A Modified Cloud Particles Differential Evolution Algorithm for Real-Parameter Optimization

The issue of exploration-exploitation remains one of the most challenging tasks within the framework of evolutionary algorithms. To effectively balance the exploration and exploitation in the search space, this paper proposes a modified cloud particles differential evolution algorithm (MCPDE) for real-parameter optimization. In contrast to the original Cloud Particles Differential Evolution (CPDE) algorithm, firstly, control parameters adaptation strategies are designed according to the quality of the control parameters. Secondly, the inertia factor is introduced to effectively keep a better balance between exploration and exploitation. Accordingly, this is helpful for maintaining the diversity of the population and discouraging premature convergence. In addition, the opposition mechanism and the orthogonal crossover are used to increase the search ability during the evolutionary process. Finally, CEC2013 contest benchmark functions are selected to verify the feasibility and effectiveness of the proposed algorithm. The experimental results show that the proposed MCPDE is an effective method for global optimization problems.


Introduction
Recently, many real-world problems which belong to optimization problems are very complex and are quite difficult to solve.Traditional optimization methods are weak in some problems which are multi-modal, high dimension, discontinuous, multi-objective, and dynamic, etc. Nature-inspired meta-heuristic algorithms which can be called artificial evolution (AE) [1] are becoming more and more popular in engineering applications by building feasible solutions.These evolutionary algorithms (EAs) which are known to be capable of finding the near-optimum solution to the real-parameter optimization problems, have been successfully applied to many optimization problems, such as optimization, scheduling, economic problems, neural network training, data clustering, large-scale, constrained, forecasting and multi-objective [2][3][4][5][6][7][8][9].
The meta-heuristic algorithms can be grouped in three main categories [10]: evolution-based, physics-based, and swarm intelligence-based methods.The evolutionary algorithms which are based on evolutionary process include Genetic Algorithm (GA) [11], Genetic Programming (GP) [12], Differential Evolution (DE) [13], Derandomized Evolution Strategy with Covariance Matrix Adaptation (CMA-ES) [14], and Biogeography-Based Optimizer (BBO) [15], et al.DE is a classical global optimization algorithm which is proposed by Storn and Price.CMA-ES, proposed by Hansen and Ostermeier, adapts the complete covariance matrix of the normal mutation distribution to solve optimization problems.Some other methods which are based on physical processes include the Simulated Annealing (SA) [16,17], Brain Storm Optimization (BSO) [18], Chemical Reaction Optimization (CRO) [19], etc. SA is a heuristic algorithm which is based on an analog of thermodynamics that describes the way metals cool and anneal [20].BSO mimics the brainstorming process in which a group of people solves a problem together [21].CRO is a chemical-reaction-inspired metaheuristic algorithm which mimics the characteristics of chemical reactions in solving optimization problems [19].Moreover, there are some swarm intelligent methods based on animal-behavior phenomena such as Artificial Bee Colony (ABC) [22], Teaching-Learning-Based Optimization (TLBO) [23,24] et al.ABC, proposed by Karaboga, simulates the foraging behavior of the honeybee swarm and has been applied to solve many engineering optimization problems [25,26].The TLBO method, proposed by Rao, is based on the effect of the influence of a teacher on the output of learners in a class [23].
The Cloud Particles Differential Evolution (CPDE) algorithm [27], which is inspired by the cloud formation and state change, is a population-based algorithm.CPDE employs phase transformation mechanism to promote the superior cloud particle to lead the swarm through the evolution.The evolutionary process is divided into three stages in CPDE.They are gaseous, liquid and solid, respectively.The cloud particles explore the searching area by condensation operation in a gaseous state.In a liquid state, the liquefaction operation is carried out to realize macro-local exploitation.In a solid state, solidification operation is carried out to realize micro-local exploitation.CPDE has been shown to perform well on many optimization problems.However, it should be noted that the new cloud particles are generated by the superior cloud particles, and then CPDE may easily trap in a local optima when solving complex problems containing multiple local optimal solutions, such as CEC2013 benchmark functions.
This paper proposes a modified cloud particles differential evolution algorithm (MCPDE).Firstly, control parameters adaptation strategies are designed by tuning the movement step and crossover factor used at different evolutionary stages.Secondly, the inertia factor is introduced to effectively balance exploration and exploitation.Superior cloud particles which are assigned with a smaller movement step guide the searching direction and exploit the area where better particles may exist, while inferior cloud particles which are assigned with a larger movement step maintain population diversity.In addition, the opposition mechanism and the orthogonal crossover are used to increase the search ability during the evolutionary process.Finally, the size of population is gradually decreased during the evolution process to result in faster convergence.
The rest of the paper is organized as follows.Section 2 reviews the basic differential evolution algorithm and variants of DE.Section 3 describes the modified cloud particles differential evolution algorithm.To evaluate the performance of MCPDE, experiments are carried out on the CEC2013 contest which includes latest 28 standard benchmark functions in Section 4. For the source code used for the compared algorithms, one may refer to http://ist.csu.edu.cn/YongWang.htm.Finally, the conclusions and possible future research are drawn up in Section 5.

Basic Differential Evolution Algorithm
DE is a well-known global optimization algorithm which includes mutation, crossover and selection.During each generation, trial vectors are produced by mutation and crossover operations.Then, vectors, which will survive to the next generation, are determined by the selection operation.

Mutation
With respect to each individual x i,G (called target vector) at generation G, a new individual , which is called the mutant vector, is produced by mutation operation and arithmetic recombination.Many mutation strategies can be found in the literature [28,29], the classical one is "DE/rand/1": The indices r1, r2, r3 are three uniformly distributed random numbers within the range [1, N].Index i is different from the indices r1, r2, r3.The control parameter F, namely mutation factor, is defined by the user for scaling the difference vector.

Crossover
To increase the population diversity, crossover operation is generally employed on the target vector . Binomial (uniform) crossover and exponential crossover are generally used in DE.In the basic version of DE, binomial crossover is used and is defined as follows: In Equation ( 2), the crossover rate Cr ∈ [0, 1] is a control parameter.rand j (0, 1) is randomly selected in the range [0, 1].j rand is randomly selected in the range [1, D].Mutant vector v i,G is generated according to Equation (1).

Selection
Selection operator determines the vectors which will survive for the next generation.If the fitness of u i,G is better than or as good as x i,G , u i,G is selected.Otherwise, x i,G is selected.The selection operation is defined as follows:

Related Works
The performance of DE is directly affected by the control parameters and related evolutionary strategies.Therefore, many variants of DE are proposed for improving the performance of the algorithm.

Adapting Control Parameters of Differential Evolution
In jDE [30], the self-adaptation of control parameters is proposed.F and Cr are encoded into the individuals and updated with some probabilities so that better control parameters are used in the next generation.In SaDE [28], promising solutions are generated with self-adapted control parameter.The parameter F is generated by N(0.5, 0.3).The crossover rate Cr is generated by N(Cr m , 0.1) with Cr m initialized to 0.5.In JADE [29], "DE/current-to-pbest" with optional external archive is introduced.The external archive stores inferior solutions to provide a promising direction for the search process and improve the population diversity.Control parameters are automatically updated according to previously successful experiences.In success-history based adaptive DE (SHADE) [31], a new parameter adaptation mechanism which is based on the successful searching experience is proposed.Many variants of parameters control such as FiADE, DMPSADE and DESSA are available in the literature [32][33][34].

Generation Strategy of Differential Evolution
DE researchers have suggested that some trial vector generation strategies and operations can improve the performance of DE.CoDE [35] combines three well-studied trial vector generation strategies with three random control parameter settings to generate trial vectors.In L-SHADE [36], the Linear Population Size Reduction (LPSR) is embedded into SHADE so that the robustness of the algorithm is improved.Swagatam [37] proposed an improvement mechanism of DE by using the concept of the neighborhood of each population member.Wenyin Gong et al. [38] proposed a crossover rate repair technique for the adaptive DE algorithms.The crossover rate in DE is repaired by its corresponding binary string which is used to replace the original crossover rate.In addition, some algorithms [39][40][41][42] are based on population initialization and population tuning strategy.

Hybridized Versions of Differential Evolution
Some useful techniques or different evolutionary algorithms are combined with DE algorithm for improving the performance of DE.A hybrid of the DE algorithm (DE/EDA) [43], proposed by Sun et al., produces new promising solutions by DE/EDA offspring generation scheme.Adam [44] proposed an adaptive memetic differential evolution algorithm.The algorithm uses Nelder-Mead algorithm as a local search method.Zheng [45] combines DE with fireworks algorithm (FA) to improve the performance of DE.Ali [46] presents a hybrid optimization approach based on DE and receptor editing property of immune system.A detailed survey of the hybrid DE algorithms can be found in [4,[47][48][49][50][51].

Modified Cloud Particles Differential Evolution Algorithm
Control parameters and evolutionary strategies can significantly influence the performance of the algorithm.Based on our previous work [27], a modified cloud particles differential evolution algorithm (MCPDE) is proposed.

The Proposed MCPDE
The relation between exploration and exploitation is an important issue in the framework of EAs.The performance of the algorithm can be effectively improved by a balance between exploration and exploitation in algorithm.Research results show that the algorithm should start with exploration and then gradually change into exploitation.Based on this analysis, inertia factor and adaptive control parameters strategies in different stage are designed to keep the balance between exploration-exploitation.The opposition mechanism and the orthogonal crossover are employed to increase the search ability during the evolutionary process.Finally, the size of population is gradually decreased during the evolution process to result in faster convergence.
Like other optimization algorithms, the proposed algorithm starts with an initial population which is composed of the cloud particles.Each cloud particle represents a feasible solution of the problem.An MCPDE population is represented as a set of real parameter vectors which is defined as follows: where D is the dimensionality of the optimization problem, and N is the population size.At each generation, in order to find better solutions, superior particles exploit the searching area with a smaller step and guide the searching direction, and inferior particles explore promising areas with a relatively large radius and maintain population diversity.The evolutionary strategy, based on DE/current-to-pbest with optional archive, is generated as follows: (5) where ω 1 and ω 2 are inertia factors, i ∈ {1, . . . ,N}, r 1 , r 2 and r 3 are mutually different random integer indices selected in the range [1,N].FES and MaxFES are the number of function evaluations and the maximum number of function evaluations, respectively.In Equation (5), 0.85 and 1.9 are achieved by experiments.The value of FES/MaxFES gradually increases as the iteration progresses.Therefore, the superior particles attract the new particle to exploit better solutions with increasing ω 1 .F i is the mutation factor that controls the speed of the algorithm process.It is used by each cloud particle x i and is generated at each generation.x best is randomly chosen as one of the top p cloud particles in the current population.p is 15% of the population size.x r3 is selected from the union of the population and the archive.If the archive size exceeds 150% of the population size, some solutions are randomly removed from the archive so that some new cloud particles can be inserted into the archive.The archive is the set of archived inferior solutions in JADE [29].However, a mathematical proof has been proposed to indicate that opposite numbers may likely to be closer to the optimal solution [52].Motivated by this, some inferior solutions of the archive are randomly selected and replaced by their opposite solutions.The opposite mechanism [39] on these inferior solutions is defined as follows: where The interchange number is N D . Figure 1 shows the curves of ω 1 and ω 2 .It can be seen that ω 1 tends to increase continually and ω 2 tends to decrease as the iteration progresses.The variation of ω 1 and ω 2 ensure that the proposed algorithm smoothly transits between exploration and exploitation.At the early evolution stage, inferior particles try to search for further areas in the solution space, and a larger ω 2 is able to maintain the diversity and exploration capability.Then, as the generation increases, ω 2 tends to decrease while ω 1 tends to increase.In this way, the new particle is strongly attracted around the current superior particles and tries to exploit better solutions which may exist in their neighborhoods.Meanwhile, the convergence speed is enhanced.
Algorithms 2016, 9, 78 5 of 21 proof has been proposed to indicate that opposite numbers may likely to be closer to the optimal solution [52].Motivated by this, some inferior solutions of the archive are randomly selected and replaced by their opposite solutions.The opposite mechanism [39] on these inferior solutions is defined as follows: where ∈ [ , ] , i = 1, …, D. = ( , , ⋯ , ) is the opposite of = ( , , ⋯ , ) .The interchange number is .
Figure 1 shows the curves of ω1 and ω2.It can be seen that ω1 tends to increase continually and ω2 tends to decrease as the iteration progresses.The variation of ω1 and ω2 ensure that the proposed algorithm smoothly transits between exploration and exploitation.At the early evolution stage, inferior particles try to search for further areas in the solution space, and a larger ω2 is able to maintain the diversity and exploration capability.Then, as the generation increases, ω2 tends to decrease while ω1 tends to increase.In this way, the new particle is strongly attracted around the current superior particles and tries to exploit better solutions which may exist in their neighborhoods.Meanwhile, the convergence speed is enhanced.

Control Parameters Assignments
In classic DE, control parameters are preset and fixed during the entire iteration process.However, it is impossible to find one constant parameter setting that can fit all problems.As pointed out in [53], the different parameter settings not only play an important role in the performance of DE, but also may be used to solve specific test problems.Thus, a novel parameter adaptation scheme is presented to adjust the parameter F and Cr at different evolutionary stage.
In MCPDE algorithm, the parameter settings are divided into three stages according to the successful mutation factors F at current generation.The initial Fi and Cri used by each cloud particle are generated independently and formulated as follows, respectively:

Control Parameters Assignments
In classic DE, control parameters are preset and fixed during the entire iteration process.However, it is impossible to find one constant parameter setting that can fit all problems.As pointed out in [53], the different parameter settings not only play an important role in the performance of DE, but also may be used to solve specific test problems.Thus, a novel parameter adaptation scheme is presented to adjust the parameter F and Cr at different evolutionary stage.
In MCPDE algorithm, the parameter settings are divided into three stages according to the successful mutation factors F at current generation.The initial F i and Cr i used by each cloud particle x i are generated independently and formulated as follows, respectively: ) + cr 0 (10) where f 0 and cr 0 are initialized to be 0.5, respectively.r 1 and r 2 are random numbers in [0, 1].
In each generation, the set S F is used to store all successful mutation factors at current generation.Similarly, the set S Cr stores all the successful crossover rates at current generation.The size of S F is recorded as |S F |.If |S F | exceeds the current population size N, randomly selected elements are deleted from S F and S Cr .Then S F and S Cr are preserved for the next generation.When the set S F is empty, it indicates that F and Cr at current generation are the proper parameters for the algorithm.Then, they are preserved for the next generation.When |S F | is less than the current population size N, new control parameters where σ (S Cr ) refers to the standard deviation of S Cr .r i is randomly selected in the range [0, 1].By the end of each generation, the parameters F and Cr are updated when |S F | is less than the current population size N, as defined by where In MCPDE algorithm, different control parameters are chosen at different stages.At the early stage of evolution, the control parameter values near f 0 and cr 0 with the randomization according to Equations ( 9) and (10).Better diversity may improve the exploration ability.Then, cloud particles try to explore further areas in the solution space.At each generation, better control parameters are preserved for the next generation.The population diversity is improved and the convergence speed is accelerated with better control parameters.However, it is difficult to find better control parameters with the increasing generation.Thus, the algorithm may hard to jump out of the local optimum because of faster convergence and poorer diversity.In order to solve these problems, some new parameters F are introduced to maintain search efficiency according to Equations ( 9) and ( 12) while some new parameters Cr are produced to improve population diversity according to Equations (11) and (13).Therefore, the performance of the algorithm MCPDE is improved by choosing different control parameters strategies at different evolutionary stages.
The size of the population used by EAs plays a significant role in controlling exploration and exploitation.Large population sizes can encourage wider exploration of the search space, while small population sizes may promote exploitation of the search space.Therefore, the population size is gradually decreased as the iteration continues.By the end of each generation, the population size N is updated and is defined by where N 0 is the initial population size.FES is the current number of fitness evaluation, and MaxFES is the maximum number of fitness evaluations.If N < N , the worst individual is deleted and the archive size is resized.Because Equation ( 7) requires at least four particles, the minimum population size N is set to 4.

Orthogonal Crossover
It is well known that crossover operation is helpful for sharing the better gene segment by exchanging the gene information of the parents.However, the quality of the offspring produced by the crossover operator is highly dependent on the characteristics of target problems, so that multiple crossover operators are employed instead of a single crossover operator for solving different optimization problems [54].As pointed out in [55], OX (orthogonal crossover) operators can conduct effective search in a region proposed by the parents.Hence, we come up with the idea that uses QOX (quantization technique with orthogonal crossover) [55] operator to enhance the search ability of MCPDE.In order to save the computational cost, we apply QOX only on a better particle which is randomly selected from P best,G .The orthogonal array used in QOX operator is often denoted by L M Q K , namely K factors (i.e., variables) with Q levels (i.e., values) and M combinations.In MCPDE, let Q = 3, M = 9, and then L 9 3 4 is used.
Q levels for the cloud particle P i,G is defined as follows: where j = 1, . . ., Q. C best i,G and C i,G are the parents which define a search range [min ] for particle P i,G .C best i,G is randomly selected from P best,G .The particle P i,G is divided into K subvectors: where t 1 , t 2 , . . ., t k−1 are randomly generated integers and 1 < t 1 < t 2 < t k−1 < . . .< D. H i is treated as a factor in QOX operator, and Q levels for are H i defined as follows:

General Experimental Setting
(1) Test Problems and Dimension Setting: For a comprehensive evaluation of MCPDE, all the CEC2013 [36] benchmark functions are used to evaluate the performance of MCPDE.The CEC2013 benchmark set consists of 28 test functions.According to their shape characteristics, these benchmark functions can be broadly classified into three kinds of optimization functions [56].

•
unimodal problems f 1 -f 5 • basic multimodal problems f 6 -f 20 , and For all of the problems, the search space is [-100,100] D .In this paper, the dimension (D) of all functions is set to 10 and 30.
(2) Experimental Platform and Termination Criterion: For all experiments, 30 independent runs are carried out on the same machine with a Celoron 3.40 GHz CPU, 4 GB memory, and windows 7 operating system with Matlab R2009b, and conducted with D × 10,000 (number of function evaluations, FES).
(3) Performance Metrics: In our experimental studies, the mean value (F mean ), standard deviation (SD), maximum value (Max) and minimum value (Min) of the solution error measure [57] which is defined as f (x)−f (x * ) are recorded for evaluating the performance of each algorithm, where f (x) is the best fitness value found by an algorithm in a run, and f (x * ) is the real global optimization value of tested problem.In order to statistically compare the proposed algorithm with its peers, Wilcoxon's rank-sum test at the 5% significance level is used to evaluate whether the median fitness values of two sets of obtained results are statistically different from each other.Three marks "−", "+" and "≈" are also used to denote that the performance of MCPDE is better than, worse than, and similar to that of the compared algorithm, respectively.

Comparison with Nine State-of-the-Art Intelligent Algorithms on 10 and 30 Dimension
In this part, MCPDE is compared with PSO, PSOcf (PSO with constriction factor) [58], TLBO, DE, JADE, CoDE, jDE, CMA-ES and CPDE.The appropriate parameters are important for the performance of the intelligent optimization algorithms.Therefore, the setting of parameters of different algorithms is given in the following: For MCPDE, The population size N is set to 13 × D. The maximum size of the archive is set to 1.5 × N.For DE, the population size N is set to 100.F and CR are set to 0.5 and 0.9, respectively.For PSO, the population size N is set to 40, the linearly decreasing inertia ω from 0.9 to 0.4 is adopted over the course of the search, and the acceleration coefficients c 1 , c 2 are both set to 1.49445.For JADE, the population size N is set to 100, p = 0.05 and c = 0.1.The parameters of other algorithms are the same as those used in the corresponding references.
The statistical results, in terms of F mean , SD, Max and Min obtained in 30 independent runs by each algorithm, are reported in Tables 1 and 2.
(1) Unimodal problems f 1 -f 5 : From the statistical results of Tables 1 and 2, we can see that MCPDE is better than other compared algorithms on unimodal problems f 1 -f 5 according to the average rank (Avg-rank) for 10 dimensions and 30 dimensions.(3) Composite problems f 21 -f 28 : As is known to all, composite problems are very time consuming for fitness evaluation compared to others because these functions combine multiple test problems into a complex landscape.Therefore, it is extremely difficult for state-of-the-art intelligent optimization algorithms to obtain relatively ideal results.Concerning the composition functions f 21 -f 28 in       All in all, MCPDE performs better than the compared algorithms on the unimodal, multimodal and composition problems with D = 10 and D = 30.Overall, Table 7 2 (D = 10) and Figure 3 (D = 30).Sixteen benchmark functions are selected to compare the performance of different algorithms in Figures 2 and 3. From Figure 2, it can be seen that MCPDE performs better than other compared algorithms on 9 out of 16 test problems.Figure 3 shows that MCPDE beats other compared algorithms on 8 out of 16 test problems.The comparison experiments indicate that MCPDE is a challenging method for these functions.Moreover, MCPDE has a higher convergence rate because of good exploration ability.
The experimental results reveal that MCPDE works well for most benchmark problems.This is due to the effective parameter adaptation approach and the inertia factors which are used in MCPDE.Better control parameters are preserved to produce new control parameters for the next generation.Therefore, the probability of finding better solutions is greater and this is helpful for improving the performance of the proposed algorithm.The inertia factors are changed during the evolution process to favor, balance, and combine the exploration with exploitation.At the beginning of the search, the inertia factor ω 1 is less than ω 2 , so it favors exploration.Then, ω 1 tends to increase continually while ω 2 tends to decrease.Accordingly, it balances the search direction.Later, the inertia factor ω 1 is greater than ω 2 , so the exploitation ability of the algorithm is dynamically adjusted.In addition, the opposition mechanism and the orthogonal crossover are helpful for increasing the search ability during the evolutionary process.Therefore, both the exploration and exploitation aspects are done in parallel during the optimization process.Accordingly, MCPDE not only can improve the convergence rate of algorithm but also can decrease the risk of premature convergence as much as possible.

Conclusions
In order to improve the exploration-exploitation dilemma in the whole search space during the evolutionary process of the optimization algorithm, a new meta-heuristic optimization algorithm MCPDE for solving real-parameter optimization problems over continuous space is proposed in this paper.An effective parameter adaptation approach and the inertia factor are introduced into the modified cloud particles differential evolution algorithm.Moreover, the opposition mechanism and the orthogonal crossover are employed to increase the search ability during the evolutionary process.Then, the proposed algorithm is applied to 28 benchmark functions from the CEC2013 benchmark suite.The experimental results indicate that MCPDE performs much better than the compared algorithms for most benchmark problems.Thus, the proposed algorithm MCPDE is effective.
Future work will focus on how to design reasonable topological structures to make the algorithm more efficient and applied to constrained and multi-objective optimization problems.Moreover, it is expected that MCPDE will be used to tackle some practical engineering problems and real word applications.

Figure 1 .
Figure 1.Variation curves of inertia factors.(a) the variation of ω1; (b) the variation of ω2.
Calcute N for the next generation according to Equations (14) and (15)15: if |S F | ≥ N then 16:delete randomly selected elements from the S F and S Cr so that the parameters size are N 17: elseif (|S F | < N and S Then, M solutions are constructed on factors H 1 , H 2 , ..., H k .The pseudo-code of MCPDE is illustrated in Algorithm 1.Algorithm 1. MCPDE Algorithm1: Initialize D (number of dimensions), N (number of population), L M Q K ; Archive A = φ; 2: Initialize population randomly 3: Generate mutation factors F and Cr according to Equations (F = φ) then 18: Update F and Cr are according to Equations (11)-(13) 19: elseif S F = φ then 20: F g+1 = F g ; Cr g+1 = Cr g ; 21: endif 22: endwhile Considering f 1 -f 5 with 10 dimensions, for f 1 , MCPDE, DE, JADE, jDE, CMA-ES and CPDE work well and obtain better results.For f 2 , MCPDE performs better than other algorithms except JADE and CMA-ES.Moreover, for f 3 , MCPDE performs significantly better than the compared algorithms.For f 4 , MCPDE, CMA-ES and CPDE beat other compared algorithms.For f 5 , MCPDE, DE, JADE, jDE and CPDE are better than other algorithms.On f 1 -f 5 , MCPDE performs better than PSO, PSOcf, TLBO, DE, JADE, CoDE, jDE, CMA-ES and CPDE on 5, 5, 5, 3, 2, 5, 3, 2, 2 test problems, respectively.The overall ranking sequences for unimodal problems are MCPDE, CMA-ES, DE, CPDE, JADE, jDE, TLBO, CoDE, PSO and PSOcf in descending direction.When the search space dimension D is set to 30, according to Table 2, MCPDE is much better than PSO, PSOcf, TLBO, DE, JADE, CoDE, jDE, CMA-ES and CPDE on 5, 5, 5, 3, 3, 5, 3, 3, 4 test functions, respectively.MCPDE performs better than all compared algorithms for the unimodal problems except f 2 and f 4 .For f 2 and f 4 , MCPDE ranks secondly.The overall ranking sequences for unimodal problems are MCPDE, jDE, DE, CMA-ES, JADE, CPDE, CoDE, PSO, TLBO and PSOcf in descending direction.The reason that MCPDE has the outstanding performance may be the use of the inertia factors, which are helpful for guiding the search direction.(2) Multimodal problems f 6 -f 20 : Considering the multimodal functions f 6 -f 20 in Table 3, MCPDE is significantly better than other algorithms on f 6 , f 7 , f 9 and f 20 .Considering f 8 , most of the compared algorithms can achieve the similar results except CMA-ES.MCPDE beats most of the compared algorithms except that JADE and jDE have a similar performance on f 11 .JADE performs best on f 12 , f 13 , f 15 and f 18 -f 19 .jDE performs best on f 14 .CMA-ES performs best on f 10 and f 16 .JADE and jDE perform best on f 17 .On these 15 multimodal problems, MCPDE performs better than PSO, PSOcf, TLBO, DE, JADE, CoDE, jDE, CMA-ES and CPDE on 14, 12, 14, 14, 4, 14, 9, 13 and 11 problems respectively.The overall ranking sequences on multimodal problems are MCPDE, JADE, jDE, CPDE, PSO, TLBO, CoDE, DE, PSOcf and CMA-ES in descending direction.When the search space dimension D is set to 30, according to the experimental results on 15 test problems from Table 4, we find that MCPDE outperforms PSO, PSOcf, TLBO, DE, JADE, CoDE, jDE, CMA-ES and on 11, 10, 13, 14, 4, 14, 6, 13, 13 testproblems, respectively.The overall ranking sequences on multimodal problems are JADE, MCPDE, jDE, CPDE, PSO, CoDE, TLBO, PSOcf, DE and CMA-ES in descending direction.
shows that MCPDE has a good performance on CEC2013 test problems.When D is set to 10, the overall ranking sequences on CEC2013 test problems are MCPDE, JADE, CPDE, jDE, DE, CoDE, TLBO, PSO, PSOcf and CMA-ES in descending direction.The overall ranking sequences on CEC2013 test functions with D = 30 are MCPDE, JADE, jDE, CPDE, DE, PSO, CoDE, CMA-ES, TLBO and PSOcf in descending direction.The convergence graphs of PSO, PSOcf, TLBO, DE, JADE, CoDE, jDE, CMA-ES, CPDE and MCPDE on different benchmark functions in terms of the mean errors (in logarithmic scale) in 30 runs are plotted in Figure