Applied Intelligence

https://doi.org/10.1007/s10489-018-1291-2

A hybrid clonal selection algorithm with modified combinatorial

recombination and success-history based adaptive mutation
for numerical optimization
Weiwei Zhang1 · Kui Gao1 · Weizheng Zhang1 · Xiao Wang1 · Qiuwen Zhang1 · Hua Wang1

© Springer Science+Business Media, LLC, part of Springer Nature 2018

Abstract
Artificial immune system is a class of computational intelligence methods drawing inspiration from biological immune
system. As one type of popular artificial immune computing model, clonal selection algorithm (CSA) has been widely
used for many optimization problems. When dealing with complex optimization problems, such as the characteristics of
multimodal, high-dimension, rotational, the traditional CSA often suffers from diversity loss, poor search ability, premature
convergence and stagnation. To address the problems, a modified combinatorial recombination is introduced to bring
diversity to the population and avoid the premature convergence. Moreover, the success-history based adaptive mutation
strategy is introduced to form a success-history based adaptive mutation based clonal selection algorithm to improve the
search ability. The mutation operator is also modified and analyzed through experimental comparison. To further improve
the precision and cope with the stagnation, the gene knockout strategy is proposed. The proposed algorithm is tested on CEC
2014 benchmarks and compared with state-of-the-art evolutionary algorithms. The experimental results show that MSHCSA
is quite competitive.

Keywords Immune system · Clonal selection algorithm · Optimization · Mutation · Adaptive

1 Introduction Where X = [x1 , x2 , · · · , xD ] is a D-dimensional vector of

decision variables in the feasible region . The optimization
Generally, an optimization problem can be described as could be either minimization problem or maximization problem.
Inspired by nature, people have developed many optimiza-
Optimize f (X), subject to X ∈  (1) tion computation methods to solve the complicated optimiza-
tion problems [1, 2]. Among them, artificial immune system
(AIS) has made great progress in both academic and indus-
 Weizheng Zhang
weizheng008@126.com trial communities recently. In view of the effectiveness and
complexity of biological immune system, various functions
Weiwei Zhang
anqikeli@163.com and principles are developed into applicable theories and
systems [3]. AISs have been applied on intrusion detec-
Kui Gao
945915123@qq.com tion [4], email spam detection [6, 7], feature selection and
prediction [8], optimization [9] and other fields [5]. Clonal
Xiao Wang
pandaxiaoxi@163.com selection algorithm (CSA) is a newly emerging compu-
tational paradigm inspired by adaptive immune response
Qiuwen Zhang
zhangqwen@126.com of immune B cells, including clonal proliferation differ-
entiation, affinity maturation and memory mechanisms, is
Hua Wang
haoyouqiezi@126.com well suited for handling optimization problems [10, 11].
However, facing more and more complicated optimization
1 Zhengzhou University of Light Industry, problem, the CSAs often suffer the premature convergence,
Zhengzhou, 450000, China poor accuracy and diversity loss.

In this paper, we propose an enhanced CSA named
MSHCSA, by introducing a new combinatorial recombina-
tion and a modified success-history based adaptive mutation
strategy. In MSHCSA, the combinatorial recombination
works for bring diversity and Success-history based adap-
tive mutation strategy is for improving the search ability.
Moreover, to further improve the accuracy and avoid the
stagnation, a gene knockout strategy is proposed. These
strategies work together to guide the evolutionary process of
MSHCSA towards the global optimum.
The remainder of the paper is organized as follows. Section 2
shows basic theory of CSA and involved strategies. Section 3
shows the proposed algorithm. Section 4 represents the
experimental simulation and analysis. The conclusion is
given in Section 5.
2 Clonal selection algorithms
In this section, we first represent the basic framework of
CSA and then give a rough review of the various variants
of the CSA in the perspective of the modified operators and
hybrid algorithms.
2.1 Basic CSA
The main idea of clonal selection theory lies in the phe-
nomenon that B cells can adaptively modify their recep-
tor termed antibodies to react the invasive non-self anti-
gens in biological immune system [12]. Inspired from the
clonal selection principle, various clonal selection algo-
rithms are proposed. The general one, known as CLONALG
[13] is one of the representatives.
In the basic CSA, antigen is represented as an optimiza-
tion problem to be solved, while the antibodies are candidate
solutions. The algorithm aims to locate the fittest solutions
to the problem. The optimization problem could be global
optimization problem, multimodal optimization problem,
multi-objective optimization problem, and etc. Fitness mea-
sures how a solution fits for the problem. There are mainly
three operations involved, which are cloning, hypermutation
and selection. Cloning operator provides candidate room
for the hypermutation to bring diversity. The introduction
of diversity helps each antibody explore the near neighbor.
Then the one with the highest fitness in the antibody pop-
ulation will be survive through selection operator, and the
others are eliminated. The framework of the basic CSA can
be described as Algorithm 1.
2.2 Improved CSAs
When implementing in practice, basic CSA gradually
evolves into various variants to meet the challenges.
W. Zhang et al.
Through literatures review, the modification to the operators
and hybridization of different strategies and algorithms
to basic CSA point out a promising direction to further
improve the search effectiveness [10].
Degeneration recognizing method is applied on clonal
selection algorithm for improving the computing speed
[9]. Baldwinian effect is introduced to CSA which forms
an improved CSA named Baldwinian Clonal Selection
Algorithm which is proposed for optimization [14]. In the
hybrid GSCSA [15], gravitational search algorithm is
employed to do the exploration while CSA is for
exploitation. Differential evolution (DE) is combined
with CSA for enhancing the operating efficiency and
convergence rate [16]. A fast clonal algorithm is proposed
which designs a parallel mutation operator comprising
of Gaussian and Cauchy mutation strategies [17]. A
quantum-inspired immune clonal algorithm (QICA) is
addressed to cope with global optimization, in which,
quantum mutation and quantum recombination are proposed
[18]. Cutello et al. introduced a real coded clonal
selection algorithm for global optimization, involving the
cloning operator, inversely proportional hypermutation and
aging operator [19]. Baldwinian learning and orthogonal
learning are incorporated with CSA generates a hybrid
learning clonal selection algorithm (HLCSA) [20], in which
Baldwinian learning is employed for exploration while the
orthogonal learning for exploitation. Yang et al. Lamarckian
learning theory is introduced to CSA, which forms the
Lamarckian Clonal Selection Algorithm (LCSA), while the
Recombination operator and tournament selection operator
are incorporated into the algorithm as well [21]. Li et al. [22]
proposed an immune clonal PSO, in which Particle swarm
optimization algorithm (PSO) is integrated with clonal
selection to enhance the performance and suppress the
degradation phenomena. Gaussian and Cauchy mutations
are introduced to CSA and Cell Repair Operator (CPO)
and Dynamic Mutation Size Operator (DMSO) are also
proposed in the improved combined mutation CSA to
handle with the unimodal and multimodal functions [23].
Vaccination and Cauchy mutation is developed into clonal
selection algorithm to accelerate the convergence rate
and improve the detection rate [24]. A mixed mutation
strategy combining with Gaussian and Cauchy mutations is
proposed in [25] and experimental results show the mixed
strategy can obtain the same or even better performance.
It can be observed that the modification to the operators
and hybridization of different strategies and algorithms are
effective improvements to basic CSA. However, the way to
modify the operators and know which strategies or algo-
rithms incorporated with CSA would benefit the performance
is worth to be studied. When the optimization problem
becomes more and more complicated, it is a delicate work
to modify the inappropriate operations and assemble the
A hybrid clonal selection algorithm with modified combinatorial recombination and success-history...
well-chosen strategies into a functional system. In this
paper, a combinatorial recombination operator is modified
and hypermutation is incorporated with success-history
based adaptive mutation strategy, together with gene
knockout strategy compose an enhanced CSA termed
MSHCSA for the complex optimization problems.
3 Proposed algorithm
3.1 Modified Recombination
During the development of B cell, the gene segments in the
libraries are combined and rearranged at the level of the
DNA. With the recombination of gene segments, recombi-
nation creates a population of cells that vary widely in their
specificity [12]. The presence of combinatorial recombina-
tion takes the responsibility of the diversification as hyper-
mutation in immunology. To simulate the arrangement of
gene segments in recombination, many recombination oper-
ators have been proposed, such as DE inspired recombi-
nation [26], binomial recombination [27] and intelligent
recombination [28] et al. The combinatorial recombina-
tion operator in RHCSA [29] is a newly proposed recom-
bination operator which could introduce diversity to the
population through fusing information between randomly
chosen parents. The new individuals are generated by
 instead of chose two individuals with high fitness form the
V A = αX VA + (1 − α)X VB
XA A B
(2)
XBV B = αXBVB + (1 − α)XA VA
where XA andXB are randomly chosen individuals from the
population, VA and VB are randomly chosen m dimensions,
m ∈ [1, D]. The recombination operator is represented
in the Fig. 1, in which two new individuals are generated
Fig. 1 Recombination process
through the combinational recombination. In the example,
m equals to 3. It needs to be known that i1 could be
different with j1 , the same to i1 , j2 and i3 , j3 as long as the
normalization has been done.
After the fitness of the new generated individuals is
evaluated, together with the original individuals, the two
ones with the higher fitness will survive, and the other two
individuals are deleted, i.e. two individuals with high fitness
form the set {XA , XB , XA  , X  } are selected.
B
It is reasonable that the diversity could be introduced
from information fusion of population according to the
above process. However, as is known to all, the distributed
population trend to fall into local optima influenced by the
“attractors” of the search area. If the population falls into
small local area, the above recombination operator cannot
bring diversity anymore and the search process is liable to
get stuck into stagnation.
To solve the problem, a modified version of the com-
binatorial recombination is proposed in this paper. During
the recombination process, instead of being abandoned
immediately, the generated candidate solutions failed in the
selection process will be kept and stored in the archive A
in the paper. Thereafter, in the proposed algorithm, one of
the parent individual XA comes from the current population
and the other parent individual XB is randomly chosen from
the union of current population and archive A. The recom-
bination is implemented as (2). Then in the selection part,
set {XA , XB , XA , X  }, only X is updated if the new gen-
B A
erated offspring have better fitness, i.e., the one with best
fitness will replace XA , and the other will be add to set A.
The size of archive A is initialized to zero, and then if the
archive size exceeds a certain threshold, say Np, then some
solutions are randomly removed from the archive to keep
the archive size at Np.

3.2 Success-history based adaptive mutation based
clonal selection
The Success-history based adaptive mutation based clonal
selection is composed of Cloning, Success-history based adap-
tive hypermutation and Selection, which is shown in Fig. 2.
i. Cloning: Each antibody Xi could generate clones as
offspring {Xi1 , Xi2 , · · · , XiNc } through Cloning, where
Nc is the clone number.
ii. Success-history based adaptive hypermutation: Hyper-
mutation brings diversity to the population by introducing

j j j j j
j Xi + Fi × (Xpbest − Xi ) + Fi × (Xr1 − Xr2 ) j ∈ randM(n)
Xi = j
Xi
j j
Where Xi is the j th dimension of ith individual, Xpbest
is the j th dimension of randomly chosen as one of the
top 100 × q% individuals in the current population with
p ∈ (0, 1]. The indices r1, r2 are randomly selected from
[1, Np] and r1 = r2  = i. Fi determines the magnitude,
randM(n) ∈ {1, · · · , n} is randomly chosen M indexes
without repetition. M is determined by
M = (α × n) + 1

j j j j j j
j Xi + Fi × (Xpbest − Xi ) + Fi × (Xr1 − Xr2 ) + Fi × (Xr3 − Xr4 ) j ∈ randM(n)
Xi = j
Xi
Where r3 and r4 are randomly selected as r1 and r2, the
added third item is expected to bring more diversity and
improve search ability. The illustrations of both current-to-
pbest/1 mutation and the modified mutation are shown in
Fig. 3. As shown in Fig. 3, the added items could bring
Fig. 2 The main steps of the
success-history based adaptive
mutation based clonal selection
W. Zhang et al.
perturbation for each clone. In view of the clonal
selection principle, the hypermutation rate follows the
inversely proportional law [13]. Each candidate solu-
tion is subject to M mutations [30] is adopted, in
which inversely proportional law is subtly used to deter-
mine the number of mutations M. Moreover, the way
Success-History based Adaptive DE (SHADE) [31, 32]
generates the mutant vector, termed current-to-pbest/1
mutation is adopted to implement the hypermutation.
j
Each clone Xi j = 1, · · · , Nc, generates the hypermu-
j
tated clones Xi as (3).
(3)
otherwise
α = exp(−ρf ∗ (Xi )) (5)
Where f ∗ (Xi ) ∈ [0, 1] is the normalized fitness of Xi , ρ
is the decay constant controls the shape of the mutation rate
· returns the lower bound integer.
Besides, we also tried to make modification on the
equation of current-to-pbest mutation i.e. (3) as (6).
(4)
j
(6)
otherwise
perturbation on the basis of current-to-pbest/1 mutation. In
view of the huge search space in handling with the high-
dimensional and rotational problems, the search process is
easily to fall into stagnation. With the introduction of the
difference from two pairs of randomly selected individuals
A hybrid clonal selection algorithm with modified combinatorial recombination and success-history...

Fig. 3 a An illustration of
current-to-pbest/1 mutation
scheme in 2D space b An
illustration of modified mutation
scheme in 2D space

is expected to bring more diversity. Moreover, the direction an element MFri , where ri ∈ [1, H ] from MF and
composition of newly introduced difference vectors and the implementing
original difference vector may somehow benefit solving the
Fi = randc(MFri , 0.1) (7)
rotation problem. The detailed discussion will be presented
in experimental part. where randc(μ, σ 2 ) is the cauchy distribution with mean
The control parameter Fi follows the success-history μ and variance σ 2 . In each generation, Fi succeeds in
based adaptation rules as [31, 32]. As shown in Table 1, a generating an individual better than the parent is recorded
historical memory archive MF which deserves H historical in SF , and at the end of generation, the contents of memory
value of Fi is set. In each generation, Fi randomly selects are updated as follows:

k meanW L (SF ) SF  = ϕ
MF = (8)
Table 1 The historical memory MF MFk otherwise
Index 1 2 ... H-1 H |SF |
k=1 ωk × (SF )
k 2

MF MF1 MF2 ... MFH −1 MFH

meanW L (SF ) = |SF |
(9)
k=1 ωk × SF
k

fk
ωk = |S | (10)
F
k=1 fk
 
fk = f (Xk ) − f (Xk ) (11)
where k ∈ [1, H ] is initialized to 1 and recursively
increases. If there is no better offspring is generated, i.e.
SF = ϕ the contents of MF will not be updated.
iii. Selection: Select the one with highest fitness among
Xi and hypermutated clones {Xi1 , Xi2 , · · · , XiNc }
3.3 Gene knockout strategy
When handling the optimization problem with multiple
optima, evolutionary algorithms always suffer premature
convergence. The problem becomes more intractable with
the dimensions of the problems increase. When facing the
huge search space in high dimensional problems, stagnation
is a common and hard to avoid situation. It is obvious that
stagnation is harmful since it cannot bring any progress to
the results but consuming computational resource. To solve
the problem, agei is introduced for each individual Xi to
count how many times it fails to update itself. If agei >
Maxage, one randomly chosen dimension of the individual
Xi will be replaces as follows.
j functions, F4-F16 are shifted and rotated multimodal
Xi = elitej + randn(0, 1) (12)
where, Maxage is a user defined threshold, j ∈ [1, D] is
a randomly chosen dimension, elite is a randomly chosen
individual from the top 100 q% best antibodies from the
population, randn(0, 1) is the Gaussian random number
with mean 0 and variance 1.
3.4 The process of MSHCSA
Combining with the new recombination and mutation
strategy, the new algorithm, denoted as MSHCSA is
presented in Algorithm 2.
W. Zhang et al.
4 Experiments
In this section, the experiments that have been done to evalu-
ate the performance of the proposed MSHCSA for a number
of benchmark functions are described. The functions were
taken from the suite of benchmark problems used in Com-
petition on Evolutionary Computation in Single Objective
Real-Parameter Numerical Optimization, held under the
2014 IEEE Congress on Evolutionary Computation (CEC)
[33], which contains 30 benchmark functions with the fea-
tures covering the unimodal, multimodal, composing test
problems by extracting features dimension-wise from sev-
eral problems, graded level of linkages, rotated trap prob-
lems. The benchmark functions are listed in Table 2. The
search range of the functions is [−100, 100]D , D is dimen-
sions, Mi is the rotation matrix, oi = [oi1 , oi2 , · · · , oiD ]T is
the shifted global optimum, which is randomly distributed in
the search range, Fi∗ = Fi (x ∗ ) is the optimal function value,
fi (x) is the basic function. In the hybrid functions, gi (x) is
the basic function used to construct the hybrid function, N is
the number of basic function, zi is randomly shuffled x −oi .
In the composition functions, λi controls the height of each
gi (x), ωi is weight value, biasi defines which optimum is
global optimum. It’s worth to address that these problems
should be treated as black-box problems, which means the
explicit equations cannot be used.
Among them, function F1-F3 are unimodal rotated
functions, F17-F22 are hybrid functions and F23-F30
are composition functions. These functions pose great
challenges. To save the space, we only give one function F12
as representative to show the landscape distribution. The
basic function of F12 is the Katsuura Function as follows.
⎛ ⎞ D101.2
32  j
10  ⎝
D 2 xi −round(2j xi )
f10 (x) = 2 1+ i ⎠ − 10
D 2 j D2
i=1 j =1
(13)
Figure 4 shows the 3D map and the contour map for 2D
function of F12 (shifted and Rotatated Katsuura Function).
Table 2 Summary of the CEC2014 benchmark functions

No. Functions Equations Fi∗ = Fi (x ∗ )

Unimodal function 1 Rotated High Conditioned Elliptic Function F1 (x) = f1 (M(x − o1 )) + F1∗ 100
2 Rotated Bent Cigar Function F2 (x) = f2 (M(x − o2 )) + F2∗ 200
3 Rotated Discus Function F3 (x) = f3 (M(x − o3 )) + F3∗ 300
4)
Simple multimodal function 4 Shifted and Rotated Rosenbrock’s Function F4 (x) = f4 (M( 2.048(x−o
100 ) + 1) + F4∗ 400
5 Shifted and Rotated Ackley’s Function F5 (x) = f5 (M(x − o5 )) + F8∗ 500
6)
6 Shifted and Rotated Weierstrass Function F6 (x) = f6 (M( 0.5(x−o
100 )) + F6∗ 600
600(x−o7 )
7 Shifted and Rotated Griewank’s Function F7 (x) = f7 (M( 100 )) + F7∗ 700
8)
8 Shifted Rastrigin’s Function F8 (x) = f8 ( 5.12(x−o
100 ) + F7∗ 800
5.12(x−o9 )
9 Shifted and Rotated Rastrigin’s Function F9 (x) = f8 (M( 100 )) + F9∗ 900
10 ) ∗
10 Shifted Schwefel’s Function F10 (x) = f9 ( 1000(x−o
100 ) + F10 1000
1000(x−o11 ) ∗
11 Shifted and Rotated Schwefel’s Function F11 (x) = f9 (M( 100 )) + F11 1100
5(x−o12 ) ∗
12 Shifted and Rotated Katsuura Function F12 (x) = f10 (M( 100 )) + F12 1200
13 ) ∗
13 Shifted and Rotated HappyCat Function F13 (x) = f11 (M( 5(x−o 100 )) + F13 1300
5(x−o14 ) ∗
14 Shifted and Rotated HGBat Function F14 (x) = f12 (M( 100 )) + F14 1400
5(x−o15 ) ∗
A hybrid clonal selection algorithm with modified combinatorial recombination and success-history...

15 Shifted and Rotated Expanded Griewank’s plus Rosenbrock’s Function F15 (x) = f13 (M( 100 ) + 1) + F15 1500
16 Shifted and Rotated Expanded Scaffer’s F6 Function F16 (x) = f14 (M(x − o16 ) + 1) + F16 ∗ 1600
Hybrid functions 17-22 Hybrid Function1-6 F (x) = g1 (M1 z1 ) + · · · +gN (MN zN ) + F ∗ (x) 1700-2200
N

Composition functions 23-30 Composition Function 1-8 F (x) = {ωi ∗ [λi gi (x) + biasi ]} + F ∗ 2300-3000
i=1

It is observed that F12 is a non-separable function which
has a huge number of local optima, and it is continuous
everywhere, yet differentiable nowhere. When handling
these kind of functions, the algorithm is very easy to be
trapped on the local optima.
In our experiments, the dimensions of these functions
are set to 10, 30, 50, and 100, respectively, which bring
more challenges. Through the simulations, each experiment
runs 30 times and during each run, the function is evaluated
for 10000*D times (D is the dimension of feasible solution
Fig. 4 a 3D map for 2D
function F12. b Contour map for
2D function F12
W. Zhang et al.
space). “−”, “+”, and “≈” indicate that the performance
of the corresponding peer algorithm is worse than, better
than, and similar to that of MSHCSA, respectively. The
best results among all the tested algorithms are marked in
bold.
4.1 Experimental parameter setting
During experimentation, involving parameters of the pro-
posed MSHCSA were set as follows based on parameter
A hybrid clonal selection algorithm with modified combinatorial recombination and success-history...
tuning simulation results: size of antibody population N is
set to 10∗D, the cloning number Nc is set to 1, and the
number of selected dimension m in recombination is set to
0.3∗D. In the gene knockout strategy, the elite rate q is set
as 0.2. The maximal age Maxage is set to 6.
4.2 Experimental investigation of MSHCSA
There are new strategies developed in the MSHCSA
algorithm. Their efficiencies are investigated through
experimental simulation in the following subsection.
4.2.1 Investigation of modified recombination
Based on the basic CSA, we compared the performance
of the basic CSA, CSA with modified recombination
strategy, and CSA with combination recombination in
RHCSA [29], which are called “Basic CSA”, “CSA+MR”
and “CSA+CR”, respectively. To save the limited space,
Tables 3 and 4 shows the experimental results on CEC2014
benchmark functions with 30 and 100 dimensions.
It is shown that all of them, including Basic CSA, and the
introduction of no matter combinatorial recombination and
modified recombination could not get satisfactory results
on the benchmark functions. It simply shows the challenge
of the benchmark functions. Even though, the results of
both CSA+MR and CSA+CR are obviously better than
Basic CSA no matter on 30 or 100 dimensions. That is
to say, the introduction of recombination to Basic CSA
could improve the performance of Basic CSA. The reason
is recombination operator could increase the diversity of
the population through the exchange of the gene segments,
and the introduced diversity could help the population
jump out of the local optima and keep searching for the
global optimum. Moreover, CSA+CR could get better
performance on unimodal functions (F1-F3) both in 30 and
100 dimension, while CSA+MR seems to be more suitable
for multimodal functions (F4-F16). It can be concluded
that the original combinatorial recombination could enhance
exploitation ability while modified recombination could
bring more diversity and motivate the optima searching
ability due to the guiding effect of failed candidate
solutions. To hybrid function (F17-F22), the CSA+CR
obtains better results than CSA+MR in 30 dimensions
and gets even results on 100 dimensions. When facing
the composition functions (F23-F30), the performance of
CSA+MR is better than CSA+CR in 30 dimension and
tie with each other in 100 dimensions. It is found that
when coping with the complicated hybrid and composition
function, neither CSA+MR nor CSA+CR has highlighting
behavior as it shows in handling simple unimodal or
multimodal functions. In view of the whole performance,
the CSA+MR is slightly better than CSA+CR.
4.2.2 Investigation of success-history based adaptive
mutation
In the proposed algorithm, the success-history based
adaptive mutation is introduced. We not only directly
introduce the current-to-pbest/1 mutation but also try a
modification version of the mutation equation. They are
denoted as C-Mutation and M-mutation respectively and
compared results are shown in Table 5. Comparing with
Table 4, it can be observed that both mutation operations
could improve the performance of the algorithm to a great
extent.
When dimension is 30D, C-Mutation achieves better
results than M mutation in F4, F6, F8-F16, F19 and F21,
but worse results in F7, F17, F18, F20 and F22. Both
of them obtain the satisfying results in the unimodal
functions (F1-F3). Obviously, C-Mutation is superior than
M mutation when handling the multimodal functions (F4-
F16). When facing the hybrid functions, the M mutation
achieves the better performance (F17-F22). They are
even to the composition functions (F23-F30). In the
30D, C Mutation shows the advantage in handling the
multimodal functions. To the other type of functions, they
perform about the same. However, C Mutation starts to
lose superiority comparing with in 30D when dimension is
50D, and the situation continues in 100D. It is addressed
that C Mutation fails to locate the global optima in F1
while M mutation succeeds in 50D, and the difference
of the error is an order of magnitude in 100D. With the
increase of the dimension, the results of M mutation shows
better results in handling unimodal functions (F1-F3), and
the overwhelming advantage of C Mutation is declining
to the multimodal functions (F4-F16). To the Hybrid and
composition functions, the M mutation is getting better and
better results when dimensions increase. The reason is that
when dimensions increase, the search space will become
huge. In this situation, the introduced diversity could help
the algorithm jumps out of the stagnation. Obviously, the
M mutation brings more diversity than C Mutation.
During the experimental comparison, the way to imple-
ment both recombination and hypermutation operators
shows their respective advantages. The modified recombi-
nation and modified success-history based adaptive muta-
tion are adopted in the paper and form the algorithm
MSHCSA through comprehensive assessment.
4.3 Comparison with the state-of-art algorithms
The performance of MSHCSA is compared with the
following five the state-of-the-art algorithms by using the
benchmark suite of CEC2014: L-SHADE [34], RSDE [35],
FWA-DM [36], CETMS [37], and OptBees [38]. For the
competitor algorithms, the best parametric setup is
Table 3 Results (mean) of basic CSA with different Recombination operators on CEC2014 with 30D

Fun F1 F2 F3 F4 F5 F6 F7 F8 F9 F10 F11

Basic CSA 2.22E+09– 1.21E+11– 1.54E+05– 3.07E+04– 2.12E+01– 4.74E+01– 9.65E+02– 4.69E+02– 5.84E+02– 8.44E+03– 8.75E+03–
CSA+CR 6.49E+08– 6.13E+10+ 6.90E+04+ 8.87E+03– 2.11E+01+ 3.73E+01– 4.91E+02+ 2.98E+02– 3.39E+02– 6.40E+03– 7.52E+03–
CSA+MR 6.42E+08 6.25E+10 7.12E+04 8.46E+03 2.11E+01 3.64E+01 4.95E+02 2.96E+02 3.33E+02 6.20E+03 7.38E+03
fun F12 F13 F14 F15 F16 F17 F18 F19 F20 F21 F22
Basic CSA 3.97E+00– 1.00E+01– 3.54E+02– 1.15E+07– 1.39E+01– 1.50E+08– 6.77E+09– 8.27E+02– 2.75E+05– 5.09E+07– 7.04E+03–
CSA+CR 2.87E+00– 7.17E+00– 1.74E+02– 5.96E+04+ 1.33E+01– 3.18E+07+ 1.68E+09+ 2.37E+02+ 6.44E+04– 8.38E+06+ 1.89E+03+
CSA+MR 2.74E+00 7.06E+00 1.68E+02 6.54E+04 1.31E+01 3.30E+07 1.73E+09 2.54E+02 4.61E+04 9.43E+06 2.12E+03
fun F23 F24 F25 F26 F27 F28 F29 F30 −/ + /≈
Basic CSA 1.46E+03– 5.12E+02– 3.63E+02– 1.81E+02– 1.66E+03– 7.79E+03– 5.39E+08– 4.34E+06– 30/0/0
CSA+CR 5.96E+02– 2.06E+02+ 2.04E+02– 1.39E+02– 7.52E+02+ 6.70E+03– 3.49E+08– 1.63E+06+ 17/13/0
CSA+MR 5.83E+02 2.07E+02 2.02E+02 1.15E+02 7.73E+02 6.70E+03 3.36E+08 1.65E+06

Table 4 Results (mean) of basic CSA with different Recombination operators on CEC2014 with 100D

Fun F1 F2 F3 F4 F5 F6 F7 F8 F9 F10 F11

Basic CSA 1.81E+10– 6.14E+11– 6.63E+05– 2.77E+05– 2.15E+01– 1.75E+02– 5.70E+03– 1.99E+03– 2.43E+03– 3.37E+04– 3.42E+04–
CSA+CR 6.40E+09+ 2.75E+11+ 3.91E+05– 7.55E+04– 2.14E+01– 1.52E+02– 2.85E+03+ 1.28E+03– 1.48E+03– 2.87E+04– 3.15E+04–
CSA+MR 6.54E+09 2.80E+11 3.53E+05 7.50E+04 2.14E+01 1.51E+02 2.85E+03 1.25E+03 1.44E+03 2.80E+04 3.07E+04
fun F12 F13 F14 F15 F16 F17 F18 F19 F20 F21 F22
Basic CSA 5.40E+00– 1.41E+01– 1.66E+03– 4.87E+08– 4.83E+01– 2.76E+09– 7.58E+10– 1.88E+04– 1.47E+07– 1.20E+09– 1.03E+06–
CSA+CR 4.36E+00– 9.02E+00+ 8.08E+02+ 1.10E+07+ 4.71E+01– 9.00E+08+ 2.99E+10+ 5.94E+03+ 7.26E+05– 1.93E+08– 2.45E+04–
CSA+MR 4.21E+00 9.04E+00 8.26E+02 1.13E+07 4.68E+01 9.42E+08 3.07E+10 6.21E+03 5.14E+05 1.91E+08 2.27E+04
fun F23 F24 F25 F26 F27 F28 F29 F30 −/ + / ≈
Basic CSA 8.90E+03– 1.90E+03– 1.76E+03– 1.51E+03– 7.13E+03– 3.93E+04– 7.01E+09– 3.77E+08– 30/0/0
CSA+CR 2.00E+02– 2.02E+02– 2.00E+02+ 2.00E+02≈ 6.02E+03+ 3.65E+04+ 5.39E+09– 1.56E+08+ 16/13/1
CSA+MR 2.00E+02 2.01E+02 2.00E+02 2.00E+02 6.16E+03 3.69E+04 5.31E+09 1.74E+08
W. Zhang et al.
Table 5 Results (mean(std)) of basic CSA with different Hypermutation operators on CEC2014 with varying dimensions

Dimension D=30 D=50 D=100

Fun C-mutation M-mutation C-mutation M-mutation C-mutation M-mutation

F1 0.00E+00(0.00E+00)≈ 0.00E+00(0.00E+00) 4.38E+04(2.85E+04)– 0.00E+00(0.00E+00) 9.17E+05(3.07E+05)– 9.95E+04(4.26E+04)

F2 0.00E+00(0.00E+00)≈ 0.00E+00(0.00E+00) 0.00E+00(0.00E+00)+ 2.51E-07(1.77E-07) 1.00E-05(1.29E-05)+ 2.46E+00(1.21E+00)
F3 0.00E+00(0.00E+00)≈ 0.00E+00(0.00E+00) 0.00E+00(0.00E+00)≈ 0.00E+00(0.00E+00) 1.12E-04(1.65E-04)– 6.59E-07(3.64E-07)
F4 5.32E-01(1.38E+00)+ 1.18E+00(5.44E-01) 2.00E+01(3.28E+00)+ 2.72E+01(8.12E-01) 1.07E+02(3.52E+01)– 8.44E+01(6.18E-01)
F5 2.00E+01(1.21E-04)≈ 2.00E+01(5.33E-05) 2.00E+01(2.28E-04)≈ 2.00E+01(2.44E-04) 2.00E+01(2.15E-03)≈ 2.00E+01(1.87E-03)
F6 4.62E-01(6.97E-01)+ 2.41E+01(9.47E-01) 3.38E+00(1.41E+00)+ 5.09E+01(1.79E+00) 2.04E+01(2.80E+00)+ 1.29E+02(2.61E+00)
F7 4.93E-04(1.88E-03)- 0.00E+00(0.00E+00) 9.86E-04(3.08E-03)– 0.00E+00(0.00E+00) 9.86E-04(2.56E-03)– 0.00E+00(0.00E+00)
F8 3.59E+00(1.62E+00)+ 5.24E+00(1.18E+00) 2.23E+01(2.82E+00)+ 3.26E+01(3.80E+00) 1.69E+02(1.22E+01)+ 2.36E+02(1.64E+01)
F9 2.09E+01(4.45E+00)+ 2.57E+01(5.10E+00) 5.00E+01(6.68E+00)+ 7.68E+01(8.82E+00) 2.52E+02(2.01E+01)+ 3.17E+02(1.82E+01)
F10 9.90E+00(4.02E+00)+ 2.17E+01(8.32E+00) 1.17E+03(2.74E+02)- 1.51E+03(2.24E+02) 1.06E+04(5.21E+02)+ 1.11E+04(5.54E+02)
F11 2.43E+03(3.07E+02)+ 2.66E+03(3.09E+02) 5.59E+03(3.87E+02)+ 5.79E+03(4.10E+02) 1.60E+04(5.55E+02)+ 1.69E+04(5.54E+02)
F12 4.03E-01(8.29E-02)+ 4.21E-01(9.78E-02) 6.89E-01(1.10E-01)– 6.82E-01(1.02E-01) 1.28E+00(1.28E-01)+ 1.33E+00(1.34E-01)
F13 1.55E-01(2.01E-02)+ 1.97E-01(3.21E-02) 2.54E-01(4.21E-02)+ 2.84E-01(2.99E-02) 3.54E-01(4.18E-02)+ 3.67E-01(3.12E-02)
F14 2.48E-01(3.96E-02)+ 1.79E-01(2.76E-02) 3.28E-01(7.26E-02)- 2.22E-01(4.02E-02) 3.57E-01(7.32E-02)– 3.12E-01(9.89E-02)
F15 3.05E+00(6.12E-01)+ 3.39E+00(6.21E-01) 8.75E+00(9.70E-01)+ 9.19E+00(1.24E+00) 3.14E+01(2.47E+00)– 3.00E+01(1.84E+00)
F16 1.04E+01(2.73E-01)+ 1.10E+01(4.18E-01) 1.92E+01(4.43E-01)+ 2.03E+01(3.24E-01) 4.21E+01(4.96E-01)+ 4.35E+01(4.63E-01)
F17 1.09E+03(3.95E+02)– 1.07E+03(1.56E+02) 2.25E+03(5.41E+02)+ 2.64E+03(2.99E+02) 9.63E+03(4.05E+02)– 5.87E+03(6.98E+02)
F18 6.96E+01(1.35E+01)– 1.78E+01(3.32E+00) 1.25E+02(2.99E+01)- 7.78E+01(1.02E+01) 2.86E+02(3.82E+01)– 2.82E+02(9.44E+01)
F19 1.08E+01(1.21E+00)+ 1.25E+01(1.01E+00) 1.52E+01(4.72E+00)+ 1.96E+01(1.56E+00) 4.90E+01(3.82E+00)– 4.31E+01(3.35E-01)
F20 1.72E+01(3.62E+00)– 1.20E+01(1.44E+00) 6.83E+01(2.81E+01)– 4.89E+01(5.76E+00) 4.40E+02(7.73E+01)– 2.18E+02(1.49E+01)
A hybrid clonal selection algorithm with modified combinatorial recombination and success-history...

F21 2.60E+02(1.07E+02)+ 3.29E+02(9.73E+01) 1.22E+03(3.83E+02)+ 1.50E+03(2.48E+02) 2.84E+03(6.88E+02)+ 4.76E+03(3.41E+02)

F22 8.21E+01(5.71E+01)– 6.23E+01(3.75E+01) 3.12E+02(1.75E+02)+ 4.63E+02(1.32E+02) 1.34E+03(1.92E+02)+ 1.53E+03(3.35E+02)
F23 3.14E+02(0.00E+00)≈ 3.14E+02(0.00E+00) 3.37E+02(1.88E-13)≈ 3.37E+02(2.85E-13) 3.45E+02(2.60E-13)≈ 3.45E+02(1.05E-12)
F24 2.24E+02(1.04E+00)– 2.10E+02(1.12E+01) 2.75E+02(2.27E+00)- 2.64E+02(3.30E-01) 3.87E+02(4.18E+00)– 3.84E+02(2.91E+00)
F25 2.00E+02(1.87E-03)≈ 2.00E+02(5.28E-04) 2.01E+02(4.86E+00)- 2.00E+02(1.59E-02) 2.00E+02(1.57E-13)+ 2.01E+02(1.70E-01)
F26 1.00E+02(3.21E-02)≈ 1.00E+02(2.81E-02) 1.00E+02(5.76E-02)≈ 1.00E+02(2.83E-02) 1.50E+02(5.06E+01)– 1.00E+02(2.89E-02)
F27 3.62E+02(1.16E+02)+ 9.05E+02(1.85E+02) 5.89E+02(3.87E+02)+ 1.73E+03(3.70E+01) 1.32E+03(9.45E+02)+ 3.71E+03(5.61E+01)
F28 3.76E+02(4.76E+00)– 3.67E+02(4.63E-01) 3.68E+02(3.75E+00)- 3.57E+02(4.41E-01) 5.10E+02(3.23E+01)– 4.04E+02(5.30E+00)
F29 2.11E+02(2.19E+00)+ 2.14E+02(6.42E-01) 2.13E+02(2.88E+00)+ 2.29E+02(8.98E-01) 2.54E+02(8.22E+00)– 2.43E+02(4.22E+00)
F30 4.64E+02(7.66E+01)– 3.25E+02(4.51E+01) 1.16E+03(2.16E+02)– 9.58E+02(1.77E+02) 3.03E+03(2.98E+02)– 2.85E+03(2.73E+02)
-/+/≈ 7/16/7 11/15/4 15/13/2
Table 6 Results (mean(std)) of MSHCSA and other algorithm on CEC2014 benchmark function with 10D

Fun L-SHADE RSDE FWA-DM CETMS OptBees MSHCSA

F1 0.00E+00(0.00E+00)≈ 0.00E+00(0.00E+00)≈ 5.01E+03(1.67E+04)– 0.00E+00(0.00E+00) ≈ 7.80E+02(6.90E+02)– 0.00E+00(0.00E+00)

F2 0.00E+00(0.00E+00)≈ 0.00E+00(0.00E+00)≈ 1.34E-04(9.49E-04)– 0.00E+00(0.00E+00) ≈ 9.80E-03(6.3E−02)– 0.00E+00(0.00E+00)
F3 0.00E+00(0.00E+00)≈ 0.00E+00(0.00E+00)≈ 1.88E-09(9.02E-09)– 0.00E+00(0.00E+00) ≈ 9.20E-01(2.90E+00)– 0.00E+00(0.00E+00)
F4 2.90E+01(1.30E+01)– 2.81E+00(2.81E+00)– 1.41E+00(1.60E+00)– 6.90E+00(1.50E+01)– 2.60E+00(2.80E+00)– 0.00E+00(0.00E+00)
F5 1.40E+01(8.80E+00)+ 1.92E+01(1.92E+01)+ 2.00E+01(4.17E-02)≈ 1.90E+01(7.6E−1)+ 1.90E+01(1.2E−04)+ 2.00E+01(2.00E+01)
F6 1.80E-02(1.30E-01)+ 5.29E-02(5.29E-02)+ 7.06E-01(6.40E-01)+ 1.70E-01(4.0E−1)+ 3.00E+00(1.20E+00)– 1.26E+00(1.26E+00)
F7 3.00E-03(6.50E-03)+ 3.55E-02(3.55E-02)– 9.48E-02(4.92E-02)– 3.10E-02(2.1E−2)– 1.50E-01(1.3E−01)– 6.94E-03(6.94E-03)
F8 0.00E+00(0.00E+00)≈ 6.61E-01(6.61E-01)– 2.54E-01(8.09E-01)– 0.00E+00(0.00E+00) ≈ 0.00E+00(0.00E+00) ≈ 4.46E-09(4.46E-09)
F9 2.30E+00(8.40E-01)+ 8.52E+00(8.52E+00)– 6.01E+00(2.45E+00)– 7.30E+00(2.90E+00)– 2.00E+01(7.50E+00)– 2.80E+00(2.80E+00)
F10 8.60E-03(2.20E-02)+ 6.84E+01(6.84E+01)– 1.59E+00(2.08E+00)– 2.20E+00(2.90E+00)– 2.10E+02(1.00E+02)– 1.94E-01(1.94E-01)
F11 3.20E+01(3.80E+01)+ 2.91E+02(2.91E+02)– 3.72E+02(1.53E+02)– 3.00E+02(2.10E+02)– 3.90E+02(1.60E+02)– 1.62E+02(1.62E+02)
F12 6.80E-02(1.90E-02)+ 2.21E-01(2.21E-01)– 4.25E-02(4.78E-02)+ 1.60E-01(4.0E−2)– 1.30E-01(7.8E−02)+ 1.45E-01(1.45E-01)
F13 5.20E-02(1.50E-02)+ 1.28E-01(1.28E-01)– 1.21E-01(7.18E-02)– 4.70E-02(2.4E−2)+ 4.10E-01(1.8E−01)– 6.81E-02(6.81E-02)
F14 8.10E-02(2.60E-02)– 1.36E-01(1.36E-01)– 2.14E-01(1.20E-01)– 1.00E-01(3.5E−2)– 3.60E-01(1.9E−01)– 7.28E-02(7.28E-02)
F15 3.70E-01(6.90E-02)+ 9.83E-01(9.83E-01)– 7.75E-01(2.63E-01)– 6.30E-01(1.3E−1)– 2.40E+00(1.20E+00)– 4.79E-01(4.79E-01)
F16 1.20E+00(3.00E-01)+ 2.23E+00(2.23E+00)– 1.76E+00(4.68E-01)+ 1.80E+00(2.8E−1)+ 2.60E+00(3.9E−01)– 1.86E+00(1.86E+00)
F17 9.80E-01(1.10E+00)+ 4.77E+01(4.77E+01)– 2.55E+02(1.77E+02)– 1.30E+02(6.60E+01)– 6.80E+02(9.40E+02)– 1.36E+01(1.36E+01)
F18 2.40E-01(3.10E-01)+ 2.00E+00(2.00E+00)– 2.52E+01(1.83E+01)– 1.80E+00(2.40E+00)– 3.30E+01(4.50E+01)– 5.07E-01(5.07E-01)
F19 7.70E-02(6.40E-02)+ 1.03E+00(1.03E+00)+ 1.30E+00(7.75E-01)– 2.90E-01(5.4E−2)+ 9.30E-01(3.7E−01)+ 1.20E+00(1.20E+00)
F20 1.80E-01(1.80E-01)+ 7.21E-01(7.21E-01)– 1.34E+01(1.16E+01)– 4.70E-01(8.8E−1)– 8.90E+00(2.20E+01)– 3.35E-01(3.35E-01)
F21 4.10E-01(3.10E-01)+ 1.21E+00(1.21E+00)+ 9.46E+01(9.80E+01)– 3.80E+00(7.70E+00)– 5.70E+01(6.20E+01)– 1.34E+00(1.34E+00)
F22 4.40E-02(2.80E-02)+ 1.17E+01(1.17E+01)– 3.41E+01(4.40E+01)– 2.90E-01(1.8E−1)+ 1.70E+01(7.40E+00) – 1.07E+01(1.07E+01)
F23 3.30E+02(0.00E+00)– 3.29E+02(3.29E+02)≈ 3.29E+02(5.59E-08)≈ 3.20E+02(0.00E+00)+ 2.70E+02(1.00E+02)+ 3.29E+02(3.29E+02)
F24 1.10E+02(2.30E+00)– 1.19E+02(1.19E+02)– 1.27E+02(2.90E+01)– 1.10E+02(3.20E+00)- 1.30E+02(9.90E+00)– 1.09E+02(1.09E+02)
F25 1.30E+02(4.00E+01)+ 1.30E+02(1.30E+02)+ 1.79E+02(2.76E+01)+ 1.20E+02(8.70E+00)+ 1.40E+02(1.30E+01)+ 1.84E+02(1.84E+02)
F26 1.00E+02(1.60E-02)≈ 1.00E+02(1.00E+02)≈ 1.00E+02(7.47E-02)≈ 1.00E+02(2.0E−2) ≈ 1.00E+02(1.9E−01) ≈ 1.00E+02(1.00E+02)
F27 5.80E+01(1.30E+02)+ 9.12E+01(9.12E+01)+ 3.21E+02(1.21E+02)– 2.20E+00(4.9E−1)+ 7.40E+00(2.40E+00)+ 2.83E+02(2.83E+02)
F28 3.80E+02(3.20E+01)– 3.87E+02(3.87E+02)– 3.47E+02(4.76E+01)– 3.60E+02(6.40E+00) – 3.00E+02(1.9E−01)+ 3.06E+02(3.06E+02)
F29 2.20E+02(4.60E-01)≈ 2.13E+02(2.13E+02)– 2.12E+02(2.08E+01)– 2.20E+02(4.20E+00) – 2.10E+02(2.20E+01)– 2.02E+02(2.02E+02)
F30 4.60E+02(1.30E+01)– 5.05E+02(5.05E+02)– 3.94E+02(1.18E+02)– 5.00E+02(1.40E+01)– 3.80E+02(7.90E+01)– 2.24E+02(2.24E+02)
-/+/≈ 6/18/6 19/6/5 23/4/3 16/9/5 19/9/2
W. Zhang et al.
Table 7 Results (mean(std)) of MSHCSA and other algorithm on CEC2014 benchmark function with 30D

Fun L-SHADE RSDE FWA-DM CETMS OptBees MSHCSA

F1 0.00E+00(0.00E+00)≈ 1.50E+03(1.70E+03)- 2.76E+05(1.82E+05)– 0.00E+00(0.00E+00) ≈ 8.50E+04(3.00E+05)– 0.00E+00(0.00E+00)

F2 0.00E+00(0.00E+00)≈ 0.00E+00(5.99E-09)≈ 1.08E-16(1.87E-16)≈ 0.00E+00(0.00E+00) ≈ 0.00E+00(0.00E+00) ≈ 0.00E+00(0.00E+00)
F3 0.00E+00(0.00E+00)≈ 4.74E-02(1.16E-01)– 4.42E-16(4.74E-16)≈ 0.00E+00(0.00E+00) ≈ 8.40E-03(3.7e−2)– 0.00E+00(0.00E+00)
F4 0.00E+00(0.00E+00)+ 3.05E+00(1.34E+01)– 2.04E+01(1.91E+01)– 0.00E+00(0.00E+00)+ 1.20E+01(1.30E+01)– 1.18E+00(5.44E-01)
F5 2.00E+01(3.70E-02)≈ 2.03E+01(9.88E-02)– 2.05E+01(5.36E-02)– 2.00E+01(3.4e−1) ≈ 2.00E+01(1.0e−5) ≈ 2.00E+01(5.33E-05)
F6 1.40E-07(9.90E-07)+ 5.16E+00(2.01E+00)+ 1.29E+01(8.25E+00)+ 1.80E+00(1.70E+00)+ 1.60E+01(3.40E+00)+ 2.41E+01(9.47E-01)
F7 0.00E+00(0.00E+00)≈ 8.46E-04(1.59E-03)– 8.55E-03(9.81E-03)– 6.50E-01(2.60E+00)– 3.70E-02(3.8e−2)– 0.00E+00(0.00E+00)
F8 0.00E+00(0.00E+00)+ 2.04E+01(7.04E+00)– 1.13E-13(4.51E-13)+ 1.30E+01(3.30E+00)- 0.00E+00(0.00E+00)+ 5.24E+00(1.18E+00)
F9 6.80E+00(1.50E+00)+ 5.80E+01(1.65E+01)– 5.66E+01(1.08E+01)– 1.20E+01(5.20E+00)+ 1.30E+02(3.20E+01)– 2.57E+01(5.10E+00)
F10 1.60E-02(1.60E-02)+ 3.29E+02(2.47E+02)– 8.53E+00(2.42E+00)+ 2.00E+03(1.20E+03)- 1.00E+03(2.50E+02)– 2.17E+01(8.32E+00)
F11 1.20E+03(1.80E+02)+ 2.74E+03(6.44E+02)– 2.63E+03(2.48E+02)+ 2.50E+03(1.00E+03)+ 2.70E+03(5.60E+02)– 2.66E+03(3.09E+02)
F12 1.60E-01(2.30E-02)+ 4.44E-01(1.66E-01)– 3.71E-01(6.66E-02)+ 2.60E-03(3.1e−3)+ 1.80E-01(6.1e−2)+ 4.21E-01(9.78E-02)
F13 1.20E-01(1.70E-02)+ 3.05E-01(5.50E-02)– 3.89E-01(5.51E-02)– 5.30E-02(1.9e−2)+ 5.60E-01(1.4e−1)– 1.97E-01(3.21E-02)
F14 2.40E-01(3.00E-02)– 2.36E-01(3.37E-02)– 2.69E-01(7.76E-02)– 5.10E-01(5.6e−2)– 3.90E-01(2.3e−1)– 1.79E-01(2.76E-02)
F15 2.10E+00(2.50E-01)+ 5.92E+00(2.59E+00)– 7.37E+00(8.46E-01)– 2.00E+01(8.30E+01)– 1.20E+01(6.90E+00)– 3.39E+00(6.21E-01)
F16 8.50E+00(4.60E-01)+ 1.06E+01(7.70E-01)– 1.10E+01(2.71E-01)≈ 1.30E+01(1.10E+00)– 1.00E+01(6.9e−1)+ 1.10E+01(4.18E-01)
F17 1.90E+02(7.50E+01)+ 1.24E+03(3.79E+02)– 6.29E+03(5.95E+03)– 1.40E+03(4.40E+02)– 2.70E+04(4.00E+04)– 1.07E+03(1.56E+02)
F18 5.90E+00(2.90E+00)+ 9.54E+01(4.34E+01)– 7.67E+01(3.66E+01)– 1.50E+02(4.70E+01)– 1.90E+02(4.70E+02)– 1.78E+01(3.32E+00)
F19 3.70E+00(6.80E-01)+ 5.65E+00(1.46E+00)+ 9.95E+00(1.93E+00)+ 7.40E+00(1.90E+00)+ 7.80E+00(1.80E+00)+ 1.25E+01(1.01E+00)
F20 3.10E+00(1.50E+00)+ 3.73E+01(2.55E+01)– 4.28E+01(2.61E+01)– 1.10E+02(1.00E+02)– 8.50E+02(7.70E+02)– 1.20E+01(1.44E+00)
F21 8.70E+01(9.00E+01)+ 4.71E+02(2.34E+02)– 7.29E+02(9.49E+02)– 1.00E+03(2.80E+03)– 1.70E+04(1.80E+04)– 3.29E+02(9.73E+01)
A hybrid clonal selection algorithm with modified combinatorial recombination and success-history...

F22 2.80E+01(1.80E+01)+ 1.91E+02(1.19E+02)– 1.46E+02(8.83E+01)– 2.00E+02(1.40E+02)– 2.30E+02(9.20E+01)– 6.23E+01(3.75E+01)

F23 3.20E+02(0.00E+00)– 3.15E+02(1.62E-06)– 3.14E+02(9.28E-14)≈ 3.10E+02(3.4e−13)+ 3.10E+02(6.6e−2)+ 3.14E+02(0.00E+00)
F24 2.20E+02(1.10E+00)– 2.24E+02(1.65E+00)– 2.26E+02(3.59E+00)– 2.20E+02(6.50E+00)– 2.30E+02(5.40E+00-) 2.10E+02(1.12E+01)
F25 2.00E+02(5.00E-02)≈ 2.03E+02(1.17E-01)– 2.01E+02(1.98E-01)– 2.00E+02(1.40E+00) ≈ 2.00E+02(1.6e−1) ≈ 2.00E+02(5.28E-04)
F26 1.00E+02(1.60E-02)≈ 1.00E+02(4.14E-02)≈ 1.00E+02(5.35E-02)≈ 1.00E+02(2.20E+01) ≈ 1.00E+02(1.7e−1) ≈ 1.00E+02(2.81E-02)
F27 3.00E+02(0.00E+00)+ 4.69E+02(9.46E+01)+ 4.01E+02(3.06E+01)+ 3.70E+02(4.90E+01)+ 4.00E+02(9.7e−1)+ 9.05E+02(1.85E+02)
F28 8.40E+02(1.40E+01)– 9.05E+02(1.21E+02)– 3.93E+02(1.46E+01)– 9.30E+02(1.20E+02)– 4.30E+02(1.50E+01)– 3.67E+02(4.63E-01)
F29 7.20E+02(5.10E+00)– 6.52E+05(2.66E+06)– 2.11E+02(2.90E+00)+ 1.50E+06(3.40E+06)– 2.10E+02(1.10E+00)+ 2.14E+02(6.42E-01)
F30 1.20E+03(6.20E+02)– 1.70E+03(8.67E+02)– 4.51E+02(1.96E+02)– 2.90E+03(1.10E+03)– 5.90E+02(9.80E+01)– 3.25E+02(4.51E+01)
−/ + / ≈ 6/17/7 24/4/2 17/8/5 15/9//6 18/8/4
Table 8 Results (mean(std)) of MSHCSA and other algorithm on CEC2014 benchmark function with 50D

Fun L-SHADE RSDE FWA-DM CETMS OptBees MSHCSA

F1 1.20E+03(1.50E+03)– 2.25E+04(1.21E+04)– 6.15E+06(2.39E+06)– 0.00E+00(0.00E+00) ≈ 1.30E+05(9.70E+04)– 0.00E+00(0.00E+00)

F2 0.00E+00(0.00E+00)+ 3.58E+03(6.56E+03)– 4.83E+03(6.06E+03)– 0.00E+00(0.00E+00) + 1.00E-03(2.9e−3)– 2.51E-07(1.77E-07)
F3 0.00E+00(0.00E+00)≈ 4.10E-01(1.38E+00)– 7.66E+01(2.46E+01)– 0.00E+00(0.00E+00) ≈ 1.70E+02(1.70E+02)– 0.00E+00(0.00E+00)
F4 5.90E+01(4.60E+01)– 6.41E+01(3.62E+01)– 5.05E+01(1.88E+01)– 7.10E+01(4.40E+01)– 3.80E+01(3.40E+01)– 2.72E+01(8.12E-01)
F5 2.00E+01(4.60E-02)≈ 2.05E+01(9.61E-02)– 2.07E+01(3.61E-02)– 2.00E+01(3.3e−1) ≈ 2.00E+01(2.5e−5) ≈ 2.00E+01(2.44E-04)
F6 2.60E-01(5.20E-01)+ 1.72E+01(4.33E+00)+ 4.39E+01(2.31E+00)+ 4.20E+00(2.60E+00) + 3.00E+01(4.50E+00)+ 5.09E+01(1.79E+00)
F7 0.00E+00(0.00E+00)≈ 2.40E-03(4.40E-03)– 2.66E-03(4.25E-03)– 4.50E-01(1.90E+00)– 2.30E-02(3.9e−2)– 0.00E+00(0.00E+00)
F8 2.60E-09(7.50E-09)+ 5.04E+01(1.24E+01)– 9.17E+00(2.39E+00)+ 2.50E+01(5.50E+00)+ 0.00E+00(0.00E+00)+ 3.26E+01(3.80E+00)
F9 1.10E+01(2.10E+00)+ 1.42E+02(3.51E+01)– 1.47E+02(1.57E+01)– 2.40E+01(8.10E+00)+ 2.40E+02(4.80E+01)– 7.68E+01(8.82E+00)
F10 1.20E-01(4.10E-02)+ 1.52E+03(9.46E+02)– 6.14E+02(1.42E+02)+ 3.90E+03(1.20E+03)– 1.80E+03(3.40E+02)– 1.51E+03(2.24E+02)
F11 3.20E+03(3.30E+02)+ 6.15E+03(1.05E+03)– 5.43E+03(4.02E+02)+ 4.20E+03(1.50E+03)+ 5.10E+03(6.80E+02)+ 5.79E+03(4.10E+02)
F12 2.20E-01(2.80E-02)+ 5.38E-01(2.02E-01)+ 6.02E-01(8.72E-02)+ 1.50E-03(1.1e−3)+ 1.60E-01(4.6e−2)+ 6.82E-01(1.02E-01)
F13 1.60E-01(1.80E-02)+ 4.23E-01(6.05E-02)– 4.88E-01(4.63E-02)– 1.10E-01(4.2e−2)+ 6.00E-01(1.1e−1)– 2.84E-01(2.99E-02)
F14 3.00E-01(2.50E-02)– 2.78E-01(2.25E-02)– 3.30E-01(1.05E-01)– 5.00E-01(6.6e−2)– 4.70E-01(2.3e−1)– 2.22E-01(4.02E-02)
F15 5.20E+00(5.10E-01)+ 9.96E+00(6.53E+00)– 2.08E+01(1.45E+00)– 1.00E+01(2.20E+01)– 2.60E+01(9.10E+00)– 9.19E+00(1.24E+00)
F16 1.70E+01(4.80E-01)+ 1.93E+01(8.42E-01)+ 2.00E+01(3.21E-01)+ 2.20E+01(1.50E+00)– 1.80E+01(8.7e−1)+ 2.03E+01(3.24E-01)
F17 1.40E+03(5.10E+02)+ 4.10E+03(1.68E+03)– 1.08E+05(6.03E+04)– 2.10E+03(5.40E+02)+ 3.20E+04(3.50E+04)– 2.64E+03(2.99E+02)
F18 9.70E+01(1.40E+01)– 3.40E+02(2.38E+02)– 3.31E+03(3.72E+03)– 2.60E+02(6.40E+01)– 6.80E+02(1.00E+03)– 7.78E+01(1.02E+01)
F19 8.30E+00(1.80E+00)+ 1.46E+01(2.03E+00)+ 2.11E+01(3.17E+00)– 1.20E+01(3.10E+00)+ 1.50E+01(2.70E+00)+ 1.96E+01(1.56E+00)
F20 1.40E+01(4.60E+00)+ 1.60E+02(7.30E+01)– 4.00E+02(8.12E+01)– 3.50E+02(1.50E+02)– 1.90E+03(1.40E+03)– 4.89E+01(5.76E+00)
F21 5.20E+02(1.50E+02)+ 1.58E+03(6.55E+02)– 2.55E+04(1.58E+04)– 1.30E+03(4.20E+02)+ 6.00E+04(4.10E+04)– 1.50E+03(2.48E+02)
F22 1.10E+02(7.50E+01)+ 4.61E+02(2.34E+02)+ 5.45E+02(1.25E+02)– 3.60E+02(1.90E+02)+ 7.50E+02(2.00E+02)– 4.63E+02(1.32E+02)
F23 3.40E+02(4.40E-13)– 3.44E+02(1.19E-05)– 3.37E+02(1.31E-05)≈ 3.40E+02(2.8e−13)– 3.30E+02(2.8e−1)+ 3.37E+02(2.85E-13)
F24 2.80E+02(6.60E-01)– 2.76E+02(2.38E+00)– 2.66E+02(3.48E+00)– 2.70E+02(1.10E+00)– 2.60E+02(3.20E+00)+ 2.64E+02(3.30E-01)
F25 2.10E+02(3.60E-01)– 2.06E+02(7.81E-01)– 2.05E+02(2.26E+00)– 2.00E+02(2.70E+00) ≈ 2.00E+02(6.0e−1) ≈ 2.00E+02(1.59E-02)
F26 1.00E+02(1.40E+01)≈ 1.12E+02(4.97E+01)– 1.00E+02(5.98E-02)≈ 1.70E+02(1.00E+02)– 1.00E+02(1.2e−1) ≈ 1.00E+02(2.83E-02)
F27 3.30E+02(3.00E+01)+ 8.04E+02(1.00E+02)+ 1.45E+03(7.27E+01)+ 4.50E+02(7.30E+01)+ 1.10E+03(2.10E+02)+ 1.73E+03(3.70E+01)
F28 1.10E+03(2.90E+01)- 1.61E+03(3.90E+02)– 4.01E+02(3.06E+01)– 1.20E+03(8.60E+01)– 4.40E+02(3.80E+01)– 3.57E+02(4.41E-01)
F29 7.90E+02(2.40E+01)- 5.28E+06(1.64E+07)– 2.23E+02(2.67E+00)+ 2.20E+06(9.10E+06)– 2.30E+02(2.30E+00)– 2.29E+02(8.98E-01)
F30 8.70E+03(4.10E+02)- 1.12E+04(1.75E+03)– 8.36E+02(3.06E+02)+ 9.80E+03(1.70E+03)– 8.10E+02(1.20E+02)+ 9.58E+02(1.77E+02)
−/ + / ≈ 10/16/4 24/6/0 19/9/2 14/12/4 17/10/3
W. Zhang et al.
Table 9 Results (mean(std)) of MSHCSA and other algorithm on CEC2014 benchmark function with 100D

Fun L-SHADE RSDE FWA-DE CETMS OptBees MSHCSA

F1 1.70E+05(5.70E+04)– 8.33E+05(2.89E+05)– 2.28E+08(4.08E+07)– 0.00E+00(0.00E+00)+ 2.90E+05(1.00E+05)– 9.95E+04(4.26E+04)

F2 0.00E+00(0.00E+00)+ 7.39E+03(9.84E+03)– 1.62E+04(1.81E+04)– 0.00E+00(0.00E+00)+ 1.00E+01(2.90E+01)– 2.46E+00(1.21E+00)
F3 0.00E+00(0.00E+00)+ 9.77E-01(2.21E+00)– 2.95E+04(3.64E+03)– 0.00E+00(0.00E+00)+ 7.50E+02(7.80E+02)– 6.59E-07(3.64E-07)
F4 1.70E+02(3.10E+01)– 1.86E+02(4.06E+01)– 1.83E+02(1.00E+02)– 1.70E+02(4.10E+01)– 1.40E+02(5.40E+01)– 8.44E+01(6.18E-01)
F5 2.10E+01(3.10E-02)– 2.08E+01(7.85E-02)– 2.10E+01(2.44E-02)– 2.00E+01(3.2e−1) ≈ 2.00E+01(2.4e−5) ≈ 2.00E+01(1.87E-03)
F6 8.70E+00(2.30E+00)+ 6.02E+01(7.48E+00)+ 1.14E+02(2.54E+00)+ 1.00E+01(4.80E+00)+ 7.10E+01(7.70E+00)+ 1.29E+02(2.61E+00)
F7 0.00E+00(0.00E+00)≈ 1.27E-03(2.71E-03)– 1.29E-01(3.14E-02)– 1.90E-01(9.6e−1)– 5.80E-03(8.0e−3)– 0.00E+00(0.00E+00)
F8 1.10E-02(7.40E-03)+ 1.94E+02(3.02E+01)+ 1.08E+02(5.40E+00)+ 5.90E+01(1.20E+01)+ 0.00E+00(0.00E+00)+ 2.36E+02(1.64E+01)
F9 3.40E+01(5.00E+00)+ 3.20E+02(5.41E+01)- 5.52E+02(4.44E+01)– 6.20E+01(8.60E+00)+ 6.60E+02(9.70E+01)– 3.17E+02(1.82E+01)
F10 2.60E+01(5.80E+00)+ 9.31E+03(2.67E+03)+ 5.67E+03(3.04E+02)+ 9.10E+03(1.60E+03)+ 4.20E+03(4.20E+02)+ 1.11E+04(5.54E+02)
F11 1.10E+04(5.60E+02)+ 1.55E+04(1.54E+03)+ 1.46E+04(6.30E+02)+ 1.00E+04(1.80E+03)+ 1.20E+04(1.10E+03)+ 1.69E+04(5.54E+02)
F12 4.40E-01(4.70E-02)+ 7.42E-01(1.97E-01)+ 1.21E+00(7.78E-02)+ 7.40E-04(3.7e−4)+ 2.30E-01(5.0e−2)+ 1.33E+00(1.34E-01)
F13 2.40E-01(2.10E-02)+ 5.44E-01(4.09E-02)– 5.61E-01(3.61E-02)– 2.00E-01(4.1e−2)+ 5.80E-01(8.1e−2)– 3.67E-01(3.12E-02)
F14 1.20E-01(7.30E-03)+ 2.09E-01(1.23E-02)+ 1.89E-01(2.06E-02)+ 2.80E-01(2.4e−2)+ 2.20E-01(2.2e−2)+ 3.12E-01(9.89E-02)
F15 1.60E+01(1.20E+00)+ 5.24E+01(1.82E+01)– 8.74E+01(5.87E+00)– 2.20E+01(4.70E+01)+ 6.50E+01(1.80E+01)– 3.00E+01(1.84E+00)
F16 3.90E+01(4.80E-01)+ 4.24E+01(1.21E+00)+ 4.35E+01(3.75E-01)≈ 4.60E+01(1.10E+00)– 4.00E+01(1.10E+00)+ 4.35E+01(4.63E-01)
F17 4.40E+03(7.10E+02)+ 9.86E+04(4.60E+04)– 2.31E+07(5.63E+06)– 5.30E+03(6.90E+02)+ 1.00E+05(6.70E+04)– 5.87E+03(6.98E+02)
F18 2.20E+02(1.70E+01)+ 1.26E+03(1.08E+03)– 5.68E+03(8.70E+03)– 5.90E+02(1.00E+02)– 1.50E+03(2.10E+03)– 2.82E+02(9.44E+01)
F19 9.60E+01(2.30E+00)– 8.16E+01(2.58E+01)– 6.34E+01(2.43E+00)– 1.00E+02(1.60E+01)– 5.20E+01(1.50E+01)– 4.31E+01(3.35E-01)
F20 1.50E+02(5.20E+01)+ 5.50E+02(1.76E+02)– 6.93E+04(1.10E+04)– 7.10E+02(1.60E+02)– 1.00E+04(4.20E+03)– 2.18E+02(1.49E+01)
F21 2.30E+03(5.30E+02)+ 3.49E+04(1.90E+04)– 9.57E+06(2.31E+06)– 3.10E+03(5.40E+02)+ 3.10E+05(1.50E+05)– 4.76E+03(3.41E+02)
A hybrid clonal selection algorithm with modified combinatorial recombination and success-history...

F22 1.10E+03(1.90E+02)+ 1.51E+03(4.63E+02)+ 1.51E+03(1.34E+02)+ 7.60E+02(3.60E+02)+ 2.00E+03(3.30E+02)– 1.53E+03(3.35E+02)

F23 3.50E+02(2.80E-13)– 3.48E+02(2.12E-03)– 3.46E+02(2.18E-01)– 3.40E+02(7.6e−11)+ 3.40E+02(9.2e−1)+ 3.45E+02(1.05E-12)
F24 3.90E+02(2.90E+00)– 4.06E+02(5.67E+00)– 3.63E+02(2.85E+00)+ 3.90E+02(5.90E+00)– 3.40E+02(1.00E+01)+ 3.84E+02(2.91E+00)
F25 2.00E+02(4.00E-13)+ 2.42E+02(7.50E+00)– 3.03E+02(1.74E+01)– 2.20E+02(6.00E+00)– 2.00E+02(1.10E+00)+ 2.01E+02(1.70E-01)
F26 2.00E+02(6.20E-13)– 1.98E+02(4.97E+01)– 1.61E+02(5.88E+01)– 2.00E+02(6.60E+01)– 1.00E+02(6.8e−2) ≈ 1.00E+02(2.89E-02)
F27 3.80E+02(3.30E+01)+ 2.01E+03(1.65E+02)+ 3.14E+03(4.13E+02)+ 5.90E+02(1.00E+02)+ 2.10E+03(1.60E+02)+ 3.71E+03(5.61E+01)
F28 2.30E+03(4.60E+01)– 4.11E+03(7.20E+02)– 1.60E+03(3.42E+02)– 2.50E+03(2.80E+02)– 6.10E+02(4.00E+01)– 4.04E+02(5.30E+00)
F29 8.00E+02(7.60E+01)– 8.29E+07(8.20E+07)– 2.70E+02(5.23E+00)– 7.40E+02(2.80E+01)– 2.70E+02(3.20E+00)– 2.43E+02(4.22E+00)
F30 8.30E+03(9.60E+02)– 1.31E+04(2.31E+03)– 2.23E+03(1.16E+03)+ 8.80E+03(1.00E+03)– 2.80E+03(2.40E+00)+ 2.85E+03(2.73E+02)
-/+/≈ 10/19/1 21/9/0 19/10/1 12/17/1 16/14/2

employed in accordance with their respective literatures.
The experiments results on all the 30 test function from
CEC2014 are performed with 10D, 30D, 50D and 100D are
displayed in Tables 6, 7, 8 and 9.
When dimension is 10D as is shown in Table 6, four
algorithms (L-SHADE, RSDE, CETMS and MSHCSA)
could achieve the global optima in all of the unimodal
functions (F1-F3). Only MSHCSA succeeds in locating the
global optimum in F4, From F5 to F11, the performance
of L-SHADE is the best among the comparing algorithms.
Comparing the other algorithms, MSHCSA does not
show advantages but is generally acceptable. To the
hybrid functions, the L-SHADE presents the overwhelming
advantage to the other algorithms. MSHCSA is better than
FWA-DM (on all hybrid functions F17-F22) and OptBees
(F17-F22 except F19). To the composition functions (F23-
F30), MSHCSA gains the best results on four functions
(F24, F26, F29, and F30). As analyzed in Section 4.2,
modified success-history based adaptive mutation is not
good at coping with multimodal functions, but shows
advantages in unimodal and composition functions.
When dimension is 30D as is shown in Table 7, only three
algorithms (L-SHADE, CETMS and MSHCSA) achieve
the global optima in all of the unimodal functions (F1-
F3). Even though L-SHADE still maintains an advantage
in most of the multimodal functions, MSHCSA gets the
best results on three functions (F5, F7 and F14). There is
not much difference with dimension is 10D in handling
Hybrid functions (F17-F22). MSHCSA achieves the best
performance on 5 out of 8 functions in composition
functions (F23-F30).
When dimension is 50D as is shown in Table 8, only
CTEMS achieve the global optima in all of the unimodal
functions (F1-F3). MSHCSA obtains the global optimum
in F1 and F3, and a little error (2.51E-07) in F2 which
is acceptable. The advantage of L-SHADE in handing the
multimodal functions continues to decline, which presents
the best results on 8 out of 16 functions. MSHCSA gets four
best results (F4, F5, F7, and F14). To the hybrid functions
(F17-F22), L-SHADE wins the comparison on 5 out of 6
function except F18 which MSHCSA gets the best results.
When comparing the composition functions, OptBees and
MSHCSA are ahead of the other algorithms.
When dimension is 100D as is shown in Table 9, CTEMS
keeps the advantage of locating the global optimum in
all of the unimodal functions (F1-F3). MSHCSA does not
obtain the satisfied results. However, it is superior to RSDE,
FWA-DE and OptBees in handling the unimodal functions.
To the multimodal functions (F4-F16), L-SHADE gets the
best results on 8 out of 16 functions, while MSHCSA
obtains best results on 3 functions. Even the performance
of all compared algorithms are unsatisfactory in handling
W. Zhang et al.
the hybrid functions, L-SHADE shows the great advantage.
MSHCSA is better than OptBees on all of the functions,
superior to RSDE and FWA-DE on 5 out of 6 functions
and gets even to CTEMS. OptBees and MSHCSA keep the
relatively better performance than the other algorithms in
coping with the composition functions.
Looking through the overall performance of the algo-
rithms on the benchmark function with different dimen-
sions, we can make a summary. As indicated by the Free
lunch theory, every algorithm presents its own advantages
on different aspect. Moreover, with the increase of dimen-
sion, the performance of most algorithms gradually dete-
riorates. Even though, the proposed MSHCSA obtains the
competitive results on most of the CEC2014 benchmark
functions. On unimodal group, MSHCSA gets the best
results with 10D and 30D, and the suboptimal results with
50D. Even when the dimension is 100D, MSHCSA is better
than RSDE, FWA-DE and OptBees. Moreover, MSHCSA
shows the potential capability of handling the composition
functions. On the whole, MSHCSA is superior to RSDE,
FWA-DE and OptBees on most of the benchmark func-
tion with different dimensions, and attains better results
than CTEMS on most of the benchmark functions with low
dimensions. The satisfactory performance of MSHCSA is
due to the introduction of the modified recombination and
success-history based adaptive mutation, where the former
brings the diversity to the population and the later improve
the searching ability. Moreover, Gene knockout strategy fur-
ther improves the precision and help jump out of stagnation.
Therefore, the proposed MSHCSA is able to solving the
complicated optimization problem effectively.
5 Conclusions
In this paper, a novel clonal selection algorithm denoted as
MSHCSA is proposed. The combinatorial recombination is
introduced and modified to bring more diversity while the
success-history based adaptive clonal selection is developed
to enhance the search ability. Moreover the Gene knockout
strategy is proposed to further improve the accuracy and
avoid stagnation. The proposed algorithm is tested on the
CEC 2014 benchmark function, which covers unimodal,
multimodal, composing test problems by extracting features
dimension-wise from several problems, graded level of
linkages, rotated trap problems with dimension ranges from
10D to 100D. Two main introduced strategies are analyzed
at first, and then the proposed algorithm is compared with
the state of the art algorithms. The experimental results
show the newly developed strategies greatly improve the
performance of the algorithm. To compare with other
algorithms, the MSHCSA is very competitive.
A hybrid clonal selection algorithm with modified combinatorial recombination and success-history...
Acknowledgements This work is supported by National Natural Sci-
ence Foundation of China (No. 61403349, 61501405, 41601418),
Funding program for key scientific research projects of universities in
Henan province (No. 18A210025), Science and technology research
key project of basic research projects in education department of
Henan province (No.15A520033, No.14B520066), doctoral founda-
tion(No. 2013BSJJ044), and Student science and technology activity
project of Zhengzhou university of light industry.
References
1. Mishra A (2017) Nature inspired algorithms: a survey of the state
of the art. Int J Adv Res Comput Sci Manag Stud 5(9):16–21
2. Boussaı̈d I, Lepagnot J, Siarry P (2013) A survey on optimization
metaheuristics. Inform Sci 237(237):82–117
3. Silva GC, Dasgupta D (2016) A survey of recent works in artificial
immune systems. Handbook on computational intelligence: vol 2:
evolutionary computation, hybrid systems, and applications, pp
547–586
4. Yin C, Ma L, Feng L (2017) Towards accurate intrusion detection
based on improved clonal selection algorithm. Multimed Tools
Appl 76(19):19397–19410
5. Anuar S, Sallehuddin R, Selamat A (2016) Implementation of arti-
ficial neural network on graphics processing unit for classification
problems. In: International conference on computational collective
intelligence, pp 303–310
6. Idris I, Selamat A, Nguyen NT, Omatu S, Krejcar O (2015) A
combined negative selection algorithm-particle swarm optimiza-
tion for an email spam detection system. Eng Appl Artif Intel
39:33–44
7. Idris I, Selamat A (2014) Improved email spam detection model
with negative selection algorithm and particle swarm optimization.
Appl Soft Comput 22(5):11–27
8. Avatefipour O, Nafisian A (2018) A novel electric load con-
sumption prediction and feature selection model based on mod-
ified clonal selection algorithm. J Intell Fuzzy Syst 34(4):2261–
2272
9. Xu N, Ding Y, Ren L, Hao K (2017) Degeneration recognizing
clonal selection algorithm for multimodal optimization. IEEE
Trans Cybern 48(3):848–861
10. Haktanirlar Ulutas B, Kulturel-Konak S (2011) A review of
clonal selection algorithm and its applications. Artif Intell Rev
36(2):117–138
11. De Castro LN, Timmis J (2002) An artificial immune network
for multimodal function optimization. In: Proceedings of IEEE
conference on evolutionary computation, pp 699–674
12. De Castro LN, Zuben FJV (1999) Artificial immune systems:
part I–basic theory and applications, Universidade Estadual de
Campinas, Dezembro de, Technical report 210
13. de Castro LN, von Zuben FJ (2002) Learning and optimization
using the clonal selection principle. IEEE Trans Evol Comput
6(3):239–251
14. Gong MM, Jiao L, Zhang L (2010) Baldwinian learning in clonal
selection algorithm for optimization. Inf Sci 180(8):1218–1236
15. Gao S, Chai H, Chen B (2013) Hybrid gravitational search and
clonal selection algorithm for global optimization, advances in
swarm intelligence, vol 7929. Springer, Berlin, pp 1–10
16. Yang G, Jin H (2013) Optimization algorithm based on differential
evolution and clonal selection mechanism. Comput Eng Appl
49(10):50–49
17. Khilwani N, Prakash A, Shankar R, Tiwari M (2008) Fast clonal
algorithm. Eng Appl Artif Intell 21(1):106–128
18. Jiao L, Li Y, Gong M (2008) Quantum-inspired immune clonal
algorithm for global optimization. IEEE Trans Syst Man Cybern
Part B Cybern Publ IEEE Syst Man Cybern Soc 38(5):1234–
1253
19. Cutello V, Nicosia G, Povene M (2006) Real coded clonal
selection algorithm for unconstrained global optimization using a
hybrid inversely proportional hypermutation operator. Acm Symp
Appl Comput 2:950–954
20. Peng Y, Lu B (2015) Hybrid learning clonal selection algorithm.
Inf Sci 296(1):128–146
21. Gong M, Jiao L, Yang J, Liu F (2010) Lamarckian learning in
clonal selection algorithm for numerical optimization. Int J Artif
Intell Tools 19(1):19–37
22. Liu Z, Li X, Zhang J (2013) Co-evolutionary particle swarm
optimization algorithm based on elite immune clonal selection.
Acta Electronica Sinica 41(11):2167–2173
23. Li R, Zhan W, Hao Z (2017) Artificial immune particle
swarm optimization algorithm based on clonal selection. Boletin
Tecnico/technical Bull 55(1):158–164
24. Zhang F, Wang C, Yang Z (2017) Clonal selection algorithm based
on vaccination and Cauchy mutation. Computer Engineering &
Applications
25. Liang L, Xu G, Liu D, Zhao S (2007) Immune clonal
selection optimization method with mixed mutation strategies. In:
International conference on bio-inspired computing: theories and
applications. Bic-Ta IEEE, pp 37–41
26. Qia Y, Houa Z, Yina M, Sunb H, Huang J (2015) An immune
multi-objective optimization algorithm with differential evolution
inspired recombination. Appl Soft Comput 29:395–410
27. Zaharie D (2009) Influence of crossover on the behavior of
differential evolution algorithms. Appl Soft Comput 9(3):1126–
1138
28. Liu R, Ma C, He F, Ma W, Jiao L (2014) Reference direction
based immune clone algorithm for many-objective optimization.
Front Comput Sci 8(4):642–655
29. Zhang W, Lin J, Jing H, Zhang Q (2016) A novel hybrid
clonal selection algorithm with combinatorial recombination
and modified hypermutation operators for global optimization.
Comput Intell Neurosci 2016:12
30. Pavone M, Narzisi G, Nicosia G (2012) Clonal selection: an
immunological algorithm for global optimization over continuous
spaces. J Glob Optim 53(4):769–808
31. Tanabe R, Fukunaga A (2013) Success-history based parameter
adaptation for differential evolution. In: 2013 IEEE Congress on
evolutionary computation, pp 71–78
32. Tanabe R, Fukunaga A (2013) Evaluating the performance of
SHADE on CEC 2013 benchmark problems. In: 2013 IEEE
Congress on evolutionary computation, pp 1952–1959
33. Liang J, Qu B, Suganthan P (2013) Problem definitions and
evaluation criteria for the CEC 2014 special session and competi-
tion on single objective real-parameter numerical optimization
34. Tanabe R, Fukunaga A (2014) Improving the search performance
of SHADE using linear population size reduction. In: 2014 IEEE
Congress on evolutionary computation, pp 1658–1665
35. Xu C, Huang H, Ye S (2014) A differential evolution with
replacement strategy for real-parameter numerical optimization.
In: 2014 IEEE Congress on evolutionary computation, pp 1617–
1624
36. Yu C, Kelley L, Zheng S, Tan Y (2014) Fireworks algorithm
with differential mutation for solving the CEC 2014 competition
problems. In: 2014 IEEE Congress on evolutionary computation,
pp 3238–324
37. Liu C, Fan L (2016) A hybrid evolutionary algorithm based on
tissue membrane systems and CMA-ES for solving numerical
optimization problems. Knowl-Based Syst 105:38–47
38. Maia RD, de Castro LN, Caminhas WM (2014) Real-parameter
optimization with optbees. In: 2014 IEEE Congress on evolution-
ary computation, pp 2649–2655

Weiwei Zhang is currently an
assistant professor in Zheng-
zhou University of Light In-
dustry. She received the M.S.
and Ph.D. degree in computer
science from Chongqing Uni-
versity, Chongqing, China, in
2008 and 2013 respectively.
Her current research interests
include evolutionary compu-
tation, dynamic optimization,
image processing, and intelli-
gent control.
Kui Gao received the B.S. deg-
ree in information technol-
ogy from Zhengzhou Univer-
sity of Light Industry in the
summer of 2018 and will be
a graduate student in He’nan
University of technology in
the coming autumn. His cur-
rent research interests include
artificial immune system and
robotic control.
Weizheng Zhang is current-
ly an assistant professor in
Zhengzhou University of Light
Industry. He received the M.S.
degree in Signal and Infor-
mation Processing from Shan-
dong University of Science
and Technology, and Ph.D.
degree in Agricultural Electri-
fication and Automation from
Zhejiang University, China, in
2012 and 2016 respectively.
His current research interests
include intelligent computing,
agricultural engineering, and
image processing.
W. Zhang et al.
Xiao Wang received the B.S.
degree from Henan Normal
University in 2004, the M.S.
degree from the University of
Electronic Science and Tech-
nology of China in 2007 and
the PhD degree from Tongji
University in 2013. He is cur-
rently an assistant professor at
the School of Computer and
Communication Engineering,
Zhengzhou University of
Light Industry, Zhengzhou,
China. His research interests
include evolutionary optimi-
zation, machine learning and
bioinformatics.
Qiuwen Zhang received the
Ph.D. degree in communica-
tion and information systems
from Shanghai University,
Shanghai, China, in 2012.
Since 2012, he has been with
the faculty of the College of
Computer and Communica-
tion Engineering, Zhengzhou
University of Light Industry,
where he is currently an Asso-
ciate Professor. His major
research interests include 3D
signal processing, pattern
recognition, evolutionary opti-
mization, and multimedia
communication.
Hua Wang is currently an assis-
tant professor at the School
of Computer and Communica-
tion Engineering, Zhengzhou
University of Light Industry,
Zhengzhou, China. He recei-
ved the B.S. and Ph.D. deg-
ree from Wuhan University,
Wuhan, China, in 2008 and
2013 respectively. His current
research interests include evo-
lutionary computation, geo-
graphic information system,
and image processing.