Received May 14, 2020, accepted May 25, 2020, date of publication May 28, 2020, date of current

version June 10, 2020.

Digital Object Identifier 10.1109/ACCESS.2020.2998324

An Improved Squirrel Search Algorithm With

Reproductive Behavior
XUNCAI ZHANG 1 , (Member, IEEE), KAI ZHAO1 , LINGFEI WANG1 ,
YANFENG WANG1 , AND YING NIU 2
1 College of Electrical and Information Engineering, Zhengzhou University of Light Industry, Zhengzhou 450002, China
2 College of Architecture Environment Engineering, Zhengzhou University of Light Industry, Zhengzhou 450002, China
Corresponding author: Xuncai Zhang ([email protected])
This work was supported in part by the National Natural Science Foundation of China under Grant 61572446, Grant 61602424, and Grant
U1804262, and in part by the Key Research and Development Program of Henan Province under Grant 202102210177 and Grant
192102210134.

ABSTRACT The squirrel search algorithm (SSA) is a recently proposed nature-inspired algorithm based on
the dynamic foraging and gliding behavior of squirrels. Because of its simplicity and stability, the squirrel
algorithm has attracted increasing research interest. However, the lack of exploration ability of the SSA
may lead to premature convergence to the local optimum. To overcome this disadvantage, an improved
SSA with reproductive behavior (RSSA) is proposed to solve the numerical optimization problem. First,
the reproductive behavior of the invasive weed algorithm (IWO) is introduced to the conventional SSA to
generate offspring individuals, and these offspring individuals are scattered into the search space by Gaussian
distribution to complete the location update. This method makes it possible for individuals with poor fitness
to enter the next generation search, improving the exploration ability of the SSA. Second, an adaptive step
strategy is proposed to adaptively adjust the search step of squirrels according to the distance between each
squirrel and other family members. This strategy effectively balances the exploration and exploitation of the
algorithm. Finally, the performance of the proposed RSSA algorithm is evaluated using Wilcoxon’s test on
unimodal, multimodal, fixed-dimensional multimodal and CEC 2014 benchmark functions. Experimental
results and statistical tests show that RSSA has better performance in terms of convergence, accuracy, and
search capability compared with other state-of-the-art algorithms.

INDEX TERMS Nature-inspired, squirrel search algorithm, exploration and exploitation, Wilcoxon’s test,
CEC 2014 benchmark functions.

I. INTRODUCTION Swarm intelligence imitates the social behavior of animal

In the field of computer science, optimization problems pose groups. Because of its simplicity, flexibility, non-derivation
the challenge of finding the optimal solution among all fea- mechanism, and avoidance of local optima, SI is widely
sible solutions. Traditional methods to solve such problems applied in feature selection [5], hardware/software co-design
include the conjugate gradient method, branch and bound [6], scheduling [7], agriculture [8], metallurgy [9], and mil-
algorithm, and dynamic programming method [1], [2]. How- itary [10]. The most representative SI algorithms are parti-
ever, with the increasing scale of optimization problems, it is cle swarm optimization (PSO) [11], ant colony optimization
becoming increasingly difficult for traditional optimization (ACO) [12], and artificial bee colony (ABC) [13]. The social
algorithms to meet the requirements of time complexity. As behavior of birds inspires PSO. It uses individual cognitive
a result, the need for more efficient algorithms is increasing. items to guide the birds towards the best position that has been
Over the past few decades, swarm intelligence (SI) has shown found. Additionally, social cognitive items are used to guide
efficient and robust performance in solving modern nonlinear all birds towards the global optimum. Because of its simplic-
numerical global optimization problems [3], [4]. ity and stability, PSO is continuously improved to solve more
complicated optimization problems [14]–[16]. Ant colony
The associate editor coordinating the review of this manuscript and optimization (ACO) is inspired by the social behavior of
approving it for publication was Muhammad Zakarya . ant colonies [17]. The pheromone matrix of the ant colony

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
101118 VOLUME 8, 2020
X. Zhang et al.: Improved SSA With Reproductive Behavior

TABLE 1. Recent techniques to improve the SSA.

evolves continuously during the iteration of the candidate structural design [30], image analysis [31], and production
solution and eventually finds the optimal solution. ACO has scheduling [32], [33].
the advantages of simple structure, ease of use, and excel- In the conventional SSA, however, the exploration abil-
lent performance. However, its convergence rate is slow in ity of SSA is regarded as weak, and it has had problems
complex optimization problems. Recently, researchers com- with premature convergence [31]. Moreover, the convergence
bined the high-speed parallel computing ability of GPUs velocity of SSA is too low, and its performance decreases
to improve the convergence speed of the ABC algorithm as the search space dimensionality increases. Therefore, a
and achieved excellent results [18]–[20]. The artificial bee considerable number of research studies have been conducted
colony algorithm simulates the intelligent foraging behavior to improve the global performance of conventional SSA. For
of a honeybee colony [21]. Owing to its outstanding global example, Wang and Du proposed an improved squirrel search
search capability, ABC has been successfully utilized in many algorithm (ISSA) [34], which utilizes the jumping search
applications, i.e., parameter estimation [22] and wireless sen- method and the progressive search method to strengthen
sor networks [23]. It has been proven that there is no algo- the diversity of global searching. The experimental results
rithm generic enough to solve all optimization problems [24]. on 22 benchmark functions demonstrated that ISSA could
Therefore, many new SI algorithms emerge every year, such improve the performance of the squirrel search algorithm.
as the grey wolf optimization (GWO) [25], firefly algorithm Walaa et al. proposed the combination of the best-fit heuristic
(FF) [26], cuckoo search (CS) [27], dragonfly algorithm (DA) (BF) [35] and squirrel swarm algorithm to improve the global
[28], and squirrel search algorithm (SSA) [29]. search ability of the algorithm named (MSBPP) [36]. In
The squirrel search algorithm (SSA) is a new swarm MSBPP, the best-fit heuristic is utilized to generate a set of
intelligent optimization algorithm initially developed by Jain initial solutions. Besides, the MSBPP introduced an operator
in 2018 [29]. It is inspired by the dynamic foraging and strategy to solve the one-dimensional bin-pacing problem.
gliding behavior of squirrels. SSA searches the global opti- However, it neglects the running time of the algorithm and
mum by gliding among different kinds of trees (normal tree, has a weak effect in high-dimension problems. To increase
acorn tree, and hickory tree), searching for food sources, population diversity, Wang et al. proposed an improved squir-
and avoiding predators. The performance of SSA is tested rel search algorithm (ISSA) with spatial variation and diffu-
on classical and modern CEC 2014 benchmark functions. sion behavior to update the positions of squirrels [30]. The
The results demonstrate the effectiveness of SSA in com- results indicated that the ISSA outperformed conventional
parison with some state-of-the-art optimization algorithms, SSA. In addition to the above improvement measures, other
such as PSO, ABC, bat algorithm (BA), and firefly algo- recent techniques to improve the SSA and their advantages,
rithm (FF). Since the development of SSA, it has been suc- disadvantages, and applications are listed in TABLE 1. How-
cessfully applied to various optimization problems such as ever, their performance is still deficient in solving problems

VOLUME 8, 2020 101119

 X. Zhang et al.: Improved SSA With Reproductive Behavior

with difficult mathematical properties, and the abilities of where FSi is the ith squirrel, rand (1, d) is a 1×d random
exploration and exploitation are not adequately balanced, matrix in the range [0,1], and d is the dimension of the
resulting in poor robustness of the algorithm. To overcome problem.
these problems, we improve a squirrel search algorithm with
reproductive behavior (RSSA) is proposed in this paper. The B. CLASSIFY THE POPULATION
main contributions of this paper are summarized as follows: The positions of the squirrels are evaluated by the fitness
1) A reproduction mechanism is proposed to increase the function f . The decision variable (position vector of the squir-
number of candidate solutions in each iteration to improve the rels) is input into the fitness function, and the corresponding
population diversity of the SSA. This mechanism improves result f (FSi ) = f (FSi1 , FSi2 , . . . , FSid ), (i = 1, 2, . . . , n)
the exploratory ability of the algorithm to avoid local opti- is the fitness value of the ith squirrel. The fitness value
mization. represents the quality of the food source sought by the ith
2) The position updating formulas are combined with the squirrel. Then, the fitness values of all the squirrels are sorted
adaptive step size strategy. In this way, the exploitation and in ascending order. The squirrel in the hickory tree (FSh )
exploration abilities of the algorithm can be properly bal- represents the individual with the minimum fitness value. The
anced. It also avoids the complex parameter adjustment of squirrels in the acorn trees (FSa ) represent the individuals
the algorithm with fitness values ranging from 2 to Nfs + 1. The remaining
The remainder of the paper is organized as follows. individuals are denoted as squirrels in the normal tress (FSn ).
Section II briefly recapitulates the basic theory of the squirrel
search algorithm. Next, the proposed RSSA with a reproduc- C. UPDATE THE POSITIONS
tion mechanism and adaptive step size strategy is presented The positions of individuals are updated by gliding to trees
in detail in Section III. Experimental results and analysis of different species. The n1 squirrels that are in normal trees
are illustrated in Section IV. Finally, Section V gives the may move towards the acorn trees, whereas the remaining n2
conclusion. squirrels may move towards the hickory tree. The n3 squirrels
that are in acorn trees may move towards the hickory tree. In
II. AN OVERVIEW OF THE SQUIRREL SEARCH
these cases, the new positions of the squirrels can be obtained
ALGORITHM
as follows:
The squirrel search algorithm imitates the dynamic foraging (
FS ti + dg ×GC × FS th − FS ti


behavior of squirrels and their efficient method of locomo- t+1 R ≥ Pdp
FS i = (2)
tion, known as gliding. It is a new swarm intelligence algo- random location otherwise
rithm proposed by Jain et al. in 2018 [29]. A squirrel is a (
FS ti + dg ×GC × FS ta − FS ti


kind of arboreal rodent that moves mainly by gliding. Flying t+1 R ≥ Pdp
FS i = (3)
squirrels are an arboreal and nocturnal type of rodent that random location otherwise
are exceptionally adapted for gliding locomotion. Squirrel
(
FS ta + dg ×GC × FS th − FS ta R ≥ Pdp


t+1
gliding is considered the most complex form of aerodynamic FS a = (4)
random location otherwise
motion and the most energy efficient [39]. Its dynamic for-
aging behavior also makes the most efficient use of food where t denotes the current iteration, R is a random number
resources [40]. In the SSA, there are four necessary assump- in the range of [0, 1], and Pdp denotes the predator presence
tions: probability. If R ≥ Pdp , the squirrels are safe and glide
(1) In a deciduous forest, there are n squirrels and n trees, through the forest in search of food. If R < Pdp , the squirrels
one squirrel in one tree. are at risk of predation and are forced to use a random walk
(2) These n trees contain a hickory tree and Nfs (1 < Nfs < (described in Section II (D)) to find a nearby hiding place. GC
n) acorn trees, and the rest are normal trees. denotes the gliding constant. In this work, the value of GC is
(3) There are only three types of trees in the forest. The considered 1.9. The dg is the random gliding distance, which
hickory tree has the best food source (hickory nuts), acorn is computed as:
trees have a general food source (acorn nuts) and normal trees hg
have no food. dg = (5)
tan (ϕ) × sf
(4) Each squirrel seeks food individually and utilizes
the available food resources by showing dynamic foraging where hg is the constant value 8; sf is the constant value
behavior. 18; and tan(ϕ) represents the gliding angle, which can be
calculated by formula (6):
A. INITIALIZE THE POPULATION D
There are n squirrels in a forest. FSU and FSL are the lower tan (ϕ) = (6)
L
and upper bounds in the search space. According to formula where D is the drag force and L is the lift force, which can be
(1), n individual squirrels are randomly generated: calculated by formulas (7) and (8):
FS i = FS L +rand(1, d)×(FS U −FS L ), (i = 1, 2,. . ., n) 1
D= (7)
(1) 2ρV 2 SCD

101120 VOLUME 8, 2020

X. Zhang et al.: Improved SSA With Reproductive Behavior

1
L = (8)
2ρV 2 SCL

where ρ (=1.204 kgm−3 ) is the air density, V (=5.25 ms−1 ) is

the gliding speed of the squirrel, S (=154 cm2 ) is the surface
area of the body, CD (= 0.6) is the frictional drag coefficient,
and CL represents a random number between 0.675 and 1.5.

D. SEASONAL TRANSITION JUDGMENT

The seasonal monitoring condition introduced in SSA
helps the algorithm to jump out of the local optimum [29].
At the beginning of each iteration, the SSA requires the
entire population to be in the winter state, which means that
all individuals are updated as described in Section II (C).
When all individuals are updated, the change in seasons is
determined according to formulas (9) and (10):
r FIGURE 1. The procedure of the conventional SSA.
Xd 2
t
Sc = (FS th,k − FS taj,k ) j = 1, 2, . . . , Nfs (9)
k=1
−6
10e A. THE REPRODUCTION MECHANISM OF SQUIRRELS
Smin = (10) In the original SSA, the optimal and suboptimal solutions
(365)t/(T /2.5)
of the current generation guide the search agent to update
where T is the maximum number of iterations and t denotes the location. However, ignoring the possibility that there are
the current number of iterations. When the seasonal constant still good individuals in the vicinity of individuals with poor
is less than the minimum (Sct < Smin ), the season changes fitness results in poor global search capabilities. Inspired
from winter to summer. When the season changes, all indi- by the IWO algorithm [41], the reproductive behavior of
viduals gliding to Fh stay at the updated positions, and all weeds is introduced into the SSA to guide the generation of
individuals gliding to Fa without encountering a predator candidate solutions.
relocate their positions according to formula (11). First, according to the reproductive ability of squirrels, it is
determined that squirrels will breed and produce offspring in
FS new
i = FS L + Lévy (n) × (FS U − FS L ) (11) better living areas. For example, the squirrels in hickory and
acorn trees have abundant food sources and are apt to survive
The Lévy distribution can be calculated by formula (12): and reproduce. The number of reproductions is determined
ra × ξ according to the fitness value. The number of offspring pro-
Lévy (x) = 0.01 × 1
(12) duced varies linearly with the fitness of the father squirrel.
|r b | β The formula is as follows:
f (FS i ) − fmin
where ra and rb are normally distributed random numbers Ns = (Smax − n) + n (i = 1, 2, . . . , n) (14)
in [0,1], β represents the constant value 0.5, and ξ can be fmax − fmin
calculated by formula (13): where f (FSi ) is the fitness value of the parent squirrel,
fmax and fmin are the maximum and minimum fitness of the
1
0 (1 + β) × sin( πβ
2 )
β population, and Smax is the maximum population.
ξ =(   β−1
) (13) Then, the numbers of generated offspring in RSSA are
0 1+β
2 ×β ×2 2 normally distributed random numbers with a mean equal to 0
and standard deviation FS Iterσ . The standard deviation of the
where 0 (x) = (x − 1). random numbers is reduced from σmax to σmin during each
The procedure of the conventional SSA is shown in iteration. The standard deviation of the population distribu-
FIGURE 1. tion FS Iter
σ is defined as:
(Iter max − Iter)h
III. THE IMPROVED SQUIRREL SEARCH ALGORITHM FS Iter
σ = (σmax − σmin ) + σmin (15)
In this section, the reproduction mechanism and adaptive Iter hmax
step size strategy are introduced to improve the conventional where σmax is the maximum standard deviation, σf is the
SSA. First, the reproduction mechanism is combined with minimum standard deviation, Iter is the maximum number of
the foraging process of squirrels to increase the diversity of iterations, and h is the nonlinear harmonic exponent. In the
the population. Furthermore, the adaptive step size strategy present work, the value of h is set to 2.
is introduced to update the position of individuals to balance It is worth noting that the squirrels with poor fitness values
the exploration and exploitation abilities. still have a chance to generate new candidate solutions, and

VOLUME 8, 2020 101121

 X. Zhang et al.: Improved SSA With Reproductive Behavior

Algorithm 1 Pseudocode of the Squirrel Reproduction where r0 is the step size factor at the distance r = 0. It has
Mechanism been mentioned above that the individuals on hickory trees,
1: Determine the initial swarm size N and number of acorn trees and normal trees represent the optimal solution,
iterations T ; suboptimal solution, and general candidate solution. In gen-
2: Randomly generate the initial population of squirrels: eral, individuals near the optimal solution should realize their
xi = FSi ; exploration ability with a smaller step size. On the other hand,
3: Initialize σ0 , σf , N0 ; individuals who are far away from the optimal or suboptimal
4: Evaluate the fitness value of each squirrel; solution should have a large step size to explore new space.
5: Determine the number of offspring squirrels Ns using The calculation formula of step size effectively balances these
Eq.(14); two cases.
6: for i = 1 to N do After the step size of all squirrels are calculated, the adap-
7: for j = 1 to Ns do tive step size operator is introduced into the position updating
8: X (i, j) = xi + randn (xi , FS Iter
σ ); process of squirrels. The new position updating formulas are
9: end for as follows:
10: if X is beyond the boundary then (
FS tn + dg GC FS th − FS tn × Stepthn R ≥ Pdp


11: Map X to the feasible region; t+1
FS n =
12: end if random location otherwise
13: end for (20)
(
t t t t


FS n + dg GC FS a − FS n × Stepan R ≥ Pdp
FS t+1
n =
one of these candidate solutions is likely to be the global random location otherwise
optimum. Therefore, the reproduction mechanism eliminates (21)
(
t t t t


the limitation of finding food only at the optimal position, FS a + dg GC FS h − FS a × Stepha R ≥ Pdp
which helps improve the exploration ability of the algorithm. FS t+1
a =
random‘location otherwise
The pseudocode of the proposed reproduction mechanism
(22)
is shown in Algorithm 1.
where the symbol description is consistent with that of for-
B. ADAPTIVE STEP SIZE STRATEGY mulas (2), (3), and (4). In addition, FIGURE 2 shows how
In the literature [29], the exploration and exploitation abilities squirrels in different positions update their position according
are balanced by adjusting the sliding constant GC . The larger the adaptive step size strategy in a two-dimensional search
the GC value, the larger the step size of the squirrel move- space. From the figure, the final position of a squirrel would
ment, and the stronger the exploration ability. For different be in a random location within a circle, which is calculated
optimization problems, however, the value of GC needs to be according to the distance from a squirrel to a food source. In
continuously adjusted to achieve the balance of search capa- other words, these squirrels on hickory nut trees, acorn trees
bility. This complex parameter adjustment is not conducive and normal trees update their position adaptively according
to solving complex optimization problems. In our study, an to the quality of food sources.
adaptive step size mechanism was introduced. The search step As shown in formulas (20), (21) and (22), in the early
of the search agent is adaptively adjusted according to the stage of the algorithm, because the relative distance between
distance between the search agent and other individuals. individuals is large and the search step size is large, the algo-
First, the distance between the current position of a squirrel rithm focuses on exploration. As the iteration progresses, the
and its target position needs to be calculated. The formulas for relative distance between individual’s decreases, the search
calculating the distance are as follows: step size also decreases, and the algorithm adaptively changes
r
XD to the development stage. In other words, it can change the
rhn = FS h − FS n = (FS h,d − FS n,d )2 (16) search capability adaptively according to the current location
d=1
r information. Therefore, the adaptive step size strategy can
XD effectively balance exploration and exploitation abilities.
ran = FS a − FS n = (FS a,d − FS n,d )2 (17)
d=1
r
XD C. ALGORITHM FLOW
rha = FS h − FS a || = (FS h,d − FS a,d )2 (18) To explain the proposed RSSA more clearly, the pseudocode
d=1
where d = 1, 2, . . . , D; D denotes the dimension; and FSh , of the algorithm is shown in Algorithm 2.
FSa , and FSn are the optimal individual, suboptimal individ- First, the algorithm initializes the positions of the indi-
uals, and general candidate solution, respectively. viduals FSi , initial population size N0 , population size N ,
Then, the adaptive step sizes are generated according to the maximum number of iterations T and other parameters (Line
distance between the current position and the target position. 1). Then, the fitness of each squirrel is evaluated and the
locations of squirrels is sorted in ascending order depending
2 2 2
Stephn = r0 e−rhi , Stepan = r0 e−ran , Stepha = r0 e−rha (19) on their fitness values. The algorithm then determines the

101122 VOLUME 8, 2020

X. Zhang et al.: Improved SSA With Reproductive Behavior

Algorithm 2 Pseudocode of RSSA Algorithm

01: Initialize T , N0 , N , σ0 , σf , h, Itermax and the population FSi with n squirrels using Eq.(1)
02: Evaluate the fitness of squirrels and then determine their locations: FSh , FSa , FSn.
03: whileIter < Itermax
04: for t1 = 1 to n1 do
05: if R ≥ Pdp
FS t+1 t t t t


06: n = FS n + dg GC FS h − FS n ×Stephn % adaptive step size factor
07: else
08: FS t+1
n = rand(1, D) × (U − L) + L
09: end
10: Determine the number of offspring squirrels (NS1 ) and FS Iter σ on normal trees according to Eq.(14) and Eq.(15);
11: for i1 = 1 to Ns1 do
12: Distribute the position of offspring according to Eq.(15)
13: FS new
i = FSi + randn (FS i , FS Iter σ )
14: end
15: end
16: for t2 = 1 to n2 do
17: if R ≥ Pdp
18:
t
FS t+1 = FS n + dg GC FS ta − FS tn × Steptan % adaptive step size factor


n
19: else
20:
FS t+1
n = rand(1, D) × (U − L) + L
21: end
22: Determine the number of offspring squirrels (NS2 ) and FS Iter
σ on normal trees according to Eq.(14) and Eq.(15);
23: for i2 = 1 to NS2 do
24: Distribute the position of offspring according to Eq.(15)
25: FS new
i = FSi + randn (FS i , FS Iter
σ )
26: end
27: end
28: for t3 = 1 to n3 do
29: if R ≥ Pdp
30:
t
FS t+1 = FS a + dg GC FS th − FS ta × Steptha % adaptive step size factor


a
31: else
32:
FS t+1
a = rand(1, D) × (U − L) + L
33: end
34: Determine the number of offspring squirrels (NS3 ) and FS Iter
σ on acorn trees according to Eq.(14) and Eq.(15);
35: for i3 = 1 to NS3 do
36: Distribute the position of offspring according to Eq.(15)
37: FS new
i = FSi + randn (FS i , FS Iter
σ )
38: end
39: end
40: if Sct < Smin % Calculate seasonal constant(Sct ) using Eq.(9)
41:
FS new
i = FS L + Lévy (n) × (FS U − FS L )
42: end
43:
finew = f ([FS new t t t
i FS h FS a FS n ])% Evaluate the fitness values of offspring and parent individual
44: Sort fitness values and then determine the locations of squirrels: FS new new new
h , FS a , FS n
45: if (The population size exceeds the upper bound N )
46: finew = finew (1: N );
47: end
48: The location of squirrel on hickory nut tree (FS new
h ) is the final optimal solution
49: Iter = Iter +1
50: End

VOLUME 8, 2020 101123

 X. Zhang et al.: Improved SSA With Reproductive Behavior

locations of squirrels on hickory nut trees (FSh ), acorn trees TABLE 2. Description of the unimodal benchmark functions.
(FSa ), and normal trees (FSn ) (Line 2). Then, the squirrel are
randomly selected so that n1 squirrels in normal trees to move
towards the hickory nut tree and n2 squirrels move towards
acorn trees. The remaining n3 squirrels, which are in acorn
trees, will move towards the hickory nut tree. The locations
of the squirrels are updated according to Eq. (20), Eq. (21) and
Eq. (22) with the proposed adaptive step size strategy (Lines
4-9, 16-21 and 28-33). Then, the offspring of the squirrels
are generated according to the proposed reproduction mech-
anism in Section III (A) and the positions of the offspring
are distributed in the surrounding search space (Lines 10-14.
Lines 22-26 and Lines 34-38). Then, the algorithm calculates
the seasonal constant and the minimum seasonal constant
and sets the constraint conditions for seasonal change. If the
conditions are met, the positions of squirrels are reset with
formula (11) (Lines 40-42). Then it is determined whether
the population size exceeds the upper bound. The individuals
with better fitness values are selected as the initial population
of the next iteration (Line 45-47). Finally, the optimal solution
(Line 48) is returned and the above steps are repeated until the
stopping criterion is satisfied.

IV. EXPERIMENTAL RESULTS AND ANALYSIS

In this section, the effectiveness of the RSSA in solving
numerical optimization problems is verified using a series of
experimental tests and statistical analysis. Twenty-two bench-
mark functions are provided in Section IV (A). The para-
metric settings of the proposed RSSA and other state of the
art algorithms are given in Section IV (B). The experimental
results are provided in Section IV (C), and the statistical anal-
ysis of the results on the benchmark functions is discussed in
Section IV (D).
FIGURE 2. The search method of RSSA with an adaptive step size strategy.
A. BENCHMARK FUNCTION
Twenty-two benchmark functions of different categories are
selected to test the performance of the proposed RSSA exploitation of the algorithm. TABLE 4 shows these CEC
including unimodal functions [42], multimodal functions 2014 functions.
[43], fixed-dimensional multimodal functions [44] and the
composition functions in CEC 2014 benchmark functions B. PARAMETER SETTING
[45]. The unimodal functions contain only one optimal solu- The proposed RSSA is compared with five other nature-
tion and are easy to solve. Therefore, they are often used to inspired algorithms, including the improved squirrel search
test the exploitation ability of algorithms. These unimodal algorithm (ISSA) [37], firefly algorithm (FF) [46], invasive
functions are shown in TABLE 2. In addition, multimodal weed optimization (IWO) [41], artificial bee colony (ABC)
functions are a class of functions that contain multiple local [47] and the conventional SSA [29]. To more clearly show
optimal solutions. Moreover, the number of local optimal the parameters used in this study, Table 5 lists the parameter
solutions increases exponentially with increasing dimension- symbols, definitions, and parameter values of all algorithms.
ality. These multimodal functions are often used to test In this table, the ‘−’ symbol indicates that the algorithm has
the exploration ability of algorithms. The fixed-dimensional no corresponding parameters. The maximum iterations (T )
multimodal functions are defined on a specific dimensional and population size (N ) (common parameters of the algo-
search space and have multiple local optimal solutions. They rithms) are set to 1000 and 50 in the first two experiments.
are necessary to evaluate the exploration ability of algorithms. Considering the complexity of the CEC 2014 functions, we
The multimodal functions and fixed-dimensional multimodal uniformly set the maximum iterations (T ) to 5000 in the third
functions used in this experiment are shown in TABLE 3. In experiment to meet the number of function evaluations. The
addition, eight CEC 2014 functions are selected to evaluate special parameters of the FF, IWO, ABC, SSA, and ISSA
the comprehensive balance ability of the exploration and algorithms are consistent with the literature [29], [37], [41],

101124 VOLUME 8, 2020

X. Zhang et al.: Improved SSA With Reproductive Behavior

TABLE 3. Description of the multimodal and fixed-dimension multimodal benchmark functions.

TABLE 4. Description of the CEC 2014 benchmark functions.

[46], [47]. The parameters of the proposed RSSA are set to not excellent in all unimodal functions. For example, it has a
Nfs = 4, GC = 1.9, Pdp = 0.1, N0 = 0.5, σ0 = 10, and worse performance on F6 in terms of the mean. In addition,
σf = 0.5. For each benchmark function, all algorithms run RSSA has the lowest standard deviation in most functions,
30 times independently from different randomly generated which shows that RSSA has good stability. To reflect the
populations. To compare the proposed RSSA fairly with other performance of RSSA on the unimodal benchmark functions
algorithms, all the algorithms are implemented in MATLAB more clearly, a convergence rate analysis is also carried out.
2016a. All computations were run on a CPU: Intel Core The convergence curves of several representative unimodal
i5-4200 M, 2.5 GHz, 4 G RAM, and Windows 7 (64-bit) functions are shown in FIGURE 3, which reflects that the
operating system. convergence rate of the RSSA is faster than that of other
algorithms on most benchmark functions. Therefore, RSSA
C. EXPERIMENTAL RESULTS has excellent convergence performance for the unimodal
1) TEST RESULTS AND ANALYSIS ON UNIMODAL
functions.
FUNCTIONS
Unimodal functions (TABLE 2) are often used to evaluate 2) TEST RESULTS AND ANALYSIS ON MULTIMODAL
the exploitation ability of algorithms. The test results for FUNCTIONS AND FIXED-DIMENSION MULTIMODAL
unimodal functions are shown in TABLE 6, including the FUNCTIONS
mean, standard deviation (Std), and best and worst fitness In this experimental test, the more complex multimodal func-
values of each algorithm after 30 runs. The mean of RSSA tions and fixed-dimension multimodal functions (TABLE 3)
is the best in most unimodal functions, such as F1 , F2 , F3 , F4 are used to evaluate the exploration ability of RSSA. In addi-
and F5 functions. However, the performance of the RSSA is tion, the corresponding statistical results of 30 independent

VOLUME 8, 2020 101125

 X. Zhang et al.: Improved SSA With Reproductive Behavior

TABLE 5. Parameter settings of the algorithms.

TABLE 6. Statistical results obtained by RSSA, ISSA, ABC, IWO, FF and SSA through 30 independent runs on functions F1 -F6 with unimodal benchmark
functions.

101126 VOLUME 8, 2020

X. Zhang et al.: Improved SSA With Reproductive Behavior

TABLE 7. Statistical results obtained by RSSA, ISSA, ABC, IWO, FF and SSA through 30 independent runs on functions F7 -F14 with multimodal and
fixed-dimension multimodal benchmark functions.

FIGURE 3. Comparison of the convergence curves of RSSA and the other algorithms on functions F1 -F6 with unimodal benchmark functions.

VOLUME 8, 2020 101127

 X. Zhang et al.: Improved SSA With Reproductive Behavior

TABLE 8. Statistical results obtained by RSSA, ISSA, ABC, IWO, FF and SSA through 30 independent runs on functions F15 -F22 with the CEC 2014
benchmark functions.

runs of RSSA and its competitors are recorded in TABLE 7. from the CEC 2014 functions, which are specially designed
These results show that the proposed RSSA has a better mean with complicated features. Therefore, this test is more chal-
and standard deviation on F7 , F9 , F10 , F11 , and F12 . It can lenging for each algorithm than the previous two experimen-
also be observed that the performance of the RSSA is similar tal studies. As seen from the test results (TABLE 8), all the
to that of the SSA and ABC algorithms on fixed-dimension algorithms failed to reach the optimal solution. However,
multimodal functions (F13 and F14 ), which is caused by RSSA achieved more acceptable results on F15 , F16 , F17 , F18 ,
the simplicity of such optimization problems. Moreover, the F20 and F22 than the other algorithms. The mean of ISSA is
convergence curves of these six algorithms on some represen- better than other algorithms on F21 . In terms of convergence
tative benchmark functions are shown in FIGURE 4, from performance, except for F16 , RSSA has the best convergence
which it can be seen that the RSSA has a better conver- performance among all the algorithms. TABLE 9 records the
gence rate and sufficient accuracy for complex multimodal CPU time of each algorithm after completing all iterations.
functions and fixed-dimensional multimodal functions. This It can be observed that the computational complexity of
experiment reflects that RSSA provides better exploration RSSA is slightly lower than that of SSA, but they are all
ability than its competitors. This is due to the enhancement higher than the other four algorithms. This is because RSSA
of the global search ability by the proposed reproduction introduces the reproduction mechanism, which increases the
mechanism. The reproduction of squirrels in the iterative computational complexity. Furthermore, the results of the
process increases the diversity of the population. Therefore, convergence analysis in FIGURE 5 show that compared with
the RSSA has greater potential to avoid local optima. the other five algorithms, RSSA still has the best conver-
gence performance. Through these comparisons, RSSA is
3) TEST RESULTS AND ANALYSIS ON THE CEC 2014 still proven to be the best algorithm in solving such complex
COMPOSITE FUNCTIONS optimization problems. The reproduction mechanism and
The purpose of this test is to evaluate the comprehensive per- adaptive step size strategy of squirrels effectively balance the
formance of the algorithms. The benchmark functions used global search and local search of the RSSA. Therefore, the
are eight high-dimensional composite functions (TABLE 4) RSSA shows stability in such complex benchmar functions.

101128 VOLUME 8, 2020

X. Zhang et al.: Improved SSA With Reproductive Behavior

TABLE 9. The CPU time required by each algorithm.

TABLE 10. Results of Wilcoxon’s test for RSSA against the other five algorithms for the 22 functions.

4) STATISTICAL ANALYSIS test [49], is applied to the statistics of the experimental results
Nonparametric statistical technology should be performed to prove the significance performance of RSSA. TABLE 10
to show the significant performance of the optimization records the P values obtained by the Wilcoxon’s test. From
algorithm. Further evidence on this can be found in an earlier the table, almost all P values are under the 95% confidence
monograph by Derrac et al. [48]. In this paper, a famous interval (α = 0.05), which proves that RSSA has significant
nonparametric statistical test method, Wilcoxon’s rank-sum difference from other algorithms. In this table, the ‘+’ and

VOLUME 8, 2020 101129

 X. Zhang et al.: Improved SSA With Reproductive Behavior

FIGURE 4. Comparison of the convergence curves of RSSA and the other algorithms on functions F7 -F14 with multimodal and fixed-dimension
multimodal benchmark functions.

‘−’ signs indicate that the reference algorithm has better TABLE 11. Average ranking of the six algorithms using MAE for 22
benchmark functions.
or worse performance than the compared algorithm, respec-
tively. The last line represents the number of ‘+’ and ‘−’
signs, which also corresponds to the number of P values
under the 95% confidence interval (α = 0.05). From the table,
RSSA has more ‘+’ signs than the other algorithms. As a
result, the RSSA has a better performance than the other five
algorithms.
Finally, to quantitatively analyze the performance of the
algorithm, the mean absolute error (MAE) values of all
algorithms are calculated and sorted. The MAE is a valid
statistical method that shows how far the result differs from
the actual value. The MAE formula is as follows:
PN
|mi − ki |
MAE = i=1 (23)
N
can be clearly observed that RSSA provides the best per-
where mi is the optimal value of an algorithm, ki repre- formance compared to its competitors. In addition, FIG-
sents the corresponding actual value of the benchmark func- URE 6 reflects the comparison results of each algorithm
tion, and N is the number of samples. In this work, N is run 660 times (30 runs for each benchmark function).
the number of benchmark functions and the MAE values The RSSA reached the optimal solution 421 times out of
and ranking of all algorithms are listed in TABLE 11. It 660 runs.

101130 VOLUME 8, 2020

X. Zhang et al.: Improved SSA With Reproductive Behavior

FIGURE 5. Comparison of the convergence curves of RSSA and the other algorithms on functions F15 -F22 with the CEC 2014 benchmark functions.

adjustment of the algorithm. The performance of RSSA is

evaluated on 22 benchmark functions including unimodal,
multimodal, fixed-dimensional multimodal, and CEC 2014
benchmark functions. Finally, a comprehensive evaluation
of RSSA is carried out by combining Wilcoxon’s statistics
and MAE analysis. The experimental results and statistical
analysis show that the accuracy, robustness and local optimal
value avoidance of RSSA are superior to those of SSA and
other advanced algorithms. In future work, the RSSA can be
developed for solving more complex problems such as the
multi-objective optimization, constrained optimization, and
FIGURE 6. Comparison of the algorithms to obtain the number of global
NP hard problems.
optimal solutions in 660 runs. The main improvement of the proposed RSSA is in the
global search ability. However, the improvement in the local
V. CONCLUSION search needs further research. Other initialization methods
In view of the low convergence accuracy and the ease with can be considered to spread the initial squirrel population
which SSA in high-dimensional problems fall into the local throughout the search space to speed up convergence. This
optimum, the RSSA with reproduction and adaptive step is also a new direction for future research.
size strategy is proposed. First, the reproduction mechanism
is applied to the search process to increase the diversity REFERENCES
of the population, thereby enhancing the exploration ability
of the algorithm. Second, the adaptive step size strategy is [1] H. Zhang, L. Cui, X. Zhang, and Y. Luo, ‘‘Data-driven robust approx-
imate optimal tracking control for unknown general nonlinear systems
combined with the position updating formulas to balance the using adaptive dynamic programming method,’’ IEEE Trans. Neural Netw.,
global and local search while avoiding the complex parameter vol. 22, no. 12, pp. 2226–2236, Dec. 2011.

VOLUME 8, 2020 101131

 X. Zhang et al.: Improved SSA With Reproductive Behavior

[2] Narendra and Fukunaga, ‘‘A branch and bound algorithm for feature subset [27] X. S. Yang and S. Deb, ‘‘Cuckoo search via Lévy flights,’’ in Proc. World
selection,’’ IEEE Trans. Comput., vol. C-26, no. 9, pp. 917–922, Sep. 1977. Congr. Nature Biologically Inspired Comput., Dec. 2009, pp. 210–214.
[3] A. Slowik and K. Halina, ‘‘Nature inspired methods and their industry [28] S. Mirjalili, ‘‘Dragonfly algorithm: A new meta-heuristic optimization
applications—Swarm intelligence algorithms,’’ IEEE Trans. Ind. Infor- technique for solving single-objective, discrete, and multi-objective prob-
mat., vol. 14, no. 3, pp. 1004–1015, Dec. 2017. lems,’’ Neural Comput. Appl., vol. 27, no. 4, pp. 1053–1073, May 2015.
[4] A. Chakraborty and A. K. Kar, ‘‘Swarm intelligence: A review of [29] M. Jain, V. Singh, and A. Rani, ‘‘A novel nature-inspired algorithm for
algorithms,’’ Nature-Inspired Comput. Optim., vol. 10, pp. 475–494, optimization: Squirrel search algorithm,’’ Swarm Evol. Comput., vol. 44,
Mar. 2017. pp. 148–175, Feb. 2019.
[5] H. Li, F. He, Y. Liang, and Q. Quan, ‘‘A dividing-based many-objective [30] P. Wang, Y. Kong, X. He, M. Zhang, and X. Tan, ‘‘An improved
evolutionary algorithm for large-scale feature selection,’’ Soft Comput., squirrel search algorithm for maximum likelihood DOA estimation and
vol. 24, no. 9, pp. 6851–6870, Sep. 2019. application for MEMS vector hydrophone array,’’ IEEE Access, vol. 7,
[6] N. Hou, F. He, Y. Zhou, and Y. Chen, ‘‘An efficient GPU-based parallel tabu pp. 118343–118358, 2019.
search algorithm for hardware/software co-design,’’ Frontiers Comput. [31] S. Agrawal, L. Samantaray, R. Panda, and L. Dora, ‘‘A new hybrid adaptive
Sci., vol. 14, no. 5, pp. 1–18, Mar. 2020. cuckoo search-squirrel search algorithm for brain MR image analysis,’’
[7] M. Le Quiniou, P. Mandel, and L. Monier, ‘‘Optimization of drinking water in Hybrid Machine Intelligence for Medical Image Analysis. Singapore:
and sewer hydraulic management: Coupling of a genetic algorithm and two Springer, 2020, pp. 85–117.
network hydraulic tools,’’ Procedia Eng., vol. 89, pp. 710–718, Nov. 2014. [32] M. S. Sanaj and P. M. Joe Prathap, ‘‘Nature inspired chaotic squirrel search
[8] Z. Sai, Y. Fan, T. Yuliang, X. Lei, and Z. Yifong, ‘‘Optimized algorithm of algorithm (CSSA) for multi objective task scheduling in an IAAS cloud
sensor node deployment for intelligent agricultural monitoring,’’ Comput. computing atmosphere,’’ Eng. Sci. Technol., Int. J., to be published, doi:
Electron. Agricult., vol. 127, pp. 76–86, Sep. 2016. 10.1016/j.jestch.2019.11.002.
[9] M. Reihanian, S. R. Asadullahpour, S. Hajarpour, and K. Gheisari, ‘‘Appli- [33] M. Basu, ‘‘Squirrel search algorithm for multi-region combined heat
cation of neural network and genetic algorithm to powder metallurgy of and power economic dispatch incorporating renewable energy sources,’’
pure iron,’’ Mater. Design, vol. 32, no. 6, pp. 3183–3188, Jun. 2011. Energy, vol. 182, pp. 296–305, Sep. 2019.
[10] A. Castiglione, R. De Prisco, A. De Santis, U. Fiore, and F. Palmieri, [34] Y. Wang and T. Du, ‘‘An improved squirrel search algorithm for global
‘‘A botnet-based command and control approach relying on swarm intelli- function optimization,’’ Algorithms, vol. 12, no. 4, p. 80, Apr. 2019.
gence,’’ J. Netw. Comput. Appl., vol. 38, pp. 22–33, Feb. 2014. [35] J. Verstichel, D. C. Patrick, and V. B. Greet, ‘‘An improved best-fit heuristic
[11] J. Kennedy and R. Eberhart, ‘‘Particle swarm optimization,’’ in Proc. for the orthogonal strip packing problem,’’ Int. Trans. Oper. Res., vol. 20,
Int. Conf. Neural Netw. (ICNN), Perth, WA, Australia, Nov./Dec. 1995, no. 5, pp. 264–271, Jun. 2013.
pp. 1942–1948. [36] W. H. El-Ashmawi and D. S. A. Elminaam, ‘‘A modified squirrel search
[12] M. Dorigo, M. Birattari, and T. Stutzle, ‘‘Ant colony optimization,’’ IEEE algorithm based on improved best fit heuristic and operator strategy for bin
Comput. Intell. Mag., vol. 1, no. 4, pp. 28–39, Nov. 2006. packing problem,’’ Appl. Soft Comput., vol. 82, Sep. 2019, Art. no. 105565.
[13] D. Karaboga and B. Basturk, ‘‘A powerful and efficient algorithm for [37] T. Zheng and W. Luo, ‘‘An improved squirrel search algorithm for opti-
numerical function optimization: Artificial bee colony (ABC) algorithm,’’ mization,’’ Complexity, vol. 2019, pp. 1–31, Jul. 2019.
J. Global Optim., vol. 39, no. 3, pp. 459–471, Apr. 2007. [38] M. Durairasan and D. Balasubramanian, ‘‘An efficient control strategy
[14] X. Yan, F. He, N. Hou, and H. Ai, ‘‘An efficient particle swarm optimization for optimal power flow management from a renewable energy source
for large-scale Hardware/Software co-design system,’’ Int. J. Cooperat. Inf. to a generalized three-phase microgrid system: A hybrid squirrel search
Syst., vol. 27, no. 1, Mar. 2018, Art. no. 1741001. algorithm with whale optimization algorithm approach,’’ Trans. Inst. Meas.
[15] J. Tian, Y. Tan, J. Zeng, C. Sun, and Y. Jin, ‘‘Multiobjective infill criterion Control, to be published, doi: 10.1177/0142331220901628.
driven Gaussian process-assisted particle swarm optimization of high- [39] K. Vernes, ‘‘Gliding performance of the northern flying squirrel (Glau-
dimensional expensive problems,’’ IEEE Trans. Evol. Comput., vol. 23, comys Sabrinus) in mature mixed forest of eastern Canada,’’ J. Mammal.,
no. 3, pp. 459–472, Jun. 2019. vol. 82, no. 4, pp. 1026–1033, Nov. 2001.
[16] T. Blackwell and J. Kennedy, ‘‘Impact of communication topology in [40] R. B. Thomas and P. D. Weigl, ‘‘Dynamic foraging behavior in the southern
particle swarm optimization,’’ IEEE Trans. Evol. Comput., vol. 23, no. 4, flying squirrel (Glaucomys Volans): Test of a model,’’ Amer. Midland
pp. 689–702, Aug. 2019. Naturalist, vol. 140, no. 2, pp. 264–271, Oct. 1998.
[17] W. Deng, H. Zhao, L. Zou, G. Li, X. Yang, and D. Wu, ‘‘A novel collabo- [41] A. R. Mehrabian and C. Lucas, ‘‘A novel numerical optimization algo-
rative optimization algorithm in solving complex optimization problems,’’ rithm inspired from weed colonization,’’ Ecol. Informat., vol. 1, no. 4,
Soft Comput., vol. 21, no. 15, pp. 4387–4398, Feb. 2016. pp. 355–366, Dec. 2006.
[18] Y. Zhou, F. He, N. Hou, and Y. Qiu, ‘‘Parallel ant colony optimiza- [42] M. Jamil and X.-S. Yang, ‘‘A literature survey of benchmark functions for
tion on multi-core SIMD CPUs,’’ Future Gener. Comput. Syst., vol. 79, global optimization problems,’’ vol. 4, no. 2, Aug. 2013, arXiv:1308.4008.
pp. 473–487, Feb. 2018. [Online]. Available: http://arxiv.org/abs/1308.4008
[19] Y. Zhou, F. He, and Y. Qiu, ‘‘Dynamic strategy based parallel ant colony [43] F. Wang, H. Zhang, K. Li, Z. Lin, J. Yang, and X.-L. Shen, ‘‘A hybrid
optimization on GPUs for TSPs,’’ Sci. China Inf. Sci., vol. 60, no. 6, particle swarm optimization algorithm using adaptive learning strategy,’’
Feb. 2017, Art. no. 068102. Inf. Sci., vols. 436–437, pp. 162–177, Apr. 2018.
[20] Y. Zhou, F. He, and Y. Qiu, ‘‘Optimization of parallel iterated local search [44] G. Kaur and S. Arora, ‘‘Chaotic whale optimization algorithm,’’ J. Comput.
algorithms on graphics processing unit,’’ J. Supercomput., vol. 72, no. 6, Design Eng., vol. 5, no. 3, pp. 275–284, Jul. 2018.
pp. 2394–2416, May 2016. [45] J. J. Liang, B. Y. Qu, and P. N. Suganthan, ‘‘Problem definitions and
[21] D. Kong, T. Chang, W. Dai, Q. Wang, and H. Sun, ‘‘An improved artificial evaluation criteria for the CEC 2014 special session and competition on
bee colony algorithm based on elite group guidance and combined breadth- single objective real-parameter numerical optimization,’’ Comput. Intell.
depth search strategy,’’ Inf. Sci., vols. 442–443, pp. 54–71, May 2018. Lab., Zhengzhou Univ., Zhengzhou, China, Tech. Rep. 201311, 2013.
[22] X. He, W. Wang, J. Jiang, and L. Xu, ‘‘An improved artificial bee colony [46] I. Fister, I. Fister, X.-S. Yang, and J. Brest, ‘‘A comprehensive review of
algorithm and its application to multi-objective optimal power flow,’’ firefly algorithms,’’ Swarm Evol. Comput., vol. 13, pp. 34–46, Dec. 2013.
Energies, vol. 8, no. 4, pp. 2412–2437, Mar. 2015. [47] Q. Yang, W. N. Chen, Z. Yu, T. Gu, Y. Li, H. Zhang, and J. Zhang,
[23] A. A. A. Ari, N. Labraoui, B. O. Yenké, and A. Gueroui, ‘‘Clustering ‘‘Adaptive multimodal continuous ant colony optimization,’’ IEEE Trans.
algorithm for wireless sensor networks: The honeybee swarms nest-sites Evol. Comput., vol. 21, no. 2, pp. 191–205, Jul. 2016.
selection process based approach,’’ Int. J. Sensor Netw., vol. 27, no. 1, [48] J. Derrac, S. García, D. Molina, and F. Herrera, ‘‘A practical tutorial on
pp. 1–13, Jun. 2018. the use of nonparametric statistical tests as a methodology for comparing
[24] D. H. Wolpert and W. G. Macready, ‘‘No free lunch theorems for optimiza- evolutionary and swarm intelligence algorithms,’’ Swarm Evol. Comput.,
tion,’’ IEEE Trans. Evol. Comput., vol. 1, no. 1, pp. 67–82, Apr. 1997. vol. 1, no. 1, pp. 3–18, Mar. 2011.
[25] S. Mirjalili, S. M. Mirjalili, and A. Lewis, ‘‘Grey wolf optimizer,’’ Adv. [49] S. García, D. Molina, M. Lozano, and F. Herrera, ‘‘A study on the use of
Eng. Softw., vol. 69, pp. 46–61, Mar. 2014. non-parametric tests for analyzing the evolutionary algorithms’ behaviour:
[26] X.-S. Yang, ‘‘Firefly algorithm, stochastic test functions and design opti- A case study on the CEC’2005 special session on real parameter optimiza-
misation,’’ vol. 2, no. 2, Mar. 2010, arXiv:1003.1409. [Online]. Available: tion,’’ J. Heuristics, vol. 15, no. 6, p. 617, May 2008.
http://arxiv.org/abs/1003.1409

101132 VOLUME 8, 2020