Received 29 October 2023, accepted 13 November 2023, date of publication 20 November 2023,

date of current version 29 November 2023.

Digital Object Identifier 10.1109/ACCESS.2023.3335235

Differential Evolution With Exponential

Crossover: An Experimental Analysis
on Numerical Optimization
JUNCAN LI AND ZHENYU MENG , (Member, IEEE)
Institute of Artificial Intelligence, Fujian University of Technology, Fuzhou 350118, China
Corresponding author: Zhenyu Meng ([email protected])
This work was supported in part by the Natural Science Foundation of Fujian Province under the Grant 2021J05227, and in part by the
Fujian University of Technology under the Grant GY-Z19013.

ABSTRACT Since the inception in 1995, Differential Evolution (DE) has gained significant attention from
researchers worldwide, and many DE variants proposed in the last decades obtained excellent performance
in many scientific and engineering applications. However, the vast majority of these well-performed DE
variants employ binomial crossover rather than exponential crossover in generating trial vector candidates
though both of the two schemes were proposed simultaneously. That may be also the reason why there
still doesn’t exist such a thorough analysis of DE with exponential crossover. Different from the majority
of DE researchers believing that DE variants with binomial crossover usually exhibit superior performance
than the ones employing exponential crossover and DE variants with exponential crossover are good at
tackling optimization problems with linkages among neighboring variables, we found that DE variants with
exponential crossover can achieve competitive performance with the ones employing binomial crossover
regardless of whether there are linkages among the variables or not after discovering the proper crossover rate
CR and its corresponding parameter control. In order to enrich research of DE on exponential crossover, this
paper presents an thorough experimental analysis of DE algorithm with exponential crossover in numerical
optimization, presenting the basic concepts of exponential crossover, giving the mathematical analysis
why they can actually achieve similar optimization between exponential crossover and binomial crossover,
experimental validation under the 100 benchmark functions from CEC2013, CEC2014, CEC2017, CEC2022
test suites as well as the tension/compression spring design problem, and summarizing and classifying
various engineering applications that exponential crossover DEs are used to solve. Furthermore, we also
look into the future challenges and potential directions for further development of DE with exponential
crossover.

INDEX TERMS Differential evolution, exponential crossover, binomial crossover, algorithms validation.

I. INTRODUCTION energy of electrons [2]. Since humans appeared on earth,

Optimization is the core of many natural processes, essen- people have been continuously working on improving various
tially following the evolutionary law of ‘‘natural selection optimization needs related to the field of human life [3]. This
and survival of the fittest’’ [1]. Over millions of years, each continuous improvement process is, in the final analysis, the
species must adapt its physical structures to its environment process of analyzing the problem and finding the optimal
in a direction that is best suited to its natural environment, solution [4]. Generally speaking, building an optimization
and even atoms in the microscopic world are always striving model is the first step in solving a real-world optimization
to build their chemical bonds in a way that minimizes the problem, however, due to the diversity and complexity
of optimization models, the optimization method becomes
The associate editor coordinating the review of this manuscript and an important approach to solve complex optimization
approving it for publication was Valerio Freschi . problems.
2023 The Authors. This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License.
For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
VOLUME 11, 2023 131677
 J. Li, Z. Meng: DE With Exponential Crossover: An Experimental Analysis

In essence, the optimization problem is a search problem, that the mutate vector contributes to the target vector. After
and the optimization method is used to solve one. The choosing n and calculating L, using the following scheme to
solution to the optimization problem typically begins with yield the trial vector:
designing the objective function of the corresponding model,
vj,i,G , for j = ⟨n⟩D , . . . , ⟨n + L − 1⟩D

and optimization is defined as the process of discovering the
uj,i,G = (2)
minimum or maximum value of this objective function [5], xj,i,G , for all other j ∈ [1, D]
[6]. If the specific formula of the objective function is known
and differentiable or second-order differentiable, it can be where ⟨·⟩D represents a modulo function with modulus D. The
solved using Newton-like methods or Newton’s method, integer L is derived from Algorithm 1, where CR represents
respectively. However, if the specific formula is known but the crossover rate. It needs to be emphasized that the L has a
not differentiable or the objective function is a black box nearly geometric distribution [6]. For each trial vector, n and
function, the Newton class method fails. In such cases, L must be reselected and recalculated as shown above. In fact,
Evolutionary Algorithms (EAs) play an important role in exponential crossover is a mixture of 1-point crossover and
solving these optimization problems [7]. EAs such as Genetic 2-point crossover, and in this crossover scheme there is
Algorithm (GA) [8], Particle Swarm Optimization (PSO) [9], a well-known ‘‘positional bias’’ [6], in which interactions
and Differential Evolution (DE) [10] are widely used due to between parameters located far apart in the D-dimension
their powerful functions, among which the DE is particularly parameters tend to be destroyed while interactions between
favored by researchers [11], [12], [13], [14], [15] because of genes located close to each other tend to be preserved.
its powerful search capacity, few control parameters and easy This position bias is then overcome by binomial crossover
implementation. that treat each parameter separately, so most researchers
From a macro perspective, like the most popular EAs, believe that the DE variant with binomial crossover has better
DE is a reliable and general function optimizer based on performance than the exponential crossover in dealing with
population, and mutation, crossover and selection operation numerical optimization problems in continuous spaces, while
are its three core operations, among which crossover scheme the DE variant with exponential crossover is good at solving
plays a role in inheriting the past and ushering in the optimization problems where there is a connection between
future. To be more specific, the mutate vector Vi,G = adjacent variables.
v1,i,G , v2,i,G , . . . , vD,i,G generated by the mutation opera-


From the perspective of space, the spatial search method
tion will be used in the crossover scheme and mixed with the of DE is to search the ‘‘solution space’’ in the way of high
target vector Xi,G = x1,i,G , x2,i,G , . . . , xD,i,G
to generate


dimensional cube. Among them, the search range of the DE
a trial vector Ui,G = u1,i,G , u2,i,G , . . . , uD,i,G . Then, the is determined by its mutation strategy, and how many steps
trial vector will be used in the selection operation to compete taken in which search direction to select which trial vector
with the target vector for the opportunity to enter the next in the search range is determined by its crossover scheme,
generation. In the DE, binomial crossover and exponential and the selection operation finally determines whether the
crossover are the two most basic crossover schemes, among search and movement carried out before is effective. Fig. 1
which for binomial crossover, the number of parameters shows the spatial search of DE on the three-dimensional
inherited from the mutate vector has a binomial distribution. view, where X represents the target vector (current position),
This scheme can be summarized as: B represents the mutate vector, and the crossover scheme is
( the point selection operation in the high-dimensional cube
vj,i,G , if rand ≤ CR or j = jrand
uj,i,G = (1) with the target vector and the mutate vector as the diagonal.
xj,i,G , otherwise According to Fig. 1, the set of alternative solutions for the
trial vector is A, B, C, D, A1 , C1 , D1 , that is, the other vertices
except the target vector X . In general, in the case that no
Algorithm 1 Calculate L prior knowledge is obtained in the evolution process, each
1: initial L ← 0. vertex may be the optimal solution, so it should be ensured
2: DO{ that the selection probabilities of these vertices are the
3: L ← L + 1. same to ensure the unbiased algorithm. Conversely, if prior
4: }WHILE ((rand(0, 1) ≤ CR) AND (L < D)) knowledge is obtained, some vertices should have higher
selection probabilities. Based on this idea, relevant scholars
where rand is a uniform random generator number between have otherwise found that although binomial crossover
0 and 1, jrand ∈ [1, 2, 3, . . . , D] ensures that Ui,G gets at least overcomes ‘‘position bias’’ in fact, there is still ‘‘selection
one component from Vi,G , and D is dimension. As for the bias’’ in this scheme [16], [17]. Later, in order to overcome
exponential crossover, an integer n is first randomly chosen the above problems, Meng et al. proposed an evolutionary
among the numbers [1, D]. This integer serves as the starting matrix implementation [18], [19], [20], [21], [22], [23] to
point for the crossover or exchange of components between eliminate the bias. However, it should be emphasized that
the target vector and the mutate vector. Then another integer from the perspective of this spatial search, binomial crossover
L is calculated, which stands for the number of components and exponential crossover are vertex-selection operations to

131678 VOLUME 11, 2023

J. Li, Z. Meng: DE With Exponential Crossover: An Experimental Analysis

also achieve competitive performance with ones employing

binomial crossover in numerical optimization problems.
From the perspective of binomial crossover and exponen-
tial crossover development, binomial crossover has been paid
more attention to and used by scholars because most people
indicate that binomial crossover is more robust and effective
than exponential crossover in dealing with real-valued
parameters of continuous spaces [24], [25]. Therefore, most
of the DE variants developed are applied binomial crossover,
FIGURE 1. The hyper-cube search manner in DE variants from a 3-D view.
and even the relevant DE review papers are mostly focused
on binomial crossover, and the exponential crossover is
only discussed and studied at a shallow level. Later, with
the development of DE, some existing literature [24], [25],
[26], [27], [28], [29] also shows that binomial crossover
achieve a high-dimensional cube, and they essentially obtain still has limitations in solving optimization problems such
the same effect. as indivisible functions and functions with strong depen-
Besides the core operations, control parameters also play dence between variables, and there are still defects in
an important role in DE. Obviously, corresponding to three processing high-dimensional complex functions. However,
core operations: mutation, crossover and selection, there will these problems can be compensated by applying some
also be three control parameters, scale factor F, crossover rate characteristics of exponential crossover, furthermore, some
CR and population size PS, where the scale factor F is used to scholars further proposed that DE with exponential crossover
compress the difference vector in the mutation operation, the is also competitive in numerical optimization. Based on this,
crossover rate CR is used to determine how many parameters many DE algorithms and applications based on exponential
are inherited from the mutate vector in the crossover scheme crossover have appeared. At the same time, much literature
and the population size PS is used to represent the overall on the theoretical analysis, experiment analysis, and formula
size, and also represents the number of function calls derivation of crossover schemes has also emerged [24], [30],
that should be consumed in each generation to perform [31], [32], [33], [34], [35], [36], [37], [38], [39], [40], where
selection operations during the entire evolution process of it is worth mentioning that theoretical analysis includes many
the algorithm. The change of each parameter will affect aspects, such as mutation probability, mutation length, the
the corresponding operation performance, and then affect population mean and variance, etc. But to the best of our
the optimization effect of the entire algorithm, similarly, the knowledge, extensive articles have not yet been published
same parameter corresponding to different strategies will that capture the entire horizon of the current based on
also form different effects. For instance, under two different exponential crossovers, including the various exponential
crossover schemes including binomial crossover and expo- crossover variants and experimental analysis, etc.
nential crossover, the same crossover rate CR will result in To fill the gap in this field, this paper attempts to provide
different numbers of parameters (mutation length) inherited a comprehensive experimental analysis of DE algorithm with
by the trial vector from the mutate vector. In general, each exponential crossover in numerical optimization, which will
crossover scheme can calculate a corresponding mutation also has an important impact on researchers who want to
length during the crossover operation, and its length affects explore this crossover scheme and the main highlights of our
the DE performance. In binomial crossover, the mutation paper can be summarized as follows:
length follows the binomial distribution, while exponential • This paper makes a mathematical analysis of the
crossover follows the geometric distribution, which is the DE algorithm with exponential crossover and mainly
main reason for the difference between binomial crossover analyzes the difference of mutation probability pm and
and exponential crossover. It is worth noting that both distri- mutation length L between binomial crossover and
butions use the crossover rate as the basic parameter, in other exponential crossover, which points out the reason why
words, different CR values will cause the distribution to most researchers prefer to apply binomial crossover
obtain different mutation length values. Therefore, although to solve optimization problems. Meanwhile, it is also
the mutation length of binomial crossover is different from pointed out that DE variants with exponential crossover
that of exponential crossover under the same crossover rate, can achieve comparable performance with the ones
we can make the two crossover schemes produce similar employing binomial crossover once discover the proper
mutation length by selecting and adjusting the crossover rate crossover rate CR and its corresponding parameter
CR value, and thus achieve equivalent optimization effect. control.
To put it another way, as long as the appropriate crossover • Some typical exponential crossover variants and bino-
rate CR and its corresponding parameter control can be mial crossover variants selected in this paper, as well
found, regardless of whether there is a relationship between as the novel exponential crossover DE algorithm,
variables, DE variants with exponential crossover can ERM-DE [41], proposed by us, are evaluated on

VOLUME 11, 2023 131679

 J. Li, Z. Meng: DE With Exponential Crossover: An Experimental Analysis

100 benchmarks of CEC2013 [42], CEC2014 [43], achieve a better approximation for the current candidate
CEC2017 [44] and CEC2022 [45] real parameter single solution. Then, on the basis of the GOBL mechanism,
objective global optimization test suites. Moreover, and equipped with adaptive control parameters similar to
a real-world optimization application will also be used the jDE algorithm [50], Rahnamayan et al. proposed a
to validate our ERM-DE algorithm. The experimental graphics processing unit-based parallel DE (GOjDE) [51],
results support our view that it is entirely feasible to which avoids manual adjustment of control parameters and
develop a DE variant with exponential crossover that improves the quality of candidate solutions, respectively.
can beat the winning DE variants, especially binomial Subsequently, the AGODE algorithm [52], a modified GODE
crossover DE variants, in real parameter single objective algorithm, was proposed, which uses an adaptive generalized
optimization competitions. GOBL mechanism to automatically adjust the opposition
• This paper also outlines and classifies the DE algorithms rate during evolution. In the AGODE algorithm, when the
with exponential crossover used to solve practical opposite operation yields more promising solutions, it may
application problems, and some important future chal- mean that need to add more opposite operations to the current
lenges and research directions in the field of ones with population by increasing the opposite rate value. On the
exponential crossover are also presented by us. contrary, if the current opposite operation does not produce a
The remaining sections of this paper are organized as good enough solution, it is necessary to decrease the opposite
follows. In Section II, the literature review of DE algorithms operation by reducing the opposite rate value.
with exponential crossover is outlined. Section III presents Compared with classical DE, the ODE, GODE, and
the mathematical analysis of DE algorithms with exponential AGODE algorithms consider both the estimate and its
crossover. The experimental analysis of DE algorithms with opposite estimate solution, thereby increasing the chances of
exponential crossover is given in Section IV. Section V finding the global optimum. However, it is worth noting that
outlines and categorizes the significant applications of DE the optimal direction of each search stage in the algorithm
algorithms with exponential crossover. Section VI, the future may be different, and the population cannot always cover
challenges and potential directions of DE with exponential the optimal solution during the evolution process. Therefore,
crossover are listed. Finally, the paper is concluded in neither OBL nor GOBL mechanisms are always effective in
Section VII. generating better opposite solutions. In order to overcome
the above problem and obtain effective opposite solutions,
II. LITERATURE REVIEW Kushida et al. improved the OBL using the archive solutions
DE is a powerful and versatile evolutionary algorithm, and proposed the AODE algorithm [53]. In this algorithm,
especially for many scientific and engineering problems, and Kushida et al. introduce an external archive, which is utilized
since its advent, it has attracted the attention of researchers in to record the best solution in each evolutionary process. This
various fields around the world as well as a wealth of DE archive is then used in the OBL operation to generate opposite
variants have emerged successively. These variants can be candidates with the aim of seeking the global optimum.
classified as DE variants that use only binomial crossover, Constrained optimization problem occurs frequently in the
exponential crossover, or a mixture of binomial crossover and real world, and there have been many studies on solving this
exponential crossover, etc. In this section, we shall engage problem using DE algorithms with traditional exponential
an in-depth discussion of the some DE papers that respect crossover and ϵ constrained method in recent years [54], [55],
exponential crossover. [56], [57]. In the early work, Takahama et al. successively
OBL is a new concept in computational intelligence as well proposed ϵDE, ϵDEg, and ϵDEag algorithms [58], [59], [60]
as an effective mechanism to enhance various optimization to solve the above problem, among which ϵDE [58] combines
methods. It evaluates a solution to a problem while simulta- ϵ constraint technology with DE to improve the optimization
neously computing its opposite solution, thereby providing performance of DE. This algorithm can run very fast and find
another opportunity to seek a candidate solution that is closer very high-quality solutions. As for the ϵDEg [59], gradient
to the global optimum. Therefore, many researchers have mutation is introduced based on ϵDE, and the gradient of
worked towards this direction and some of them also combine this constraint is used to deal with problems with difficult
it with the traditional exponential crossover, proposing constraints. As for ϵDEag [60], not only gradient-based
many DE variants. Rahnamayan et al. [46] proposed an mutation is applied, but also external archives are introduced
opposition-based learning DE variant, the ODE algorithm, to maintain the diversity of individuals, and a new method
which was used for population initialization and evolution. is proposed to select the ϵ level control parameters in the
Later, to further improve the ODE algorithm performance, algorithm. It is worth noting that the first three constrained
a Generalized Opposition-based DE (GODE) [47] was raised algorithms all combine DE/rand/1/exp operation. In addition,
by Rahnamayan et al. In the GODE algorithm, a global Takahama also borrowed the ϵDE algorithm [58] and brought
opposition based learning (GOBL) has also been developed, forward the ϵ constrained rank-based DE (ϵRDE) [61],
which can accelerate the convergence rate of DE, as demon- in which the ϵ constrained method uses the ϵ level
strated in the papers [47], [48], [49]. This GOBL considers comparison to convert an unconstrained problem algorithm
the estimate and its opposite estimate at the same time to to a constrained problem algorithm. In fact, in addition
131680 VOLUME 11, 2023
J. Li, Z. Meng: DE With Exponential Crossover: An Experimental Analysis

to the ϵRDE algorithm, Takahama later proposed an DE factor of the sub-populations. It is worth mentioning that
variant with exponential crossover, ϵDEkr, which combines the DE/rand/1/exp operation will also be performed for each
the ϵ-constraint method with kernel regression estimation sub-population. Weber et al. [68] also mainly investigated
comparison [62]. There are also numerous scholars who methods for embedding exponential crossover and multiple
have proposed new mechanisms and incorporated them into scale factor values in a distributed DE framework such
ϵ-constrained DE algorithms for verification, such as the as PDE [69], IBBDE [70], and DDE [71], [72], [73],
SRS-ϵDEag algorithm proposed by Bu et al. [63]. Unlike and compared with results from previous studies using the
previous gradient-based restoration algorithms that utilize binomial crossover. The experimental results indicate that
gradient information from the constraint set, this paper [63] using multiple scale factors is not beneficial to the system
proposed a Species-based Repair Strategy (SRS) to select when associated with the exponential crossover, and in some
representative infeasible individuals for restoration instead cases even detrimental to the performance of the algorithm.
of randomly selecting certain infeasible individuals as in the Different from the initialization population of DE in
ϵDEag algorithm. By integrating the SRS strategy into the the past, the MDEA algorithm proposed by Ramezani and
ϵDEag algorithm and conducting experimental verification, Lotfi [74] divides the variable range into equal sub-ranges
the results [63] demonstrate superior performance of the according to the population size, so as to carry out biased
proposed algorithm in most benchmark tests, with a reduction initialization population, which is defined as shown in Eq. 3,
in the number of repaired infeasible individuals. Furthermore, where i ∈ [1, 2, . . . , K ], j ∈ [1, 2, . . . , D], K is the
to address the constraint problem more effectively, Yi et al. number of individuals in population set, D is dimension, PS is
introduced a constraint-based mutation operator that takes population size, Xmax and Xmin are the ranges of the variables
into account both constraint violation and the objective respectively.
function. They also proposed a novel adaptive setting method
Xj,i,G = (i − 1) · (Xmax − Xmin )/(PS − 1) + Xmin (3)
and combined these approaches with the JADE algorithm,
resulting in the constraint-based JADE algorithm known as Meanwhile, the MDEA algorithm uses exponential crossover
IϵJADE [64]. and two different selection approaches, MDEA-1 and
Hybrid algorithms have achieved considerable success in MDEA-2, to improve the performance of DE in minimiz-
solving various optimization problems in the past, and based ing continuous spaces. Among them, MDEA-1 performs
on this idea, Singh and Ray [65] proposed an algorithm that a fitness comparison between the trial vector and the
combines the advantages of EA, exponential crossover DE, current vector, and MDEA-2 selects a specified number
and MA, which uses a set of random solutions as a starting of vectors from the current generated and sorted popula-
point and generates subpopulations through EA-based or tion. Garcia-Martinez et al. proposed a role differentiation
DE-based equal-probability evolution. A local search is mechanism and a malleable mating mechanism to improve
then performed to further improve the object value of the DE performance in continuous optimization problems [75].
solution in the population. Yasuda et al. proposed a new Among them, the role difference mechanism creates four
optimization method based on DE and Evolution Strategy different groups, and each operation only involves the vector
(ES) to supplement the insufficient searching capability of of the corresponding role, including the placing, leading,
DE [66]. In this algorithm, the (1+1)-ES operation is applied correcting, and receiving vectors, and the malleable mating
to the best vector Xbest obtained after running an exponential operation is a method to ensure that the mutation operation
crossover DE and uses the 1/5-success rule to fit the step performed on the similarity relation of vectors, identified as
size σ . The performance of the algorithm combining the DE/malleable-mating/1 in this paper, is defined as follows:
two optimization methods of DE and ES has been verified P L C
Vi,G = Xi,G + F · (Xi,G − Xi,G ) (4)
through numerical experiments, and the experiment results
demonstrate that, in comparison to previous optimization where Xi,GP is the placing vector, which is used to point to

methods, this approach enhances the local search capability the search area where the mutate vector Vi,G will be sampled.
of DE and improves performance on multi-modal and noisy L and X C are the leading vector and correcting vector,
Xi,G i,G
problems [66]. respectively, which are used to guide the neighborhood
Weber et al. proposed the Shuffle or Update Paral- sampling of Xi,G P to specific directions. X R is receiving
i,G
lel DE (SOUPDE) [67], which incorporates shuffle and vector used to provide basic parameter values for the
update mechanisms to improve the effectiveness of solving generated trial vector. This mutation strategy allows placing
large-scale optimization problems. The method is a structured vectors to adapt to the evolutionary mating trend, ensuring
population algorithm, which divides the population into that the vectors used for mutation operations maintain a
subpopulations, and each subpopulation rapidly exploits certain similarity relationship. In contrast to using random
certain regions of the decision space, thereby dramatically variables for mutation operations, the proposed mutation
and rapidly reducing fitness values in a highly multivariate strategy in this paper utilizes an adaptive method [75], which
fitness environment. And the role of the shuffling mechanism can dynamically adjust the mating trend based on the problem
is to randomly rearrange individuals in the sub-populations, requirements and the algorithm’s search progress. In other
while the update mechanism is used to update the scale words, the mating can be adapted between random vectors,
VOLUME 11, 2023 131681
 J. Li, Z. Meng: DE With Exponential Crossover: An Experimental Analysis

similar vectors, or dissimilar vectors. Furthermore, after evolutionary framework based on DE logic, supplemented by
generating the mutate vector, exponential crossover will be the shallow utilization of local search. Within the DE-based
employed to generate the trial vector. evolution framework, two collaborative search operators are
Based on the idea of minimizing control parameters, employed. The first operator is based on mutation and
Wang et al. proposed a new variant of DE, Gaussian exponential crossover, similar to the standard DE. However,
Bare-Bones DE (GBDE) [76], [77]. Different from the in this paper [83], the definition of exponential crossover
mutation strategy used by DE in the past, the GBDE in the standard DE is slightly modified by fixing the
algorithm adopts a Gaussian mutation strategy, which is approximate proportion of mutate genes in the offspring,
defined as follows: instead of using the crossover rate CR, as follows:
Xbest,G + Xi,G Dm
Vi,G ∼ N ( , |Xbest,G − Xi,G |) (5) αm ≈ , (7)
2 D
where N (·) represents Gaussian distribution with mean where the dimensions are denoted by D, and the number
(Xbest,G + Xi,G )/2 and standard deviation |Xbest,G − Xi,G |, of mutate genes expected to be replicated in the progeny is
Xbest,G is the current best individual, Xi,G is ith individual, expressed by Dm . The other is a search operator based on
Vi,G is its corresponding mutate vector. In addition, the trigonometric function mutation. At the same time, in order to
algorithm also adopts a dynamic mechanism to adaptively counteract the over-exploitation behavior of search operators
update the parameter CR based on Gaussian distribution. and prevent the algorithm from converging prematurely,
Kamiyama et al. proposed the LDDVDE [78] algorithm, this algorithm also proposed a periodic random perturbation
which borrows the DSM [79] mechanism and uses a local mechanism for virtual populations. Subsequently, Poiko-
descent direction vector the basis of the operation vector. lainen and Neri also raised a new implementation of the
In this algorithm, all parent vectors need to be sorted before memetic structure method, namely DE with Concurrent
generating local descent direction vectors, and then the DSM Fitness Based Local Search (DEcfbLS) [84], which consists
mechanism is embedded into the DE mutation operation. of a local search operator and DE/rand/1/exp framework to
Next, three vectors, Xr1 ,G , Xr2 ,G , and Xr3 ,G are randomly solve continuous optimization problems. From an overall
selected, where the best score is recorded as Xrbest,G and perspective, based on fitness adaptive rules, promising
the worst score is recorded as Xrworst,G , and these are then individuals will be selected for local searches that move along
used to generate local descent direction vectors. Therefore, the axis and replenish the normal search movements of the DE
the mutation strategy can be expressed using the following structure. And from a detailed point of view, the local search
formula: is applied to all individuals in the population with a small
local budget after the population is initialized. In addition,
Vi,G = Xr1 ,G + F · (Xrbest,G − Xrworst,G ) (6)
the improved exponential crossover applied in this paper [84]
where Xr1 ,G represents the first vector chosen at random. is the same as DEcDE [83], which is a slight change to the
After generating the mutate vector, the exponential crossover original exponential crossover, and the crossover rate CR
will be applied to generate the trial vector, and the is replaced by fixing the approximate proportion of mutate
whole algorithm exhibits better search performance than genes in the offspring.
traditional DE. Furthermore, Kamiyama et al. also raised Later, in order to overcome the adverse effects of the
the MOLDDVDE algorithm [80], which is largely similar to unscaled behavior of the exponential crossover operator
LDDVDE. It employs the same strategy of generating trial on solving high-dimensional problems, Zhao and Sugan-
vectors based on exponential crossovers as well as uses local than [26] came up with a linearly scalable exponential
descent direction vector based on the operation vector, and the crossover based on multiple consecutive dimensions, whose
key difference is that MOLDDVDE redefines the generation mutation length L is calculated according to the Eq. 8,
of local descent direction vectors by incorporating the HV where ⌈·⌉ is the function for upward rounding. Moreover,
sorting method [81], [82]. Zhao et al. combined this linearly scalable exponential
Whether it is binomial or exponential crossover, the crossover operator with SaDE-MMTS [28] for solving
essential contribution lies in determining how many com- 19 numerical optimization problems in the Special Issue of
ponents the trial vector inherits from the mutate vector. Soft Computing on Scalability of Evolutionary Algorithms
This concept has inspired various studies on mutation for Large Scale Continuous Optimization Problems and
length by some relevant scholars, resulting in different showed strong competitive results.
exponential crossover variants. These variants will be further L = ⌈(D/100)⌉ (8)
discussed in the following. Neri et al. presented an algorithm,
called Disturbed Exploitation compact DE (DEcDE) [83], In fact, the computing performance of DE can not be ignored
to solve a complex optimization problem using a memetic when solving larger problem scales, inspired by this, the
computational approach. Instead of relying solely on the performance analysis and improvement of parallel DE were
evolutionary framework for global search and the local search carried out by Zibin [85]. Firstly, this paper [85] analyzes
algorithm for local search, this meme scheme incorporates the the basic flow of DE, pointing out the design of DE parallel

131682 VOLUME 11, 2023

J. Li, Z. Meng: DE With Exponential Crossover: An Experimental Analysis

computing that can be easily executed in Math Kernel Algorithm 2 Multiple Exponential Recombination
Library (MKL) and Compute Unified Device Architecture 1: Target Vector ← Xi,G = (x1,i,G , x2,i,G , . . . , xD,i,G ).
(CUDA). The execution order of mutation operation and 2: Mutate Vector ← Vi,G = (v1,i,G , v2,i,G , . . . , vD,i,G ).
crossover scheme is adjusted by matrix calculation, and other 3: Trial Vector ← Ui,G = (u1,i,G , u2,i,G , . . . , uD,i,G ).
operations can be executed in parallel under MKL/CUDA, 4: Parameter to control the mutation probability ← CR.
which reduces the running time of the algorithm. Next, 5: The lengths of the exchanged segments ← T .
by analyzing the distribution of the length L of the crossover 6: Em ← T ∗ CR.
bits in the trial vector, the L value is reset from the 7: Es ← T ∗ (1 − CR).
perspective of probability theory, and a new exponential 8: CRm ← Em /(Em + 1).
crossover scheme was raised by Zibin to make its calculation 9: CRs ← Es /(Es + 1).
performance faster. The formula for generating L is as shown 10: n ← randomly generated from the interval [1, D].
in Eq. 9 and Eq. 10, where N is equal to 1 − FX (x), FX (x) is 11: k ← 1.
the distribution function of the exponential distribution, D is 12: Mutation_Enable ← 1.
the dimension, and the crossover rate is denoted by CR. 13: while k ≤ D do
14: if Mutation_Enable = 1 then
FX (x) = 1 − CRx (9) 15: while k ≤ D and rand(0, 1) ≤ CRm do
d ← ⟨n⟩D , where ⟨n⟩D equals to n if n ≤ D

 1, CR = 0 16:
and equals to n − D if n > D.

lnN

17:
L = min(⌈ ⌉, D), CR ∈ (0, 1) (10)

 lnCR 18: ud,i,G ← vd,i,G .
D, CR = 1 n ← n + 1.

19:
20: k ← k + 1.
Moreover, this paper [85] utilizes MKL/CUDA to generate 21: end while
L for all DE individuals in parallel to speed up the process 22: Mutation_Enable ← 0.
of exponential crossover. In conclusion, through matrix 23: else
calculation combined with a new exponential crossover 24: while k ≤ D and rand(0, 1) ≤ CRs do
scheme, DE programs demonstrate superior computational 25: d ← ⟨n⟩D .
performance. 26: ud,i,G ← xd,i,G .
In addition, Qiu et al. demonstrated that only using 27: n ← n + 1.
consecutive or nonconsecutive crossover schemes is not 28: k ← k + 1.
enough to correctly handle different types of variable interre- 29: end while
lations [25]. To fill this problem niche, a multiple exponential 30: Mutation_Enable ← 1.
recombination [25], also called a semi-continuous crossover 31: end if
operation, was proposed. In this proposed scheme, the trial 32: end while
vector is constructed by multiple fragments of the mutate
vector and the target vector through parameters CRm and
CRs , respectively. For a detailed process, please refer to
Algorithm 2. It is worth mentioning that this method has observing that in the experimental part, Tanabe and Fukunaga
stronger dependent variable management capabilities and applied SEC to the SHADE algorithm and obtained better
exhibits strong competitiveness in handling different types performance [86].
of variable correlations when one is used for six mutation Competitive strategy is a commonly used method in
strategies, including DE/rand/1, DE/rand/2, DE/current-to- DE with hybrid crossover, which mainly makes various
rand/2, DE/best/1, DE/best/2, DE/current-to-best/2, as well strategies compete with each other in the evolution process
as one adaptive DE variant, the SaDE algorithm. Besides of each generation, so as to improve the optimization
mutation length, another research direction for improving algorithm performance. It is worth mentioning that each
exponential crossover is component ordering. Some related strategy will have a corresponding probability value, and the
scholars have made significant efforts in this area and have probability value will be adaptively updated according to the
developed new exponential crossover variants. Tanabe and learning experience of previous generations. There are many
Fukunaga indicated that exponential crossover is overrated papers here that could fall into this category, for example,
and proposed the Shuffled Exponential Crossover (SEC) [86] Zhao et al. combined the JADE mutation strategy with
that does not depend on arbitrary dependencies between binomial crossover, exponential crossover, and non-crossover
adjacent variables, as shown in Algorithm 3. They indicated to improve the SaDE algorithm [87], [88], [89] and mixed
that the SEC should be substituted for the exponential it with an improved multi-track search (MMTS) algorithm,
crossover because it can yield a much more stable search thereby proposing the SaDE-MMTS algorithm [28], which
performance whether the function is adjacent or distributed is used to solve the continuous optimization problem.
unless there is some prior knowledge that there is a Among them, the trial vector generation strategy under the
dependency between continuous variables [86]. It is worth SaDE-MMTS algorithm and its related control parameter

VOLUME 11, 2023 131683

 J. Li, Z. Meng: DE With Exponential Crossover: An Experimental Analysis

Algorithm 3 Shuffled Exponential Crossover (SEC) optimal area, so as to finally converge to the global optimum.
1: Dimension ← D. Ali et al. proposed a Cluster-based DE algorithm numerical
2: Population size ← PS. optimization with Heterogeneous Influence (CDE-HI) [95],
3: Target vector Xi,G ← (x1,i,G , xj,i,G , . . . , xD,i,G ), i ← which uses the K -means clustering method to divide the
[1, 2, . . . , PS]. population of the DE into many clusters with three different
4: Mutate vector Vi,G ← (v1,i,G , vj,i,G , . . . , vD,i,G ), i ← mutation strategies, DE/best_group/1/exp, DE/rand/1/exp,
[1, 2, . . . , PS]. and DE/rand/1/bin. Among them, the DE/best_group /1
5: Trial vector Ui,G ← ∅. randomly selects a portion of individuals N from the
6: Ui,G ← Xi,G . population PS, and selects the best individual from this group
7: k ← 1. as Xbest,G , while r1 , r2 , and i are all different individuals from
8: S ← randperm(D), randperm is a function [1, PS], and the formula is as follows:
9: of random permutation. Vi,G = Xi,G +F · (Xbest,G − Xi,G )+F · (Xr1 ,G − Xr2 ,G ) (11)
10: repeat
11: j ← S(k). The CDE-HI algorithm effectively balances the exploration
12: uj,i,G ← vj,i,G . and exploitation abilities by incorporating different mutation
13: k ← k + 1. strategies for different clusters, and demonstrates strong com-
14: until rand[0, 1) < CR and k < D petitiveness in 25 complex benchmark functions on single
objective real-parameter numerical optimization. Ali et al.
posed a new variant of DE, sTDE-dR [96], which divides
the population into multiple tribes and assigns a set of
values are gradually adaptive by learning their previous different mutation and crossover schemes to improve the
experience in generating promising solutions. Next, the quality of the search solution and avoid the occurrence of
MMTS, which is applied constantly to refine some solutions premature and stagnant phenomena. Among them, strategies
with different distributions at different search phases to meet based on competitive success will be used to determine
global and local search requirements, is used to improve the life cycle of each tribe and its participation rate in
solutions generated by SaDE-MMTS. the next generation. Sallam et al. proposed an improved
Bujok and Tvrdk introduced a competitive crossover multi-operator DE (IMODE) algorithm [97], which mainly
strategy based on the SHADE algorithm [90], which is uses a variety of mutation strategies and crossover schemes
the competitive crossover alternation between binomial and to improve the algorithm performance. Meanwhile, the SQP
exponential crossover based on probability values to handle method is applied for local search, and the best operator is
some difficult problems defined by the combination of obtained according to the quality of the solution and the sub-
multiple multimodal functions. Hui and Suganthan put population diversity. Moreover, Aggarwal and Mishra also
forward ensemble DE with dynamic subgroups and adaptive proposed a multi-operator DE variant, the MOEA/D-MODE
clearing (EDESAC) [91] to maintain population diversity algorithm [98], in which the entire population is divided into
and retain information about past good solutions, thereby multiple sub-populations, and different mutation strategies
solving dynamic optimization problems. Takahama and and crossover schemes are assigned to the sub-populations
Sakai proposed a large-scale DE optimization algorithm for evolution to solve multi-objective optimization problems.
for landscape morphology detection and diversity archives
(LMDEa) [92], which uses small population size and III. MATHEMATICAL ANALYSIS OF DE WITH
large diversity archives to solve large-scale optimization EXPONENTIAL CROSSOVER
problems and simply controlled scale factors by observing the Crossover scheme plays an indispensable role in DE perfor-
landscape morphology of search points to maintain diversity. mance optimization, and the analysis of crossover scheme
Additionally, the use of the DE/rand/1 trial vector generation mainly starts from two aspects of mutation probability pm and
strategy in combination with exponential crossover and mutation length L. The mutation probability pm refers to the
binomial crossover enhances the ability to explore the optimal probability that any a component in the trial vector inherits
solution. Afterward, based on the LMDEa variant [92], from the mutate vector, and the mutation length L refers to the
the concept of ranking-based DE (RDE) [62] and the idea number of components trial vector inherited from the mutate
of multiple mutation strategies were introduced to propose vector. These two aspects are different in different crossover
an improved LMDEa variant, LMRDEa [93], which is a schemes and different crossover rate CR values will lead to
population-based search method for large-scale optimization different mutation probabilities pm and mutation lengths L.
problems in continuous spaces. Binomial crossover and exponential crossover are two
Lynn et al. proposed a population-based DE [94], which basic crossover schemes of DE, and in binomial crossover,
divides the entire population into two sub-populations to the outcome of each trial is derived from either a component
balance the exploration and exploitation capabilities of the of the mutate vector or the target vector, which is equivalent
algorithm. The algorithm explores the problem search space to a simulation based on D independent Bernoulli trials.
through these two populations, and then fine-tunes the And according to Eq. 1, the condition ‘‘rand ≤ CR or

131684 VOLUME 11, 2023

J. Li, Z. Meng: DE With Exponential Crossover: An Experimental Analysis

j = jrand ’’ ensures that at least one component in the binomial Then, according to the Eq. 15, the average number of
crossover is inherited from the mutate vector. In this case, the mutation components E(L) can also be calculated [24], [32],
successful event in each Bernoulli experiment is the union [99], as shown below:
of two independent events consisting of ‘‘rand ≤ CR’’ D−1
and ‘‘j = jrand ’’, where the first event has probability CR
X 1−CRD
E(L) = (1−CR)· h·CRh−1 +D·CRD−1 = (16)
and the other event has probability 1/D. Since for any two 1−CR
h=1
independent events A and B, the probability of their union
For calculating the value of the mutation probability pm under
is Prob(A ∪ B) = Prob(A) + Prob(B) − Prob(A) · Prob(B),
exponential crossover, it is necessary to consider those two
the probability that a component in the binomial crossover
parameters, the starting point n and the mutation length L at
inherits from the mutate vector is pm = CR·(1−1/D)+1/D.
the same time. For any component j, it will only mutate when
As for the mutation length L, its values are in {1, . . . , D}
d(j, n) < L, where the d(A, B) function [32] is defined as
and have the property L = L ′ + 1, where L ′ is a binomial
follows:
distribution with parameters D − 1 and CR. Therefore, (
Prob(Ui,GL=h ) = Prob(U L ′ =h−1 ) = C h−1 · CRh−1 · (1 − A − B, if A ≥ B
i,G D−1 d(A, B) = (17)
CR)D−h , h ∈ {1, . . . , D}, and the E(L) = E(L ′ ) + 1 = D + A − B, otherwise
(D − 1) · CR + 1 is the average number of mutate components.
Here, a vector in the trial vector set with h parameters Also, since the starting point n is randomly selected in
L=h , and {1, . . . , D}, in other words, n can take any number of
inherited from the mutant vector Vi,G is given by Ui,G
the probability of generating such a trial vector candidate them with the probability of 1/D, the probability that any
by inheriting h parameters from the mutant vector Vi,G is component j inherits from the mutate vector can be defined
given by Prob(Ui,G L=h ). These characteristics of the binomial as:
D
crossover can be summarized as follows, which is the result 1 X
derived also in [24], [32], and [99]. pm = · Prob(d(j, n) < L)
D
n=1
1 1 1
D−1
pm = CR · (1 − ) + (12) = ·
X
Prob(L > d) (18)
D D D
L=h h−1
Prob(Ui,G ) = CD−1 · CRh−1 · (1 − CR)D−h (13) d=0
D−1
E(L) = D · pm = (D − 1) · CR + 1 (14) X
Prob(L > d) = (1 − CR) · CRh−1 + CRD−1 = CRd
h=d+1
In the exponential crossover, first determine two points,
where the first point n is an integer randomly selected (19)
from {1, . . . , D} to represent the starting point of the Thus, the calculation formula of the mutation probability pm
exchange, and the other integer L is used to represent the under exponential crossover can be organized as shown in
number of components inherited from the mutate vector, Eq. 20. This is also the result derived of [24], [32], and [99].
which is calculated according to Algorithm 1. According to D−1
Algorithm 1, there are also two constraints when calculating 1 X d 1 − CRD
pm = · CR = (20)
L: ‘‘rand ≤ CR’’ and ‘‘L ≤ D’’. Assuming that if there D D · (1 − CR)
d=0
is only the condition ‘‘rand ≤ CR’’, then L will obey the
geometric distribution with the parameter 1 − CR (CR at this From Eq. 12, Eq. 14, Eq. 16 and Eq. 20, it can be seen that
time represents the probability of success) and take its value the relationship between the crossover rate CR, the mutation
in {1, . . . , D}, and the probability of generating such a trial probability pm , and the mutation length L, is different in the
vector candidate by inheriting h parameters from the mutant binomial crossover and exponential crossover, in other words,
vector Vi,G will be expressed by Prob(Ui,G L=h ) = (1 − CR) · under the same CR value, the optimization performance
h−1
CR . At this time, coupled with the boundary constraint of binomial crossover and exponential crossover on DE is
condition of ‘‘L ≤ D’’, L can be considered to obey the different. To be more specific, according to Eq. 12 and
truncated geometric distribution, and when L = h = D, Eq. 20, in the binomial crossover, there is a linear relationship
Prob(Ui,G L=h ) = CRD−1 . It is worth noting that exponential between the mutation probability pm and the crossover rate
crossover also ensures that the trial vector can inherit at CR, but it is nonlinear in the exponential crossover, and at the
least one component from the mutate vector. To sum up, the same CR value, the pm value of the exponential crossover is
probability of mutation length L = h can be expressed by the smaller than the pm value of the binomial crossover, which
following formula, which is the result derived also in [24], becomes more obvious as the problem scale increases. For
[32], and [99]. the mutation length L, it can be seen from Eq. 14 that
the mutation length of binomial crossover is in a linear
relationship with the crossover rate CR, in which the mutation
(
L=h (1 − CR) · CRh−1 , if 1 ≤ h < D
Prob(Ui,G ) = (15) length increases with the increase of CR value, meanwhile,
CRD−1 , if h = D through the derivation analysis of Eq. 16, as shown in Eq. 21

VOLUME 11, 2023 131685

 J. Li, Z. Meng: DE With Exponential Crossover: An Experimental Analysis

and Eq. 22, it can also be seen that the mutation length of negligible when the number of components mutated is equal.
exponential crossover also increases with the increase of CR Based on this, according to Eq. 14 and Eq. 16, we find that by
value. assigning different CR to binomial crossover and exponential
crossover, it is possible to obtain similar mutation lengths
CRD−1 · [(D − 1) · CR − D] + 1
E ′ (L) = (21) for binomial crossover and exponential crossover, and similar
(1 − CR)2 algorithm performance can also be produced. In other words,
(
D−1
Y (CR) = CR · [(D − 1) · CR − D] assuming a binomial crossover rate CRb is given, then there
′ (22)
Y (CR) = (D − 1) · CRD−2 · [D · (CR − 1)] is always an exponential crossover rate CRe , CRe ∈ [0, 1],
which satisfies Eq. 25, and vice versa.
Specifically, when the CR value is between [0,1], Y ′ < 0,
so Y is monotonically decreasing, and its maximum value 1 − CRD
e
(D − 1) · CRb + 1 = (25)
Y (CR = 0) = 0, the minimum value Y (CR = 1) = −1. 1 − CRe
Next, by putting the value of Y into E ′ (L), it can be seen that
Here is an example to explain the feasibility of this theory:
when the CR value falls between [0,1], E ′ (L) > 0, so E(L)
If the value of the mutation length is fixed at D2 , using Eq.
is monotonically increasing. Although the mutation length
14 we calculate the crossover rate CR for the DE variant with
of binomial crossover and exponential crossover increases D−2
binomial crossover is 2·(D−1) . Similarly, we can also calculate
monotonically with the crossover rate CR, further formula 1
analysis, such as Eq. 23 and Eq. 24, shows that when the another approximate crossover rate CR value 21 D−1 through
integer dimension D is greater than or equal to 2 (since both Eq. 16, so that the DE variant with exponential crossover can
binomial crossover and exponential crossover ensure that the also produce a mutation length of D2 .
trial vector inherits at least one component from the mutate Further, we found that both binomial crossover and expo-
vector, the mutation length is both 1 when D=1) and CR is nential crossover are essentially vertex selection operations
between [0,1], the mutation length of binomial crossover is for high-dimensional cubes, and even some schemes that
always greater or equal to that of exponential crossover. cannot be covered by exponential crossover in theoretical
2D − 1 schemes can also be obtained through continuous
D · CR − D · CR2 + CR2 − 2 · CR + CRD
1E(L) = (23) adjacent point search. As a result, DE variants with expo-
1 − CR nential crossover can achieve comparable performance with
W (CR) = D·CR−D·CR2 +CR2 −2·CR+CRD


 the ones employing binomial crossover once discover the
 W ′ (CR) = D−2·D·CR+2·CR−2+D·CRD−1

proper crossover rate CR and its corresponding parameter
(24) control. Although this task is difficult, we draw on the idea
 W ′′ (CR) = (D−1)·(D · CRD−2 −2)


 ′′′ of evolutionary matrix proposed in QUATRE algorithm [18],
W (CR) = D·(D − 1)·(D − 2)·CRD−3
[19], [20], [21], [22], [23] to propose a novel parameter
To put it in more detail, 1E(L) represents the difference control technique for scale factor F and crossover rate CR,
between the mutation length of the binomial crossover and which is different from the control technique based on fitness
the exponential crossover, W (CR) is the molecular value of value. Under this parameter control technique, the adaptive
1E(L), and W ′ (CR), W ′′ (CR), and W ′′′ (CR) are the first, parameters F and CR have the characteristics of fitness
second and third derivative formulas of W (CR), respectively. value independence, in which the crossover rate CR value is
It is worth noting that in D ≥ 2 and CR value falls in the automatically generated not only in the initial stage of the
interval of [0,1], W ′′′ (CR) ≥ 0, then W ′′ (CR) monotone algorithm, but also in the whole process of the algorithm
increasing. By analogy, by analyzing W ′′ (CR), and W ′ (CR) evolution. At the same time, a novel population diversity
step by step, it can be concluded that W (CR) ≥ 0, and then mechanism combining archived global opposition learning
it is obtained that 1E(L) ≥ 0. Based on this, we can see that with a mutation strategy based on historical information,
under the same CR, the mutation length of binomial crossover as well as a timestamp mechanism are also proposed by us
is longer than that of exponential crossover, in other words, and used to further optimize the performance of our ERM-
the degree of increase in population diversity of binomial DE algorithm, so that it has achieved comparable or even
crossover will be stronger than exponential crossover, and better performance than the advanced DE variants in the
the algorithm performance will be better than exponential experiment. For the core idea of the algorithm, we present
crossover. Therefore, in order to obtain a longer exponential it in the form of pseudo code in the supplementary material,
crossover mutation length to increase population diversity, as shown in Algorithm S1. If relevant scholars want to learn
the CR value should be selected closer to 1. Perhaps this is more about the specific core content of the algorithm, you can
why exponential crossover is more sensitive to CR values refer to our next paper [41], which is in the review stage.
falling between 0.9 and 1 [24], [31], [32].
It needs to be emphasized that some papers [32], [34] IV. EXPERIMENT ANALYSIS OF DE WITH EXPONENTIAL
also show that although the mutation probability between CROSSOVER
binomial and exponential crossover differs for the same CR In this section, we first perform experiments on the
value, the difference in performance between the two is E(L) of binomial and exponential crossover when

131686 VOLUME 11, 2023

J. Li, Z. Meng: DE With Exponential Crossover: An Experimental Analysis

TABLE 1. Recommended parameter settings of all these contrasted algorithms.

FIGURE 2. E(L) versus CR when D=10. FIGURE 3. E(L) versus CR when D=30.

D = 10, 30, and 50 with respect to CR. Next, we have CEC2022 test suite has 12 benchmark functions, which helps
selected some DE variants with exponential crossover: the to avoid overfitting problems caused by using only one test
SaDE-MMTS algorithm [28], SOUPDE algorithm [67], suite. Moreover, a real-world optimization application will
GODE algorithm [47], AGODE algorithm [52], AODE also be used to validate our ERM-DE algorithm.
algorithm [53], SHADE algorithm with embedded shuf- All experiments were conducted on Windows 10 operating
fle exponential crossover operator (denoted here by system on a personal computer with Intel (R) Core (TM)
SHADE_sec) [86], SHADE algorithm in which the binomial i7-7500U CPU @ 2.70GHz. Each benchmark function of
crossover operator and exponential crossover operator com- the algorithm was run 51 times, and the fitness error 1f =
pete with each other (expressed here by SHADE_cc) [90], f (X ) − f (X ∗ ) of the algorithm was recorded firstly on each
recent DE-EXP algorithm [99] and DE variants with binomial benchmark of the total 51 runs, and then the mean value and
crossover: JADE algorithm [100], SHADE algorithm [101], standard deviation of the fitness error were calculated. The
LSHADE algorithm [102], cooperative strategy based DE symbols >, ≈, < indicate that a given algorithm performed
with population diversity enhancement (CS-DE) [103], significantly better (>), not significantly different better or
compared with our proposed ERM-DE algorithm [41] for worse (≈), or significantly worse (<) compared to another
optimal performance, and the convergence speed comparison algorithm. The maximum number of function evaluations
plots of these DE variants are also presented in this part, is 10000*D, where D is the dimension. By the way, this
thereby proving that it is feasible to develop a powerful DE article identifies functions value that is less than or equal to
variant with exponential crossover, which can beat the DE eps = e − 16 as 0.
variants with classical exponential crossover and binomial
crossover in the single objective real parameter optimization A. EFFECT OF CR ON E (L): A MULTIDIMENSIONAL
competition. The parameter settings of all algorithms for PERSPECTIVE
comparison are shown in Table 1. And the test suite used Here we perform experiments on the E(L) of binomial and
for testing these algorithms consists of the CEC2013 test exponential crossover when D=10, 30, and 50 with respect to
suite [42], CEC2014 test suite [43], CEC2017 test suite [44], CR, as shown in Figs. 2-4, respectively. As can be seen from
and CEC2022 test suite [45]. This large test suite has a total Figs. 2- 4, regardless of dimension D=10, 30 or 50, when CR
of 100 functions, CEC2013 test suite has 28 benchmark value falls between 0.9 and 1, the mutation length obtained
functions, the CEC2014 test suite has 30 benchmark by exponential crossover is considerable, and it can also be
functions, CEC2017 test suite has 30 benchmark functions, seen from these figures that the allocation of different CR to

VOLUME 11, 2023 131687

 J. Li, Z. Meng: DE With Exponential Crossover: An Experimental Analysis

TABLE 2. Algorithm comparison between SaDE-MMTS, SOUPDE, GODE, AGODE, AODE, SHADE_sec, SHADE_cc, DE-EXP and our algorithm under 2013-10D
optimization.

TABLE 3. Algorithm comparison between SaDE-MMTS, SOUPDE, GODE, AGODE, AODE, SHADE_sec, SHADE_cc, DE-EXP and our algorithm under 2013-30D
optimization.

we summarize the results of Tables 2-10 into Table 11.

From Table 11, we can see that our ERM-DE algorithm
obtains significant performance improvement on 10D, 30D,
and 50D CEC2013, CEC2014, and CEC2017 test suites in
comparison with the other eight DE variants with exponential
crossover. Besides the CEC2013, CEC2014, and CEC2017
test sets, we also evaluate these exponential crossover DE
variants and our ERM-DE algorithm on the CEC2020 and
CEC2022 test sets and the results are listed in Tables S1-S6
(see Supplementary Material). It can still be seen from Tables
S1-S6 that our proposed algorithm is competitive with these
FIGURE 4. E(L) versus CR when D=50. exponential crossover variants.

C. OUR ERM-DE VS BINOMIAL CROSSOVER VARIANTS

binomial crossover and exponential crossover can make the The optimization results of the DE variants with bino-
two crossover schemes obtain similar mutation length. mial crossover (JADE, SHADE, LSHADE, and CS-DE
algorithm) and our ERM-DE algorithm are compared under
B. OUR ERM-DE VS EXPONENTIAL CROSSOVER VARIANTS the CEC2013, CEC2014, and CEC2017 test sets for real
The optimization results of the DE variants with exponen- parameter single objective optimization on 10D, 30D, 50D,
tial crossover (SaDE-MMTS, SOUPDE, GODE, AGODE, and the results are listed in Tables 12-20, respectively. Also,
AODE, SHADE_sec, SHADE_cc, DE-EXP algorithm) and we summarize the results of Tables 12-20 into Table 21. From
our proposed ERM-DE algorithm are compared under Table 21, we can see that our ERM-DE algorithm obtains
the CEC2013, CEC2014, and CEC2017 test sets for real significant performance improvement on 10D, 30D, and 50D
parameter single objective optimization on 10D, 30D, 50D, CEC2013, CEC2014, and CEC2017 test suites in comparison
and the results are listed in Tables 2-10, respectively. Also, with the other four DE variants with binomial crossover.

131688 VOLUME 11, 2023

J. Li, Z. Meng: DE With Exponential Crossover: An Experimental Analysis

TABLE 4. Algorithm comparison between SaDE-MMTS, SOUPDE, GODE, AGODE, AODE, SHADE_sec, SHADE_cc, DE-EXP and our algorithm under 2013-50D
optimization.

TABLE 5. Algorithm comparison between SaDE-MMTS, SOUPDE, GODE, AGODE, AODE, SHADE_sec, SHADE_cc, DE-EXP and our algorithm under 2014-10D
optimization.

TABLE 6. Algorithm comparison between SaDE-MMTS, SOUPDE, GODE, AGODE, AODE, SHADE_sec, SHADE_cc, DE-EXP and our algorithm under 2014-30D
optimization.

Besides the CEC2013, CEC2014, and CEC2017 test sets, 1) OPTIMIZATION COMPARISON ON CONVERGENCE SPEED
we also evaluate these binomial crossover DE variants and Besides validating the performance of our proposed ERM-DE
our ERM-DE algorithm on the CEC2022 test set for 10D, algorithm in terms of convergence accuracy, in this part,
20D real-parameter single-objective optimization, and the the convergence of all these advanced DE variants under
results are listed in Tables S7 and S8 (see Supplementary 28 benchmarks is also compared by using convergence curves
Material). It can still be seen from Tables S7 and S8 that for the median of 51 runs on each benchmark. Here we only
our ERM-DE algorithm is competitive with these binomial show the convergence curves for the 30D CEC2013 test set
crossover variants. optimization due to page constraints, and these comparisons

VOLUME 11, 2023 131689

 J. Li, Z. Meng: DE With Exponential Crossover: An Experimental Analysis

TABLE 7. Algorithm comparison between SaDE-MMTS, SOUPDE, GODE, AGODE, AODE, SHADE_sec, SHADE_cc, DE-EXP and our algorithm under 2014-50D
optimization.

TABLE 8. Algorithm comparison between SaDE-MMTS, SOUPDE, GODE, AGODE, AODE, SHADE_sec, SHADE_cc, DE-EXP and our algorithm under 2017-10D
optimization.

TABLE 9. Algorithm comparison between SaDE-MMTS, SOUPDE, GODE, AGODE, AODE, SHADE_sec, SHADE_cc, DE-EXP and our algorithm under 2017-30D
optimization.

of the 28 benchmark functions are summarized in the four fa1 , fa2 , fa4 and fa10 . Therefore, our ERM-DE algorithm is
plots from Figs. 5-8. From these figures, we can note that the also strongly competitive with these advanced DE variants
our ERM-DE algorithm outperforms the other state-of-the-art in terms of convergence speed.
DE variants on benchmarks fa7 , fa8 , fa16 , and fa18 , and it still
competitive with other advanced algorithms on benchmark D. APPLICATION
competitiveness on fa3 , fa5 , fa6 , fa15 , fa17 , fa23 −fa28 . Moreover, In this section, a classical engineering design problem, ten-
it can be seen from the figures that our ERM-DE algorithm sion/compression spring design, is used to test the ability of
can even find global optimal solutions in the benchmarks our proposed ERM-DE algorithm to solve practical problems.

131690 VOLUME 11, 2023

J. Li, Z. Meng: DE With Exponential Crossover: An Experimental Analysis

TABLE 10. Algorithm comparison between SaDE-MMTS, SOUPDE, GODE, AGODE, AODE, SHADE_sec, SHADE_cc, DE-EXP and our algorithm under
2017-50D optimization.

TABLE 11. Summary of the comparison results between our algorithm and the other contrasted DE variants with exponential crossover under the
CEC2013, CEC2014 and CEC2017 benchmarks on 10D, 30D and 50D optimization respectively.

TABLE 12. Algorithm comparison between JADE, SHADE, LSHADE, CS-DE, and our algorithm under 2013-10D optimization.

This problem has been solved with various best-performing way is used to define the problem:
algorithms, including sCMAgES [104], EnMODE [105] and
min f (x) = x12 · x2 · (x3 + 2) (26)
COLSHADE [106]. Therefore, these algorithms mentioned
above are chosen for comparison in this subsection. The problem contains three variables: the diameter of the
Optimizing the weight of the stretched or compressed wire, x1 ∈ [0.05, 2.00], the mean value of the coil
spring is the main goal of this problem, and the following diameter, x2 ∈ [0.25, 1.30], and the number of active coils,

VOLUME 11, 2023 131691

 J. Li, Z. Meng: DE With Exponential Crossover: An Experimental Analysis

TABLE 13. Algorithm comparison between JADE, SHADE, LSHADE, CS-DE, and our algorithm under 2013-30D optimization.

TABLE 14. Algorithm comparison between JADE, SHADE, LSHADE, CS-DE, and our algorithm under 2013-50D optimization.

TABLE 15. Algorithm comparison between JADE, SHADE, LSHADE, CS-DE, and our algorithm under 2014-10D optimization.

x3 ∈ [2.00, 15.0], are used to calculate the weights. It is worth mentioning that each of the algorithms
Meanwhile, the four constraints contained in this problem, mentioned above for solving this engineering design problem
as shown in Eq. 27, also need to be satisfied. is run 25 in this subsection, and the function value f (X )

131692 VOLUME 11, 2023

J. Li, Z. Meng: DE With Exponential Crossover: An Experimental Analysis

TABLE 16. Algorithm comparison between JADE, SHADE, LSHADE, CS-DE, and our algorithm under 2014-30D optimization.

TABLE 17. Algorithm comparison between JADE, SHADE, LSHADE, CS-DE, and our algorithm under 2014-50D optimization.

of the algorithm is recorded firstly on each benchmark of V. APPLICATIONS OF DE ALGORITHMS WITH

the total 25 runs, and then the best result (‘‘Best’’), worst EXPONENTIAL CROSSOVER
result (‘‘Worst’’), and mean value and standard deviation Optimization problems are ubiquitous in our daily life,
(‘‘Mean/Std’’) of the fitness error are calculated, and they are as small as the atoms in the microcosm, they always construct
summarized in Table 22. From the Table 22, we can notice their chemical bonds in a way that minimizes the electron
that our ERM-DE algorithm is competitive in solving the energy, while as large as each species is constantly adjusting
real-world constrained optimization. their structures in the process of biological evolution to adapt
to their environment. Therefore, these real-life optimization
x23 · x3


 g1 (x) = 1 − ≤0 problems have been solved all the time since the dawn of
71785 · x14



 humankind. During the exploration of solutions for these
 2
 g2 (x) = 4 · x2 − x1 · x2 + 1

 optimization problems, computational intelligence optimiza-
−1≤0

12566(x2 · x1 − x1 ) 5108 · x12
3 4 tion algorithms, particularly DE, have gained increasing

 140.45 · x1 recognition and attention from the majority.
 g3 (x) = 1 −
 ≤0


 x22 · x3 The DE was proposed by R. Storn and K. V. Price in
 g (x) = x1 + x2 − 1 ≤ 0


 1995 [107]. A year later, DE featured prominently at the
4
1.5 First International Competition on Evolutionary Optimiza-
(27) tion, which was held in May 1996 in conjunction with

VOLUME 11, 2023 131693

 J. Li, Z. Meng: DE With Exponential Crossover: An Experimental Analysis

FIGURE 5. Here we present the convergence comparison by employing the median value of 51 runs on 30D
optimization under fa1 -fa8 of our test suite.

131694 VOLUME 11, 2023

J. Li, Z. Meng: DE With Exponential Crossover: An Experimental Analysis

FIGURE 6. Here we present the convergence comparison by employing the median value of 51 runs on 30D
optimization under fa9 -fa16 of our test suite.

VOLUME 11, 2023 131695

 J. Li, Z. Meng: DE With Exponential Crossover: An Experimental Analysis

FIGURE 7. Here we present the convergence comparison by employing the median value of 51 runs on 30D
optimization under fa17 -fa24 of our test suite.

131696 VOLUME 11, 2023

J. Li, Z. Meng: DE With Exponential Crossover: An Experimental Analysis

FIGURE 8. Here we present the convergence comparison by employing the median value of 51 runs on 30D optimization
under fa25 -fa28 of our test suite.

TABLE 18. Algorithm comparison between JADE, SHADE, LSHADE, CS-DE, and our algorithm under 2017-10D optimization.

the 1996 IEEE International Conference on Evolutionary in 1997 proved to be one of the best among competing
Computing (CEC) [108]. Then, the DE proposed by Price in algorithms. Articles detailing its algorithm [109] were
the Second International Evolutionary Optimization Contest published followed immediately in quick succession. When

VOLUME 11, 2023 131697

 J. Li, Z. Meng: DE With Exponential Crossover: An Experimental Analysis

TABLE 19. Algorithm comparison between JADE, SHADE, LSHADE, CS-DE, and our algorithm under 2017-30D optimization.

TABLE 20. Algorithm comparison between JADE, SHADE, LSHADE, CS-DE, and our algorithm under 2017-50D optimization.

TABLE 21. Summary of the comparison results between our algorithm and the other contrasted DE variants with binomial crossover under the CEC2013,
CEC2014 and CEC2017 benchmarks on 10D, 30D and 50D optimization respectively.

it comes to solving optimization problems, DE is designed strategies. Among them, the crossover scheme applied by
to efficiently search for the optimal solution to a given this algorithm: binomial crossover or exponential crossover
problem by iteratively adjusting a set of parameters and schemes, are used to combine the best features of two or more

131698 VOLUME 11, 2023

J. Li, Z. Meng: DE With Exponential Crossover: An Experimental Analysis

TABLE 22. The optimization results comparison among sCMAgES, VI. FUTURE CHALLENGES AND POTENTIAL DIRECTIONS
EnMODE, COLSHADE and our ERM-DE for the real-world
tension/compression spring design optimization. OF DE WITH EXPONENTIAL CROSSOVER
Although research on and with DE has reached an impressive
state, research on DE with exponential crossover is relatively
sparse and there are still many open problems. Below,
we unfold some important future research challenges and
directions in the DE with exponential crossover field.
solutions in order to create a new, potentially better solution 1) Most researchers of DE believe that exponential
that further approximates the optimal solution region of the crossover cannot solve the real-parameter optimiza-
applied problem. tion problems in continuous space, however, from
Due to the DE being a kind of intelligent optimization the perspective of high dimensional space, expo-
algorithm with strong global optimization ability, researchers nential crossover and binomial crossover are both
from several fields of science and engineering have been vertex-selection operations to realize a high dimen-
applying DE to solve optimization problems in their fields. sional cube and should be able to solve real-parameter
Therefore, among the existing DE literature, the literature optimization problems in continuous space. Therefore,
on the application is also numerous and multifaceted. These it will be an important research challenge and direction
include papers on solving various application problems in the future to further prove the equality of expo-
based on exponential crossover DE variants. Although the nential crossover and binomial crossover in solving
previous experimental studies of most scholars under the real-parameter optimization problems in continuous
general test set show that binomial crossover can solve space theoretically.
the continuous space parameter optimization problem well, 2) Nowadays, some proposed advanced DE variants,
and exponential crossover cannot achieve an equivalent including DE variants with binomial crossover and DE
optimization effect, optimization problems in real life are variants with exponential crossover, all use parameter
complex and diverse, such as multi-objective, constrained, adaptation strategies based on fitness values. In gen-
large-scale, uncertain optimization problems, discrete space eral, by taking the fitness difference as the weight of
optimization problems, etc. thus some binomial crossover the control parameter adaptation, the DE variant can
variants are not necessarily better than exponential crossover generate more accurate control parameters, leading to
variants at solving these problems. This paper focuses on the better optimization performance. However, in some
DE variants with exponential crossover, and this section lists optimization cases, exact fitness values cannot be
some of its major applications in years order, including the obtained. In these cases, all these fitness-based DE
trial vector generation strategy with exponential crossover variants cannot work, therefore, further developing
used and the main issues involved, as shown in Table 23. a fitness-independent DE variant with exponential
A keen observation of Table 23 reveals that most crossover will also be a direction that relevant scholars
practitioners prefer to apply DE/rand/1/exp, DE/rand/2/exp, can study in the future, and it will also be a challenge.
DE/best/1/exp, DE/best/2/exp, DE/rand-to-best/1/exp, and 3) Most of the DE variants are used to solve single-
DE/current-to-best/1/exp trial vector generation strategy to objective optimization problems, while in real life
solve some specific problems. These optimization appli- multi-objective optimization problems usually deal
cations can be roughly classified into these categories: with more than three objective functions. Therefore,
electronic engineering, multimodal optimization, discrete extending multi-objective DE variants with exponential
optimization, mechanical engineering, communication engi- crossover to solve multi-objective problems remains a
neering, fuzzy logic, mathematical problems, aerospace challenging area of future research so far.
engineering, agronomy, geology, structural optimization, 4) It would also be a worthwhile and challenging idea
water resources management, graphics processing, elec- to develop a parameter-free DE variant with expo-
tricity, optics, production management, machine learning, nential crossover to eliminate the sensitivity of DE
biomedicine, combinatorial optimization, chemistry, data search performance to algorithm-specific parameter
processing, emergency management, routing and trans- changes. Compared to most existing DE variants, the
portation planning, banking, and these categories are also parameter-free DE variant with exponential crossover
displayed by us in tabular form, as shown in Table 24. It is does not demand any human intervention to determine
worth mentioning that Table 24 presents the general field the best combination of algorithm-specific parameters
of application of DE algorithms with exponential crossover, for solving any given optimization problems, specif-
while Table 23 shows that ones with exponential crossover ically, it can automatically and adaptively fine-tune
are applied to solve a specific problem, in other words, the parameters to the problem being solved based on
Table 23 shows that ones with exponential crossover are used the currently searched target landscape and population
to solve specific problems in various fields. Readers who diversity, etc.
are interested in a particular issue can be redirected to the 5) Opposite based learning is a method of introducing
corresponding references to learn more about the application. opposite ideas into the learning process, aiming to

VOLUME 11, 2023 131699

 J. Li, Z. Meng: DE With Exponential Crossover: An Experimental Analysis

TABLE 23. Summary of some specific applications of DE algorithms with exponential crossover.

131700 VOLUME 11, 2023

J. Li, Z. Meng: DE With Exponential Crossover: An Experimental Analysis

TABLE 24. Different application areas of DE algorithms with exponential crossover.

improve learning efficiency and performance. It intro- spaces, and the existing literature consistently demonstrates
duces opposite phenotypes into the individual learning that DE exhibits remarkable performance in optimizing a
mechanism, allowing learning to proceed from the wide range of optimization problems.
opposite direction, thereby more comprehensively The overall performance of a certain DE algorithm is
exploring the solution space. Therefore, further intro- often influenced by the choice of crossover scheme, as dif-
ducing this mechanism into the DE with exponential ferent schemes exhibit distinctive characteristics. Binomial
crossover to enhance population diversity and thereby crossover and exponential crossover are two fundamental
improve algorithm performance may also be a research schemes used in DE, each displaying unique performance
focus for relevant scholars in the future. behaviors. When it comes to addressing real parameter
optimization problems, most DE researchers believe that
VII. CONCLUSION exponential crossover is not well-suited for such problems,
Life is filled with an ever-increasing complexity of real-world while binomial crossover is effective in tackling them, which
optimization problems, and there is a growing demand among is the main reason why binomial crossover has gained
researchers from various fields for robust, fast, and accurate widespread popularity. As a result, during the development
optimizers. Since its inception, DE has emerged as a simple of DE, the focus of researchers has predominantly been
and efficient approach for global optimization in continuous on studying DE variants with binomial crossover, with less

VOLUME 11, 2023 131701

 J. Li, Z. Meng: DE With Exponential Crossover: An Experimental Analysis

attention given to DE variants using exponential crossover, [12] Z. Meng and C. Yang, ‘‘Hip-DE: Historical population based mutation
let alone the papers on thorough experimental analysis strategy in differential evolution with parameter adaptive mechanism,’’
Inf. Sci., vol. 562, pp. 44–77, Jul. 2021.
of DE algorithm with exponential crossover on numerical [13] Z. Meng, ‘‘Dimension improvements based adaptation of control param-
optimization. eters in differential evolution: A fitness-value-independent approach,’’
For the sake of filling the above deficiencies in the DE area, Expert Syst. Appl., vol. 223, Aug. 2023, Art. no. 119848.
[14] Z. Meng, Z. Song, X. Shao, J. Zhang, and H. Xu, ‘‘FD-DE: Differential
let novice and expert researchers have a deep understanding evolution with fitness deviation based adaptation in parameter control,’’
of the DE algorithm with exponential crossover in recent ISA Trans., vol. 139, pp. 272–290, Aug. 2023.
years, this paper attempts to provide an experimental analysis [15] Z. Meng and Q. Zhang, ‘‘HPDE: A dynamic hierarchical population
based differential evolution with novel diversity metric,’’ Eng. Appl. Artif.
of DE algorithm with exponential crossover on numerical Intell., vol. 126, Nov. 2023, Art. no. 106989.
optimization. Starting with a comprehensive literature review [16] Z. Meng, J.-S. Pan, and L. Kong, ‘‘Parameters with adaptive learning
of the some DE algorithms with exponential crossover. Next, mechanism (PALM) for the enhancement of differential evolution,’’
it provides a mathematical analysis of the DE algorithm with Knowl.-Based Syst., vol. 141, pp. 92–112, Feb. 2018.
[17] Z. Meng, J.-S. Pan, and K.-K. Tseng, ‘‘PaDE: An enhanced differential
exponential crossover, and it is proved from the mathematical evolution algorithm with novel control parameter adaptation schemes
point of view that the ones with exponential crossover can for numerical optimization,’’ Knowl.-Based Syst., vol. 168, pp. 80–99,
compete with the ones with binomial crossover in numerical Mar. 2019.
[18] Z. Meng, J.-S. Pan, and H. Xu, ‘‘QUasi-affine TRansformation evo-
optimization problems as long as the appropriate crossover lutionary (QUATRE) algorithm: A cooperative swarm based algorithm
rate CR and its parameter control can be found. Then, several for global optimization,’’ Knowl.-Based Syst., vol. 109, pp. 104–121,
typical DE variants with exponential crossover and ones with Oct. 2016.
[19] Z. Meng and J.-S. Pan, ‘‘QUasi-affine TRansformation Evolution-
binomial crossover are selected and compared with the DE ary (QUATRE) algorithm: A parameter-reduced differential evolution
with exponential crossover, ERM-DE, proposed by us under algorithm for optimization problems,’’ in Proc. IEEE Congr. Evol.
the test sets CEC2013, CEC2014, CEC2017 and CEC2022, Comput. (CEC), Jul. 2016, pp. 4082–4089.
[20] Z. Meng and J.-S. Pan, ‘‘QUasi-affine TRansformation evolution with
so as to verify the views proposed in the above mathematical external ARchive (QUATRE-EAR): An enhanced structure for differen-
analysis. Furthermore, our ERM-DE is also validated under tial evolution,’’ Knowl.-Based Syst., vol. 155, pp. 35–53, Sep. 2018.
real-world optimization applications, and the results also [21] Z. Meng, J.-S. Pan, and F. Lin, ‘‘The QUATRE structure: An efficient
approach to tackling the structure bias in differential evolution,’’ in Proc.
prove its superiority as well as the various applications of DE IEEE Int. Conf. Syst. Man Cybern. (SMC), Oct. 2019, pp. 3074–3079.
algorithms with exponential crossover are summarized and [22] Z. Meng, Y. Chen, X. Li, C. Yang, and Y. Zhong, ‘‘Enhancing QUasi-
classified. It unfolds a number of future challenges, potential affine TRansformation evolution (QUATRE) with adaptation scheme
research directions, and conclusions as well. The content of on numerical optimization,’’ Knowl.-Based Syst., vol. 197, Jun. 2020,
Art. no. 105908.
the paper indicates the fact that the study of DE variants with [23] Z. Meng and J. Zhang, ‘‘QUATRE-EMS: QUATRE algorithm with novel
exponential crossover has a lot of room for improvement and adaptation of evolution matrix and selection operation for numerical
may be a good direction for research in the next few years. optimization,’’ Inf. Sci., vol. 651, Dec. 2023, Art. no. 119714.
[24] C. Lin, A. Qing, and Q. Feng, ‘‘A comparative study of crossover
in differential evolution,’’ J. Heuristics, vol. 17, no. 6, pp. 675–703,
REFERENCES Dec. 2011.
[25] X. Qiu, K. C. Tan, and J.-X. Xu, ‘‘Multiple exponential recombination
[1] H. Spencer, The Principles of Biology, vol. 1. New York, NY, USA:
for differential evolution,’’ IEEE Trans. Cybern., vol. 47, no. 4,
D. Appleton, 1896.
pp. 995–1006, Apr. 2017.
[2] L. Pauhng, The Nature of the Chemical Bond. New York, NY, USA:
[26] S.-Z. Zhao and P. N. Suganthan, ‘‘Empirical investigations into the
Cornell Univ. Press, 1960.
exponential crossover of differential evolutions,’’ Swarm Evol. Comput.,
[3] Z. Meng and J.-S. Pan, ‘‘Monkey king evolution: A new memetic vol. 9, pp. 27–36, Apr. 2013.
evolutionary algorithm and its application in vehicle fuel consumption
[27] A. LaTorre, S. Muelas, and J.-M. Peña, ‘‘A MOS-based dynamic memetic
optimization,’’ Knowl.-Based Syst., vol. 97, pp. 144–157, Apr. 2016.
differential evolution algorithm for continuous optimization: A scalability
[4] T. Bäck and H.-P. Schwefel, ‘‘An overview of evolutionary algorithms test,’’ Soft Comput., vol. 15, no. 11, pp. 2187–2199, Nov. 2011.
for parameter optimization,’’ Evol. Comput., vol. 1, no. 1, pp. 1–23, [28] S.-Z. Zhao, P. N. Suganthan, and S. Das, ‘‘Self-adaptive differential
Mar. 1993. evolution with multi-trajectory search for large-scale optimization,’’ Soft
[5] S. S. Rao, Engineering Optimization: Theory and Practice. Hoboken, NJ, Comput., vol. 15, no. 11, pp. 2175–2185, Nov. 2011.
USA: Wiley, 2009. [29] J. Brest and M. S. Maučec, ‘‘Self-adaptive differential evolution algorithm
[6] K. Price, R. M. Storn, and J. A. Lampinen, Differential Evolution: using population size reduction and three strategies,’’ Soft Comput.,
A Practical Approach to Global Optimization. Cham, Switzerland: vol. 15, no. 11, pp. 2157–2174, Nov. 2011.
Springer, 2006. [30] E. Mezura-Montes, J. Velázquez-Reyes, and C. A. C. Coello, ‘‘A compar-
[7] A. P. Engelbrecht, Computational Intelligence: An Introduction. Hobo- ative study of differential evolution variants for global optimization,’’ in
ken, NJ, USA: Wiley, 2007. Proc. 8th Annu. Conf. Genetic Evol. Comput., 2006, pp. 485–492.
[8] J. Holland, Adaptation in Natural and Artificial Systems: An Intro- [31] D. Zaharie, ‘‘A comparative analysis of crossover variants in differential
ductory Analysis With Applications to Biology, Control, and Artificial evolution,’’ in Proc. IMCSIT, 2007, pp. 171–181.
Intelligence. 1975. [32] D. Zaharie, ‘‘Influence of crossover on the behavior of differential
[9] J. Kennedy and R. Eberhart, ‘‘Particle swarm optimization,’’ in Proc. evolution algorithms,’’ Appl. Soft Comput., vol. 9, no. 3, pp. 1126–1138,
IEEE ICNN, vol. 4, Nov./Dec. 1995, pp. 1942–1948. Jun. 2009.
[10] R. Storn and K. Price, ‘‘Differential evolution—A simple and efficient [33] G. Jeyakumar and C. S. Velayutham, ‘‘An empirical performance analysis
heuristic for global optimization over continuous spaces,’’ J. Global of differential evolution variants on unconstrained global optimization
Optim., vol. 11, no. 4, pp. 341–359, 1997. problems,’’ Int. J. Comput. Inf. Syst. Ind. Manag. Appl., vol. 2, pp. 77–86,
[11] Z. Meng and J.-S. Pan, ‘‘HARD-DE: Hierarchical ARchive based muta- Jan. 2010.
tion strategy with depth information of evolution for the enhancement of [34] J. Montgomery, ‘‘Crossover and the different faces of differential
differential evolution on numerical optimization,’’ IEEE Access, vol. 7, evolution searches,’’ in Proc. IEEE Congr. Evol. Comput., Jul. 2010,
pp. 12832–12854, 2019. pp. 1–8.

131702 VOLUME 11, 2023

J. Li, Z. Meng: DE With Exponential Crossover: An Experimental Analysis

[35] S. Thangavelu, G. Jeyakumar, and C. S. Velyautham, ‘‘Population [56] T. Takahama and S. Sakai, ‘‘Constrained optimization by applying the
variance based empirical analysis of the behavior of differential evolution α constrained method to the nonlinear simplex method with mutations,’’
variants,’’ Appl. Math. Sci., vol. 9, no. 66, pp. 3249–3263, 2015. IEEE Trans. Evol. Comput., vol. 9, no. 5, pp. 437–451, Oct. 2005.
[36] S. A. Sharma and P. P. Prajapati, ‘‘Implementation of differential [57] E. Mezura-Montes and C. A. C. Coello, ‘‘Constraint-handling in nature-
evolutionary algorithm for different approaches,’’ Int. J. Electron. inspired numerical optimization: Past, present and future,’’ Swarm Evol.
Commun., vol. 4, no. 4, pp. 23–28, 2016. Comput., vol. 1, no. 4, pp. 173–194, Dec. 2011.
[37] R. Salkoski and I. Chorbev, ‘‘Influence of weighting factor and crossover [58] T. Takahama and S. Sakai, ‘‘Constrained optimization by the ε
constant on the behavior of differential evolution algorithms with a constrained differential evolution with gradient-based mutation and
penalty function approach,’’ in Proc. ICT Innov., Web, 2018, pp. 1–15. feasible elites,’’ in Proc. IEEE Int. Conf. Evol. Comput., Jul. 2006,
[38] P. Bujok, ‘‘On the performance and complexity of crossover in differential pp. 1–8.
evolution algorithm,’’ in Proc. Int. Conf. Artif. Intell. Soft Comput., [59] T. Takahama and S. Sakai, ‘‘Solving difficult constrained optimization
Zakopane, Poland, Oct. 2020, pp. 363–375. problems by the ε constrained differential evolution with gradient-based
[39] C. Zhang, A. K. Qin, W. Shen, L. Gao, K. C. Tan, and X. Li, mutation,’’ in Constraint-Handling in Evolutionary Optimization. Berlin,
‘‘E-constrained differential evolution using an adaptive e-level control Germany: Springer, 2009, pp. 51–72.
method,’’ IEEE Trans. Syst. Man, Cybern. Syst., vol. 52, no. 2, [60] T. Takahama and S. Sakai, ‘‘Constrained optimization by the ε
pp. 769–785, Feb. 2022. constrained differential evolution with an archive and gradient-
[40] Z. Meng and C. Yang, ‘‘Two-stage differential evolution with novel based mutation,’’ in Proc. IEEE Congr. Evol. Comput., Jul. 2010,
parameter control,’’ Inf. Sci., vol. 596, pp. 321–342, Jun. 2022. pp. 1–9.
[41] J. Li and Z. Meng, ‘‘Elite-guided resampling and mutli-mutation [61] T. Takahama and S. Sakai, ‘‘Efficient constrained optimization by the
based differential evolution with exponential crossover for numerical ε constrained rank-based differential evolution,’’ in Proc. IEEE Congr.
optimization,’’ Appl. Soft Comput., to be published. Evol. Comput., Jun. 2012, pp. 1–8.
[42] J. J. Liang, B. Qu, P. N. Suganthan, and A. G. Hernández-Díaz, ‘‘Problem [62] T. Takahama and S. Sakai, ‘‘Efficient constrained optimization by the
definitions and evaluation criteria for the CEC 2013 special session ε constrained differential evolution with rough approximation using
on real-parameter optimization,’’ Comput. Intell. Lab., Zhengzhou kernel regression,’’ in Proc. IEEE Congr. Evol. Comput., Jun. 2013,
Univ., Zhengzhou, China, Nanyang Technol. Univ., Singapore, pp. 1334–1341.
Tech. Rep. 201212, 2013, pp. 281–295, vol. 201212, no. 34. [63] C. Bu, W. Luo, and T. Zhu, ‘‘Differential evolution with a species-based
[43] J. J. Liang, B. Y. Qu, and P. N. Suganthan, ‘‘Problem definitions and repair strategy for constrained optimization,’’ in Proc. IEEE Congr. Evol.
evaluation criteria for the CEC 2014 special session and competition on Comput. (CEC), Jul. 2014, pp. 967–974.
single objective real-parameter numerical optimization,’’ Comput. Intell. [64] W. Yi, H. Qiu, Y. Chen, J. Lu, Z. Pei, and C. Zhang, ‘‘An improved
Lab., Zhengzhou Univ., Zhengzhou China, Nanyang Technol. Univ., adaptive differential evolution approach for constrained optimization
Singapore, Tech. Rep. 635, 2013, p. 490, vol. 635. problems,’’ in Proc. IEEE 24th Int. Conf. Comput. Supported Cooperat.
[44] G. Wu, R. Mallipeddi, and P. N. Suganthan, ‘‘Problem definitions and Work Design (CSCWD), May 2021, pp. 696–701.
evaluation criteria for the CEC 2017 competition on constrained real- [65] H. K. Singh and T. Ray, ‘‘Performance of a hybrid EA-DE-memetic
parameter optimization,’’ Nat. Univ. Defense Technol., Changsha, Hunan, algorithm on CEC 2011 real world optimization problems,’’ in Proc. IEEE
PR China Kyungpook Nat. Univ., Daegu, South Korea Nanyang Technol. Congr. Evol. Comput. (CEC), Jun. 2011, pp. 1322–1326.
Univ., Singapore, Tech. Rep. 201709, 2017.
[66] K. Yasuda, K. Makise, and K. Tamura, ‘‘A study on combination of
[45] A. Kumar, K. V. Price, A. W. Mohamed, A. A. Hadi, and P. N. Suganthan,
differential evolution and evolution strategy,’’ in Proc. IEEE Int. Conf.
‘‘Problem definitions and evaluation criteria for the CEC 2022 special
Syst., Man, Cybern. (SMC), Oct. 2012, pp. 587–592.
session and competition on single objective bound constrained numerical
[67] M. Weber, F. Neri, and V. Tirronen, ‘‘Shuffle or update parallel differential
optimization,’’ Nanyang Technol. Univ. Singapore, Tech. Rep. 202112,
evolution for large-scale optimization,’’ Soft Comput., vol. 15, no. 11,
2022.
pp. 2089–2107, Nov. 2011.
[46] S. Rahnamayan, H. R. Tizhoosh, and M. M. A. Salama, ‘‘Opposition-
[68] M. Weber, F. Neri, and V. Tirronen, ‘‘A study on scale factor/crossover
based differential evolution,’’ IEEE Trans. Evol. Comput., vol. 12, no. 1,
interaction in distributed differential evolution,’’ Artif. Intell. Rev., vol. 39,
pp. 64–79, Feb. 2008.
no. 3, pp. 195–224, Mar. 2013.
[47] H. Wang, Z. Wu, and S. Rahnamayan, ‘‘Enhanced opposition-based dif-
ferential evolution for solving high-dimensional continuous optimization [69] D. K. Tasoulis, N. G. Pavlidis, V. P. Plagianakos, and M. N. Vrahatis,
problems,’’ Soft Comput., vol. 15, no. 11, pp. 2127–2140, Nov. 2011. ‘‘Parallel differential evolution,’’ in Proc. Congr. Evol. Comput.,
[48] H. Wang, Z. Wu, Y. Liu, J. Wang, D. Jiang, and L. Chen, ‘‘Space Jun. 2004, pp. 2023–2029.
transformation search: A new evolutionary technique,’’ in Proc. 1st [70] J. Apolloni, G. Leguizamón, J. García-Nieto, and E. Alba, ‘‘Island
ACM/SIGEVO Summit Genetic Evol. Comput., 2009, pp. 537–544. based distributed differential evolution: An experimental study on hybrid
[49] H. Wang, Z. Wu, S. Rahnamayan, and L. Kang, ‘‘A scalability test for testbeds,’’ in Proc. 8th Int. Conf. Hybrid Intell. Syst., Sep. 2008,
accelerated DE using generalized opposition-based learning,’’ in Proc. pp. 696–701.
9th Int. Conf. Intell. Syst. Design Appl., Nov. 2009, pp. 1090–1095. [71] I. D. Falco, A. D. Cioppa, D. Maisto, U. Scafuri, and E. Tarantino,
[50] J. Brest, S. Greiner, B. Boskovic, M. Mernik, and V. Zumer, ‘‘Self- ‘‘Satellite image registration by distributed differential evolution,’’ in
adapting control parameters in differential evolution: A comparative Proc. Workshops Appl. Evol. Comput. Cham, Switzerland: Springer,
study on numerical benchmark problems,’’ IEEE Trans. Evol. Comput., 2007, pp. 251–260.
vol. 10, no. 6, pp. 646–657, Dec. 2006. [72] I. De Falco, D. Maisto, U. Scafuri, E. Tarantino, and A. D. Cioppa,
[51] H. Wang, S. Rahnamayan, and Z. Wu, ‘‘Parallel differential evolution ‘‘Distributed differential evolution for the registration of remotely sensed
with self-adapting control parameters and generalized opposition-based images,’’ in Proc. 15th EUROMICRO Int. Conf. Parallel, Distrib. Netw.-
learning for solving high-dimensional optimization problems,’’ J. Parallel Based Process. (PDP), Feb. 2007, pp. 358–362.
Distrib. Comput., vol. 73, no. 1, pp. 62–73, Jan. 2013. [73] I. De Falco, U. Scafuri, E. Tarantino, and A. Della Cioppa, ‘‘A distributed
[52] H. Liu, Z. Wu, H. Wang, S. Rahnamayan, and C. Deng, ‘‘Improved differential evolution approach for mapping in a grid environment,’’
differential evolution with adaptive opposition strategy,’’ in Proc. IEEE in Proc. 15th EUROMICRO Int. Conf. Parallel, Distrib. Netw.-Based
Congr. Evol. Comput. (CEC), Jul. 2014, pp. 1776–1783. Process. (PDP), Feb. 2007, pp. 442–449.
[53] J.-i. Kushida, A. Hara, and T. Takahama, ‘‘An improvement of opposition- [74] F. Ramezani and S. Lotfi, ‘‘The modified differential evolution algorithm
based differential evolution with archive solutions,’’ in Proc. Int. Conf. (MDEA),’’ in Proc. Asian Conf. Intell. Inf. Database Syst., 2012,
Adv. Mech. Syst., Aug. 2014, pp. 463–468. pp. 109–118.
[54] Z. Michalewicz, ‘‘A survey of constraint handling techniques in [75] C. García-Martínez, F. J. Rodríguez, and M. Lozano, ‘‘Role differen-
evolutionary computation methods,’’ Evol. Program., vol. 4, pp. 135–155, tiation and malleable mating for differential evolution: An analysis on
Mar. 1995. large-scale optimisation,’’ Soft Comput., vol. 15, no. 11, pp. 2109–2126,
[55] C. A. C. Coello, ‘‘Theoretical and numerical constraint-handling Nov. 2011.
techniques used with evolutionary algorithms: A survey of the state of [76] H. Wang, S. Rahnamayan, H. Sun, and M. G. H. Omran, ‘‘Gaussian
the art,’’ Comput. Methods Appl. Mech. Eng., vol. 191, nos. 11–12, bare-bones differential evolution,’’ IEEE Trans. Cybern., vol. 43, no. 2,
pp. 1245–1287, Jan. 2002. pp. 634–647, Apr. 2013.

VOLUME 11, 2023 131703

 J. Li, Z. Meng: DE With Exponential Crossover: An Experimental Analysis

[77] H. Wang, Z. Xiao, Y. Zhang, X. Liao, Y. Ruan, J. Fu, and D. Jiang, [99] Z. Meng and Y. Chen, ‘‘Differential evolution with exponential crossover
‘‘A scalability test of Gaussian bare-bones differential evolution on high- can be also competitive on numerical optimization,’’ Appl. Soft Comput.,
dimensional optimization problems,’’ in Proc. 9th Int. Conf. Natural vol. 146, Oct. 2023, Art. no. 110750.
Comput. (ICNC), Jul. 2013, pp. 374–378. [100] J. Zhang and A. C. Sanderson, ‘‘JADE: Adaptive differential evolution
[78] D. Kamiyama, K. Tamura, and K. Yasuda, ‘‘Down-hill simplex method with optional external archive,’’ IEEE Trans. Evol. Comput., vol. 13, no. 5,
based differential evolution,’’ in Proc. SICE Annu. Conf., Aug. 2010, pp. 945–958, Oct. 2009.
pp. 1641–1646. [101] R. Tanabe and A. Fukunaga, ‘‘Success-history based parameter adap-
[79] J. A. Nelder and R. Mead, ‘‘A simplex method for function minimiza- tation for differential evolution,’’ in Proc. IEEE Congr. Evol. Comput.,
tion,’’ Comput. J., vol. 7, no. 4, pp. 308–313, Jan. 1965. Jun. 2013, pp. 71–78.
[80] D. Kamiyama, K. Tamura, and K. Yasuda, ‘‘Local descent direction [102] R. Tanabe and A. S. Fukunaga, ‘‘Improving the search performance of
vector-based differential evolution for multiobjective optimization,’’ SHADE using linear population size reduction,’’ in Proc. IEEE Congr.
Electron. Commun. Jpn., vol. 97, no. 5, pp. 36–42, May 2014. Evol. Comput. (CEC), Jul. 2014, pp. 1658–1665.
[81] E. Zitzler and L. Thiele, ‘‘Multiobjective optimization using evolutionary [103] Z. Meng, Y. Zhong, and C. Yang, ‘‘CS-DE: Cooperative strategy based
algorithms—A comparative case study,’’ in Proc. Int. Conf. Parallel differential evolution with population diversity enhancement,’’ Inf. Sci.,
Problem Solving From Nature, Amsterdam, The Netherlands, Sep. 1998, vol. 577, pp. 663–696, Oct. 2021.
pp. 292–301. [104] A. Kumar, S. Das, and I. Zelinka, ‘‘A modified covariance matrix
[82] E. Zitzler, L. Thiele, M. Laumanns, C. M. Fonseca, and V. G. da Fonseca, adaptation evolution strategy for real-world constrained optimization
‘‘Performance assessment of multiobjective optimizers: An analysis problems,’’ in Proc. Genetic Evol. Comput. Conf. Companion, Jul. 2020,
and review,’’ IEEE Trans. Evol. Comput., vol. 7, no. 2, pp. 117–132, pp. 11–12.
Apr. 2003. [105] K. M. Sallam, S. M. Elsayed, R. K. Chakrabortty, and M. J. Ryan,
[83] F. Neri, G. Iacca, and E. Mininno, ‘‘Disturbed exploitation compact ‘‘Multi-operator differential evolution algorithm for solving real-world
differential evolution for limited memory optimization problems,’’ Inf. constrained optimization problems,’’ in Proc. IEEE Congr. Evol.
Sci., vol. 181, no. 12, pp. 2469–2487, Jun. 2011. Comput., Sep. 2020, pp. 1–8.
[84] I. Poikolainen and F. Neri, ‘‘Differential evolution with concurrent fitness [106] J. Gurrola-Ramos, A. Hernández-Aguirre, and O. Dalmau-Cedeño,
based local search,’’ in Proc. IEEE Congr. Evol. Comput., Jun. 2013, ‘‘COLSHADE for real-world single-objective constrained optimization
pp. 384–391. problems,’’ in Proc. IEEE Congr. Evol. Comput. (CEC), Jul. 2020,
[85] P. Zibin, ‘‘Performance analysis and improvement of parallel differential pp. 1–8.
evolution,’’ 2021, arXiv:2101.06599. [107] R. Stom and K. Price, ‘‘Differential evolution—A simple and efficient
[86] R. Tanabe and A. Fukunaga, ‘‘Reevaluating exponential crossover in adaptive scheme for global optimization over continuous spaces,’’ ICSI,
differential evolution,’’ in Proc. Int. Conf. Parallel Problem Solving From New Delhi, India, Tech. Rep. TR-95.012, 1995.
Nature. Cham, Switzerland: Springer, 2014, pp. 201–210. [108] R. Storn and K. Price, ‘‘Minimizing the real functions of the ICEC’96
[87] A. K. Qin, V. L. Huang, and P. N. Suganthan, ‘‘Differential evolution contest by differential evolution,’’ in Proc. IEEE Int. Conf. Evol. Comput.,
algorithm with strategy adaptation for global numerical optimization,’’ May 1996, pp. 842–844.
IEEE Trans. Evol. Comput., vol. 13, no. 2, pp. 398–417, Apr. 2009. [109] R. Storn and K. Price, ‘‘Differential evolution—A simple evolution
[88] A. K. Qin and P. N. Suganthan, ‘‘Self-adaptive differential evolution strategy for fast optimization,’’ Dr. Dobb’s J., vol. 22, no. 4, pp. 18–24,
algorithm for numerical optimization,’’ in Proc. IEEE Congr. Evol. 1997.
Comput., Sep. 2005, pp. 1785–1791. [110] R. Joshi and A. C. Sanderson, ‘‘Minimal representation multisensor
[89] V. L. Huang, A. K. Qin, and P. N. Suganthan, ‘‘Self-adaptive differential fusion using differential evolution,’’ IEEE Trans. Syst. Man, Cybern. A,
evolution algorithm for constrained real-parameter optimization,’’ in Syst. Humans, vol. 29, no. 1, pp. 63–76, Jan. 1999.
Proc. IEEE Int. Conf. Evol. Comput., Jul. 2006, pp. 17–24. [111] R. Storn, ‘‘System design by constraint adaptation and differential
[90] P. Bujok and J. Tvrdík, ‘‘Adaptive differential evolution: SHADE with evolution,’’ IEEE Trans. Evol. Comput., vol. 3, no. 1, pp. 22–34,
competing crossover strategies,’’ in Proc. Int. Conf. Artif. Intell. Soft Apr. 1999.
Comput. Cham, Switzerland: Springer, 2015, pp. 329–339. [112] R. Thomsen, ‘‘Multimodal optimization using crowding-based differen-
[91] S. Hui and P. N. Suganthan, ‘‘Ensemble differential evolution with tial evolution,’’ in Proc. Congr. Evol. Comput., Jun. 2004, pp. 1382–1389.
dynamic subpopulations and adaptive clearing for solving dynamic [113] T. Gong and A. L. Tuson, ‘‘Differential evolution for binary encoding,’’ in
optimization problems,’’ in Proc. IEEE Congr. Evol. Comput., Jun. 2012, Soft Computing in Industrial Applications. Cham, Switzerland: Springer,
pp. 1–8. 2007, pp. 251–262.
[92] T. Takahama and S. Sakai, ‘‘Large scale optimization by differential [114] F. A. C. Viana, F. C. S. G. de Oliveira, J. A. F. Borges, and V. Steffen Jr.,
evolution with landscape modality detection and a diversity archive,’’ in ‘‘Differential evolution applied to the design of a three-dimensional
Proc. IEEE Congr. Evol. Comput., Jun. 2012, pp. 1–8. vehicular structure,’’ in Proc. Int. Design Eng. Tech. Conf. Comput. Inf.
[93] J.-i. Kushida, A. Hara, and T. Takahama, ‘‘Rank-based differential Eng. Conf., vol. 48078, 2007, pp. 1321–1330.
evolution with multiple mutation strategies for large scale global [115] S. Almeida-Luz, M. A. Vega-Rodriguez, J. A. Gomez-Pulido, and
optimization,’’ in Proc. IEEE Congr. Evol. Comput. (CEC), May 2015, J. M. Sanchez-Perez, ‘‘A differential evolution algorithm for location
pp. 353–360. area problem in mobile networks,’’ in Proc. 15th Int. Conf. Softw.,
[94] N. Lynn, R. Mallipeddi, and P. N. Suganthan, ‘‘Differential evolution with Telecommun. Comput. Netw., Sep. 2007, pp. 1–5.
two subpopulations,’’ in Proc. International Conf. Swarm, Evol., Memetic [116] B. V. Babu and S. A. Munawar, ‘‘Differential evolution strategies for
Comput., Bhubaneswar, India, Dec. 2014, pp. 1–13. optimal design of shell-and-tube heat exchangers,’’ Chem. Eng. Sci.,
[95] M. Z. Ali, N. Awad, R. Duwairi, J. Albadarneh, R. G. Reynolds, and P. N. vol. 62, no. 14, pp. 3720–3739, Jul. 2007.
Suganthan, ‘‘Cluster-based differential evolution with heterogeneous [117] S. Paul, S. Kumar, and L. Singh, ‘‘A novel evolutionary TSK-subsethood
influence for numerical optimization,’’ in Proc. IEEE Congr. Evol. model and its parallel implementation,’’ in Proc. IEEE Int. Conf. Fuzzy
Comput. (CEC), Mar. 2015, pp. 393–400. Syst. (IEEE World Congr. Comput. Intell.), Jun. 2008, pp. 1880–1885.
[96] M. Z. Ali, N. H. Awad, P. N. Suganthan, and R. G. Reynolds, ‘‘An [118] L. Singh, S. Kumar, and S. Paul, ‘‘Automatic simultaneous architecture
adaptive multipopulation differential evolution with dynamic population and parameter search in fuzzy neural network learning using novel
reduction,’’ IEEE Trans. Cybern., vol. 47, no. 9, pp. 2768–2779, variable length crossover differential evolution,’’ in Proc. IEEE Int.
Sep. 2017. Conf. Fuzzy Syst. (IEEE World Congr. Comput. Intelligence), Jun. 2008,
[97] K. M. Sallam, S. M. Elsayed, R. K. Chakrabortty, and M. J. Ryan, pp. 1795–1802.
‘‘Improved multi-operator differential evolution algorithm for solving [119] B. Sarkar, ‘‘An efficient method for near-optimal polygonal approxima-
unconstrained problems,’’ in Proc. IEEE Congr. Evol. Comput. (CEC), tion based on differential evolution,’’ Int. J. Pattern Recognit. Artif. Intell.,
Jul. 2020, pp. 1–8. vol. 22, no. 6, pp. 1267–1281, Sep. 2008.
[98] S. Aggarwal and K. K. Mishra, ‘‘Multi-operator differential evolution [120] C. Ampatzis and D. Izzo, ‘‘Machine learning techniques for approxima-
with MOEA/D for solving multi-objective optimization problems,’’ tion of objective functions in trajectory optimisation,’’ in Proc. IJCAI
J. Telecommun. Inf. Technol., vol. 3, no. 2022, pp. 85–95, Sep. 2022. Workshop Artif. Intell. Space, 2009, pp. 1–6.

131704 VOLUME 11, 2023

J. Li, Z. Meng: DE With Exponential Crossover: An Experimental Analysis

[121] J. Adeyemo and F. Otieno, ‘‘Differential evolution algorithm for solving [143] T. Takahama and S. Sakai, ‘‘Differential evolution with dynamic strategy
multi-objective crop planning model,’’ Agricult. Water Manage., vol. 97, and parameter selection by detecting landscape modality,’’ in Proc. IEEE
no. 6, pp. 848–856, Jun. 2010. Congr. Evol. Comput., Jun. 2012, pp. 1–8.
[122] J. A. Adeyemo and F. O. Otieno, ‘‘Maximum irrigation benefit using [144] G. Iacca, F. Caraffini, F. Neri, and E. Mininno, ‘‘Robot base disturbance
multiobjective differential evolution algorithm (MDEA),’’ OIDA Int. J. optimization with compact differential evolution light,’’ in Proc. Eur.
Sustain. Develop., vol. 1, no. 2, pp. 39–44, 2010. Conf. Appl. Evol. Comput., Málaga, Spain, Apr. 2012, pp. 285–294.
[123] Z. Salam and N. Bahari, ‘‘Selective harmonics elimination PWM [145] W. Wisittipanich and V. Kachitvichyanukul, ‘‘Two enhanced differential
(SHEPWM) using differential evolution approach,’’ in Proc. Joint Int. evolution algorithms for job shop scheduling problems,’’ Int. J. Prod.
Conf. Power Electron., Drives Energy Syst. Power India, Dec. 2010, Res., vol. 50, no. 10, pp. 2757–2773, May 2012.
pp. 1–5. [146] L. Wang and L.-P. Li, ‘‘An effective differential harmony search algorithm
[124] Y. Hajizadeh, M. Christie, and V. Demyanov, ‘‘Comparative study of for the solving non-convex economic load dispatch problems,’’ Int. J.
novel population-based optimization algorithms for history matching and Electr. Power Energy Syst., vol. 44, no. 1, pp. 832–843, Jan. 2013.
uncertainty quantification: PUNQ-S3 revisited,’’ in Proc. Abu Dhabi Int. [147] G. Iacca, F. Caraffini, and F. Neri, ‘‘Memory-saving memetic computing
Petroleum Exhib. Conf., 2010, pp. 1–18. for path-following mobile robots,’’ Appl. Soft Comput., vol. 13, no. 4,
[125] M. Ruciński, D. Izzo, and F. Biscani, ‘‘On the impact of the migration pp. 2003–2016, Apr. 2013.
topology on the island model,’’ Parallel Comput., vol. 36, nos. 10–11, [148] P. Ren and N. Li, ‘‘Optimal estimation for fault sections in power systems
pp. 555–571, Oct. 2010. based on differential harmony search,’’ in Proc. 9th Int. Conf. Natural
[126] Y. Hajizadeh, M. Christie, and V. Demyanov, ‘‘History matching with Comput. (ICNC), Jul. 2013, pp. 817–821.
differential evolution approach; a look at new search strategies,’’ in Proc. [149] K. Andriosopoulos, M. Doumpos, N. C. Papapostolou, and
SPE Europec Featured EAGE Conf. Exhib., 2010, Art. no. 130253. P. K. Pouliasis, ‘‘Portfolio optimization and index tracking for the
[127] J. A. Adeyemo and F. O. Otieno, ‘‘Maximization of hydropower using shipping stock and freight markets using evolutionary algorithms,’’
strategies of differential evolution,’’ OIDA Int. J. Sustain. Develop., vol. 1, Transp. Res. E, Logistics Transp. Rev., vol. 52, pp. 16–34, Jun. 2013.
no. 2, pp. 33–37, 2010. [150] Y. Liu, W.-N. Chen, Z.-H. Zhan, Y. Lin, Y.-J. Gong, and J. Zhang, ‘‘A set-
[128] M. S. Khan, A. F. M. Ayob, A. Isaacs, and T. Ray, ‘‘A novel evolutionary based discrete differential evolution algorithm,’’ in Proc. IEEE Int. Conf.
approach for 2D shape matching based on B-spline modeling,’’ in Proc. Syst., Man, Cybern., Oct. 2013, pp. 1347–1352.
IEEE Congr. Evol. Comput. (CEC), Jun. 2011, pp. 655–661. [151] R. Arunkumar and V. Jothiprakash, ‘‘Chaotic evolutionary algorithms
[129] A. Qing, ‘‘Comment on ‘differential evolution as applied to electro- for multi-reservoir optimization,’’ Water Resour. Manage., vol. 27,
magnetics,’’’ IEEE Antennas Propag. Mag., vol. 53, no. 4, pp. 169–171, pp. 5207–5222, Oct. 2013.
Aug. 2011. [152] S. Nonsiri, F. Christophe, E. Coataneé;a, and F. Mokammel, ‘‘A combined
[130] R. Arunkumar and V. Jothiprakash. (2012). Optimal Reservoir Operation design structure matrix (DSM) and discrete differential evolution (DDE)
for Hydropower Generation Using Differential Evolution Algorithm. approach for scheduling and organizing system development tasks
[Online]. Available: https://api.semanticscholar.org/CorpusID: modelled using SysML,’’ J. Integr. Design Process Sci., vol. 18, no. 3,
221664526 pp. 19–40, 2014.
[131] Y. Zhao, F. Chen, Q. Shen, Q. Liu, and L. Zhang, ‘‘Optimizing low [153] M. Amjad, Z. Salam, and K. Ishaque, ‘‘A model parameter extraction
loss negative index metamaterial for visible spectrum using differential method for dielectric barrier discharge ozone chamber using differential
evolution,’’ Opt. Exp., vol. 19, no. 12, p. 11605, 2011. evolution,’’ Meas. Sci. Rev., vol. 14, no. 2, pp. 78–86, Apr. 2014.
[132] W. Wisittipanich and V. Kachitvichyanukul, ‘‘Differential evolution [154] A. Akutekwe, H. Seker, and S. Iliya, ‘‘An optimized hybrid dynamic
algorithm for job shop scheduling problem,’’ Ind. Eng. Manage. Syst., Bayesian network approach using differential evolution algorithm for
vol. 10, no. 3, pp. 203–208, Sep. 2011. the diagnosis of hepatocellular carcinoma,’’ in Proc. IEEE 6th Int. Conf.
[133] Y. Zhao, F. Chen, Q. Shen, and L. Zhang, ‘‘Light-trapping design of Adapt. Sci. Technol. (ICAST), Oct. 2014, pp. 1–6.
graphene transparent electrodes for efficient thin-film silicon solar cells,’’ [155] S. Ziabakhsh, H. Alavi-Rad, M. A. Ahandani, and M. C. E. Yagoub,
Appl. Opt., vol. 51, no. 25, p. 6245, 2012. ‘‘Design and optimization of a fully differential CMOS variable-gain
[134] F. Neri, M. Weber, F. Caraffini, and I. Poikolainen, ‘‘Meta-Lamarckian LNA with differential evolution algorithm for WLAN applications,’’
learning in three stage optimal memetic exploration,’’ in Proc. 12th U.K. J. Circuits, Syst. Comput., vol. 23, no. 9, Oct. 2014, Art. no. 1450124.
Workshop Comput. Intell. (UKCI), Sep. 2012, pp. 1–8. [156] L. Unzueta, W. Pimenta, J. Goenetxea, L. P. Santos, and F. Dornaika,
[135] X. Zhang, H. Tong, and X. Chen, ‘‘Multi-scale segmentation algorithm ‘‘Efficient generic face model fitting to images and videos,’’ Image Vis.
parameters optimization based on evolutionary computation,’’ in Proc. Comput., vol. 32, no. 5, pp. 321–334, May 2014.
Int. Symp. Intell. Comput. Appl. Cham, Switzerland: Springer, 2012, [157] J. Moravec and P. Pošík, ‘‘A comparative study: The effect of the
pp. 347–358. perturbation vector type in the differential evolution algorithm on the
[136] G. Iacca, F. Neri, E. Mininno, Y.-S. Ong, and M.-H. Lim, ‘‘Ockham’s accuracy of robot pose and heading estimation,’’ Evol. Intell., vol. 6, no. 3,
Razor in memetic computing: Three stage optimal memetic exploration,’’ pp. 171–191, Feb. 2014.
Inf. Sci., vol. 188, pp. 17–43, Apr. 2012. [158] N. S. Tung and S. Chakravorty, ‘‘Active power dispatch planning using
[137] Y. Zhao, F. Chen, Q. Shen, and L. Zhang, ‘‘Optimal design of light differential evolution algorithm,’’ J. Ind. Intell. Inf., vol. 2, no. 3,
trapping in thin-film solar cells enhanced with graded SiNx and SiOxNy pp. 200–204, 2014.
structure,’’ Opt. Exp., vol. 20, no. 10, pp. 11121–11136, 2012. [159] M. G. Villarreal-Cervantes, D. D. L. Cruz-Mucino, and
[138] Y. Zhao, F. Chen, Q. Shen, and L. Zhang, ‘‘Optimizing low loss E. A. Portilla-Flores, ‘‘Optimum design of parallelogram five-bar
silver nanowires structure metamaterial at yellow light spectrum with manipulator for dexterous workspace by using ELEMAEF in differential
differential evolution,’’ Phys. Lett. A, vol. 376, no. 4, pp. 252–256, evolution,’’ Appl. Math. Inf. Sci., vol. 8, no. 5, pp. 2129–2140, Sep. 2014.
Jan. 2012. [160] S. Iliya, E. Goodyer, M. Gongora, J. Shell, and J. Gow, ‘‘Optimized
[139] I. Poikolainen, G. Iacca, F. Neri, E. Mininno, and M. Weber, ‘‘Shrinking artificial neural network using differential evolution for prediction of
three stage optimal memetic exploration,’’ in Proc. 5th Int. Conf. RF power in VHF/UHF TV and GSM 900 bands for cognitive radio
Bioinspired Optim. Methods their Appl., 2012, pp. 24–25. networks,’’ in Proc. 14th U.K. Workshop Comput. Intell. (UKCI),
[140] S. Nonsiri, E. Coatanea, and M. Bakhouya, ‘‘A discrete differen- Sep. 2014, pp. 1–6.
tial evolution algorithm for product development scheduling,’’ in [161] N. Panagant and S. Bureerat, ‘‘Solving partial differential equations using
Proc. 32nd Comput. Inf. Eng. Conf., A B, vol. 2, Aug. 2012, a new differential evolution algorithm,’’ Math. Problems Eng., vol. 2014,
pp. 1229–1236. pp. 1–10, Jan. 2014.
[141] U. K. Chakraborty, T. E. Abbott, and S. K. Das, ‘‘PEM fuel cell modeling [162] N. Khan, F. Neri, and S. Ahmadi, ‘‘Adaptive differential evolution applied
using differential evolution,’’ Energy, vol. 40, no. 1, pp. 387–399, to point matching 2D GIS data,’’ in Proc. IEEE Symp. Ser. Comput. Intell.,
Apr. 2012. Dec. 2015, pp. 1719–1726.
[142] N. D. Jana and J. Sil, ‘‘Protein structure prediction in 2D hp lattice [163] L. Unzueta, W. Pimenta, J. Goenetxea, L. P. Santos, and F. Dornaika,
model using differential evolutionary algorithm,’’ in Proc. Int. Conf. Inf. ‘‘Efficient deformable 3D face model fitting to monocular images,’’
Syst. Design Intell. Appl. (INDIA) Held, Visakhapatnam, India, Jan. 2012, in Face Image Analysis: Theory and Applications. Berlin, Germany:
pp. 281–290. Springer, 2016, pp. 77–102.

VOLUME 11, 2023 131705

 J. Li, Z. Meng: DE With Exponential Crossover: An Experimental Analysis

[164] H. V. H. Ayala, F. M. D. Santos, V. C. Mariani, and L. D. S. Coelho, [184] K. Sethanan, W. Wisittipanich, N. Wisittipanit, K. Nitisiri, and
‘‘Image thresholding segmentation based on a novel beta differential K. Moonsri, ‘‘Integrating scheduling with optimal sublot for parallel
evolution approach,’’ Expert Syst. Appl., vol. 42, no. 4, pp. 2136–2142, machine with job splitting and dependent setup times,’’ Comput. Ind.
Mar. 2015. Eng., vol. 137, Nov. 2019, Art. no. 106095.
[165] A. Maesani, G. Iacca, and D. Floreano, ‘‘Memetic viability evolution for [185] N. Xue, I. Triguero, G. P. Figueredo, and D. Landa-Silva, ‘‘Evolving deep
constrained optimization,’’ IEEE Trans. Evol. Comput., vol. 20, no. 1, CNN-LSTMs for inventory time series prediction,’’ in Proc. IEEE Congr.
pp. 125–144, Feb. 2016. Evol. Comput. (CEC), Jun. 2019, pp. 1517–1524.
[166] M. Li, Y. Zhao, L. He, W. Chen, and X. Xu, ‘‘The parameter calibration [186] M. Mayo and M. Daoud, ‘‘Data normalisation using differential evolution
and optimization of social force model for the real-life 2013 Ya’an and aggregated logistic functions,’’ in Proc. IEEE Congr. Evol. Comput.
earthquake evacuation in China,’’ Saf. Sci., vol. 79, pp. 243–253, (CEC), Jun. 2019, pp. 920–927.
Nov. 2015. [187] M. F. Fateh, A. Zameer, S. M. Mirza, N. M. Mirza, M. S. Aslam, and
[167] N. L. Tsakiridis, J. B. Theocharis, and G. C. Zalidis, ‘‘DECO3R: Dif- M. A. Z. Raja, ‘‘Differential evolution based computation intelligence
ferential evolution based COoperative-COmpeting learning of COmpact solver for elliptic partial differential equations,’’ Frontiers Inf. Technol.
fuzzy rulebased classification systems,’’ in Proc. IEEE Int. Conf. Fuzzy Electron. Eng., vol. 20, no. 10, pp. 1445–1456, Oct. 2019.
Syst. (FUZZ-IEEE), Aug. 2015, pp. 1–8. [188] W. Xu, W. K. Wong, K. C. Tan, and J.-X. Xu, ‘‘Finding high-dimensional
[168] N. L. Tsakiridis, J. B. Theocharis, R. Pérez, A. González, and D-optimal designs for logistic models via differential evolution,’’ IEEE
G. C. Zalidis, ‘‘Extensions of the DECO3R algorithm for generating Access, vol. 7, pp. 7133–7146, 2019.
compact and cooperating fuzzy rule-based classification systems,’’ [189] A. Nasim, L. Burattini, M. F. Fateh, and A. Zameer, ‘‘Solution of linear
in Proc. IEEE Int. Conf. Fuzzy Syst. (FUZZ-IEEE), Jul. 2016, and non-linear boundary value problems using population-distributed
pp. 1429–1436. parallel differential evolution,’’ J. Artif. Intell. Soft Comput. Res., vol. 9,
[169] H. Wanli, ‘‘Parameters optimization of region growing segmentation no. 3, pp. 205–218, Jul. 2019.
based on differential evolution algorithm,’’ in Proc. Int. Conf. Adv. Mech. [190] J. Su, D. V. Vargas, and K. Sakurai, ‘‘Attacking convolutional neural
Eng. Ind. Informat., 2015, pp. 153–158. network using differential evolution,’’ IPSJ Trans. Comput. Vis. Appl.,
[170] S. Kunnapapdeelert and V. Kachitvichyanukul, ‘‘Modified de algorithms vol. 11, no. 1, pp. 1–16, Dec. 2019.
for solving multi-depot vehicle routing problem with multiple pickup and [191] G. J. Krishna and V. Ravi, ‘‘Feature subset selection using adaptive
delivery requests,’’ in Toward Sustainable Operations of Supply Chain differential evolution: An application to banking,’’ in Proc. ACM India
and Logistics Systems. Berlin, Germany: Springer, 2015, pp. 361–373. Joint Int. Conf. Data Sci. Manage. Data, 2019, pp. 157–163.
[171] X.-Q. Li, Z. Wang, and L.-H. Fu, ‘‘A laser-based measuring system for [192] H. Dong, L. Gao, X. Li, H. Zhong, and B. Zeng, ‘‘Differential evolution
online quality control of car engine block,’’ Sensors, vol. 16, no. 11, with better and nearest option for function optimization,’’ in Proc. IEEE
p. 1877, Nov. 2016. Congr. Evol. Comput. (CEC), Jun. 2019, pp. 641–645.
[172] N. L. Tsakiridis, J. B. Theocharis, and G. C. Zalidis, ‘‘A fuzzy rule-based [193] X. Song, Y. Wu, and M. Zhao, ‘‘Differential evolution algorithm based
system utilizing differential evolution with an application in vis-NIR on sort mutation operation for vehicle routing problem with time
soil spectroscopy,’’ in Proc. IEEE Int. Conf. Fuzzy Syst. (FUZZ-IEEE), windows,’’ IOP Conf. Ser., Mater. Sci. Eng., vol. 790, no. 1, Mar. 2020,
Jul. 2017, pp. 1–7. Art. no. 012083.
[173] A. K. Bhandari, A. Kumar, S. Chaudhary, and G. K. Singh, ‘‘A new beta [194] I. M. Ali, D. Essam, and K. Kasmarik, ‘‘A novel design of differential
differential evolution algorithm for edge preserved colored satellite image evolution for solving discrete traveling salesman problems,’’ Swarm Evol.
enhancement,’’ Multidimensional Syst. Signal Process., vol. 28, no. 2, Comput., vol. 52, Feb. 2020, Art. no. 100607.
pp. 495–527, Apr. 2017. [195] A. Qing, ‘‘Benchmark electromagnetic inverse scattering by using differ-
[174] T. Miyao, K. Funatsu, and J. Bajorath, ‘‘Exploring differential evolution ential evolution—A big data perspective,’’ Appl. Comput. Electromagn.
for inverse QSAR analysis,’’ FResearch, vol. 6, p. 1285, Jul. 2017. Soc. J. (ACES), vol. 35, no. 4, pp. 375–382, 2020.
[175] R. Arunkumar and V. Jothiprakash, ‘‘Optimal crop plans for a multi- [196] G. J. Krishna and V. Ravi, ‘‘Mining top high utility association rules using
reservoir system having intra-basin water transfer using multi-objective binary differential evolution,’’ Eng. Appl. Artif. Intell., vol. 96, Nov. 2020,
evolutionary algorithms coupled with chaos,’’ Comput. Electron. Art. no. 103935.
Agricult., vol. 140, pp. 34–47, Aug. 2017. [197] S. Atta and G. Sen, ‘‘Multiple allocation p-hub location problem for
[176] T. Vivekanandan and N. C. S. N. Iyengar, ‘‘Optimal feature selection content placement in VoD services: A differential evolution based
using a modified differential evolution algorithm and its effectiveness for approach,’’ Int. J. Speech Technol., vol. 50, no. 5, pp. 1573–1589,
prediction of heart disease,’’ Comput. Biol. Med., vol. 90, pp. 125–136, May 2020.
Nov. 2017. [198] J. Rajalakshmi and S. Durairaj, ‘‘Application of multi-objective opti-
[177] Y. Zhao, M. Li, X. Lu, L. Tian, Z. Yu, K. Huang, Y. Wang, and mization algorithm for siting and sizing of distributed generations in
T. Li, ‘‘Optimal layout design of obstacles for panic evacuation using distribution networks,’’ J. Combinat. Optim., vol. 41, no. 2, pp. 267–289,
differential evolution,’’ Phys. A, Stat. Mech. Appl., vol. 465, pp. 175–194, Feb. 2021.
Jan. 2017. [199] K. Chantarasamai and O.-U. Lasunon, ‘‘Modified differential evolution
[178] M. F. Fateh, A. Zameer, N. M. Mirza, S. M. Mirza, and M. A. Z. Raja, algorithm for U-shaped assembly line balancing type 2,’’ Int. J. Intell.
‘‘Biologically inspired computing framework for solving two-point Eng. Syst., vol. 14, no. 4, pp. 452–462, Aug. 2021.
boundary value problems using differential evolution,’’ Neural Comput. [200] A. Rey and C. Delrieux, ‘‘Multithresholding techniques in SAR image
Appl., vol. 28, no. 8, pp. 2165–2179, Aug. 2017. classification,’’ Remote Sens. Appl., Soc. Environ., vol. 23, Aug. 2021,
[179] M. Ohenoja, M. Ruusunen, and K. Leiviskä, ‘‘Hierarchical control of an Art. no. 100540.
integrated fuel processing and fuel cell system,’’ Materials, vol. 12, no. 1, [201] H. Wang, B. Sheng, Q. Lu, X. Yin, F. Zhao, X. Lu, R. Luo, and
p. 21, Dec. 2018. G. Fu, ‘‘A novel multi-objective optimization algorithm for the integrated
[180] A. Bhuvanesh, S. T. J. Christa, S. Kannan, and M. K. Pandiyan, scheduling of flexible job shops considering preventive maintenance
‘‘Multistage multiobjective electricity generation expansion planning for activities and transportation processes,’’ Soft Comput., vol. 25, no. 4,
Tamil Nadu considering least cost and minimal GHG emission,’’ Int. pp. 2863–2889, Feb. 2021.
Trans. Electr. Energy Syst., vol. 29, no. 2, Feb. 2019, Art. no. e2708. [202] G. J. Krishna and V. Ravi, ‘‘High utility itemset mining using binary
[181] T.-J. Zhang, H.-X. Shen, Z. Li, H. Qie, J. Cao, H.-N. Li, and Y.-K. Yang, differential evolution: An application to customer segmentation,’’ Expert
‘‘Restricted constellation design for regional navigation augmentation,’’ Syst. Appl., vol. 181, Nov. 2021, Art. no. 115122.
Acta Astronautica, vol. 150, pp. 231–239, Sep. 2018. [203] H. Peng, Y. Han, C. Deng, J. Wang, and Z. Wu, ‘‘Multi-strategy co-
[182] Y. Zhang, L. Peng, G. Dai, and M. Wang, ‘‘Enhanced hybrid differential evolutionary differential evolution for mixed-variable optimization,’’
evolution for Earth-moon low-energy transfer trajectory optimization,’’ Knowl.-Based Syst., vol. 229, Oct. 2021, Art. no. 107366.
Int. J. Aerosp. Eng., vol. 2018, pp. 1–17, Jan. 2018. [204] Y. Han, H. Peng, A. Jiang, C. Wang, F. Kong, and M. Li, ‘‘A differential
[183] P. Sriboonchandr, N. Kriengkorakot, and P. Kriengkorakot, ‘‘Improved evolution algorithm for multi-objective mixed-variable optimization
differential evolution algorithm for flexible job shop scheduling prob- problems,’’ in Proc. Int. Conf. Bio-Inspired Comput., Theories Appl.
lems,’’ Math. Comput. Appl., vol. 24, no. 3, p. 80, Sep. 2019. Cham, Switzerland: Springer, 2021, pp. 145–159.

131706 VOLUME 11, 2023

J. Li, Z. Meng: DE With Exponential Crossover: An Experimental Analysis

[205] C. E. D. S. Santos, R. C. Sampaio, L. D. S. Coelho, G. A. Bestard, ZHENYU MENG (Member, IEEE) received the
and C. H. Llanos, ‘‘Multi-objective adaptive differential evolution for B.S. degree in computer science from Shandong
SVM/SVR hyperparameters selection,’’ Pattern Recognit., vol. 110, Normal University, in 2008, the M.Phil. degree
Feb. 2021, Art. no. 107649. in computer science from the Harbin Institute of
[206] G. Cheng, L. Si, P. Tang, Y. Zhuang, H. Sun, and J. Ding, ‘‘Topology Technology, in 2011, and the Ph.D. degree in
optimization of the azimuth-rotation-independent polarization conversion computer science from the Shenzhen Graduate
metasurface for bandwidth enhancement,’’ Opt. Exp., vol. 30, no. 23, School, Harbin Institute of Technology, Shenzhen,
pp. 41340–41349, 2022.
in 2018.
[207] H. Wang, W. Jiang, Q. Hu, J. Zhang, and Y. Jia, ‘‘Differential evolution
After completing the M.Phil. degree, he was a
algorithm-aided time-varying carrier frequency offset estimation for
Research Assistant with the Guangzhou Institute
OFDM underwater acoustic communication,’’ J. Mar. Sci. Eng., vol. 10,
no. 12, p. 1826, Nov. 2022. of Advanced Technology, Chinese Academy of Sciences, from 2012 to 2013,
[208] Q. Luo, W. Peng, G. Wu, and Y. Xiao, ‘‘Orbital maneuver optimization before pursuing the Ph.D. degree. He is currently the Vice Dean of
of Earth observation satellites using an adaptive differential evolution computer science and mathematics, the Director of the Institute of
algorithm,’’ Remote Sens., vol. 14, no. 9, p. 1966, Apr. 2022. Artificial Intelligence, and a Full Professor with the Fujian Key Provincial
[209] D. Buakum and W. Wisittipanich, ‘‘Self-learning differential evolution Key Laboratory of Data Mining and Application, Fujian University of
algorithm for scheduling of internal tasks in cross-docking,’’ Soft Technology. His research interests include evolutionary computation, swarm
Comput., vol. 26, no. 21, pp. 11809–11826, Nov. 2022. intelligence, and vehicle navigation.
[210] D. Buakum and W. Wisittipanich, ‘‘Selective strategy differential
evolution for stochastic internal task scheduling problem in cross-docking
terminals,’’ Comput. Intell. Neurosci., vol. 2022, pp. 1–17, Nov. 2022.

JUNCAN LI received the bachelor’s degree in

computer science and technology from Minnan
Normal University, in 2021. He is currently
pursuing the master’s degree with the Fujian
University of Technology. His research interests
include evolutionary computation and intelligent
algorithms.

VOLUME 11, 2023 131707