Expert Systems With Applications 149 (2020) 113338

Contents lists available at ScienceDirect

Expert Systems With Applications

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

Chimp optimization algorithm

M. Khishe a, M.R. Mosavi b,∗
a
Department of Electrical Engineering, Imam Khomeini Marine Science University, Nowshahr, Iran
b
Department of Electrical Engineering, Iran University of Science and Technology, Narmak, Tehran 16846-13114, Iran

a r t i c l e i n f o a b s t r a c t

Article history: This paper proposes a novel metaheuristic algorithm called Chimp Optimization Algorithm (ChOA) in-
Received 11 April 2019 spired by the individual intelligence and sexual motivation of chimps in their group hunting, which is
Revised 11 January 2020
different from the other social predators. ChOA is designed to further alleviate the two problems of slow
Accepted 24 February 2020
convergence speed and trapping in local optima in solving high-dimensional problems. In this paper, a
Available online 29 February 2020
mathematical model of diverse intelligence and sexual motivation of chimps is proposed. In this regard,
Keywords: four types of chimps entitled attacker, barrier, chaser, and driver are employed for simulating the diverse
Chimp intelligence. Moreover, four main steps of hunting, i.e. driving, chasing, blocking, and attacking, are im-
Mathematical model plemented. The proposed ChOA algorithm is evaluated in 3 main phases. First, a set of 30 mathematical
Metaheuristic benchmark functions is utilized to investigate various characteristics of ChOA. Secondly, ChOA was tested
Optimization by 13 high-dimensional test problems. Finally, 10 real-world optimization problems were used to evaluate
the performance of ChOA. The results are compared to several newly proposed meta-heuristic algorithms
in terms of convergence speed, the probability of getting stuck in local minimums, and exploration, ex-
ploitation. Also, statistical tests were employed to investigate the significance of the results. The results
indicate that the ChOA outperforms the other benchmark optimization algorithms.
© 2020 Elsevier Ltd. All rights reserved.

1. Introduction Nature-inspired MOAs solve optimization problems by imitating

physical or biological phenomena. They can be divided into four
Metaheuristic Optimization Algorithms (MOAs) have become major categories: physics-based, evolution-based, swarm-based,
very popular in engineering applications. As the complicacy of and human-based methods. Evolutionary Algorithms (EAs) are usu-
problems increases, the need for new MOAs becomes obvious more ally inspired by the concepts of natural evolution. The search pro-
than before. The reasons for this demand, can be summarized cess initiates with a stochastically generated population which is
into five major motivations: (i) simple concepts and structure, evolved over following generations. The prominent feature of EAs
which assists scientists to learn MOAs quickly and apply them to is that the best individuals are always merged together to create
their problems; (ii) derivation-free mechanisms: this makes MOAs the subsequent generation of individuals. This let the population
highly suitable for real-world engineering problems with costly or be optimized over the course of generations. The most popular
unknown gradient information; (iii) local optima avoidance: they EA is Genetic Algorithms (GA) (Holland, 1992) that simulates the
have greater abilities to avoid local minima compared to conven- Darwinian evolution concepts. Some other popular EAs are Dif-
tional optimization algorithms; (iv) flexibility, which refers to the ferential Evolution (DE) (Storn & Price, 1997), Evolution Strategy
pertinence of MOAs to different problems without any specific (ES) (Beyer & Schwefel, 2002), and Biogeography-Based Optimizer
changes in their structure (they assume problems as black boxes); (BBO) (Khishe, Mosavi, & Kaveh, 2017; Kaveh et al., 2019).
(v) relatively simple and entirely effective hardware implementa- Physics-based methods mimic the physical concepts in the
tion, the majority of MOAs have parallel structures, therefore, hard- world so that a stochastic set of search agents communi-
ware implementation and parallel computing (e.g. via Filed Pro- cate and move entire search space pursuant physical concepts.
grammable Gate Array (FPGA)) can strongly increase their perfor- Some of the most popular methods are Simulated Annealing
mances. (SA) (Kirkpatrick, Gelatt, & Vecchi, 1983), Big-Bang Big-Crunch
(BBBC) (Erol & Eksin, 2006), Gravitational Search Algorithm (GSA)
(Rashedi, Nezamabadi-Pour, & Saryazdi, 2009), Chaotic Fractal Walk
∗ Trainer (Khishe, Mosavi, & Moridi, 2018), and Adaptive Best-mass
Corresponding author.
E-mail addresses: m_khishe@alumni.iust.ac.ir (M. Khishe), m_mosavi@iust.ac.ir Gravitational Search Algorithm (Mosavi, Khishe, Parvizi, Naseri, &
(M.R. Mosavi). Ayat, 2019).

https://doi.org/10.1016/j.eswa.2020.113338
0957-4174/© 2020 Elsevier Ltd. All rights reserved.
2 M. Khishe and M.R. Mosavi / Expert Systems With Applications 149 (2020) 113338

Fig. 1. Two different plot of relationship between body size and brain size in various mammals.

Fig. 2. Phylogeny of super-family Hominoid.

The third group of MOAs includes algorithms inspired by

human behaviours in the literature. Some of the most popu-
lar techniques are Tabu (Taboo) Search (TS) (Osman, 1993), Im-
perialist Competitive Algorithm (ICA) (Atashpaz-Gargari & Lucas,
2007), Teaching Learning Based Optimization (TLBO) (Rao, Savsani,
& Vakharia, 2011), Interior Search Algorithm (ISA) (Ravakhah,
Fig. 3. Human and chimp DNA.
Khishe, Aghababaee, & Hashemzadeh, 2017), Innovative Gunner
(AIG) (Pijarski & Kacejko, 2019).
The forth group of MOAs includes Swarm Intelligence-based Al-
gorithms (SIAs) that originates from natural behaviour of animals Other recently proposed SIAs are Cuckoo Search (CS) (Yang &
in their herds, flock, colonies, and schools. The most popular algo- Deb, 2009), Bat-inspired Algorithm (BA) (Yang, 2010), Firefly Algo-
rithm, in this category, is Particle Swarm Optimization (PSO) (Han, rithm (FA) (Yang, 2010), Krill Herd (KH) (Gandomi & Alavi, 2012),
Lu, Hou, & Qiao, 2016). Two other popular swarm-based algorithms Grey Wolf Optimizer (GWO) (Emary, Zawbaa, & Grosan, 2017),
are Ant Colony Optimization (ACO) (Dorigo, Birattari, & Stutzle, GWO with LevyF́light (Heidari & Pahlavani, 2017), chaotic GWO
2006) and Artificial Bee Colony (ABC) (Basturk & Karaboga, 2006). (Heidari & Abbaspour, 2017), evolutionary population dynamics
 M. Khishe and M.R. Mosavi / Expert Systems With Applications 149 (2020) 113338 3

Fig. 4. The first phase of hunting process (exploration).

Fig. 6. Mathematical models of dynamic coefficients (f) related to independent

groups for (a) ChOA1 and (b) ChOA2.

Table 1
The dynamic coefficient of f vector.

Groups ChOA1 ChOA2

Group1 1.95−2t1/4 /T1/3 2.5−(2log(t)/log(T))

Fig. 5. The second phase of hunting process (exploitation). Group2 1.95−2t1/3 /T1/4 (−2t3 /T3 ) + 2.5
Group3 (−3t3 /T3 ) + 1.5 0.5 + 2exp[−(4t/T)2 ]
Group4 (−2t3 /T3 ) + 1.5 2.5 + 2(t/T)2 − 2(2t/T)

and Grasshopper Optimization Approaches (GOA) (Mafarja et al.,

2017), binary Salp Swarm Algorithm (BSSA) with crossover scheme • SIAs memorize search space information over the course of it-
(Farisa et al., 2018), hybrid GOA and MLP (Heidari, Farisa, Aljarah, eration while EAs discard any information of the prior genera-
& Mirjalili, 2019), hybrid binary Ant Lion Optimizer (ALO) with tions.
rough set and approximate entropy (Mafarja & Mirjalili, 2019), hy- • SIAs almost use memory to keep the best solution acquired so
brid MLP and Salp Swarm Algorithm (MLP-SSA) (Khishe & Moham- far.
madi, 2019), improved Monarch Butterfly Optimization (MBO) al- • SIAs generally have fewer parameters to adjust compare to
gorithm (Sun, Chen, Xu, & Tian, 2019), Improved Whale Trainer other MOAs.
(IWT) (Khishe & Mosavi, 2019), and hybrid Dragonfly Optimization • SIAs have fewer operators compared to EAs (crossover, muta-
Algorithm and MLP (DOA-MLP) (Khishe & Saffari, 2019). tion, immigration, and so on).
This category of MOAs started to be interesting since PSO • SIAs are easier to implement than the other MOA groups.
was proven to be very competitive with EAs, human-based, and
physical-based methods. Totally, SIAs have some advantages over In spite of the demand for more function evaluation, the lit-
other MOAs that are listed below: erature shows that SIAs are highly appropriate for solving real
4 M. Khishe and M.R. Mosavi / Expert Systems With Applications 149 (2020) 113338

world engineering problems since they are able to elude local min-
ima, exploring the search space more complete, and exploiting the
global optimum more reliable than any other MOAs. In addition,
the No Free Lunch (NFL) theorem shows that all the MOAs execute
equally on all optimization problems (Wolpert & Macready, 1997).
Hence, there are still problems that have not been solved, or they
can be resolved better by new MOAs. The main motivations of this
article are these two reasons, in which a novel SIAs is proposed
and compared to the current well-known MOAs in the literature.
In spite of the considerable number of recently proposed publi-
cations in this field, there are yet other intelligent swarming be-
haviours in nature that have not obtained merit attention. One
of the amazing swarming behaviours in nature is the Intelligent
Group Haunting (IGH) of chimps. Since there is no research in the
literature to simulate the IGH of chimps, this article aims to first
discover the main characteristics of chimps’ IGH. An MOA is then
proposed based on the modelled IGH called Chimp Optimization
Algorithm (ChOA). In addition to NFL theorem underpinning work
motivation, the main reasons for choosing chimps from among nu-
merous swarming behaviour are individual intelligence and sexual
motivation. These two vary from the other hunters in nature.
Irrespective of the differences between the MOAs, a common
characteristic is the division of the search producer into two
phases: exploration and exploitation. The exploration phase refers
to the producer of investigating the search space as widely as pos-
sible. A MOA needs to have random operators to stochastically and
globally discover the search space in order to reinforce this phase.
Nevertheless, exploitation refers to the local search ability around
the promising areas gained in the exploration phase. Having strong
operators within these two phases or finding a proper balance
between them is considered a challenging point in the literature
(Mirjalili, 2015, 2016).
The main differences between the social behaviour of chimps
Fig. 7. Two and three-dimensional position vectors and their possible next loca-
tions. and any other flocking behaviours are:

1) Individuals’ Diversity: In a group of chimps, individuals are not

basically quite similar in terms of ability and intelligence, but
they all perform their tasks as members of a hunting group.
Each individual’s ability can be useful in a special phase of the
hunting event. Therefore, a chimp according to his special abil-
ity takes responsibility for a part of hunt (Stanford, 1996). In
this article, a mathematical model of diverse chimps called in-
dependent chimps is proposed. In other words, various models
with diverse curvatures, slopes, and interception points are uti-
lized to give chimps different behaviours as in natural hunting
duties. Independent chimps can improve the exploration phase
by discovering the searching space more thoroughly.
2) Sexual Motivation: As well as nutritional advantages of group
hunting, it has also been proved that chimps’ hunting is af-
fected by the probable social benefits of obtaining meat (Stan-
ford et al., 1994). Acquiring meat provides an opportunity to
trade it in return for social favours, e.g. sex and grooming. This
incentive in the final stage causes chimps to forget their re-
sponsibilities in hunting process. Therefore, they try to obtain
meat chaotically. This unconditional behaviour in final stage
lead to improve exploitation phase and convergence rate.

To sum up, the main contribution of the paper can be catego-

rized as follow:

 Stage 1: According to the comprehensive background study

in literature, MOAs are categorized into four main groups
as physics-based, evolution-based, swarm-based, and human-
based. The result of this stage was choosing a swarm-based al-
gorithm based on their ability and our target.
Fig. 8. Position updating in ChOA.  Stage 2: A comprehensive study has been done to choose a spe-
cial creature that has not been previously modelled and also
 M. Khishe and M.R. Mosavi / Expert Systems With Applications 149 (2020) 113338 5

Fig. 9. Position updating mechanism of chimps and effects of |a| on it.

Fig. 10. The chaotic maps used in the article.

Table 2
Chaotic maps.

No Name Chaotic map Range

1 Quadratic xi + 1 = x2 i − c, c = 1 (0,1)
1 xi = 0
2 Gauss/mouse xi+1 = { 1 (0,1)
mod (xi ,1 )
otherwise
3 Logistic xi + 1 = α xi (1 − xi ),α = 4 (0,1)
4 Singer xi+1 = μ(7.86xi − 23.31x2i + 28.75x3i − 13.302875x4i ), μ = 1.07 (0,1)
5 Bernoulli xi + 1 = 2xi (mod1) (0,1)
xi
xi < 0.7
6 Tent xi+1 = { 10 0.7 (0,1)
3
( 1 − xi ) 0.7 ≤ xi

having special intelligent behaviour. So, the result of this stage  Stage 5: Evaluation of the proposed ChOA algorithm by 30
was choosing chimp and it’s Intelligent Group Haunting (IGH). mathematical benchmark functions, 13 high-dimensional test
 Stage 3: Discovering and modelling the main characteristics of problems, and 10 real-world optimization problems.
Intelligent Group Haunting (IGH) of chimps (i.e.: diverse intelli-
gence and sexual motivation)
 Stage 4: The implementation of four main steps of hunting as The rest of the paper is structured as follows. Section 2 de-
(driving, chasing, blocking, and attacking) scribes the chimp optimization algorithm developed in the arti-
cle. Optimization problems and their experimental results are pre-
6 M. Khishe and M.R. Mosavi / Expert Systems With Applications 149 (2020) 113338

Table 3 As shown in Fig. 3, these two species share a 98.8% of their DNAs
Unimodal benchmark function.
(Tomkins & Bergman, 2012).
Function Dim Range fmin The chimp’s colony is a fission-fusion society. This kind of soci-
 ety is one in which the combination or size of the colony changes
F1 (x ) = ni=1 x2i 30, 100 [ − 100, 100] 0
 
F2 (x ) = ni=1 |xi | + ni=1 |xi | 30, 100 [ − 10, 10] 0 as time passes and members move throughout the environment.
n i 2
F3 (x ) = i=1 ( j−1 x j ) 30, 100 [ − 100, 100] 0 For chimps that live in fission-fusion colonies, group composition
F4 (x) = max i {|xi |,1 ≤ i ≤ n} 30, 100 [ − 100, 100] 0 is a dynamic property (Couzin & Laidre, 2009). Considering these
 −1 2
F5 (x ) = ni=1 [100(xi+1 − x2i ) + (xi − 1 ) ]
2
30, 100 [ − 30, 30] 0 issues, the independent group concept is proposed. In this tech-

F6 (x ) = ni=1 ([xi + 0.5] )
2
30, 100 [ − 100, 100] 0 nique, each group of the chimps independently attempts to dis-

F7 (x ) = ni=1 ix4i + random[0, 1 ) 30, 100 [ − 1.28, 1.28] 0
cover the search space with its own strategy. In each group, chimps
are not quite similar in terms of ability and intelligence, but they
are all doing their duties as a member of the colony. The ability of
each individual can be useful in a particular situation.
In a chimp colony, there are four types of chimps entitled
driver, barrier, chaser, and attackers. They all have different abil-
ities, but these diversities are necessary for a successful hunt.
Drivers follow the prey without attempting to catch up with it.
Barriers place themselves in a tree to build a dam across the pro-
gression of the prey. Chasers move rapidly after the prey to catch
up with it. Finally, attackers prognosticate the breakout route of
the prey to infliction it (the prey) back towards the chasers or
down into the lower canopy. These steps of hunting process are
shown in Fig. 4. Attackers are thought to need much more cog-
nitive endeavour in prognosticating the subsequent movements of
the prey, and they are thus remunerated with a larger piece of
meat after a successful hunt. This important role (attacking) corre-
lates positively with the age, smartness, and physical ability. More-
over, chimps can change duties during the same hunt or keep their
same duty during the entire process (Boesch, 2002).
It has been proven that chimps hunt to obtain meat for trad-
ing in social favours such as coalitionary support, sex or grooming
(Stanford et al., 1996). So, by opening up a new realm of privileges,
smartness may have an indirect effect on hunting. To the best of
our knowledge, in addition to humans, this social incentives has
been proposed only for chimps. Hence, it would represent a criti-
cal difference between chimps and other social predators that de-
pend on cognitive ability. This social incentive (sexual motivation)
causes the chimps to act chaotically in the final stage of hunting
process so that all chimps abandon their special duties and they
try to get meat, frantically. Generally speaking, the hunting pro-
cess of chimps is divided into two main phases: Exploration which
Fig. 11. Presents the pseudo-code of ChOA. consists of driving, blocking and chasing the prey and Exploitation
which consists of attacking the prey. These two phases are shown
in Figs. 4 and 5, respectively. Then, all of these concepts of ChOA
sented and discussed in Sections 3. Finally, Section 4 concludes the are mathematically formulated in the following section.
work and suggests directions for further research.
2.2. Mathematical model and algorithm

2. Chimp optimization algorithm

In this section, mathematical models of independent group,
driving, blocking, chasing and attacking are presented. Correspond-
This section presents and discusses the inspiration of ChOA
ing ChOA algorithm is then specified.
method. Afterwards, it provides the mathematical model of the
proposed algorithm.
2.2.1. Driving and chasing the prey
As mentioned above, the prey is hunted during the explo-
2.1. Inspiration ration and exploitation phases. To mathematically model driving
and chasing the prey, Eqs. (1) and (2) are proposed.
 
d = c.xprey (t ) − m.xchimp (t )
Chimps (sometimes called Chimpanzees) are one of two merely
(1)
African species of great ape. They are as much as the closest to
the humans’ living relatives. As shown in Fig. 1, the chimps, as
well as the dolphins, have the most similar Brain to Body Ratio xchimp (t + 1 ) = xprey (t ) − a.d (2)
(BBR) to humans. As discussed in Roth and Dicke (2005) mammals
Where t indicates the number of current iteration, a, m, and c
with relatively larger BBR are mostly assumed to be smarter. The
are the coefficient vectors, Xprey is the vector of prey position and
chimp and the human DNA are so similar because they are de-
Xchimp is the position vector of a chimp. a, m, and c vectors are
scended from a single ancestor species (Hominoid) that lived seven
calculated by the Eqs. (3)–(5), respectively.
or eight million years ago. Fig. 2 indicates the phylogeny of super-
family Hominoid (Israfil, Zehr, Mootnick, Ruvolo, & Steiper, 2011). a = 2.f.r1 − f (3)
M. Khishe and M.R. Mosavi / Expert Systems With Applications 149 (2020) 113338 7

Fig. 12. Convergence curve of algorithms on the unimodal test functions.

8 M. Khishe and M.R. Mosavi / Expert Systems With Applications 149 (2020) 113338

Fig. 12. Continued

the individuals can be considered as a single group with one com-

c = 2.r2 (4) mon search strategy. However, theoretically, in every population-
based optimization algorithm, different independent groups that
have a common goal can be used to have a direct and ran-
m = Chaotic_value (5)
dom search result at the same time. In the following, indepen-
In which, f is reduced non-linearly from 2.5 to 0 through dent groups of chimp using different strategies to update f will
the iteration process (in both exploitation and exploration phase). be modelled mathematically. Updating the independent groups can
Where r1 and r2 are the random vectors in the range of [0,1]. Fi- be implemented by any continuous function. These functions must
nally, m is a chaotic vector calculated based on various chaotic be chosen in such a way that during each iteration f is reduced
map so that this vector represents the effect of the sexual moti- (Mirjalili, Lewis, & Sadiq, 2014).
vation of chimps in the hunting process. A full description of this These four independent groups use their own patterns to search
vector will be described in detail in the following subsections. In the problem space locally and globally. Also among various strate-
the conventional population-based optimization algorithm, all par- gies which have been tested, two different versions of ChOA with
ticles have similar behaviour in local and global searches so that
M. Khishe and M.R. Mosavi / Expert Systems With Applications 149 (2020) 113338 9

Fig. 13. Convergence curve of algorithms on the multimodal test functions.

10 M. Khishe and M.R. Mosavi / Expert Systems With Applications 149 (2020) 113338

Fig. 13. Continued

various independent groups called ChOA1 and ChOA2 are selected allowed to access any position between the points shown in Fig.
to have the best performance in the benchmark optimization prob- 7 through the random vectors r1 and r2 . So, any chimp can ran-
lems. The dynamic coefficients of f have been proposed in Table domly change its location within the space surrounding the prey
1 and Fig. 6. In this table, T represents the maximum number of using Eqs. (1) and (2).
iterations, and t indicates the current iteration. These dynamic co- This concept can be generalized to an n-dimensional search
efficients have been chosen with various curves and slopes so that space. As mentioned in the previous section, the chimps also attack
each independent group has specific searching behaviour for the the prey with the chaotic strategy. This method is mathematically
sake of improving the performance of ChOA. formulated in the following section.
Some points may be considered to understand how indepen-
dent groups are effective in ChOA: 2.2.2. Attacking method (exploitation phase)
To mathematically model attacking behaviour of chimps, two
• Independent groups have different strategies to update f, so approaches are designed as follows: The chimps are capable of ex-
chimps could explore the search space with different capabil- ploring the prey’s location (by driving, blocking and chasing) and
ity. then encircling it. The hunting process is usually conducted by at-
• Diverse and dynamic strategies of f cause balancing between tacker chimps. Driver, barrier and chaser chimps are occasionally
global and local search. participate in the hunting process. Unfortunately in an abstract
• Independent groups contain non-linear strategies such as loga- search space there is no information about the optimum location
rithmic and exponential functions for f, so ChOA could be ef- (prey). In order to mathematically simulate the behaviour of the
fective in solving complex optimization problems. chimps, it is assumed that the first attacker (best solution avail-
• ChOA with independent groups could be adaptable in solving a able), driver, barrier and chaser are better informed about the lo-
wider range of optimization problems. cation of potential prey. So, four of the best solutions yet obtained
is stored and other chimps are forced to update their positions ac-
To understand the effects of Eqs. (1) and (2), a two-dimensional
cording to the best chimps locations. This relationship is expressed
representation of the position vector and a number of possible
by the Eqs. (6)–(8).
neighbours are shown in Fig. 7a. As can be observed, a chimp in
position (x,y) can change its position with respect to prey’s (x∗ , dAttacker = |c1 xAttac ker − m1 x|, dBarrier = |c2 xBarrier − m2 x|,
(6)
y∗ )location. Various locations around the most suitable agent can dChaser = |c3 xChaser − m3 x|, dDriver = |c4 xDriver − m4 x|.
be taken considering its current location and changing and setting
the values of a and c vectors. For instance the location of (x∗ − x, x1 = xAttac ker − a1 (dAttac ker ), x2 = xBarrier − a2 (dBarrier ),
(7)
y∗ ) is obtained by setting a = (1,0), m = (1,1) and c = (1,1). Up- x3 = xChaser − a3 (dChaser ), x4 = xDriver − a4 (dDriver ).
dated possible locations of a chimp in a three-dimensional space x1 + x2 + x3 + x4
are indicated in Fig. 7b. It should be noted that the chimps are x(t + 1 ) = (8)
4
 M. Khishe and M.R. Mosavi / Expert Systems With Applications 149 (2020) 113338 11

Fig. 14. Convergence curve of algorithms on the fixed-dimension multimodal benchmark functions.
12 M. Khishe and M.R. Mosavi / Expert Systems With Applications 149 (2020) 113338

Fig. 14. Continued

 M. Khishe and M.R. Mosavi / Expert Systems With Applications 149 (2020) 113338 13

Fig. 14. Continued

Table 4
Multimodal benchmark function.

Function Range Dim fmin

 
F8 (x ) = ni=1 −xi sin( |xi | ) [ − 500, 500] 30,100 -418.9829 × Dim
n
F9 (x ) = i=1 [xi − 10cos (2πxi ) + 10] [ − 5.12, 5.12]
2
30,100 0
F10 (x ) = −20 exp(−0.2 1n ni=1 x2i )
 [ − 32, 32] 30,100 0
− exp( 1n ni=1 cos(2π xi ) ) + 20 + e
 
F11 (x ) = 40100 ni=1 x2i − ni=1 cos( √xi ) + 1 [ − 600, 600] 30,100 0
i
2
 −1 (yi − 1 )2 [1 + 10sin2 (π yi+1 )]
F12 (x ) = πn {10sin(π y1 ) + ni=1 }
+ ( yn − 1 ) [ − 50, 50] 30,100 0
n
+ i=1 u(xi , 10, 100, 4 )
yi = 1 + xi 4+1
k ( xi − a )
m
xi > a
u(xi , a, k, m ) = { 0 −a < xi < a
k(−xi − a )
m
xi < −a
 (x − 1 )2 [1 + sin2 (3π xi + 1 )]
F13 (x ) = 0.1{sin2 (3π x1 ) + ni=1 i }
+ (xn − 1 )2 [1 + sin2 (2π xn )] [ − 50, 50] 30,100 0
n
+ i=1 u(xi , 5, 100, 4 )

Table 5
Fixed-dimension multimodal benchmark function.

Function Range Dim fmin

25
F14 (x ) = ( 1
+ j=1 2 1 −1
) [ − 65, 65] 2 1
j+ i=1 (xi −ai j )
500 6

11 x1 (b2i +bi x2 )

2
F15 (x ) = i=1 [ai − b2 +bi x3 +x4 ] [ − 5, 5] 4 0.00030
i
F16 (x ) = 4x21 − 2.1x41 + 13 x61 + x1 x2 − 4x22 + 4x42 [ − 5, 5] 2 − 1.0316
F17 (x ) = ( 5.1 2
) (
x2 − 4π 2 x1 + π5 x1 − 6 2 + 10 1 − 81π cosx1 + 10 ) [ − 5, 5] 2 0.398
F18 (x ) = ( ) (
[1 + x1 + x2 + 1 2 19 − 14x1 + 3x21 − 14x2 + 6x1 x2 + 3x22 )]
[ − 2, 2] 2 3
×[30 + ( ) (
2x1 − 3x2 2 × 18 − 32x1 + 12x21 + 48x2 − 36x1 x2 + 27x22 )]
  2
F19 (x ) = (
− 4i=1 ci exp − 3j=1 ai j x j − pi j ( ) ) [1, 3] 3 − 3.86
  2
F20 (x ) = (
− 4i=1 ci exp − 6j=1 ai j x j − pi j ( ) ) [0, 1] 6 − 3.32
5 −1
F21 (x ) = ( )( )
T
− i=1 [ X − ai X − ai + ci ] [0, 10] 4 − 10.1532
 −1
F22 (x ) = ( )( )
T
− 7i=1 [ X − ai X − ai + ci ] [0, 10] 4 − 10.4028
10 −1
F23 (x ) = ( )( )
T
− i=1 [ X − ai X − ai + ci ] [0, 10] 4 − 10.5363
14 M. Khishe and M.R. Mosavi / Expert Systems With Applications 149 (2020) 113338

Table 6
The rotated and shifted benchmark functions.

Function Range Dim fmin

[ (1−cos(x2 )) ]
2
F24 (x ) = sin(x1 )e
[ − 2π , 2π ] 4 -106.764537
+ cos(x2 )e[(1−sin(x1 ))
2

√ + ( x1 − x2 )
] 2

sin2 [ x1 2 +x2 2 ]−0.5

F25 (x ) = 0.5 + 2 [ − 100, 100] 40 0.5
[1+0.001(x1 2 +x2 2 )]
m
−1 √
sin2 [ x21 +x22 ]−0.5
F26 (x ) = ( 0.5 + 2 ) [ − 100, 100] 40 0.5
[1+0.001(x21 +x22 )]
i=1
F27 (x ) = −4|sin(x1 ) cos(x2 )e|cos((x1 +x2 )/200| |
2 2
[ − 10, 10] 10 − 10.8723
2 0.5
F28 (x ) = −0.0 0 01[|sin(x1 ) sin(x2 )e|100−[(x1 +x2 ) ]/π|
| + 1]0.1
2
[ − 10, 10] 20 0.1
0.5 −1
F29 (x ) = − exp[−|cos(x1 ) cos(x2 )e|1−[( x21 +x22 ) ]/ π| |] [ − 11, 11] 30 − 0.96354
m
−1  
F30 (x ) = (−(xi+1 + 47 ) sin( |xi+1 + xi /2 + 47| ) + sin( |xi − xi+1 + 47| )(−xi )) [ − 512, 512] 30 959.64
i=1

Table 7
The naming style for ChOAs.

Chaotic map Updating Strategies

Quadratic Gauss/Mouse Logistic Singer Bernoulli Tent

Type 1 (ChOA1) ChOA11 ChOA12 ChOA13 ChOA14 ChOA15 ChOA16

Type 2 (ChOA2) ChOA21 ChOA22 ChOA23 ChOA24 ChOA25 ChOA26

Table 8
Parameters and initial values of the benchmark algorithm.

Algorithm Parameter Value

ChOA f Table I
r1, r2 Random
m Chaotic
Number of Chimps 50
Maximum number of iterations 250
BBO Habitat modification probability 1
Immigration probability bounds per gene [0,1]
Step size for numerical integration of probabilities 1
Max immigration (I) and Max emigration (E) 1
Mutation probability 0.005
Population size 50
Maximum number of generations 250
GWO Number of wolf 50
Upper bound 5
Lower bound −5
Maximum number of iterations 250
LGWO a0 2
β ~U(0,2)
p ~U(0,1)
ALO w [2,6]
Number of search agent 50
Modified bound [−100,100]
Maximum number of iterations 250
BH a [0,1]
Number of stars 100
Maximum number of iterations 250
PSO Cognitive constant (C1 ) 1
Social constant (C2 ) 1
Local constant (W) 0.3
Population size 50
Maximum number of iterations 250
GA Type Real coded
Selection Roulette wheel
Recombination Single-point (1)
Mutation Uniform (0.01)
Layout Full connection
Population size 50
Maximum number of iterations 250
GSA Population size 50
Number of masses 30
Gravitational constant 1
Maximum number of iterations 250
pa 0.25
CS Population size 50
Maximum number of iterations 250
 M. Khishe and M.R. Mosavi / Expert Systems With Applications 149 (2020) 113338 15

Table 9
The results of unimodal benchmark functions.

Algorithm F1 F2 F3 F4 F5 F6 F7

ChOA11 Ave 5.9216e-33 1.0792e-19 1.9616e-08 1.0878e-08 27.1256 0.78715 0.0011101

Std 0.0000507 0.00017244 0.0015376 0.00029664 0.001221 0.00066241 0.0004353
p-value 0.0001 0.0001 0.0001 0.0001 0.0001 0.0001 0.0001
ChOA12 Ave 6.8573e-49 2.1821e-28 1.3912e-08 1.4402e-12 27.1546 0.2159 0.0011056
Std 0.00000003 0.00000035 0.0000141 0.00001211 0.0016241 0.00091769 0.683e-05
p-value N/A N/A N/A N/A 0.0001 0.0001 N/A
ChOA13 Ave 5.793e-25 2.6344e-15 0.0016344 2.0177e-06 27.1812 0.59441 0.0014983
Std 0.00044803 0.0025339 0.0008674 0.00036884 0.0012275 0.00022243 0.000892
p-value 0.0001 0.0001 0.0001 0.0001 0.0001 0.0001 0.0001
ChOA14 Ave 2.2761e-07 1.337e-06 2.3752 0.029591 29.0001 0.92924 0.0020332
Std 0.00068042 0.00028847 0.00061586 9.2999e-05 0.00070309 0.00028788 0.0012945
p-value 0.0047 0.0047 0.0047 0.0047 0.0047 0.0047 0.0047
ChOA15 Ave 1.7851e-09 1.0453e-07 1.9476 0.028648 28.0388 1.2473 0.0019176
Std 0.0023071 0.00058469 0.0014968 0.00034511 0.00048538 0.0024164 0.0001202
p-value 0.00797 0.00797 0.00797 0.00797 0.00797 0.00797 0.00797
ChOA16 Ave 2.0825e-05 9.8232e-05 2.02153 2.4511 33.9835 3.9568 0.010535
Std 0.00088144 0.00020824 0.00048348 0.00014341 0.00076807 0.0008887 0.0009381
p-value 0.00797 0.0047 0.0047 0.0047 0.00797 0.00797 0.0001
BBO Ave 0.013011 0.2334 3.9745 1.7185 40.8825 0.39002 0.081996
Std 0.00028037 0.0013183 0.00057292 0.00036264 0.00034323 0.0007505 3.456e-05
p-value 6.39e-05 6.39e-05 0.0047 6.39e-05 0.00797 0.00797 0.0001
BH Ave 0.27385 1.132 3.9664 1.0276 50.9993 0.065608 0.11347
Std 0.00093598 6.3233e-05 0.0016358 0.00026678 0.0021593 0.00073555 2.793e-05
p-value 0.00797 6.39e-05 6.39e-05 6.39e-05 0.00797 0.00797 0.0001
ALO Ave 1.2444 2.3692 3.8894 33.7621 63.3616 2.0344 0.79273
Std 0.00092526 7.7181e-05 0.0010751 0.0011727 0.0007689 0.00073103 0.001886
p-value 0.0057 6.39e-05 6.39e-05 6.39e-05 6.39e-05 6.39e-05 6.39e-05
GWO Ave 5.0555e-13 4.6933e-08 0.7213 0.00074374 28.8595 1.2504 0.0014836
Std 0.00040636 0.00013043 0.00078211 7.9584e-06 7.3315e-05 0.00063077 0.0005685
p-value 0.0057 0.0049 6.39e-05 6.39e-05 6.39e-05 0.0049 0.0057
ChOA21 Ave 1.6375e-26 6.8737e-17 3.8819e-07 1.4583e-06 27.1736 0.50289 0.0016598
Std 0.00057013 0.00068177 0.00016291 0.00029769 0.00010131 0.00098584 0.0015441
p-value 0.00797 0.0049 0.0049 0.0057 0.0049 0.0049 0.0057
ChOA22 Ave 1.552 1.1836 1.759 0.9313 58.5251 1.2803 3.3205
Std 0.00114 0.0013475 9.319e-05 0.00077073 0.0011949 0.0019343 0.0011812
p-value 0.00797 0.0049 0.0049 0.00797 0.0057 0.0049 0.0057
ChOA23 Ave 0.0042069 0.46036 2.06 0.7323 20.8532 0.058577 0.10077
Std 0.00031728 8.4022e-05 0.0013834 0.00052112 0.00010311 0.00026976 7.668e-05
p-value 0.0057 0.00797 0.0049 0.00797 N/A 0.0057 0.00797
ChOA24 Ave 0.43608 0.65876 2.7305 1.8589 39.7043 0.18706 0.072983
Std 0.00087541 0.001156 0.00011975 0.0030722 0.00026654 0.0011506 0.0004988
p-value 0.0057 0.0057 0.0049 0.0049 0.0057 0.00797 0.00797
ChOA25 Ave 0.02323 0.58268 3.3941 1.9377 40.7174 0.0064761 0.063093
Std 0.00027643 0.00016033 0.0012325 0.0011574 0.00032931 0.00000431 0.0010881
p-value 0.0057 0.0049 0.0049 0.0057 0.00797 N/A 0.00797
ChOA26 Ave 0.10797 0.16594 2.6906 1.3117 38.7458 0.070298 0.11899
Std 0.00036936 0.00084257 0.00013504 0.00023687 0.0015624 0.0016968 0.0005133
p-value 0.0057 0.00797 0.0049 0.0049 0.0057 0.0049 0.0049

Fig. 8 shows the process of updating the search chimp’s location its current position and the position of the prey. Fig. 9 shows that
in two-dimensional search space regarding the position of other the inequality forces the chimps to attack the prey.
chimp positions. As it can be seen, the final position is located According to the operators that have already been presented,
randomly in a circle which is defined by attacker, barrier, chaser ChOA allows the chimps to update their positions according to the
and driver chimp positions. In other words, the prey position is positions of attacker, barrier, chaser, and driver chimps and attack
estimated by four best groups and other chimps randomly update the prey. However, ChOAs may still be at the risk of trapping in
their positions within its vicinity. local minima, so other operators are required to avoid this issue.
Although, the proposed driving, blocking, and chasing mechanism
2.2.3. Prey attacking (utilization) somehow shows exploration process, ChOA requires more opera-
As mentioned previously, in the final stage, the chimps will at- tors to emphasize exploration phase.
tack the pray and finish the hunt as soon as the prey stops mov-
ing. To mathematically model the attacking process, the value of f 2.2.4. Searching for pray (exploration)
should be reduced. Note that the variation range of the a is also As previously mentioned, the exploration process among the
reduced by f. In other words, a is a random variable in the interval chimps is mainly done considering the location of attacker, bar-
of [−2f,2f], whereas the value of f reduces from 2.5 to 0 in the pe- rier, chaser, and driver chimps. They diverge to seek for the prey
riod of iterations. When the random values of a lie in the range of and aggregate to attack prey. In order to mathematically model
[−1,1], the next position of a chimp can be in any location between the divergence behaviour, the a vector with a random value bigger
16 M. Khishe and M.R. Mosavi / Expert Systems With Applications 149 (2020) 113338

Table 10
The results of multimodal benchmark functions.

Algorithm F8 F9 F10 F11 F12 F13

ChOA11 Ave −6432.073 5.6843e-14 3.9968e-14 0 0.03789 0.59045

Std 3.2605 0.0007579 0.014542 0 0.00062651 0.018694
p-value 0.0001 N/A 0.0057 0.0057 0.0049 0.0049
ChOA12 Ave −3150.5985 2.738 7.9936e-15 0 0.29035 1.7768
Std 21.9845 0.008412 0.000851 0 0.015873 0.0097053
p-value 0.0057 0.0057 0.0049 0.0049 0.0001 0.0001
ChOA13 Ave −3628.8022 5.6843e-14 1.0036e-13 0 0.043508 0.53169
Std 5.1249 0.0012031 0.0051235 0 0.010383 0.003692
p-value 0.0057 0.0057 0.0049 0.0049 0.0001 N/A
ChOA14 Ave −5652.3897 1.1596 2.9619 0.036661 0.16291 3.0763
Std 7.6746 0.0030101 0.010755 0.0015965 0.014936 0.018627
p-value 0.0057 0.0057 0.0049 0.0049
ChOA15 Ave −5594.9085 1.3417 2.9668 0.014122 0.07438 2.8352
Std 4.5861 0.0010454 0.0075506 0.013896 0.00086557 0.019978
p-value 0.00747 0.0001 0.0057 0.0057 0.0049 0.0049
ChOA16 Ave −5588.3064 5.5591e-05 2.9668 0.42516 0.4587 2.2013
Std 21.4912 0.0064053 0.013362 0.0076654 0.007336 0.018548
p-value 0.00747 0.0057 0.0057 0.0049 0.0049 0.0001
BBO Ave −9952.6876 1.0027 0.5129 0.29361 2.1009 4.8418
Std 7.2952 0.012332 0.01008 0.0074866 0.0089069 0.0047195
p-value 6.39e-05 6.39e-05 6.39e-05 0.0057 0.0057 0.0049
BH Ave −9230.6623 2.0007 0.6284 0.31144 1.1078 0.39991
Std 5.5351 0.017319 0.011713 0.0096344 0.014307 2.8186e-05
p-value 6.38e-05 6.39e-05 6.39e-05 6.39e-05 6.38e-05 6.38e-05
ALO Ave −7167.2698 2.4653 0.8945 2.5676 1.5463 5.7996
Std 10.5861 0.012509 0.0092374 0.016871 0.0043937 0.00586957
p-value 6.38e-05 6.38e-05 6.39e-05 6.39e-05 6.39e-05 0.0001
GWO Ave −5665.3886 1.4001 2.9667 0.014955 0.081592 2.316
Std 24.838 0.0066279 0.0073361 0.0084009 0.0056174 0.010415
p-value 6.39e-05 6.39e-05 6.39e-05 6.39e-05 6.39e-05 0.0001
ChOA21 Ave −6738.8454 0.0026 1.2168e-15 0 0.027962 0.8424
Std 6.4067 0.010056 0.0000505 0 0.0000267 0.023382
p-value 6.39e-05 6.39e-05 N/A N/A N/A 0.0045
ChOA22 Ave −2609.1446 2.2198 0.7874 0.13063 0.1821 0.52586
Std 3.81561 0.013331 0.00065341 0.0036728 0.0076626 0.000825
p-value 0.00747 0.0057 0.0057 0.0049 0.0049 N/A
ChOA23 Ave −1023.9291 1.0698 1.2264 0.12105 3.0047 2.513
Std 1.8642 0.0032444 0.010339 0.016238 0.0046195 0.0013249
p-value N/A 0.00747 0.0057 0.0057 0.0049 0.0049
ChOA24 Ave −9655.5997 1.42 0.7238 0.010202 0.283 5.1047
Std 10.4246 0.0070324 0.002375 0.011905 0.016156 0.02809
p-value 0.00747 0.0057 0.0057 0.0049 0.0049 0.0001
ChOA25 Ave −9146.7686 1.6997 1.7298 0.025172 7.4374 4.3983
Std 9.9677 0.017173 0.014912 0.010107 0.013853 0.0070056
p-value 0.0001 0.00747 0.0057 0.0057 0.0049 0.0049
ChOA26 Ave −9026.9166 1.7645 1.3173 0.01028 0.89062 3.9773
Std 4.356 0.006938 0.014907 0.013683 0.00010712 0.00455227
p-value 0.00747 0.00747 0.00747 0.0057 0.0057 0.0049

than 1 or smaller than −1 is used, so that the search agents are especially in the final iterations. c vector is also considered as the
forced to diverge and get distant from prey. This procedure shows influence of the obstacles which prevent chimps from approach-
the exploration process and allows the ChOA to search globally. Fig. ing the prey in nature. In general, natural obstacles in the path
9 shows that the inequality |a| > 1 forces the chimps to scatter in of chimps prevent them from approaching the prey with proper
the environment to find a better prey. This section is inspired from speed. This is the precise expression of the c vector effect. Depend-
GWO (Mirjalili, 2013). ing on chimp’s position, the c vector can assign a random weight
Another ChOA component that affects the exploration phase is to prey in order to make the hunt harder or easier.
the value of c. As in Eq. (4), c vector elements are random variables
in the interval of [0,2]. This component provides random weights
for prey to reinforce (c > 1) or lessen (c < 1) the effect of prey lo- 2.2.5. Social incentive (sexual motivation)
cation in the determination of the distance in Eq. (5). It also helps As mentioned previously, acquiring meet and subsequent social
ChOA to enhance its stochastic behaviour along the optimization motivation (sex and grooming) in the final stage causes chimps to
process and reduce the chance of trapping in local minima. c is release their hunting responsibilities. Therefore, they try to obtain
always needed to generate the random values and execute the ex- meat forcefully chaotic.
ploration process not only in the initial iterations, but also in the This chaotic behaviour in final stage helps chimps to further al-
final iterations. This factor is very useful for avoiding local minima, leviate the two problems of entrapment in local optima and slow
convergence rate in solving high-dimensional problems.
 M. Khishe and M.R. Mosavi / Expert Systems With Applications 149 (2020) 113338 17

Table 11
The results of fixed-dimension multimodal benchmark functions.

Algorithm F14 F15 F16 F17 F18

ChOA11 Ave 0.998 0.020364 −1.0316 0.39792 3

Std 0.00059718 0.010885 0.0095395 0.0047829 0.010803
p-value 0.00747 0.0057 0.0057 0.0049 0.0049
ChOA12 Ave 0.998 0.00034398 −1.0316 0.39792 3
Std 0.0057686 0.00000001 0.021147 0.00026104 0.009066
p-value 0.00747 N/A 0.0057 0.0049 0.0049
ChOA13 Ave 0.99801 0.00067708 −1.0316 0.39865 3.0001
Std 0.0047493 0.0068029 0.0096524 0.015434 0.03319
p-value 0.0001 0.0001 0.00747 0.0057 0.0057
ChOA14 Ave 0.998 0.0012896 −1.0316 0.39833 3.0002
Std 0.00037144 0.011627 0.0052976 0.0020581 0.0067462
p-value 0.00747 0.0057 0.0057 0.0049 0.0049
ChOA15 Ave 0.99802 0.00125 −1.0316 0.39796 3.0001
Std 0.014758 0.0077204 0.016366 0.0023049 0.00043324
p-value 0.00747 0.0057 0.0057 0.0049 0.0049
ChOA16 Ave 0.99809 0.0013562 −1.0316 0.39805 3.0001
Std 0.0024767 0.0011969 0.00089862 0.0062243 0.0070154
p-value 0.00747 0.0057 0.0057 0.0049 0.0049
BBO Ave 0.998 0.00042263 −1.0316 0.39789 3
Std 0.0090912 0.01693 0.014821 0.0065533 0.01335
p-value 6.39e-05 6.39e-05 6.39e-05 6.39e-05
BH Ave 0.998 0.020363 −1.0316 0.39789 3
Std 0.0038131 0.00066429 0.012439 0.014056 0.021488
p-value 6.39e-05 6.39e-05 6.39e-05 6.39e-05 6.39e-05
ALO Ave 0.998 0.020363 −1.0316 0.39789 3
Std 0.021358 0.0017314 0.012602 0.0096567 0.007996
p-value 6.39e-05 6.39e-05 6.39e-05 6.39e-05 6.39e-05
GWO Ave 0.998 0.0012849 −1.0316 0.39842 3.0001
Std 0.018609 0.020989 0.021182 0.0097607 0.0099344
p-value 6.38e-05 6.39e-05 6.38e-05 6.39e-05 6.38e-05
ChOA21 Ave 0.998 0.00035113 −1.0316 0.39789 3.0004
Std 0.0063754 0.0028921 0.0011222 0.013257 0.011011
p-value 0.00747 0.0057 0.0057 0.0049 0.0049
ChOA22 Ave 2.9821 0.00099085 −1.0316 0.39789 3
Std 0.0078901 0.0042495 0.010545 0.00093272 0.016413
p-value 0.00747 0.0057 0.0057 0.0049 0.0001
ChOA23 Ave 0.998 0.00034349 −1.0316 0.39789 3
Std 0.0001241 0.00000013 0.0000741 0.000793 0.0001531
p-value N/A N/A N/A N/A N/A
ChOA24 Ave 0.998 0.00033082 −1.0316 0.39789 3
Std 0.0092874 0.0071116 0.0061951 0.0045872 0.01198
p-value 0.00747 0.0057 0.0057 0.0049 0.0049
ChOA25 Ave 0.998 0.0016554 −1.0316 0.39789 3
Std 0.0056774 0.006215 0.0010265 0.0086782 0.0020994
p-value 0.0047 0.00757 0.0057 0.0049 0.0049
ChOA26 Ave 0.998 0.00030769 −1.0316 0.39789 3
Std 0.010431 0.0032917 0.0047039 0.0044647 0.00028328
p-value 0.00547 0.00457 0.0057 0.0049 0.0049

The chaotic maps which have been used to improve the perfor- Where μ is a random number in [0,1].
mance of ChOA are explained in this section. Six chaotic maps have In brief, the searching process in ChOA begins with generat-
been used in this article as shown in Table 2 and Fig. 10. These ing a stochastic population of chimps (candidate solutions). Then,
chaotic maps are deterministic processes which also have random all chimps are randomly divided into four predefined independent
behaviour. In this article, value 0.7 has been considered as the pri- groups entitled attacker, barrier, chaser and driver. Each chimp up-
mary point of all the maps in accordance with reference (Saremi, dates its f coefficients using the group strategy. During the iter-
Mirjalili, & Lewis, 2014). To model this simultaneous behaviour, we ation period, attacker, barrier, chaser and driver chimps estimate
assume that there is a probability of 50% to choose between either the possible prey locations. Each candidate solution updates its dis-
the normal updating position mechanism or the chaotic model to tance from the prey. Adaptive tuning the c and m vectors cause
update the position of chimps during optimization. The mathemat- local optima avoidance and faster convergence curve, simultane-
ical model is expressed by Eq. (9). ously. The value of f is reduced from 2.5 to zero, to enhance the
 process of exploitation and attacking the prey. The inequality |a| >
xprey (t ) − a.di f μ< 0.5 1 results in divergence of the candidate solutions, otherwise, they
xchimp (t + 1 ) = (9)
Chaotic_valuei f μ ≥ 0.5 eventually converge toward the prey. Fig. 11 presents the pseudo-
code of ChOA.
18 M. Khishe and M.R. Mosavi / Expert Systems With Applications 149 (2020) 113338

Table 12
The results of fixed-dimension multimodal benchmark functions (continued).

Algorithm F19 F20 F21 F22 F23

ChOA11 Ave −3.8622 −3.1969 −2.593 −10.2537 −9.3837

Std 0.019162 0.018342 0.022841 0.014715 0.0066907
p-value 0.00747 0.0057 0.0057 0.0049 0.0049
ChOA12 Ave −3.8619 −3.1825 −6.7593 −9.2651 −7.9056
Std 0.0071951 0.00030283 0.0070818 0.0043589 0.0081338
p-value 0.001 0.00747 0.0057 0.0057 0.0001
ChOA13 Ave −3.8614 −3.3106 −7.9664 −8.6936 −10.0206
Std 0.012393 0.0093514 0.016142 0.0085863 0.0021619
p-value 0.00452 0.00747 0.0057 0.0057 0.0049
ChOA14 Ave −3.8547 −2.0591 −4.8606 −5.0383 −5.0358
Std 0.006271 0.0001319 0.01722 0.0076048 0.013176
p-value 0.0045 N/A 0.00747 0.0057 0.0057
ChOA15 Ave −3.8548 −2.6329 −5.0189 −0.91158 −5.1029
Std 0.021395 0.014033 0.013209 0.0041865 0.0017119
p-value 0.0001 0.00747 0.0057 0.0057 0.0049
ChOA16 Ave −3.8624 −2.2492 −4.8537 −4.8401 −4.8898
Std 0.00026981 0.0075036 0.00034583 0.008309 0.0075659
p-value 0.00747 0.0057 0.0057 0.0049 0.0049
BBO Ave −3.8628 −3.322 −5.1008 −2.7519 −10.5364
Std 0.0051 0.0044809 0.010935 0.014868 0.014644
p-value 6.38e-05 6.39e-05 6.38e-05 0.00005 6.39e-05
BH Ave −3.8628 −3.2031 −2.6829 −10.4029 −10.5364
Std 0.0034586 0.00073532 0.0091199 0.0021605 0.022491
p-value 6.38e-05 6.39e-05 6.38e-05 0.00005 6.39e-05
ALO Ave −3.8628 −3.1974 −10.1532 −10.4029 −10.5364
Std 0.0077814 0.0075256 0.0051117 0.0067976 0.016077
p-value 6.38e-05 6.39e-05 6.38e-05 0.00005 6.39e-05
GWO Ave −3.8541 −3.0731 −0.88288 −5.0532 −5.0601
Std 0.0027469 0.00016015 0.0067064 0.0016984 0.0034639
p-value 6.38e-05 6.39e-05 6.38e-05 0.00005 6.39e-05
ChOA21 Ave −3.8627 −3.322 −10.1505 −10.4028 −10.5336
Std 0.011721 0.011144 0.003524 0.0017081 0.0044508
p-value 0.00747 0.0057 0.0057 0.0049 0.0049
ChOA22 Ave −3.8628 −3.322 −2.6829 −10.4029 −10.5364
Std 0.0053146 0.00084038 0.0039454 0.0004916 0.0018552
p-value 0.0045 0.00747 0.0057 N/A 0.0001
ChOA23 Ave −3.8628 −3.2031 −10.1532 −2.7659 −10.5364
Std 0.000045 0.0001641 0.0000534 0.0035001 0.0015218
p-value N/A N/A N/A N/A
ChOA24 Ave −3.8628 −3.322 −5.1008 −10.4029 −10.5364
Std 0.013575 0.0039439 0.010463 9.5041e-05 0.022171
p-value 0.00747 0.0057 0.0057 0.0049 0.0049
ChOA25 Ave −3.8628 −3.2031 −10.1532 −10.4029 −10.5364
Std 0.003998 0.011924 0.0079751 8.3701e-05 0.0023049
p-value 0.00747 0.0057 0.0057 0.0049 0.0049
ChOA26 Ave −3.8628 −3.322 −10.1532 −10.4029 −10.5364
Std 0.012586 0.01132 0.018268 0.02068 0.012983
p-value 0.0045 0.00747 0.0057 0.0057 0.0049

3. Simulation results and discussion modal, multimodal benchmark functions have more than one min-
imum, making them more challenging than unimodal benchmarks.
In this section, the ChOA is tested on 30 benchmark functions. Therefore, exploration and local minima avoidance of optimizers
The first 23 test functions are the classical benchmark functions can be tested by the multimodal benchmark functions. It should be
used in the many kinds of research (Digalakis & Margaritis, 2001; mentioned that the difference between fixed-dimensional multi-
Molga & Smutnicki, 2005; Yang, 2010). Generally, these functions modal benchmarks in Table 4 and multimodal benchmarks in Table
are divided into three groups: unimodal, multimodal, and fixed- 5 is the ability to define the desired number of design variables.
dimension multimodal which are reported in Tables 3–5, respec- The mathematical models of fixed-dimensional benchmark func-
tively. In these tables, dim indicates the dimension of the problem, tions do not let us tune the number of design variables, but they
fmin is the minimum reported in the literature, and Range is the prepare various search space compared to multimodal benchmark
boundary of the problem’s search space. functions in Table 4.
The three aforementioned groups of benchmark functions are In the following, in order to have a comprehensive compari-
utilized with different characteristics to test the performance of son, we use other newly proposed rotated and shifted benchmark
the ChOA from different aspects. As their names imply, unimodal function defined in the IEEE CEC 2013 special session and Compe-
benchmark functions have a single minimum so they can test the tition on Niching Methods for Multimodal Function Optimization
exploitation and convergence rate of ChOA. In contrast to uni- (Li, Engelbrecht, & Epitropakis, 2013) and also (Mishra, 2007). The
 M. Khishe and M.R. Mosavi / Expert Systems With Applications 149 (2020) 113338 19

Table 13
The results of shifted and rotated benchmark functions.

Algorithm F24 F25 F26 F27 F28 F29 F30

ChOA11 Ave −104.6332 2.0711e-5 1.9616e-9 −8.2221 0.9256 −0.78715 1000.254

Std 0.0000607 0.0011244 0.0001376 0.000114 0.001221 0.00054 10.2254
p-value 0.007937 0.0001 0.0001 0.0047 0.0047 0.0001 0.00747
ChOA12 Ave −103.2584 1.1821e-4 1.4412e-8 −7.0022 0.9946 −0.5214 999.547
Std 0.004101 0.0000122 0.0001457 0.000012 0.00162 0.00091 11.2547
p-value 0.007937 0.0001 0.0001 0.0047 0.0047 0.0001 0.00747
ChOA13 Ave −102.2589 2.6224e-5 2.0041e-7 −7.1245 0.8812 −0.59441 1001.0254
Std 0.0022803 0.0021139 0.005574 0.00022 0.0099275 0.00044 12.2547
p-value 0.007937 0.0001 0.0001 0.0047 0.0047 0.0001 0.00747
ChOA14 Ave −102.2369 0.0012 1.22335 −6.029591 0.8001 10.9291 985.635
Std 0.068042 0.0022847 0.00001586 9.2999e-05 0.000703 0.00028 9.2514
p-value 0.007937 0.0001 0.0001 0.0047 0.0047 0.0001 0.00747
ChOA15 Ave −102.3251 1.0453e-4 1.33254 −6.0286 1.0888 10.2473 998.254
Std 0.03071 0.001845 0.0014448 0.00011 0.000485 0.00241 2.2547
p-value 0.007937 0.0001 0.0001 0.0047 0.0047 0.0001 0.00747
ChOA16 Ave −102.3214 0.0058 3.12451 −6.0011 1.0005 15.9568 1000.0214
Std 0.000224 0.00980824 0.00048348 0.000121 0.00076 0.00089 5.5454
p-value 0.007937 0.0001 0.0001 0.0047 0.0047 0.0001 0.0047
BBO Ave −100.8888 3.2584 4.9745 −5.7122 1.8825 2.39002 1005.888
Std 0.0002803 0.01183 0.0005112 0.00364 0.000343 0.0007505 12.547
p-value 6.38e-05 6.39e-05 6.38e-05 0.00005 6.39e-05 6.39e-05 6.39e-05
BH Ave −99.2233 4.2154 4.1214 −5.3621 1.0093 1.065608 1009.214
Std 0.002298 0.001254 0.0011118 0.00026 0.00215 0.000735 14.2147
p-value 6.38e-05 6.39e-05 6.38e-05 0.00005 6.39e-05 6.39e-05 6.39e-05
ALO Ave −100.2589 5.2104 3.2147 −6.3254 1.3644 8.0344 999.357
Std 0.0011126 0.000114 0.0020711 0.00117 0.0007689 0.00073 3.2546
p-value 6.38e-05 6.39e-05 6.38e-05 0.0001 6.39e-05 6.39e-05 6.39e-05
GWO Ave −102.9999 0.14572 0.00013 −6.00074 1.8595 10.2504 984.2145
Std 0.0002236 0.0002243 0.001111 0.000123 7.3315e-05 0.000639 4.5454
p-value 6.38e-05 6.39e-05 6.38e-05 0.00047 6.39e-05 6.39e-05 6.39e-05
ChOA21 Ave −104.2312 1.8737e-17 0.004e-09 −7.2145 0.1006 −0.50289 968.245
Std 0.0001013 0.0002177 9.319e-05 0.000297 0.000007 0.00098584 3.2458
p-value 0.007937 0.4429 N/A 0.0081 N/A N/A 0.0047
ChOA22 Ave −105.2587 0.1245e-20 3.8819e-07 −7.5214 21.5251 −0.2803 960.999
Std 0.0000114 0.0000075 0.00088791 0.0000707 5.22119 0.0019343 1.25478
p-value N/A N/A 0.0047 N/A 6.39e-05 0.04785 N/A
ChOA23 Ave −104.8555 0.00036 3.7777 −6.1323 0.8532 1.058577 987.245
Std 0.0003172 8.4022e-05 0.0013834 0.00012 0.000125 0.00026976 3.6547
p-value 0.007937 0.0001 0.0001 0.00014 0.0074 0.0001 0.0047
ChOA24 Ave −103.2555 0.00876 4.2305 −7.0009 0.7043 2.18706 984.258
Std 0.0012541 0.001156 0.0001175 0.00307 0.000266 0.0011506 5.2145
p-value 0.007937 0.0001 0.0001 0.00047 0.0047 0.00047 0.00747
ChOA25 Ave −103.2411 1.58268 5.3941 −6.9377 159.7174 3.0064761 982.021
Std 0.0017643 0.0011603 0.0012325 0.00174 0.00032931 0.0008431 4.25896
p-value 0.007937 0.0001 0.0001 0.000041 0.0001 0.0047 0.00747
ChOA26 Ave −102.0541 0.16594 4.2211 −6.3117 111.7458 2.088298 1000.024
Std 0.0016936 0.003257 0.00013504 0.000287 0.0015624 0.00169 7.2145
p-value 0.007937 0.0001 0.00098 0.00021 0.0001 0.0001 0.00747

remaining benchmark functions are more complex and follow the For verifying the results, the ChOAs are compared to ALO
paradigm of composition functions. The mathematical models of (Mirjalili, 2015) as a kind of SIAs, BBO (Simon, 2008) as a pow-
these benchmark functions are shown in Table 6. erful kind of EAs and BH (Hatamlou, 2013) as a physics-based al-
The ChOAs have been divided into and named with regard to gorithm. In addition, the ChOAs are compared with GWO (Mirjalili
the type of the dynamic strategies for independent groups (illus- et al., 2014) as the most famous hunting-based benchmark algo-
trated in Table 1) and the number of the chaotic maps (illustrated rithm. The parameters of these algorithms are presented in Table
in Table 2). For instance, if the dynamic strategy number one- 8.
(from Table 1) and tent map (from Table 2) has been used to en- For these experiments, each test is carried out a Windows 10
hance ChOA, the name of that algorithm will be ChOA16 in such system using Intel Core i7, 3.8 GHz, 16 G RAM and, Matlab R2016a.
a way that 1 refers to the dynamic strategy type 1 and 6 refers to The ChOA algorithms were run 30 times on each benchmark func-
the row number of the temp map in Table 2. This type of naming tion. The statistical results (Average (Ave), Standard Deviation (Std),
has been thoroughly shown in Table 7. and p-value) are reported in Tables 9 to 13. The best results are il-
Figs. 12 to 15 draw a comparison between the convergence lustrated in bold type. The concepts of Ave and Std can be used
curves of different algorithms for unimodal, multimodal, fixed- to show the algorithms capability of avoiding the local minima.
dimension multimodal, and rotated and shifted benchmark func- The lower the value of Ave, the greater the algorithms capability of
tions, respectively. finding a solution near the global optimum. Although the Ave value
20 M. Khishe and M.R. Mosavi / Expert Systems With Applications 149 (2020) 113338

Table 14
The results of unimodal benchmark functions (100-dimensional).

Function ChOA PSO GSA BH

Ave Std Ave Std Ave Std Ave Std

F1 2.8e-09 1.1e-09 2.799 0.721 86.213 0.04243 9.34 2.0731

F2 44.16 18.17 23.87 2.432 152.44 0.00122 320.82 38.28
F3 194.11 42.18 391.34 41.57 169.88 0.03780 900.75 409.06
F4 2.18 0.0846 3.131 0.0879 10.265 0.0013 5.6670 0.8293
F5 13.94 1.71 75.2342 5.245 321.13 0.05329 117.80 49.07
F6 1.60e-07 1.09e-08 3.421 1.206 207.13 0.0003 4.2056 1.005
F7 0.000546 0.004407 1.721 4.0133 2.309 1.99e-05 0.4344 0.0127
GWO CS LGWO GA

Ave Std Ave Std Ave Std Ave Std

F1 4. 989 2. 678 3.80e-05 1.85e-05 6.1218 0.5744 18.7475 15.500

F2 28.647 0.9384 33.10 0.0656 22.191 1.0191 526.612 92.127
F3 2219 85.540 1,9527 52.7511 1852 418.4 933.21 332.1
F4 3.689 0.4572 2.9362 0.6877 2.736 0.0947 8.5144 0.5321
F5 262.7 11.490 27.672 13.88 11.171 0.76 91.4449 62.706
F6 13.99 2.1091 1.17e-05 1.55e-05 1.11e-05 1.05e-05 40.56 23.681
F7 0.8391 0.05354 0.00131 0.0087 0.00273 0.00041 9.56 5.1061

of the two algorithms can be equal, their performance in finding tively and they also utilize the chaotic maps biasing chimps to
the global optimum may differ in each iteration. Thus, Std is used move quickly toward the global optimum (prey). As can be seen
to make a better comparison. To have a lower dispersion of results, from Fig. 12 and Table 9, among the two proposed group sets, the
the Std should have a small value. first group set indicates much superior results for unimodal bench-
According to Derrac, García, Molina, and Herrera (2011), statis- mark function. This better result can be well justified by Fig. 6. Ac-
tical tests are required to evaluate the performance of MOAs ad- cording to this figure, ChOA1 has an excellent local search ability
equately. Comparing MOAs according to their Ave and Std values because the forms of updating strategies were chosen in such a
is not enough (Garcia, Molina, Lozano, & Herrera, 2009) and a sta- way that different groups tend to converge faster than ChOA2 and
tistical test is needed to indicate a remarkable improvement of a they search more locally than globally. In other words, the reduc-
new MOA in comparison to the other existing MOAs to solve a tion rate of the independent groups’ coefficient of ChOA1 is faster
particular optimization problem (Mirjalili & Lewis, 2013). In order than those coefficients of ChOA2. So, ChOA1 allows chimps to dis-
to see whether the results of ChOA differ from other benchmark cover the search space more locally than globally, because the am-
algorithms in a statistically significant way, Wilcoxon’s rank-sum plitude of the searching coefficient decreases severely after almost
(Wilcoxon, 1945), which is a non-parametric statistical test, was one-quarters of the allowed iterations.
performed and the significance level of 5% accomplished. The cal- It should be noted that this considerable improvement has not
culated p-values of the Wilcoxon’s rank-sum are given in the re- been made only with categorizing chimps to independent groups
sults as well. The N/A in tables is the abbreviation of ‘‘Not Appli- but also by utilizing the new chaotic map in such a way that this
cable’’ which means that the corresponding MOA cannot be com- chaotic behaviour in final stage helps chimps to further decline
pared with itself in the rank-sum test. Conventionally, p-values less the problems of slow convergence rate. As can be seen from Fig.
than 0.05 are considered as strong evidence against the null hy- 12 and Table 4, the chaotic map number two i.e., Gauss/mouse
pothesis. Note that p-values greater than 0.05 are underlined in the map has the most significant effect on the global minima finding
tables. and convergence speed so that ChOA12 has the best results in the
five out of seven unimodal benchmark functions and at least the
3.1. Evaluation of exploitation ability second best optimizer in other benchmark functions. The afore-
mentioned algorithm can hence provide fair exploitation ability.
Functions F1, F2,…, F7 have only one global optimum since they These superior results of ChOA12 are based on the special form
are unimodal. These benchmark functions permit to evaluate the of the Gauss/mouse map. This chaotic map has very special shape
exploitation capability of the investigated MOAs. Table 9 illustrates in such a way that in early stage it has large and extremely
ChOAs is very competitive with other MOAs. It can be seen from variable amplitude, while its amplitude and variableness decrease
Table 9 that ChOA12 has the best results in five out of seven uni- severely in the final stages. This special shape of Gauss/mouse map
modal test functions. causes chimps to behave both very extensively in early stage and
Fig. 12 shows the convergence curves of the algorithms. As can in focus of the final stages. Generally speaking, chaotic maps pro-
be seen from these curves, ChOA12 has the best convergence rates vide soft transition between global and local search ability. These
for most of the benchmark functions, followed by ChOA11 and maps prevent chimps from quickly becoming trapped in local min-
ChOA21. ima because chimps have stochastically movement even in the fi-
It is worth mentioning that unimodal test functions have no lo- nal stages. This stochastic movement in the final stage may be con-
cal minima and there is only one global minimum in the search sidered as sexual motivations. This is the main reason for the supe-
space. So these kinds of benchmark functions are quite appro- rior results of the proposed maps (specially the Gauss/mouse map).
priate for evaluating the convergence capability of MOAs. Conse- In this way, chimps tend to broadly discover promising regions
quently, the results of the ChOAs show that independent groups of search space and exploit the best one. Chimps change abruptly
could improve significantly the convergence ability of the ChOAs. in the early stages of the hunting process and then gradually con-
The two main reasons for the superior results is that the chimps verge. However, there is no additional computational cost for the
have diversity in their fission-fusion societies and are able to ex- proposed algorithm.
ploit knowledge of the position of near optimal solutions effec-
 M. Khishe and M.R. Mosavi / Expert Systems With Applications 149 (2020) 113338 21

Fig. 15. Convergence curve of algorithms on the rotated and shifted multimodal benchmark functions.
22 M. Khishe and M.R. Mosavi / Expert Systems With Applications 149 (2020) 113338

Fig. 15. Continued

Table 15
The results of multimodal benchmark functions (100-dimensional).

p ChOA PSO GSA BH

Ave Std Ave Std Ave Std Ave Std

F8 −44,426 144.5 −18,136 4962.4 −35,969 1876 −25,632 869.47

F9 11.89 1.005 62.58 2.301 12.01 0.12365 60.38 7.96
F10 0.3058 0.00542 1.183 0.07627 1.293 0.0974 1.159 0.0077
F11 0.0014 0.00021 270.2 11.49 400.5 0.8532 411 22.42
F12 0.3982 0.09591 2.07e + 05 2.77e + 05 1.00e + 08 1.99e+05 1.09e + 09 2.28e + 08
F13 0.13915 0.22199 1.24e + 06 3.82e + 05 1.00e + 08 1.99e + 05 1.25e + 09 3.85e + 08

GWO CS LGWO GA

Ave Std Ave Std Ave Std Ave Std

F8 −55,771 3097.8 −52,600 156.04 −39,753 649.69 −28,660 1011

F9 58.95 6.653 45.58 7.889 13.45 1.058 137.8 3.155
F10 1.544 0.06684 17.654 2.982 2.4297 0.038545 1.361 0.0421
F11 0.0011 0.00011 0.001191 0.001148 1.7048 0.014301 175.8 7.3
F12 2.37e + 07 1.22e + 07 1.00e + 10 0.0045 2.426 0.05985 3.14e + 09 2.54e + 08
F13 3.87e + 07 1.80e + 07 1.00e + 10 0.0568 0.0014 0.0568 1.38e + 10 1.45e + 09
 M. Khishe and M.R. Mosavi / Expert Systems With Applications 149 (2020) 113338 23

Table 16
The general description of real-world constrained optimization problem. D is the total number of decision variables of
the problem, g is the number of inequality constraints and h is the number of equality constraints.

No ID Problems D g h

1 RC01 Heat Exchanger Network Design (case 1) 9 0 8

2 RC04 Reactor Network Design (RND) 6 1 4
3 RC11 Two-reactor Problem 7 4 4
4 RC14 Multi-product batch plant 10 10 0
5 RC16 Optimal Design of Industrial refrigeration System 14 15 0
6 RC23 Optimal Design of Industrial refrigeration System 5 8 3
7 RC35 Optimal Sizing of Distributed Generation for Active Power Loss Minimization 153 0 148
8 RC37 Optimal Power flow (Minimization of Active Power Loss) 126 0 116
9 RC45 SOPWM for 3-level Inverters 25 24 1
10 RC51 Beef Cattle(case 1) 59 14 1

3.2. Evaluation of exploration ability not only of avoiding local optima but also improved convergence
speed. The results of the ChOA show that the proposed indepen-
In contrast to unimodal benchmark functions, multimodal prob- dent groups permit chimps to have various patterns for following
lems include many local minima whose number increases drasti- the social behaviour of the whole society, resulting in higher local
cally with the number of design variables (problem size). Hence, minima avoidance ability.
this kind of benchmark functions turns very helpful if the inten- Finally, for comprehensive comparison, newly proposed ChOAs
tion is to evaluate the exploration capability of a MOA and avoid- were tested using some complex, rotated and shifted version of
ing local minima. Table 10 and Fig. 13 show the results for multi- multimodal benchmark functions.
modal benchmark functions (F8-F13). As the results show, ChOAs Fig. 15 and Table 13 show the results obtained using afore-
have also fair exploration capability. In fact, ChOAs always is the mentioned algorithms and benchmark functions. As can be seen,
most efficient algorithm in the all of the benchmark problems. This the convergence rate of ChOAs were significantly better than for
is due to the four different mechanisms of exploration in ChOAs other MOAs, which is even better than results obtained in previ-
leading this algorithm into the global optimum in the early stages ous tests (unimodal, fixed-dimension and multimodal). Because the
and chaotic mechanism guaranteeing to reach the best result in the complexity of these benchmark function is more than other bench-
final stages. marks. Therefore, the ability of ChOAs in these complex problem is
The results of Table 10 indicate that the independent groups in- more evident than other experiment.
creased the performance of ChOA in terms of avoiding local min-
ima. As may be observed in Fig. 13, similar to the results of uni-
modal benchmark functions, the convergence speed of ChOAs is
3.3. Optimization of high-dimensional problems using ChOA
almost better than the other MOAs. The ChOA group set 2 (espe-
cially ChOA21) have the best convergence rates among the ChOAs.
To further confirm the capability of ChOA in working with
The group set 2 has the special updating coefficients so that these
high-dimensional problems, this subsection investigates the 100-
special updating forms give ChOA more randomized search ability
dimensional versions of the unimodal (F1 to F7) and multimodal
in comparison with ChOA group set 1; therefore, the chimps are
(F8 to F13) optimization test functions introduced in the preced-
not easily trapped in local minima. In other words, the reduction
ing subsections. 50 search agents (candidate solutions) are utilized
rate of the independent groups’ coefficient of ChOA2 is less than
to solve these benchmark optimization problems over 20 0 0 iter-
those coefficients of ChOA1. As a result, the updating coefficients
ations. Finally, the results are illustrated in Tables 14 and 15 for
of ChOA2s allows chimps to discover globally the search space. Be-
unimodal and multimodal test functions, respectively.
cause the amplitudes of the searching coefficient decrease gradu-
As results are shown in Table 14, the ChOA outperforms all the
ally after almost three-quarters of the allowed iterations.
other algorithms on five of the unimodal optimization benchmark
In addition, the ChOA21 outperform other ChOA2s through its
functions (F1, F3, F4, F6, and F7). Besides, Table 15 indicates that
special chaotic map (Quadratic map). In this map, a very slight
this algorithm provides the best results on five of the six mul-
change of the input value can lead to significantly various be-
timodal optimization benchmark functions (F8, F9, F10, F12, and
haviour of the map’s amplitude. This particular behaviour of the
F13). For the rest of the unimodal and multimodal optimization
quadratic map causes chimps to explore the search space com-
benchmark functions, the ChOA is ranked as the second-best af-
pletely even in the final stages in such a way that this map
ter LGWO (F2 and F5), and GWO (F11). Poor performances of the
enhances the exploration capability of ChOA21 more than other
majority of algorithms in Tables 14 and 15 show that such high-
ChOAs combined with the other chaotic map.
dimensional optimization benchmark functions can be very chal-
Unlike the multimodal test functions, the fixed-dimension mul-
lenging. These results highly evidence that the ChOA algorithm can
timodal benchmark functions have few local minima. As shown in
be very effective for solving high-dimensional optimization prob-
Tables 11 and 12, the results of all MOAs are similar on six of
lems as well.
the functions. However, the ChOA outperform the other MOAs on
To sum up, the results of this subsection indicates that ChOAs
F20 to F22. ChOA23 has the best results in the almost all of these
propose high exploitation and exploration. First, the proposed
benchmark functions. Fig. 14 shows the convergence rate of the
individual intelligence (autonomous group in initial iteration)
algorithms dealing with fixed-dimension benchmark functions. All
and sexual motivation (chaotic behaviour in final iteration) of
the MOAs have close convergence curves, somewhat better for the
chimps in their group hunting promote exploration, enhance the
ChOAs. The analogy of results and convergence rate is owing to the
ChOA algorithm to avoid local optima stagnation when solving
low dimensional characteristic of these benchmark problems; the
high-dimensional optimization problems. Secondly, the decreasing
effect of independent groups is more apparent for the high dimen-
shape of f for each independent group of chimps, emphasizes ex-
sional problems. To sum up, the results indicate that the indepen-
ploitation as iteration increases, which results in a very precise es-
dent groups and the chaotic maps are profitable for ChOA in terms
timation of the global optimum.
24 M. Khishe and M.R. Mosavi / Expert Systems With Applications 149 (2020) 113338

Table 17
The results of ChOA in comparison with benchmark algorithms for real-world problems.

Problem Optimization Algorithms

ID metric GA GSA PSO BH CS GWO LGWO ChOA

RC01 Ave 2.18E + 02 3.12E + 02 2.23E + 02 3.36E + 02 4.02E + 02 2.12E + 02 2.01E + 02 1.92E + 2
STD 5.34E−02 2.18E−02 2.81E−02 3.21E−02 3.18E−02 2.85E-02 1.67E-03 1.45E-03
RC04 Ave −1.38E-01 −3.86E-01 −1.15E-01 −3.60E-00 −3.86E-00 −3.52E-01 −3.56E-01 −3.87E-01
STD 7.98E−01 1.92E−01 2.84E−01 3.41E−01 3.80E−01 3.96E + 00 2.57E + 00 1.02E−01
RC11 Ave 12.63E + 01 11.93E + 01 11.62E + 01 10.56E + 01 10.13E + 01 10.17E + 01 10.74E + 01 9.99E + 01
STD 2.76E + 01 1.67E + 01 2.21E + 01 1.79E + 01 1.97E + 01 1.54E + 01 2.34E + 01 1.34E + 01
RC14 Ave 9.45E + 04 7.24E + 04 9.82E + 04 9.24E + 04 9.23E + 04 6.12E + 04 5.89E + 04 5.92E + 04
STD 5.84E + 00 5.62E + 02 1.91E + 00 9.68E + 02 8.10E + 00 1.87E + 01 9.59E + 00 1.01E + 00
RC16 Ave 4.59E-02 4.09E-02 4.52E-02 4.58E-02 4.89E-02 4.09E-02 4.98E-02 3.99E-02
STD 2.41E−04 8.58E-02 5.32E−03 2.51E−02 6.98E−03 8.58E-02 3.71E-01 1.58E-02
RC23 Ave 3.18E + 01 3.75E + 01 2.88E + 01 2.85E + 01 2.27E + 01 2.82E + 01 1.98E + 01 2.02E + 01
STD 6.78E + 00 4.73E + 00 4.21E + 00 6.73E + 00 1.21E + 00 6.27E + 00 1.02E-02 1.27E + 00
RC35 Ave 9.78E-02 9.56E-02 9.36E-02 9.74E-02 9.56E-02 9.24E-02 9.81E-02 9.01E-02
STD 8.64E-02 9.47E-02 9.42E-01 9.64E-01 4.73E+00 6.56E-01 6.15E-01 5.42E-01
RC37 Ave 2.85E-02 3.11E-02 3.24E-02 2.94E-02 2.45E-02 3.11E-02 2.92E-02 2.20E-02
STD 1.64E-02 1.24E-02 1.79E-02 2.02E-02 1.40E-02 1.78E-02 1.68E-02 1.11E-02
RC45 Ave 4.38E-02 3.92E-02 4.49E-02 4.75E-02 4.12E-02 4.02E-02 3.89E-02 3.99E-02
STD 1.36E-03 1.07E-03 1.54E-03 1.50E-03 1.81E-02 1.47E-03 1.03E-03 1.53E-03
RC51 Ave 4.59E + 03 4.91E + 03 5.24E + 03 5.29E + 03 4.89E + 03 4.62E + 03 4.59E + 03 4.55E + 03
STD 1.49E + 00 2.94E + 00 2.84E + 00 2.75E + 00 2.21E + 00 2.83E + 00 1.90E + 00 1.09E + 00

3.4. Results and analysis of real-world problems viding chimps into independent groups and allowing them to have
different searching behaviour.
In this section, the effectiveness of ChOA is investigated using Finally, the special decreasing shapes of various f parameter
Ten real-world problems from IEEE CEC2020 (Kumar et al., 2019). evidenced that the ChOA requires chimps to move suddenly in
It’s worth mentioning that these evaluations are carried out ac- the initial steps of the algorithm and locally in the final steps of
cording to the guidelines of CEC2020. Note that the complete de- the algorithm, which leads to a gentle transition and balance be-
scription of these real-world problems is described in CEC2020 tween exploitation and exploration. The ChOA was compared to
(Kumar et al., 2019). However, the general description of the real- nine well-known optimization algorithms in the literature: PSO,
world problems, used in this section, can be obtained from Table GA, GSA, BH, GWO, CS, BBO, ALO, and LGWO. Wilcoxon statisti-
16. Also, the results of these evaluations are shown in Table 17. cal tests were also conducted when comparing other optimiza-
Based on Table 17, it is seen that the conventional LGWO rep- tion algorithms. The results indicated that the ChOA provides very
resents the best performance in two cases Optimal Design of In- competitive results and outperforms other optimization algorithms
dustrial refrigeration System (RC23) and SOPWM for 3-level In- in the majority of benchmark functions. The statistical test also
verters (RC45). The proposed ChOA provides the best results in proved that the results were statistically significant for the ChOA.
the remaining real-world optimization test cases (RC01, RC04, RC11, Therefore, it may be concluded from the comparative results that
RC14, RC16, RC35, RC37, and RC51). Therefore, In comparison with the proposed ChOA is able to be employed as an alternative opti-
GA, GSA, PSO, BH, CS, GWO and LGWO algorithms, the ChOA’s sta- mizer for optimizing various high-dimensional optimization prob-
tistical results indicate that it can be considered as the best op- lems.
timization algorithm in working with real-world optimization test The paper also considered solving ten real-world optimization
problems. The LGWO is also the second best algorithm in dealing problems using the ChOA. The results of the ChOA on these real-
with these test cases. world problems were compared to a wide range of other optimiza-
tion algorithms. The comparative results indicated that the ChOA
is able to solve real-world optimization problems with unknown
4. Conclusion search spaces as well. Other conclusion remarks that can be made
from the results of this study are as follows:
This paper proposed a novel hunting-based optimization algo-
rithm called ChOA. The proposed ChOA mimicked the social di- • Dividing chimps in independent groups guarantees exploration
versity and hunting behaviour of chimps. Four hunting behaviors of the search space, particularly for problems of higher dimen-
(driving, chasing, blocking, and attacking), several operators such sionality.
as diverse intelligence and sexual motivation, and also four kinds • The proposed semi-deterministic feature of chaotic maps em-
of chimps were proposed and mathematically modelled for supply- phasizes the exploitation ability of the ChOA.
ing the ChOA with high exploitation and exploration. The perfor- • The use of chaotic maps assists the ChOA algorithm to resolve
mance of ChOA was benchmarked on 30 mathematical test func- local optima stagnations.
tions, 13 high-dimensional test problems, and 10 real-world opti- • Local optima avoidance is very high since the ChOA algorithm
mization problems in terms of exploration, exploitation, local op- employs a four kind of population of search agents to approxi-
tima avoidance, and convergence rate. As per the superior results mate the global optimum.
of the ChOA on the majority of the unimodal test functions and • The special decreasing shapes of various f parameter promotes
convergence curves, it can be concluded that the proposed algo- exploitation and convergence rate as the iteration counter in-
rithm benefits from convergence rate and high exploitation. The creases.
main reason for the high exploitation and convergence speed is • Chimps memorize search space information over the course of
due to the proposed semi-deterministic feature of chaotic maps. iteration.
High exploration of ChOA can be concluded from the results of • ChOA almost uses memory to keep the best solution acquired
multimodal and composite test functions, which is because of di- so far.
 M. Khishe and M.R. Mosavi / Expert Systems With Applications 149 (2020) 113338 25

• ChOA generally has a few parameters to adjust. Khishe, M., Mosavi, M. R., & Kaveh, M. (2017). Improved migration models of bio-
• Considering the parallel structure of independent groups and geography-based optimization for sonar data set classification using neural net-
work. Applied Acoustic, 118, 15–29.
the simplicity of ChOA, it is very easy to implement the pro- Khishe, M., Mosavi, M. R., & Moridi, A. (2018). Chaotic fractal walk trainer for sonar
posed algorithm. data set classification using multi-layer perceptron neural network and its hard-
• Chimps are not quite similar in terms of ability and intelligence, ware implementation. Applied Acoustics, 137, 121–139.
Khishe, M., & Saffari, A. (2019). Classification of sonar targets using an MLP neu-
but they all perform their tasks as members of a hunting group. ral network trained by dragonfly algorithm. Wireless Personal System, 108(4),
So, each individual’s ability can be useful in a special phase of 2241–2260.
the hunting event. Kirkpatrick, S., Gelatt, C. D., & Vecchi, M. P. (1983). Optimization by simulated an-
nealing. Science (New York, N.Y.) (New York, N.Y.), 13(220), 671–680.
Several research directions can be recommended for future Kumar, A., Wu, G., Ali, M. Z., Mallipeddi, R., Suganthan, P. N., & Das, S. (2019). A
test-suite of non-convex constrained optimization problems from the real-world
studies with the proposed algorithm. Utilizing the ChOA to tackle and some baseline results. Swarm and Evolutionary Computation, 2019, 1–15.
different optimization problems in different industrial tasks. Also, Li, X., Engelbrecht, A., & Epitropakis, M. G. (2013). Benchmark functions for CEC’2013
modifying ChOA to solve multi- and many-objective optimization special session and competition on niching methods for multimodal function
optimization. In Evolutionary Computation and Machine Learning Group Tech-
problems can be investigated as a good contribution. Besides, the
nical Report. RMIT University Available from: Http://goanna.cs.rmit.edu.au/
effectiveness of ChOA can be compared with other hunting-based ∼xiaodong/cec13-niching/competition/ .

optimizers for solving different optimization problems. Another re- Mafarja, M. M., Aljarah, I., Heidari, A. A., Hammouri, A. I., Farisa, H., Al-Zoubia, A. M.,
et al. (2017). Evolutionary population dynamics and grasshopper optimization
search direction is to investigate the effectiveness of other chaotic
approaches for feature selection problems. Knowledge-Based Systems, 145, 25–45.
maps in improving the performance of the ChOA algorithm. Finally, Mafarja, M. M., & Mirjalili, S. A. (2019). Hybrid binary ant lion optimizer with rough
it is possible to design a discrete extension of ChOA. set and approximate entropy reducts for feature selection. Soft Computing, 23,
6249–6265.
References Mirjalili, S. A. (2015). The ant lion optimizer. Advances in Engineering Software, 83,
80–98.
Mirjalili, S. A. (2016). Dragonfly algorithm: A new meta-heuristic optimization tech-
Atashpaz-Gargari, E., & Lucas, C. (2007). Imperialist competitive algorithm: An algo- nique for solving single-objective, discrete, and multi-objective problems. Neural
rithm for optimization inspired by imperialistic competition. In Proceedings of Computing and Application, 27(4), 1053–1073.
the 2007 IEEE congress on evolutionary computation (pp. 4661–4667). Mirjalili, S., Lewis, A., & Sadiq, A. S. (2014). Autonomous particles groups for par-
Basturk, B., & Karaboga, D. (2006). An artificial bee colony (ABC) algorithm for nu- ticle swarm optimization. Arabian Journal for Science and Engineering, 39(6),
meric function optimization. In Proceedings of the 2006 IEEE Swarm Intelligence 4683–4697.
Symposium (pp. 12–14). Mishra, S. (2007). Some new test functions for global optimization and performance
Beyer, H. G., & Schwefel, H. P. (2002). Evolution strategies—A comprehensive intro- of repulsive particle swarm method. MPRA Article, no. 2718, posted 13, Available
duction. Natural Computing, 1(1), 3–52. from: Https://mpra.ub.uni-muenchen.de/2718/
Boesch, C. (2002). Cooperative hunting roles among taï chimpanzees. Human Nature, Molga, M., & Smutnicki,.C. (.2005). Test functions for optimization needs. Available
13, 27–46. from: Http://www.zsd.ict.pwr.wroc.pl/files/docs/functions.pdf
Couzin, I. D., & Laidre, M. E. (2009). Fission-fusion populations. Current Biology, Mosavi, M. R., Khishe, M., Parvizi, G. R., Naseri, M. J., & Ayat, M. (2019). Training
19(15), 633–635. multi-layer perceptron utilizing adaptive best-mass gravitational search algo-
Derrac, J., García, S., Molina, D., & Herrera, F. (2011). A practical tutorial on the use rithm to classify sonar dataset. Archive of Acoustics, 44(1), 137–151.
of nonparametric statistical tests as a methodology for comparing evolutionary Osman, I. H. (1993). Metastrategy simulated annealing and tabu search algorithms
and swarm intelligence algorithms. Swarm and Evolutionary Computation, 1(1), for the vehicle routing problem. Annals of Operations Research, 41(4), 421–451.
3–18. Pijarski, P., & Kacejko, P. (2019). A new metaheuristic optimization method: The
Digalakis, J. G., & Margaritis, K. G. (2001). On benchmarking functions for genetic algorithm of the innovative gunner (AIG). Engineering Optimization, 51(12),
algorithms. International Journal of Computer Mathematics, 77(4), 481–506. 2049–2068.
Dorigo, M., Birattari, M., & Stutzle, T. (2006). Ant colony optimization. IEEE Compu- Rao, R. V., Savsani, V. J., & Vakharia, D. P. (2011). Teaching-learning-based optimiza-
tational Intelligence Magazine, 1(4), 29–39. tion: A novel method for constrained mechanical design optimization problems.
Emary, E. H., Zawbaa, M., & Grosan, C. (2017). Experienced grey wolf optimization Computer-Aided Design, 43(3), 227–330.
through reinforcement learning and neural networks. IEEE Transaction on Neural Rashedi, E., Nezamabadi-Pour, H., & Saryazdi, S. (2009). GSA: A gravitational search
Network Learning System, 99, 1–14. algorithm. Information Science, 13(179), 2232–2248.
Erol, O. K., & Eksin, I. (2006). A new optimization method: Big bang-big crunch. Ravakhah, S., Khishe, M., Aghababaee, M., & Hashemzadeh, E. (2017). Sonar false
Advances in Engineering Software, 37(2), 106–111. alarm rate suppression using classification methods based on interior search
Farisa, H., Mafarja, M. M., Heidari, A. A., Aljarah, I., Al-Zoubia, A. M., Mirjalili, S. A., algorithm. International Journal of Computer Science and Network Security, 17(7),
et al. (2018). An efficient binary salp swarm algorithm with crossover scheme 58–65.
for feature selection problems. Knowledge-Based Systems, 154, 43–67. Roth, G., & Dicke, U. (2005). Evolution of the brain and intelligence. Trends in Cog-
Gandomi, A. H., & Alavi, A. H. (2012). Krill herd: A new bio-inspired optimization nitive Sciences, 9(5), 250–257.
algorithm. Communications in Nonlinear Science and Numerical Simulation, 17(12), Saremi, S., Mirjalili, S., & Lewis, A. (2014). Biogeography-based optimization with
4831–4845. chaos. Neural Computing and Applications, 25(5), 1077–1097.
Garcia, S., Molina, D., Lozano, M., & Herrera, F. (2009). A study on the use of non– Stanford, C. B. (1996). The hunting ecology of wild chimpanzees: Implications for
parametric tests for analysing the evolutionary algorithms’ behaviour: A case the evolutionary ecology of pliocene hominids. American Anthropologist, 98(1),
study on the CEC’2005 special session on real parameter optimization. Journal 96–113.
of Heuristics, 15(6), 617–644. Storn, R., & Price, K. (1997). Differential evolution – A Simple and efficient heuristic
Han, H. G., Lu, W., Hou, Y., & Qiao, J. F. (2016). An adaptive-pso-based self-organiz- for global optimization over continuous spaces. Journal of Global Optimization,
ing rbf neural network. IEEE Transaction on Neural Network Learning System, 99, 11, 341–359.
1–14. Sun, L., Chen, S., Xu, J., & Tian, Y. (2019). Improved monarch butterfly optimization
Heidari, A. A., & Abbaspour, R. A. (2017). Enhanced chaotic grey wolf optimizer for algorithm based on opposition-based learning and random local perturbation.
real-world optimization problems: A comparative study. Handbook of Research Complexity, 2019, 20.
on Emergent Applications of Optimization Algorithms, 693–727. Tomkins, J., & Bergman, J. (2012). Genomic monkey business-estimates of nearly
Heidari, A. A., Farisa, H., Aljarah, I., & Mirjalili, S. A. (2019). An efficient hybrid multi- identical human-chimp dna similarity re-evaluated using omitted data. Journal
layer perceptron neural network with grasshopper optimization. Soft Computing, of Creation, 26(1), 94–100.
23, 1432–7643. Wilcoxon, F. (1945). Individual comparisons by ranking methods. Biometrics Bulletin,
Heidari, A. A., & Pahlavani, P. (2017). An efficient modified grey wolf optimizer with 1(6), 80–83.
levy ´flight for optimization tasks. Applied Soft Computing, 60, 115–134. Wolpert, D. H., & Macready, W. G. (1997). No free lunch theorems for optimization.
Holland, J. H. (1992). Genetic algorithms. Scientific America, 267, 66–72. IEEE Transaction on Evolutionary Computing, 1, 67–82.
Israfil, H., Zehr, S. M., Mootnick, A. R., Ruvolo, M., & Steiper, M. E. (2011). Unresolved Yang, X. S. (2010). A new metaheuristic bat-inspired algorithm. In Proceedings of the
molecular phylogenies of gibbons and siamangs (Family: Hylobatidae) based on 2010 workshop on nature inspired cooperative strategies for optimization (NICSO
mitochondrial, Y-linked, and X-linked loci indicate a rapid miocene radiation or 2010) (pp. 65–74). Springer.
sudden vicariance event. Molecular Phylogenetics and Evolution, 58(3), 447–455. Yang, X. S., & Deb, S. (2009). Cuckoo search via lévy flights. In Proceedings
Khishe, M., & Mohammadi, H. (2019). Sonar target classification using multi-layer of the 2009 IEEE world congress on nature & biologically inspired computing
perceptron trained by salp swarm algorithm. Ocean Engineering, 181, 98–108. (pp. 210–214).
Khishe, M., & Mosavi, M. R. (2019). Improved whale trainer for sonar datasets clas-
sification using neural network. Applied Acoustic, 154, 176–192.
26 M. Khishe and M.R. Mosavi / Expert Systems With Applications 149 (2020) 113338

Mohammad Khishe received his B.Sc. degree from Uni- Mohammad-Reza Mosavi (Corresponding Author) re-
versity of Nowshahr Marine Sciences, Nowshahr, Iran, ceived his B.S., M.S., and Ph.D. degrees in Electronic En-
M.Sc. degree from Islamic Azad University, Qazvin Branch, gineering from Iran University of Science and Technology
and Ph.D. degree from Iran University of Science and (IUST), Tehran, Iran in 1997, 1998, and 2004, respectively.
Technology in 2007, 2011, and 2018, respectively. He is He is currently faculty member (full professor) of the De-
currently faculty member (assistant professor) of the De- partment of Electrical Engineering of IUST. He is the au-
partment of Electrical Engineering of University of Now- thor of more than 350 scientific publications in journals
shahr Marine Sciences. His research interests include neu- and international conferences in addition to 10 academic
ral networks, meta-heuristic algorithms and digital de- books. His research interests include circuits and systems
sign. design. He is also editor-in-chief of “Iranian Journal of
Marine Technology” and editorial board member of “Ira-
nian Journal of Electrical and Electronic Engineering”.