, R. Kumar and S. Das / Applied Soft Computing Journal 75 (2019) 84–105

Fig. 16. Rivals bid behavior modeled for different energy blocks.

3. Obtain profit winning bids by executing the optimization it is able to exercise the market power in a dominant way as the
routine with the application of β -GWO for each sample of EMP values are highest for the β -GWO. It is worth to mention
rival bids. here that for system demand 3800 MW, the EMP values are lowest
4. Calculate means of all the winning bids and obtain the final for GGWO variant and for 4000 MW, EMP values is lowest for
value of MCP and profit by subtracting the operating cost mGWO. For the sake of clarification, all these results are plotted.
from the revenue obtained from the bidding. The Figs. 17 and 18 show the EMP, mean MCP, convergence curve
of a randomly selected sample of the profit maximization process.
6.2.3. Results With the observation reported in Table 12 and these figures, it
The application of proposed β -GWO is tested over a simulation is empirical to state that the proposed β -GWO shows promising
case, where Generator-1 and its four rival generating plants par- results for this stochastic, constrained maximization problem. This
ticipate in the auction. For this study, we generated 100 random application advocates the effective applicability of β -GWO on hard
sample bids for three capacity blocks. Hence, the optimization optimization problem.
process involves 12 stochastic and 3 unknown variables. To obtain
fair results maximum iteration (1000), 30 search agents are kept 7. Conclusion
the same for all participating variants of GWO. To study the impact
of optimizers on market simulations, important indicators are The bridging mechanism between the exploration and exploita-
calculated and reported such as MCP, optimal bid prices and profit. tion phase holds the major responsibility of the performance of
Results obtained from the optimization process is averaged over the population-based metaheuristic algorithm. In this paper au-
30 runs. The results in terms of the mean of the optimal bid prices thors performed an experiment of embedding β -Chaotic sequence
for all three capacity blocks, MCP and profit obtained for 3800 MW in monotonically decreasing linear mechanism of a well-known
and 4000 MW demand are shown in Table 12. In this study, we have metaheuristic GWO. Following are the major findings of the work:
compared the results of β -GWO with the other variants of GWO. To 1. In this work, authors have experimented and changed the
study the impact of optimizers on the profit maximization process, known linearly decreasing bridging mechanism to random
we calculate the index which is named as Exercise of Market Power unknown adaptive, yet linearly decreasing mechanism, and
(EMP), which is nothing but the ratio of the difference of MCP observe the performance of the optimization algorithm by
and operating cost to the operating cost of the Generator-1. The calculating statistical attributes (mean and SD) of the inde-
parameter is an important indicator of the ability of a bidder to use pendent runs along with the p-values of the rank-sum test.
its power to raise the prices and its courage to bid aggressively. The The proposed chaotic mechanism assists wolves to transit
higher EMP indicates the ability of a supplier to set the prices above slowly from searching, hunting and attacking phase. Hence
the marginal cost. in a way, it provides assistance to avoid local minima trap.
Inspecting the results of Table 12, we observe that the for 2. β -GWO variant is tuned on the basis of the sensitivity of
both demand levels results of proposed β -GWO is competitive. different initial starting points of chaos. After careful ob-
The optimal results are shown in boldface. From the results, it is servations of different calculated statistical parameters ob-
empirical to say that mean MCP obtained from the 100 samples are tained from the application of proposed variant on 12 shifted
maximum when the process is handled by β -GWO. MCP and profit and biased benchmark functions, it is observed that initial
obtained through the solution, both of these factors are aligned value x1 = 0.8 provides better results. Authors have also
with each other. It is observed that profits obtained for these demonstrated the impact of initial conditions on optimiza-
particular demand levels are highest for β -GWO as compared to tion performance. It is concluded that high vulnerability
other variants of GWO. From the analysis of the EMP, it is observed exists between starting points and performance of chaotic
that when a supplier bid strategically with the proposed β -GWO optimization algorithms.
A. Saxena, R. Kumar and S. Das / Applied Soft Computing Journal 75 (2019) 84–105 103

Fig. 17. MCP, EMP, profit, and convergence: Comparison for 3800-MW demand level.

Fig. 18. MCP, EMP, profit and convergence: Comparison for 4000-MW demand level.

3. The proposed variant has been compared with other vari- a scalability analysis has also been presented by consider-
ants of GWO and recently published popular and contem- ing both 10-D and 30-D problems. It is observed that the
performance of β -GWO is competitive as compared with
porary metaheuristic algorithms. It is observed that the per-
the GWO variants, some contemporary algorithms and some
formance of β -GWO is competitive and in some cases it
front rankers and participants of CEC. The statistical test
outperforms. advocates the suitability of the proposed approach.
4. The proposed variant has also benchmarked on CEC-2017 5. Last but not the least, the applicability of the proposed β -
functions. To exhibit the suitability of the proposed variant, GWO has been demonstrated over two real engineering
104 A. Saxena, R. Kumar and S. Das / Applied Soft Computing Journal 75 (2019) 84–105

Table 12
Block bids, MCP and profit for single trading hour.
System demand (MW) Algorithms Bid-1 ($/MWh) Bid-2($/MWh) Bid-3($/MWh) MCP ($/MWh) EMP Profit ($)
GWO [34] 1.18E+01 2.25E+01 3.27E+01 3.27E+01 2.49E+00 1.03E+04
OGWO [38] 1.49E+01 2.50E+01 3.29E+01 3.29E+01 2.51E+00 1.04E+04
GGWO [42] 1.18E+01 2.24E+01 3.25E+01 3.25E+01 2.46E+00 1.01E+04
β -GWO 1.15E+01 2.26E+01 3.29E+01 3.29E+01 2.51E+00 1.04E+04
3800
mGWO [39] 1.14E+01 2.26E+01 3.28E+01 3.28E+01 2.49E+00 1.03E+04
IGWO [44] 1.12E+01 2.23E+01 3.27E+01 3.27E+01 2.48E+00 1.02E+04
MGWO [43] 1.15E+01 2.26E+01 3.26E+01 3.26E+01 2.48E+00 1.02E+04
InGWO [36] 1.50E+01 2.49E+01 3.26E+01 3.26E+01 2.47E+00 1.02E+04
GWO [34] 1.16E+01 2.23E+01 3.32E+01 3.32E+01 1.07E+00 1.05E+04
OGWO [38] 1.48E+01 2.48E+01 3.30E+01 3.30E+01 1.06E+00 1.03E+04
GGWO [42] 1.15E+01 2.27E+01 3.30E+01 3.30E+01 1.06E+00 1.04E+04
β -GWO 1.16E+01 2.24E+01 3.32E+01 3.32E+01 1.07E+00 1.05E+04
4000
mGWO [39] 1.13E+01 2.25E+01 3.29E+01 3.29E+01 1.05E+00 1.03E+04
IGWO [44] 1.16E+01 2.25E+01 3.32E+01 3.32E+01 1.07E+00 1.05E+04
MGWO [43] 1.14E+01 2.21E+01 3.31E+01 3.31E+01 1.07E+00 1.05E+04
InGWO [36] 1.47E+01 2.48E+01 3.30E+01 3.30E+01 1.06E+00 1.04E+04

problems. It has been observed that the proposed variant [21] A.H. Gandomi, G.J. Yun, X.-S. Yang, S. Talatahari, Chaos-enhanced accelerated
shows better performance as compared with other variants particle swarm optimization, Commun. Nonlinear Sci. Numer. Simul. 18 (2)
(2013) 327–340.
of GWO and GWO itself.
[22] K.T. Alligood, T.D. Sauer, J.A. Yorke, Chaos, Springer, 1996.
Application of the proposed β -GWO on more challenging problems [23] E. Lorenz, The butterfly effect, World Sci. Ser. Nonlinear Sci. Ser. A 39 (2000)
91–94.
like controller design for Automatic Generation Control (AGC) of
[24] S. Šajić, N. Maletić, B.M. Todorović, M. Šunjevarić, Random binary sequences
the deregulated power system and protein folding optimization in telecommunications, J. Electr. Eng. 64 (4) (2013) 230–237.
will be investigated in future. Simultaneously, it will be very in- [25] G.I. Sayed, A.E. Hassanien, A.T. Azar, Feature selection via a novel chaotic crow
teresting to observe the impact of other chaotic sequences on the search algorithm, Neural Comput. Appl. (2017) 1–18.
mechanism and the performance of GWO. [26] A. Gandomi, X.-S. Yang, S. Talatahari, A. Alavi, Firefly algorithm with chaos,
Commun. Nonlinear Sci. Numer. Simul. 18 (1) (2013) 89–98.
[27] M. Kohli, S. Arora, Chaotic grey wolf optimization algorithm for constrained
References optimization problems, J. Comput. Des. Eng. (2017).
[28] A.H. Gandomi, X.-S. Yang, Chaotic bat algorithm, J. Comput. Sci. 5 (2) (2014)
[1] I. Fister Jr, X.-S. Yang, I. Fister, J. Brest, D. Fister, A brief review of nature- 224–232.
inspired algorithms for optimization, arXiv preprint arXiv:1307.4186, 2013. [29] S. Mirjalili, A.H. Gandomi, Chaotic gravitational constants for the gravitational
[2] T. Weise, Global optimization algorithms-theory and application, Self- search algorithm, Appl. Soft Comput. 53 (2017) 407–419.
published 2, 2009. [30] R. Zahmoul, R. Ejbali, M. Zaied, Image encryption based on new Beta chaotic
[3] A. Arabali, M. Ghofrani, M. Etezadi-Amoli, M.S. Fadali, Y. Baghzouz, Genetic- maps, Opt. Lasers Eng. 96 (2017) 39–49.
algorithm-based optimization approach for energy management, IEEE Trans. [31] R. Hamza, A novel pseudo random sequence generator for image-
Power Deliv. 28 (1) (2013) 162–170. cryptographic applications, J. Inf. Secur. Appl. 35 (C) (2017) 119–127.
[4] C. Zhang, J. Jiang, Y. Gao, W. Zhang, Q. Liu, X. Hu, Charging optimization in [32] X. Shu, J. Wang, H. Wang, X. Yang, Chaotic direct sequence spread spectrum for
lithium-ion batteries based on temperature rise and charge time, Appl. Energy secure underwater acoustic communication, Appl. Acoust. 104 (2016) 57–66.
194 (2017) 569–577. [33] C.B. Amar, M. Zaied, A. Alimi, Beta wavelets. synthesis and application to lossy
[5] J.H. Holland, Genetic algorithms, Sci. Amer. 267 (1) (1992) 66–73.
image compression, Adv. Eng. Softw. 36 (7) (2005) 459–474.
[6] I. Rechenberg, Evolution strategy: Natures way of optimization, in: Optimiza-
[34] S. Mirjalili, S.M. Mirjalili, A. Lewis, Grey wolf optimizer, Adv. Eng. Softw. 69
tion: Methods and Applications, Possibilities and Limitations, Springer, 1989,
(2014) 46–61.
pp. 106–126.
[35] E. Emary, H.M. Zawbaa, A.E. Hassanien, Binary grey wolf optimization ap-
[7] X. Yao, Y. Liu, G. Lin, Evolutionary programming made faster, IEEE Trans. Evol.
proaches for feature selection, Neurocomputing 172 (2016) 371–381.
Comput. 3 (2) (1999) 82–102.
[36] A. Saxena, B.P. Soni, R. Kumar, V. Gupta, Intelligent grey wolf optimizer-
[8] D. Simon, Biogeography-based optimization, IEEE Trans. Evol. Comput. 12 (6)
development and application for strategic bidding in uniform price spot
(2008) 702–713.
energy market, Appl. Soft Comput. (2018).
[9] R. Eberhart, J. Kennedy, A new optimizer using particle swarm theory, in:
[37] E. Gupta, A. Saxena, Grey wolf optimizer based regulator design for automatic
Micro Machine and Human Science, 1995. MHS’95., Proceedings of the Sixth
generation control of interconnected power system, Cogent Eng. 3 (1) (2016)
International Symposium on, IEEE, 1995, pp. 39–43.
1151612.
[10] X.-S. Yang, A new metaheuristic bat-inspired algorithm, Nature inspired
cooperative strategies for optimization (NICSO 2010), 2010, pp. 65–74. [38] M. Pradhan, P.K. Roy, T. Pal, Oppositional based grey wolf optimization al-
[11] J.C. Bansal, H. Sharma, S.S. Jadon, M. Clerc, Spider monkey optimization gorithm for economic dispatch problem of power system, Ain Shams Eng. J.
algorithm for numerical optimization, Memetic Comput. 6 (1) (2014) 31–47. (2017).
[12] X.-S. Yang, S. Deb, Cuckoo search via lévy flights, in: Nature & Biologically [39] N. Mittal, U. Singh, B.S. Sohi, Modified grey wolf optimizer for global engineer-
Inspired Computing, 2009. NaBIC 2009. World Congress on, IEEE, 2009, pp. ing optimization, Appl. Comput. Intell. Soft Comput. 2016 (2016) 8.
210–214. [40] T. Jayabarathi, T. Raghunathan, B. Adarsh, P.N. Suganthan, Economic dispatch
[13] W.-T. Pan, A new fruit fly optimization algorithm: taking the financial distress using hybrid grey wolf optimizer, Energy 111 (2016) 630–641.
model as an example, Knowl.-Based Syst. 26 (2012) 69–74. [41] L.K. Panwar, S. Reddy, A. Verma, B. Panigrahi, R. Kumar, Binary grey wolf
[14] X.-S. Yang, Firefly algorithms for multimodal optimization, in: International optimizer for large scale unit commitment problem, Swarm Evol. Comput.
Symposium on Stochastic Algorithms, Springer, 2009, pp. 169–178. (2017).
[15] E. Rashedi, H. Nezamabadi-Pour, S. Saryazdi, Gsa: a gravitational search algo- [42] B. Yang, X. Zhang, T. Yu, H. Shu, Z. Fang, Grouped grey wolf optimizer for
rithm, Inf. Sci. 179 (13) (2009) 2232–2248. maximum power point tracking of doubly-fed induction generator based
[16] O.K. Erol, I. Eksin, A new optimization method: big bang–big crunch, Adv. Eng. wind turbine, Energy Convers. Manage. 133 (2017) 427–443.
Softw. 37 (2) (2006) 106–111. [43] A. Khandelwal, A. Bhargava, A. Sharma, H. Sharma, Modified grey wolf opti-
[17] A. Hatamlou, Black hole: A new heuristic optimization approach for data mization algorithm for transmission network expansion planning problem,
clustering, Inf. Sci. 222 (2013) 175–184. Arab. J. Sci. Eng. (2017) 1–10.
[18] D.H. Wolpert, W.G. Macready, No free lunch theorems for optimization, IEEE [44] W. Long, X. Liang, S. Cai, J. Jiao, W. Zhang, A modified augmented Lagrangian
Trans. Evol. Comput. 1 (1) (1997) 67–82. with improved grey wolf optimization to constrained optimization problems,
[19] Z. Guo, H. Yang, S. Wang, C. Zhou, X. Liu, Adaptive harmony search with best- Neural Comput. Appl. 28 (1) (2017) 421–438.
based search strategy, Soft Comput. 22 (4) (2018) 1335–1349. [45] K. Nuaekaew, P. Artrit, N. Pholdee, S. Bureerat, Optimal reactive power
[20] Z. Guo, X. Yue, H. Yang, K. Liu, X. Liu, Enhancing social emotional optimization dispatch problem using a two-archive multi-objective grey wolf optimizer,
algorithm using local search, Soft Comput. 21 (24) (2017) 7393–7404. Expert Syst. Appl. (2017).
A. Saxena, R. Kumar and S. Das / Applied Soft Computing Journal 75 (2019) 84–105 105

[46] A. Saxena, S. Shekhawat, Ambient air quality classification by grey wolf opti- [59] R. Tanabe, A. Fukunaga, Evaluating the performance of shade on cec
mizer based support vector machine, J. Environ. Public Health 2017 (2017). 2013 benchmark problems, in: Evolutionary Computation (CEC), 2013 IEEE
[47] R.-E. Precup, R.-C. David, E.M. Petriu, Grey wolf optimizer algorithm-based Congress on, IEEE, 2013, pp. 1952–1959.
tuning of fuzzy control systems with reduced parametric sensitivity, IEEE [60] R. Tanabe, A.S. Fukunaga, Improving the search performance of shade using
Trans. Ind. Electron. 64 (1) (2017) 527–534. linear population size reduction, in: Evolutionary Computation (CEC), 2014
[48] R. Kumar, A. Kumar, R.K. Pandey, Electrocardiogram signal compression using IEEE Congress on, IEEE, 2014, pp. 1658–1665.
beta wavelets, J. Math. Modell. Algorithms 11 (3) (2012) 235–248. [61] S. Das, P.N. Suganthan, Problem definitions and evaluation criteria for cec
[49] S. Saremi, S. Mirjalili, A. Lewis, Biogeography-based optimisation with chaos, 2011 competition on testing evolutionary algorithms on real world opti-
Neural Comput. Appl. 25 (5) (2014) 1077–1097. mization problems, Jadavpur University, Nanyang Technological University,
[50] C. Muro, R. Escobedo, L. Spector, R. Coppinger, Wolf-pack (canis lupus) hunt- Kolkata, 2010.
ing strategies emerge from simple rules in computational simulations, Behav. [62] A. Saxena, S. Shekhawat, R. Kumar, Application and development of en-
Process. 88 (3) (2011) 192–197. hanced chaotic grasshopper optimization algorithms, Modell. Simul. Eng.
[51] J.G. Digalakis, K.G. Margaritis, On benchmarking functions for genetic algo- 2018 (2018).
rithms, Int. J. Comput. Math. 77 (4) (2001) 481–506. [63] S. Saremi, S. Mirjalili, A. Lewis, Grasshopper optimisation algorithm: Theory
[52] X.-S. Yang, Appendix a: test problems in optimization, Eng. Optim. (2010) and application, Adv. Eng. Softw. 105 (2017) 30–47.
261–266. [64] J.J. Liang, A.K. Qin, P.N. Suganthan, S. Baskar, Comprehensive learning particle
[53] M. Molga, C. Smutnicki, Test functions for optimization needs, Test functions swarm optimizer for global optimization of multimodal functions, IEEE Trans.
for optimization needs, 2005, 101. Evol. Comput. 10 (3) (2006) 281–295.
[54] N. Awad, M. Ali, J. Liang, B. Qu, P. Suganthan, Problem definitions and evalua- [65] F. Van den Bergh, A.P. Engelbrecht, A cooperative approach to particle swarm
tion criteria for the cec 2017 special session and competition on single objec- optimization, IEEE Trans. Evol. Comput. 8 (3) (2004) 225–239.
tive bound constrained real-parameter numerical optimization, in: Technical [66] A. Auger, N. Hansen, A restart cma evolution strategy with increasing popu-
Report, NTU, Singapore, 2016. lation size, in: Evolutionary Computation, 2005. the 2005 IEEE Congress on,
[55] F. Wilcoxon, Individual comparisons by ranking methods, Biometrics Bull. 1 Vol. 2, IEEE, 2005, pp. 1769–1776.
(6) (1945) 80–83. [67] J. Kumpiene, A. Lagerkvist, C. Maurice, Stabilization of pb-and cu-
[56] A. Askarzadeh, A novel metaheuristic method for solving constrained engi- contaminated soil using coal fly ash and peat, Environ. Pollut. 145 (1)
neering optimization problems: crow search algorithm, Comput. Struct. 169 (2007) 365–373.
(2016) 1–12. [68] S. Gupta, K. Deep, A novel random walk grey wolf optimizer, Swarm Evol.
[57] X.-S. Yang, Flower pollination algorithm for global optimization, in: UCNC, Comput. (2018).
Springer, 2012, pp. 240–249. [69] P. Bajpai, S. Singh, Fuzzy adaptive particle swarm optimization for bidding
[58] R. Storn, K. Price, Differential evolution–a simple and efficient heuristic for strategy in uniform price spot market, IEEE Trans. Power Syst. 22 (4) (2007)
global optimization over continuous spaces, J. Global Optim. 11 (4) (1997) 2152–2160.
341–359. [70] E.J. Anderson, A.B. Philpott, Estimation of electricity market distribution func-
tions, Ann. Oper. Res. 121 (1–4) (2003) 21–32.

You might also like