Literature DB >> 35035007

Differential evolution-assisted salp swarm algorithm with chaotic structure for real-world problems.

Guoxi Liang1, Huiling Chen2, Zhifang Pan3, Hongliang Zhang2, Tong Liu4, Xiaojia Ye5, Ali Asghar Heidari2.   

Abstract

There is a new nature-inspired algorithm called salp swarm algorithm (SSA), due to its simple framework, it has been widely used in many fields. But when handling some complicated optimization problems, especially the multimodal and high-dimensional optimization problems, SSA will probably have difficulties in convergence performance or dropping into the local optimum. To mitigate these problems, this paper presents a chaotic SSA with differential evolution (CDESSA). In the proposed framework, chaotic initialization and differential evolution are introduced to enrich the convergence speed and accuracy of SSA. Chaotic initialization is utilized to produce a better initial population aim at locating a better global optimal. At the same time, differential evolution is used to build up the search capability of each agent and improve the sense of balance of global search and intensification of SSA. These mechanisms collaborate to boost SSA in accelerating convergence activity. Finally, a series of experiments are carried out to test the performance of CDESSA. Firstly, IEEE CEC2014 competition fuctions are adopted to evaluate the ability of CDESSA in working out the real-parameter optimization problems. The proposed CDESSA is adopted to deal with feature selection (FS) problems, then five constrained engineering optimization problems are also adopted to evaluate the property of CDESSA in dealing with real engineering scenarios. Experimental results reveal that the proposed CDESSA method performs significantly better than the original SSA and other compared methods.
© The Author(s), under exclusive licence to Springer-Verlag London Ltd., part of Springer Nature 2022.

Entities:  

Keywords:  Chaotic initialization; Engineering optimization problems; Feature selection; Global optimization; Salp swarm algorithm

Year:  2022        PMID： 35035007      PMCID： PMC8743356          DOI： 10.1007/s00366-021-01545-x

Source DB:  PubMed          Journal:  Eng Comput        ISSN： 0177-0667            Impact factor:   8.083


Introduction

It is generally known that population-based optimization algorithms can find high-quality solutions based on random operations [1–5, 112–116]. Being a new population optimizer, the salp swarm algorithm (SSA) [6] was created by Mirjalili in 2017. In literature [6], it can be found that SSA is superior to well-known algorithms: harmony search (HS) [7], particle swarm optimization (PSO) [8], gravitational search algorithm (GSA) [9], and flower pollination algorithm (FPA) [10] on CEC-BBOB-2015 test functions, and the multi-objective SSA (MSSA) outperforms the classical multi-objective algorithms: multiobjective particle swarm optimization (MOPSO) [11] and multiobjective evolutionary algorithm based on decomposition (MOEA/D) [12] on IEEE CEC2009 test suits, meanwhile, SSA and MSSA have been used to deal with seven engineering optimization problems. As a swarm algorithm, it has a high potential to approximate the optimal solutions with a satisfying convergence rate and coverage of solution space. In addition, it is much easier to comprehend and implement due to its simple mathematical model and framework. So it has been employed frequently to realize many real-world requests, for instance, feature selection (FS), energy, and environmental protection [13-15]. Lately, Faris et al. have summarized the work on SSA and they adopted it to train extreme learning machine (ELM) significantly [16]. One of the most crucial applications of SSA is FS. Faris et al.[17] built a binary SSA integrating a crossover scheme to deal with FS; 22 well-known UCI datasets tested its performance. Sayed et al. [18] presented a hybrid algorithm by embedding chaos theory into SSA. Mohammad et al. [19] integrated singer chaotic map and a new local search algorithm (LSA) into SSA for FS. Although all the experiment results clearly show that SSA shows promising performance in dealing with FS, the improved SSA variants are aimed at specific datasets. We see they perform badly on some datasets, so the algorithms need to be enhanced further and expand the field of application. Another critical application of SSA is parameter estimation and result forecast. El-Fergany [20] presented an SSA-based method to evaluate parameters of polymer exchange membrane fuel cells (PEMFCs) to guarantee cognitive modeling, simulation, and control. Zhang et al. [21] applied SSA in parameter estimation for the soil water retention curve. Hussien et al. [22] introduced a K-nearest neighbor (KNN) to SSA for predicting chemical compound activities. Zhao et al. [23] established a novel model named SSA-LSSVM for CO2 prediction; SSA was used to optimize two parameters in the least-squares support vector machine (LSSVM). Although SSA variants get better results in dealing with parameter identification and result forecast problems, due to the model’s performance depends on the given datasets, the precision needs to be further enhanced to be hired in the real application fields. The SSA was also employed to tackle and recognize solutions to other engineering optimization. Ateya et al. [24] used chaotic maps to enhance SSA and applied it to optimize controllers’ numbers and connections between switches and controllers in large scale SDN networks. Yang et al. [25] proposed a memetic SSA (MSSA) for the PV systems in the condition of partial shading. Wang et al. [26] introduced bound estimation into multi-objective SSA to deal with short-term load forecasting problems. Abbassi et al. [27] utilized an SSA-based strategy to obtain the optimal configurations of photovoltaic cell models. Abadi et al. [28] combined SSA and genetic algorithm (GA) for nurses’ scheduling and designation in COVID-19 pandemic. Salma et al. [29] adopted both a random mutation operator and an adaptive process to boost the search capability of SSA for solving optimal power flow (OPF) problems. Ewees et al. [30] combined firefly algorithm (FA) with SSA to deal with unrelated parallel machine scheduling problems (UPMSP). Rohit et al. [31] brought the self-adaptive parameters strategy in SSA to boost its convergence performance. Although SSA gets better results in solving real engineering optimization problems, we see the accuracy and efficiency requires to be augmented for satisfying actual needs. It is observed that the SSA performs better in solving a lot of optimization problems. However, it needs to be improved in terms of convergence precision and rate in dealing with complex optimization problems. A range of scholars applied different local and global search enhanced mechanisms to boost its exploration and exploitation drifts. Ibrahim et al. [32] combined the characteristic of SSA with the feature of PSO, which is named SSAPSO, to improve the global and local search capability of basic SSA. In this framework, if the random value is bigger than 0.5, the SSA’s mechanisms are adopted. Otherwise, PSO’s mechanisms are applied; they reported that the SSAPSO is better than the simple SSA in dealing with 15 benchmark functions. Furthermore, the SSAPSO is employed to deal with FS problems on the UCI datasets. Rizk-Allah et al. [33] presented five versions of binary SSA with different transformations, and they all have been tested on 18 benchmark functions. The results show that the binary SSA embedding a modified Arctan transformation called BBSA-V has competitive performance. Divya et al. [34] embedded random opposition-based learning, multiple leadership, and simulated annealing into SSA to enhance its performance in solving complex multi-modal problems. Malik et al. [35] employed a novel lifetime scheme to accelerate the convergence performance of SSA. Bhaskar et al. [36] employed Gaussian, Cauchy, and levy-flight mutation schemes to enhance the performance of SSA. Fatima et al. [37] introduced the local best information and differential evolution to improve SSA named MSSA. Nibedan et al. [38] adopted both normal distributed mutation operator and oppositional learning concept to enhance the performance of SSA. Ren et al. [39] used the random replacement strategy and double adaptive weight to improve the performance of SSA. Saafan et al. [40] integrated improved whale optimization algorithm (IWOA) into SSA to boost its performance. Although the above mechanisms have enhanced SSA comparatively, these strategies are usually for the given problems, at the same time the convergence precision and rate of the process requires to be still enhanced. As mentioned above, though many scholars have done a lot of valuable efforts to improve the property of SSA, we observe its exploitation and exploration propensities need to be further improved. As we experienced, the chaotic map has a good random performance, so it is usually utilized in improving the algorithm’s parameters, intensification, and the initialization of the population. Zhang et al. [14] proposed a new SSA variant named CMSSA, three mechanisms: Gaussian mutation, Cauchy mutation, and chaotic exploitative mechanism (CEM) were combing with SSA to boost its performance. The results verify that CMSSA is better than PSO with an aging leader and challengers (ALCPSO) [41], comprehensive learning particle swarm optimizer (CLPSO) [42], self-adaptive differential evolution (SaDE) [43], jDE [44], hybridizing grey wolf optimization (HGWO) [45], and improved grasshopper optimization algorithm (IGOA) [46] on 23 well-known benchmark functions. Emre et al. [47] used a chaotic map to update the control parameter, mutualistic relationship to enhance search performance, and a random technique to enhance the diversity. Liu et al. [48] integrated the chaotic exploitative trends and multi-population structure to enhance the performance of SSA. DE method has a superior searching performance, and it has widely been used to uphold the performance of other algorithms. Rohit et al. [49] adopted the mutation operators to boost the performance of SSA. Zhang et al. [50] combined the improved mutation operation of DE and restart mechanism with SSA to improve its optimization performance. In this paper, a chaotic SSA with DE (CDESSA) is proposed by adopting chaotic initialization and DE-based mechanisms to help the SSA jump out of the local optimum (LO) and accelerate its convergence rapidity too. As far as we know, these strategies are firstly utilized to advance the effectiveness of SSA regarding exploratory and exploitative drifts. This paper has contributions as follow: For one aspect, chaotic initialization is utilized to provide a better initial population; it is conducive to finding better global optima (GO). In another aspect, DE is employed to boost the searching capability of each individual to increase the convergence potential and convergence accuracy of SSA. Chaotic initialization and DE can help the exploration and exploitation competencies of SSA to make a more suitable balance. We provide an efficient solver, called the CDESSA, to deal with functions of IEEE CEC2014, FS problems, constrained engineering optimization to not only solve high-dimensional and multimodal problems but also deal with complex practical application problems. This work is conducted as below: The basic theory of original SSA was elaborated in Sect. 2. The proposed CDESSA was described in detail in Sect. 3. The experiment results and discussions were explained in Sect. 4. At last, the conclusions and future research plans were given briefly in Sect. 5.

Salp swarm algorithm (SSA)

SSA is a bio-inspired optimizer firstly established by Mirjalili in 2017 [6], which is inspired by the swarming activities of salps when they are navigating and foraging for foods in oceans. Interestingly, salps often come together in a chain. This phenomenon can be illustrated in Fig. 1. Though this phenomenon cannot be interpreted by the theory, this interesting phenomenon can be ascribed to gain better movement through both rapid cooperation movement and foraging by Anderson [51].
Fig. 1

A demonstration of salp chains

A demonstration of salp chains To establish a mathematical model for the swarm, the salp swarm is divided into two groups: the leader and followers. The salp in the first place of the chain is called leader, so the rest of the salps are treated as followers. Taken literally, the leader leads the direction of the swarm, and the followers follow the leader one after another. Like other swarm optimizers, the location of salps is defined as a D-dimensional vector where D is the dimension of the given problem. Hence, the position of all salps forms a two-dimension matrix called X. It is also supposed that the food source called F in the search space is the foraging target of the swarm. The following equation updates the position of the leader:where shows the position of leader in the jth dimension, F is the position of the food source in the jth dimension, ub denotes the jth dimension of the upper bound, lb indicates the jth dimension of lower bound, c1, c2, and c3 are control parameters. The rule in Eq. (1) shows that the position of the leader is updated only based on the food source. The parameters c2 and c3 are random real numbers that are uniformly generated in the region of [0,1]. The most critical parameter c1 used to keep the poise between exploration and exploitation drifts in SSA is calculated as follows:where l means current iteration, meanwhile, L is the maximum iteration. The location of followers is renewed ed by:where means the location of the ith follower salp in the jth dimension. The flowchart of SSA is demonstrated in Fig. 2. The implemented process of SSA can be depicted as the below steps:
Fig. 2

The flowchart of SSA

The flowchart of SSA Step 1: Produce the initial salp population randomly within the boundaries of the solution space and evaluate the population and gain the best salp, as the global best solution (FoodPosition), and attain its fitness (FoodFitness), l = 1. Step 2: If the current iteration l is not more than the maximum iteration L, go to Step 3; otherwise, go to Step 6. Step 3: Update the parameter c1 according to Eq. (2). Step 4: if i == 1, replace the location of the leading salp based on Eq. (1); else renew the location of the follower salp based on Eq. (3). Step 5: Amend the salps according to the upper and lower boundaries of the problem, and evaluate the salp population, if the best salp in the current population is better than the global best solution (FoodPosition), then, update the global best solution (FoodPosition) by the best salp in the current population; otherwise, the global best solution (FoodPosition) keeps unchanged, l = l + 1, and go to Step 2. Step 6: Return the best solution (FoodPosition) and its fitness (FoodFitness).

Proposed chaotic SSA with DE (CDESSA)

The original SSA has some advantages such as quick convergence speed and simple structure. However, when solving more complicated practical optimization problems, it may easily face stagnation problems. The performance of SSA mainly depends on the interactions between the leader agent and follower agents. When a single agent falls into the LO, the agent can jump out of the LO by the full effect of the leading agent. But if most agents sink into the same condition, the whole algorithm will slow down and lock down around the LO finally. This condition provides the opportunity for other strategies to be embedded in SSA, so the performance of SSA will be improved.

Chaotic initialization mechanism

Literature [52, 53] has pointed out that a well-distributed initial population is always useful. For instance, it will assist the population to converge to a better optimal solution, and it may also augment the convergence speed and quickness of the optimization process. The chaotic map has been widely utilized throughout the initialization period of the optimizers due to the following characteristics: closeness to the initial condition, semi-randomness, and ergodicity. In the initialization phase, for one aspect, the chaotic map is utilized to produce random parameters to replace the uniform or gaussian parameters [54-56]. For another aspect, the chaotic map is adopted to generate a chaotic solution [57-59], as well as the role of opposition-based learning [60]. In this paper, the logistic map is exploited to generate a chaotic solution of the agent in the initialization; the distribution of the logistic map can be denoted by Eq. (4).where is coefficient, it is often equal to 4, at this time chaotic map will be produced; is the index of current chaotic number; is population size; indicates a random number in the range that is not equal to 0.25, 0.5, and 0.75; chaotic numbers will be generated based on Eq. (4). The chaotic solution of the ith agent can be attained by Eq. (5).where is the tth chaotic numbers; denotes the ith agent; indicates the chaotic solution of the ith agent.

Differential evolution mechanism

The DE algorithm is a population-based algorithm and it has been embedded into many optimizers [61, 117]. Its optimization process mainly includes mutation, crossover, and selection, in which mutation is the key factor of DE; it determines the optimization adaptation of DE mainly. Aim to prevent SSA from dropping into premature convergence. The well-known DE is introduced into SSA to generate a new, better candidate solution for each search agent as an intensification tool. The population has search agents. Each agent is a vector with dimensions denoted as .

Mutation

In the mutation process, the mutation operator is carried out by selecting three diverse agents , , and from the current population, where , so a mutant vector can be attained. The mutation operator can be denoted as Eq. (6).where is the mutation scale factor, it is a vector of dimensions. Each dimension distributes in the range . In this study, mutation scale factor can be generated by Eq. (7).where stands for uniform distribution, it means generating a vector with dimensions. Each dimension is uniformly distributed in the range . In this paper, and .

Crossover

In this process, a trial vector can be generated by carrying out a crossover operator. The crossover operator can be modeled as Eq. (8):where is a probability factor in determining the diversity of the swarm and prevents the algorithm from dwindling into local minima. In this paper, . is a random number in the range ., is an index, which is randomly generated between . It ensures that one element of the trial vector comes from mutant vector .

Selection

Selection operator means that the better offspring is chosen between the trial vector and the search agent based on fitness. It can guarantee that the better candidate solution will be retained, so the convergence speed and accuracy of the algorithm will be enhanced. This operator can be denoted as Eq. (9):

The implementation of CDESSA

The implementation details of CDESSA are introduced in this section. In the original SSA, the search agent moves to the next location based on the leader agent or its previous agent; therefore, a promising solution cannot always be obtained in each time so that the algorithm will trap into early convergence; its convergence speed and accuracy will be prevented to be further improved. To further develop the exploitation and exploration trends of SSA, chaotic initialization mechanism and DE-based mechanism are combined with SSA to form a new enhanced SSA, named CDESSA. In CDESSA, a chaotic initialization mechanism is utilized to build a better opening population for the subsequent optimization process. Because the quality of the initial population regulates the convergence rapidity and accuracy of the population-based algorithms, a chaotic initialization mechanism guarantees that we keep an initial population as random as possible. This gives the algorithm more chance to escape from the LO and find a more promising solution. A DE-based mechanism is utilized to intensify the exploitation and exploration of the search agent as an intensification technique. This gives the search agents more opportunities to find a more satisfactory solution. Hence, the convergence speed and accuracy of the algorithm will be further enhanced. The flowchart of CDESSA is shown in Fig. 3. The implemented process of the CDESSA method can be depicted as below:
Fig. 3

Flowchart of CDESSA

Flowchart of CDESSA Step 1: Generate the location of the search agent in the initial population randomly by Eq. (10), .where means the jth dimension of the position of the ith agent; and denote the jth dimension of the lower bound and upper bound, respectively; produces a random number in the range. Step 2: Evaluate the fitness of each individual in the initial population; calculate the chaotic solution of the individual in the initial population by Eqs. (4) and (5), thus, the chaotic solutions can be obtained and form a chaotic population; choose the best solutions from both the initial population and the chaotic population to update the initial population, repeat this process times and obtain the best initial population to participate in the next iteration process; get the optimal solution (FoodPosition), . Step 3: Initialize the algorithm parameters such as , , , , in which can be generated by Eq. (11):where is the current fitness evolution number; is the maximum fitness evolution number. Step 4: If , refresh the location of the leader agent by Eq. (1); else renew the location of the follower agents by Eq. (3). Step 5: Select three different agents from the current population, then carry out a DE mechanism and generate the candidate solution of the search agent by Eqs. (6), (7), (8), update the search agent by Eq. (9), . Step 6: Amend the position of the agent within the lower and upper bound, evaluate the fitness of each agent in the current population, ; if the optimal agent is better than the global optimal solution (FoodPosition), update the global optimal solution (FoodPosition) by the optimal agent in the current population; else the global optimal solution (FoodPosition) keeps unchanged. Step 7: If the termination condition is met (), repeat Steps 4–6; otherwise, go to Step 8. Step 8: Output the global optimal solution (FoodPosition) and its fitness (FoodFitness).

The time complexity of CDESSA

The proposed algorithm is made of chaotic initialization mechanism, DE mechanism, and basic SSA, so its computational complexity depends on the chaotic initialization mechanism, DE mechanism, and original SSA. As mentioned earlier, N and MaxFEs signify group size and maximum fitness evolution number, respectively. Similarly, D means the dimension of the given problem, respectively. The computing time complexity of initialization process is O(N × D), The computing time complexity of chaotic initialization mechanism is O((N − 1) × N × D). Because the initial population and chaotic initialization consume N × N fitness evolutions. There are two fitness evolutions in each iteration, so the rest of the iterations are (MaxFEs − N × N)/2. In each iteration, the computing time complexity of updating the position by basic SSA is O(N × D), the computing time complexity of the DE mechanism is O(N × D), the computing time complexity of updating the global optimal solution is O(N). So, the computing time complexity of the proposed algorithm is O((N × D + (N − 1) × N × D) + (N × D + N × D + N) × (MaxFEs − N × N)/2), which can be simplified to O((2 × N × D + N) × ((MaxFEs − N × N)/2 + N × N × D)).

Experimental results

To estimate the performance of CDESSA, a series of experiments is conducted, firstly, the IEEE CEC2014 benchmark suite is used to test the performance of CDESSA in solving high-dimensional and multimodal problems; secondly, FS problems are adopted to evluate the capability of CDESSA in dealing with discrete optimization problems; finally, CDESSA is used to solve four constrained engineering optimization problems. All the experiments can evaluate the performance of CDESSA in handling global optimization problems comprehensively and objectively. To be fair, as per standards in machine learning works [98-101] all the experiments were performed on the server with Intel (R) Xeon(R) E5-2650 v3 (2.30 GHz) CPU and 20 GB RAM, it runs Windows Server 2008 R2 version operating system, all the codes are implemented by MATLAB R2014b.

Experiment 1: results for IEEE CEC2014 problems

In this section, CDESSA is compared with the state-of-the-art algorithms on the IEEE CEC2014 benchmark suit. IEEE CEC2014 benchmark suite can be divided into four types such as unimodal, multimodal, hybrid, and composition functions. The summary of the CEC2014 benchmark suite is demonstrated in Table 1. In this experiment, to be fair, the swarm size N was 40, the maximum evaluations (MaxFEs) were set to D × 10,000, in which D is problem dimensions, at the same time each method runs 30 times on each function independently.
Table 1

IEEE CEC2014 benchmark suit

IndexFunction nameCategoryRangeOptimum
F1Rotated high conditioned elliptic functionUnimodal[− 100, 100]100
F2Rotated bent cigar functionUnimodal[− 100, 100]200
F3Rotated discus functionUnimodal[− 100, 100]300
F4Shifted and rotated Rosenbrock’s functionMultimodal[− 100, 100]400
F5Shifted and rotated Ackley’s functionMultimodal[− 100, 100]500
F6Shifted and rotated Weierstrass functionMultimodal[− 100, 100]600
F7Shifted and rotated Griewank’s functionMultimodal[− 100, 100]700
F8Shifted Rastrigin’s functionMultimodal[− 100, 100]800
F9Shifted and rotated Rastrigin’s functionMultimodal[− 100, 100]900
F10Shifted Schwefel’s functionMultimodal[− 100, 100]1000
F11Shifted and rotated Schwefel’s functionMultimodal[− 100, 100]1100
F12Shifted and rotated Katsuura functionMultimodal[− 100, 100]1200
F13Shifted and rotated HappyCat functionMultimodal[− 100, 100]1300
F14Shifted and rotated HGBat functionMultimodal[− 100, 100]1400
F15Shifted and rotated expanded Griewank’s plus Rosenbrock’s functionMultimodal[− 100, 100]1500
F16Shifted and rotated expanded Scaffer’s F6 functionMultimodal[− 100, 100]1600
F17Hybrid function 1 (N = 3)Hybrid[− 100, 100]1700
F18Hybrid function 2 (N = 3)Hybrid[− 100, 100]1800
F19Hybrid function 3 (N = 4)Hybrid[− 100, 100]1900
F20Hybrid function 4 (N = 4)Hybrid[− 100, 100]2000
F21Hybrid function 5 (N = 5)Hybrid[− 100, 100]2100
F22Hybrid function 6 (N = 5)Hybrid[− 100, 100]2200
F23Composition function 1 (N = 5)Composition[− 100, 100]2300
F24Composition function 2 (N = 3)Composition[− 100, 100]2400
F25Composition function 3 (N = 3)Composition[− 100, 100]2500
F26Composition function 4 (N = 5)Composition[− 100, 100]2600
F27Composition function 5 (N = 5)Composition[− 100, 100]2700
F28Composition function 6 (N = 5)Composition[− 100, 100]2800
F29Composition function 7 (N = 3)Composition[− 100, 100]2900
F30Composition function 8 (N = 3)Composition[− 100, 100]3000
IEEE CEC2014 benchmark suit At the same time, seven competitive improved evolutionary algorithms are chosen to participate in this competition; the parameters of all the involved algorithms are listed in Table 2 in detail. The Wilcoxon singed-rank test [62] was utilized with a 0.05 significance level to measure whether there is a statistical significance between CDESSA and other competitors. In the results, “w’’ sign identifies that CDESSA is superior to the competitors, “t” symbol denotes a tie and so there is not a statistically significant difference, and “l” sign signifies CDESSA is inferior to the peer algorithms, and hence CDESSA exhibits the worse performance.
Table 2

Parameters settings for the involved algorithms

AlgorithmParameters
ALCPSO [41]w = 0.4; c1 = c2 = 2.0; Θ0 = 60; T = 2
CLPSO [42]wmin = 0.2; wmax = 0.9; c = 1.496
DECLS [63]Fmin = 0.2; Fmax = 0.8; CR = 0.2; m = 1500
WDE [64]No parameters
ESSA [65]r1 = randi(50); c1 = 2 × exp(− (4 × r1/MaxFEs)^2); c2 = rand(0, 1); c3 = rand(0, 1)
CMSSA [14]

\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mu = 4$$\end{document}μ=4; C1 = rand(0, 1) and C1 ≠ 0, 0.25, 0.5, 0.75, 1.00; m = 1000

c1 = 2 × exp(− (4 × FEs/MaxFEs)^2); c2 = rand(0, 1); c3 = rand(0, 1)

CSSA [18]c1 = 2 × exp(− (4 × FEs/MaxFEs)^2); c2 = C × Os × 1- Os); C = 4; O0 = 0.7; c3 = rand(0, 1)
CDESSA (ours)Fmin = 0.2; Fmax = 0.8; CR = 0.1; c1 = 2 × exp(− (4 × FEs/MaxFEs)^2); c2 = rand(0, 1); c3 = rand(0, 1)
Parameters settings for the involved algorithms ; C1 = rand(0, 1) and C1 ≠ 0, 0.25, 0.5, 0.75, 1.00; m = 1000 c1 = 2 × exp(− (4 × FEs/MaxFEs)^2); c2 = rand(0, 1); c3 = rand(0, 1)

The impact of chaotic initialization and DE

As it was mentioned in Sect. 3, chaotic initialization and DE mechanism are integrated with the original SSA. To measure the influences of each strategy and acquire the best mechanism combinations for the next experiments. Among these combinations, CSSA indicates only a chaotic initialization mechanism is embedded in the original SSA; DESSA means only the DE mechanism is introduced in basic SSA; CDESSA denotes both chaotic initialization, and DE mechanism are applied to enhance the performance of SSA. All these mechanism combinations are evluated on the CEC2014 benchmark 30D test. Then, the Friedman test is employed to measure the differences between all strategies according to the average ranking value. The results of the Friedman test are displayed on Table 3. In Table 3, it can be observed that CDESSA gets the minimum average value of 1.8233 and the best rank No. 1. All these results indicate that the performance of CDESSA outperforms CSSA and DESAA. Based on the results, CDESSA with the best performance comparing with other mechanism combinations is chosen to take part in the following experiments as the proposed algorithm.
Table 3

Average ranking values of different mechanisms combinations using the Friedman test

AlgorithmAverageRankw/l/t
SSA3.3511425/0/5
CSSA2.7211316/0/14
DESSA2.104427/0/23
CDESSA1.82331NA
Average ranking values of different mechanisms combinations using the Friedman test

Diversity analysis and balance analysis

To measure the population diversity and the balance between exploration and exploitation of CDESSA and SSA in the search process, the diversity analysis and balance analysis experiment is conducted based on the IEEE CEC2014 benchmark 30D test, the population is formed by 40 agents, when iteration reachs 1000 the search process is terminated. The diversity and balance analysis convergence plots are indicated in Figs. 4 and 5, respectively.
Fig. 4

Diversity analysis of CDESAA and SSA

Fig. 5

Balance analysis of CDESAA and SSA

Diversity analysis of CDESAA and SSA Balance analysis of CDESAA and SSA In Fig. 4, the horizontal axis indicates the iteration number; the vertical axis denotes the average distance between the individuals in the population. It indicates the individual distribution in the current population. And then, it reveals the population diversity. In Fig. 4, at the initialization phase the mean distance between agents is much bigger because the agents are distrbuted randomly in the solution space, the curve will go down as the iteration increases, it will reach stabilization finally. At the initialization stage, the average distance between the agents of the CDESSA is much bigger than basic SSA owing to the chaotic initialization mechanism performed in the initialization. The chaotic initialization mechanism makes the initial population distributes more uniformly so that the population can search for more potential regions. It indicates that the proposed CDESSA has more diversity at the initial phase. The time at which the average distance between the agents in the CDESSA population reaches the steady-state is later than the original SSA. It can be ascribed to the DE mechanism, which enhances the exploration performance of SSA. It increases the diversity of the population and avoids the population being premature convergence to a certain extent. In Fig. 5, the horizontal ordinate means the iteration number, the ordinate menas the percentage, the red curve denotes the exploration, the blue curve indicates the exploitation, an cyan incremental-decremental curve is appended in the chart, firstly, it increases from 0 to 100%, when the exploration and exploitation of the algorithm reach balance condition it reaches the biggest value 100%. After this moment, the exploitation curve rises to 100%, but the exploration curve falls to 0, the incremental-decremental curve also drops to 0, this phenomenon matches the optimization cycle of the population-based method, most of the population-based methods mainly performs the global searching stage to accelerate convergence rate in the base phase, in the later stage it carries out the exploitation step to boost the convergence accuracy. From Fig. 5, it can be found that the difference between the exploration and the exploitation in CDESSA is much higher than that in basic SSA, because that the chaotic initialization mechanism enhances the exploration at the initial stage. The moment that the exploration and exploitation of the algorithm archive balance are later in the CDESSA than that in the original SSA, this indicates that the DE mechanism boosts the exploration ability and makes it last longer, so the stable moment will be delayed. In the same way, the average exploration of CDESSA is bigger than the original SSA. This also demonstrates that the chaotic initialization mechanism and the DE mechanism promote the exploration performance of SSA.

Comparison with other state-of-the-art algorithms on 30D test

In this section, CDESSA is in comparison to other state-of-the-art methods at D = 30. The results of all the competitors for IEEE CEC2014 problems at D = 30 are shown in Appendix Table 19, which include mean and std values for each function. The overall ranking of all the methods based on the average value is listed in Table 4, the last row of Table 4 shows the total number that CDESSA wins, tie with, and loses to the corresponding competitor, respectively. The results of the Friedman test are exhibited in Table 5. The average execution time of each method on each function is demonstrated in Table 6. The convergence trend pictures of all the involved algorithms are demonstrated in Fig. 6.
Table 19

The comparison of the statistical results obtained by all the algorithms for IEEE CEC2014 benchmark set at D = 30

FunctionCDESSAALCPSOCLPSODECLSWDEESSACMSSACSSA
F1Mean5.1178E+063.8183E+068.6948E+063.5424E+071.2191E+064.4948E+081.7684E+081.4524E+09
Std2.4857E+064.4221E+062.8766E+067.2489E+063.8228E+059.1378E+077.9380E+073.1774E+08
F2Mean1.1616E+042.7288E+033.3033E+022.8621E+021.9791E+052.8886E+107.0948E+098.0997E+10
Std4.9199E+034.0431E+032.8546E+023.7650E+025.5043E+045.5721E+093.6092E+098.6653E+09
F3Mean2.4218E+035.0318E+025.1807E+028.2948E+023.0678E+025.8719E+046.5821E+048.6033E+04
Std1.7644E+034.7619E+024.1198E+025.4118E+022.0947E+006.5709E+039.2802E+034.2412E+03
F4Mean5.0427E+025.2013E+024.6732E+025.1888E+024.7405E+024.5246E+031.1231E+031.7201E+04
Std3.0615E+014.3643E+012.1343E+011.4877E+012.1154E+011.2952E+033.1329E+023.0831E+03
F5Mean5.2000E+025.2084E+025.2035E+025.2066E+025.2010E+025.2094E+025.2063E+025.2100E+02
Std5.4914E−044.1263E−023.3276E−026.4879E−021.4489E−026.6935E−021.0610E−017.5178E−02
F6Mean6.0451E+026.1605E+026.1304E+026.2352E+026.1752E+026.3816E+026.2971E+026.4361E+02
Std2.0009E+003.2295E+001.3128E+001.1726E+001.4102E+001.0527E+003.7748E+001.9596E+00
F7Mean7.0000E+027.0002E+027.0000E+027.0000E+027.0022E+029.8088E+027.4964E+021.4290E+03
Std5.3021E−032.5217E−021.0773E−044.7851E−054.2756E−024.4217E+011.9734E+019.0238E+01
F8Mean8.2965E+028.2590E+028.0000E+028.0057E+028.2039E+021.0915E+039.9710E+021.1486E+03
Std1.2340E+011.1249E+012.1111E−146.8025E−012.4843E+001.4627E+012.3029E+012.6091E+01
F9Mean9.3058E+029.9528E+029.5484E+021.0284E+039.9459E+021.2162E+031.1313E+031.2881E+03
Std8.7011E+002.8439E+019.1513E+009.8285E+001.0061E+012.0212E+012.6263E+012.5424E+01
F10Mean2.5326E+031.6146E+031.0002E+031.0592E+031.1743E+037.8095E+035.5847E+038.8347E+03
Std5.3345E+022.8425E+022.0865E−013.0006E+017.2388E+012.5732E+026.6503E+024.5561E+02
F11Mean3.3447E+034.1266E+033.2931E+036.2535E+033.3239E+038.4011E+036.1996E+039.4236E+03
Std5.9149E+026.8264E+022.3655E+022.6771E+023.7297E+023.4310E+027.0523E+024.6617E+02
F12Mean1.2000E+031.2013E+031.2003E+031.2011E+031.2003E+031.2026E+031.2014E+031.2029E+03
Std2.6773E−027.0762E−015.0381E−021.6030E−013.5121E−022.9375E−014.3493E−015.8209E−01
F13Mean1.3002E+031.3006E+031.3003E+031.3004E+031.3003E+031.3043E+031.3014E+031.3088E+03
Std4.3694E−028.8522E−024.0031E−025.1422E−023.4639E−023.0038E−018.3066E−016.3670E−01
F14Mean1.4002E+031.4005E+031.4003E+031.4003E+031.4002E+031.4956E+031.4213E+031.6598E+03
Std2.9654E−022.9438E−012.9997E−024.3346E−021.6199E−021.2457E+011.2326E+013.4827E+01
F15Mean1.5033E+031.5102E+031.5080E+031.5137E+031.5142E+033.9173E+042.3612E+032.0237E+05
Std7.1186E−014.2158E+001.2449E+001.0237E+001.8570E+001.7312E+041.0336E+035.7712E+04
F16Mean1.6117E+031.6117E+031.6101E+031.6118E+031.6107E+031.6129E+031.6121E+031.6130E+03
Std5.1713E−014.6002E−013.8732E−013.3162E−014.2076E−012.1211E−015.7342E−012.7261E−01
F17Mean3.5523E+056.1469E+059.5440E+051.9938E+066.4131E+031.2874E+071.0017E+072.0851E+08
Std2.3001E+059.8434E+053.5717E+058.2585E+051.1857E+035.1141E+065.5887E+068.7289E+07
F18Mean3.0982E+039.2892E+031.8912E+032.8229E+041.8849E+035.2829E+084.1053E+056.9709E+09
Std1.1338E+036.8859E+034.3581E+012.0136E+041.2122E+011.9103E+081.5381E+062.3652E+09
F19Mean1.9073E+031.9099E+031.9075E+031.9091E+031.9084E+032.0875E+031.9772E+032.4720E+03
Std1.2202E+002.3585E+007.3849E−016.5911E−018.8385E−014.6144E+014.3889E+011.1743E+02
F20Mean2.5654E+033.1253E+034.5585E+035.8161E+032.0615E+033.5192E+043.6802E+048.3994E+05
Std3.6691E+026.4931E+021.3313E+032.0172E+031.2106E+011.4011E+041.7025E+049.4277E+05
F21Mean1.3905E+059.8407E+041.0113E+053.6529E+053.3914E+034.0021E+061.3701E+061.0476E+08
Std1.1740E+051.0701E+055.9635E+041.4589E+051.9269E+021.4150E+069.9496E+056.7098E+07
F22Mean2.5501E+032.6415E+032.4015E+032.3905E+032.3561E+033.2611E+032.9349E+034.1500E+04
Std1.5187E+021.6510E+028.7300E+018.6361E+018.2231E+012.0342E+022.5402E+029.0925E+04
F23Mean2.5000E+032.6152E+032.6152E+032.5000E+032.6153E+032.5000E+032.5000E+032.6315E+03
Std4.8880E−103.4930E−031.9961E−051.4838E−036.5050E−036.6696E−040.0000E+002.9750E+02
F24Mean2.6000E+032.6346E+032.6242E+032.6000E+032.6305E+032.6000E+032.6000E+032.6008E+03
Std8.2131E−076.9787E+002.8352E+001.2568E−022.3898E+001.9544E−030.0000E+003.0240E−01
F25Mean2.7000E+032.7105E+032.7080E+032.7000E+032.7070E+032.7000E+032.7000E+032.7000E+03
Std1.0097E−112.9481E+009.0397E−011.8428E−058.6860E−017.8508E−060.0000E+001.3814E−02
F26Mean2.7002E+032.7396E+032.7004E+032.7003E+032.7003E+032.7033E+032.7005E+032.7891E+03
Std4.8157E−026.3586E+016.5513E−025.4189E−023.5895E−023.1199E−013.0667E−012.5344E+01
F27Mean2.9000E+033.4269E+033.1133E+032.9829E+033.1159E+032.9000E+032.9000E+034.6425E+03
Std4.0005E−062.3071E+024.7968E+001.7252E+023.9843E+009.5188E−050.0000E+003.5225E+02
F28Mean3.0000E+034.3573E+033.7004E+033.0649E+033.7869E+033.0000E+033.0000E+031.0600E+04
Std8.0742E−065.4016E+023.2056E+011.9809E+023.0236E+011.7517E−040.0000E+001.6905E+03
F29Mean3.1000E+031.5025E+063.8571E+037.1376E+034.1467E+033.8087E+033.1000E+037.1757E+07
Std1.0010E−023.9131E+061.0846E+029.9666E+031.3439E+024.7289E+020.0000E+001.6145E+08
F30Mean3.2000E+031.3056E+046.5089E+037.3542E+035.8091E+033.2530E+033.2000E+039.9014E+06
Std4.2556E−031.4091E+048.4703E+021.2939E+037.1908E+025.6418E+010.0000E+005.9535E+06
Table 4

The overall ranking of all the methods based on average fitness results on IEEE CEC2014 benchmark set at D = 30

FunctionCDESSAALCPSOCLPSODECLSWDEESSACMSSACSSA
F132451768
F243215768
F352341678
F435142768
F516352748
F613254768
F734215768
F854123768
F914253768
F1054123768
F1134162758
F1215342768
F1315342768
F1415342768
F1513245768
F1634152768
F1723451768
F1834251768
F1915243768
F2023451678
F2142351768
F2245321768
F2326547318
F2428647315
F2528746315
F2617432658
F2727546318
F2827546318
F2927465318
F3027564318
Mean2.40004.73333.10004.06673.16675.96674.76677.8000
Rank15243768
w/l/tNA21/4/518/8/423/0/117/11/230/0/023/3/430/0/0
Table 5

The average ranking of the peers based on the Friedman test on IEEE CEC2014 benchmark set at D = 30

AlgorithmAverage rankingRank
CDESSA2.36221
ALCPSO4.52445
CLPSO3.23562
DECLS4.07224
WDE3.36333
ESSA6.09007
CMSSA4.71226
CSSA7.64008
Table 6

The average time of the involved algorithms on IEEE CEC2014 benchmark set at D = 30

FunctionCDESSAALCPSOCLPSODECLSWDEESSACMSSACSSA
F12.6158E+016.0934E+001.1700E+011.6362E+012.1543E+015.3274E+001.8102E+025.1896E+00
F22.2375E+014.3384E+008.1048E+001.3437E+011.8264E+013.9453E+001.9535E+023.8491E+00
F31.5266E+013.1044E+006.0965E+009.6674E+001.3985E+012.6635E+002.1915E+022.8954E+00
F41.3598E+012.8574E+005.1829E+008.4885E+001.2169E+012.3400E+002.2011E+022.6520E+00
F51.3838E+013.2516E+005.7315E+008.3320E+001.1959E+012.5413E+002.2083E+022.9313E+00
F64.2349E+013.2175E+013.6824E+013.9893E+014.4447E+013.1578E+012.8048E+023.2221E+01
F72.6257E+015.5006E+009.4708E+001.5121E+012.1206E+015.2697E+001.7777E+024.9978E+00
F82.4877E+014.5495E+009.4355E+001.4783E+012.0580E+014.2885E+001.7900E+024.0966E+00
F92.6303E+015.4866E+001.1082E+011.5582E+012.1088E+014.8043E+001.8371E+024.6613E+00
F102.8528E+016.4959E+001.2195E+011.6164E+012.1264E+015.8870E+001.7950E+026.0851E+00
F112.7532E+016.9623E+001.3032E+011.7792E+012.3706E+016.9639E+001.8715E+026.5983E+00
F123.1903E+011.3420E+011.8772E+012.1715E+012.7090E+011.2545E+012.1665E+021.2839E+01
F131.4432E+012.9775E+005.7122E+008.6773E+001.2850E+012.4669E+002.2100E+022.7149E+00
F142.1869E+014.4502E+008.8395E+001.2944E+011.8247E+014.1678E+001.8874E+023.9905E+00
F152.7019E+015.1387E+009.9705E+001.5121E+012.1627E+015.3982E+001.7925E+025.1044E+00
F162.6777E+015.8162E+001.1726E+011.5796E+012.1190E+015.1272E+001.8321E+024.9712E+00
F172.7632E+016.2031E+001.1265E+011.6324E+012.2225E+015.9239E+001.8220E+025.8209E+00
F182.0563E+014.5141E+008.5281E+001.2830E+011.7317E+013.8979E+002.1149E+023.8558E+00
F192.0580E+019.4599E+001.2648E+011.5767E+012.0220E+018.7995E+002.3282E+029.2223E+00
F202.5844E+015.4694E+001.1372E+011.6251E+012.0170E+014.8901E+001.7055E+025.0206E+00
F212.6183E+016.0882E+001.1521E+011.6770E+012.1628E+015.6972E+001.7961E+025.4658E+00
F222.4760E+016.1683E+001.1559E+011.5975E+012.0077E+015.6296E+001.9657E+025.3347E+00
F233.3324E+011.4152E+011.8854E+012.3892E+012.8908E+011.4030E+012.0033E+021.3120E+01
F241.9648E+017.6456E+001.1050E+011.4429E+011.8700E+017.4444E+002.2525E+027.4948E+00
F251.9876E+018.8245E+001.2054E+011.5019E+011.8616E+017.6232E+002.3020E+027.9997E+00
F265.3570E+014.1491E+014.5466E+014.9744E+015.2069E+013.8414E+013.1190E+023.9700E+01
F271.7857E+001.3830E+001.5155E+001.6581E+001.7356E+001.2805E+001.0397E+011.3233E+00
F285.9522E−024.6101E−025.0518E−025.5271E−025.7855E−024.2682E−023.4655E−014.4111E−02
F291.9841E−031.5367E−031.6839E−031.8424E−031.9285E−031.4227E−031.1552E−021.4704E−03
F306.6135E−055.1224E−055.6131E−056.1412E−056.4283E−054.7425E−053.8506E−044.9012E−05
Fig. 6

The convergence graphs of F6, F9, F12, F13, F14, F15, F19, F26 on CEC2014 benchmark set at D = 30

The overall ranking of all the methods based on average fitness results on IEEE CEC2014 benchmark set at D = 30 The average ranking of the peers based on the Friedman test on IEEE CEC2014 benchmark set at D = 30 The average time of the involved algorithms on IEEE CEC2014 benchmark set at D = 30 The convergence graphs of F6, F9, F12, F13, F14, F15, F19, F26 on CEC2014 benchmark set at D = 30 Based on the results in Appendix Table 19 and the rangking table Table 4, it can be found that CDESSA outperforms ESSA and CSSA on all 30 functions; CDESSA wins ALCPSO on 21 functions, it loses to ALCPSO on 4 functions (F1, F2, F3, F10), it ties with ALCPSO on other 5 functions; CDESSA is superior to CLPSO on 18 functions, it is inferior to CLPSO on 8 functions (F2, F3, F4, F8, F10, F16, F18, F22); CDESSA outperforms DECLS on 23 functions, it is worse than DECLS on 6 functions (F2, F3, F7, F8, F10, F22), there is no statistically significant difference between CDESSA and DECLS on other one function; Compared with WDE, CDESSA gains more promising results on 17 functions, it loses to WDE on 11 functions (F1, F3, F4, F8, F10, F16, F17, F18, F20, F21, F22), it shows a tie on other 2 functions; Compared to CMSSA, CDESSA shows more competitive performance on 23 functions, it is behind CMSSA on 3 functions (F27, F29, F30), on other 4 functions CDESSA and CMSSA go to deuce. Based on the above analysis, it comes to a conclusion that on most of the Unimodal functions CDESSA does not demonstrate more competitive performance than ALCPSO, CLPSO, DECLS, and WDE, at the same time CDESSA does not acquire more promising results on Multimodal functions than WDE, but our proposed algorithm CDESSA shows more competitive performance on Hybrid, and Composition functions, these functions are more difficult to solve than Unimodal and Multimodal functions. In a word, our CDESSA gains the lowest rank value 2.4000 and ranks first, which means CDESSA can beat all the competitors on the IEEE CEC2014 benchmark suit. The average ranking of all the peers obtained by the Friedman test is shown in Table 5, CDESSA gets the lowest average ranking 2.3622 and No. 1 rank, but the average ranking of No.2 rank CLPSO is 3.2356 who is more than CDESSA, which means CDESSA shows more disadvantage than other competitors. The average execution time of all the compared algorithms on each function of IEEE CEC2014 problems at D = 30 is demonstrated in Table 6. From Table 6, it can be found that the time consumption of CMSSA is much bigger than other algorithms due to the chaotic exploitative mechanism (CEM), Gaussian and Cauchy mutation consume more computing time, CSSA and ESSA consume less computing time than other methods owing to it only modifies the parameters and does not introduce extra time consumption, the time consumption of the proposed CDESSA is similar to WDE, it costs less computing time than CMSSA. However, it spends more computing time than CSSA, ESSA, ALCPSO, CLPSO, DECLS. These results indicate that the time complexity of CDESSA is higher than CSSA, ESSA, ALCPSO, CLPSO, DECLS, it is lower than CMSSA; it is equal to WDE. These results verify that the performance of CDESSA is improved significantly in comparison with the original SSA and other competitors. It also states clearly that the improvements between the CDESSA and other competitors are outstanding. The presented CDESSA performs better due to the following reasons. During the initialization, the chaotic initialization mechanism can obtain a better initial population through the less simulated stochastic process. It ensures that the initial population distributes as uniformly and randomly as possible, literature [52, 53] states clearly that a well-distributed initial population will promote the population to find a better optimal solution faster; during each iteration, the DE mechanism will enhance the exploitation and exploration of each salp as an intensification tool by mutation, crossover, and selection operators, more potential areas will be traversed. All the embedded mechanisms strengthen the exploitation and exploration capacity of the original SSA, so the proposed CDESSA has better convergence accuracy and faster convergence speed, as revealed in Fig. 6 the presented algorithm CDESSA can find a more promising solution quickly.

Scalability test for CDESSA

The scalability test for CDESSA was organized in this section. The dimensionality D is rose to 50 and 100 to estimate the capability of CDESSA in dealing with the benchmark problems with higher dimensions. These state-of-the-art algorithms were utilized as competitors as in Sect. 4.1.3. The comparison results of all the competitors on the IEEE CEC2014 benchmark set at D = 50 and 100 are exhibited in Appendix Tables 20 and 21, respectively. The overall ranking of all the methods based on the average value on the IEEE CEC2014 benchmark set at D = 50 and 100 are illustrated in Tables 7 and 8, respectively. From the last row in Tables 7 and 8, it can be seen that the victory function total number of the presented algorithm CDESSA is bigger than the lost function number and the tied function number compared with the competitors. It indicates that the proposed CDESSA outperforms all the peers. From the average Friedman ranking in Table 9, the presented CDESSA achieves the best average ranking value, which is 2.1433 and 1.9211 for D = 50 and 100, respectively. All the comparison results demonstrate that CDESSA holds distinct advantages in solving higher dimensions to optimize the problem. The CDESSA is far advance than the second one in dealing with higher dimensions problems.
Table 20

The comparison of the statistical results obtained by all the algorithms for IEEE CEC2014 benchmark set at D = 50

FunctionCDESSAALCPSOCLPSODECLSWDEESSACMSSACSSA
F1Mean4.9717E+061.3860E+071.6105E+071.7817E+085.1869E+061.2768E+092.8028E+088.0607E+09
Std1.1071E+061.2575E+074.8790E+062.5782E+071.2849E+062.6815E+081.1331E+082.0410E+09
F2Mean2.4341E+031.1335E+042.3598E+028.8604E+034.5509E+067.9764E+101.7121E+101.7177E+11
Std1.7794E+039.8814E+033.1860E+017.5909E+031.0651E+067.3787E+094.1677E+091.4287E+10
F3Mean2.6810E+033.9672E+031.9200E+034.7623E+037.1866E+021.1283E+051.4386E+052.0410E+05
Std1.3086E+033.2273E+039.2032E+021.6114E+037.7945E+019.6557E+032.2301E+045.5604E+04
F4Mean5.1857E+025.5072E+025.0210E+024.9769E+025.2376E+021.6895E+042.5573E+034.8780E+04
Std3.2124E+015.1645E+016.9195E+001.2985E+001.3693E+014.2758E+039.6743E+026.2749E+03
F5Mean5.2000E+025.2099E+025.2044E+025.2094E+025.2018E+025.2112E+025.2069E+025.2116E+02
Std8.7611E−056.8421E−022.9670E−026.6778E−022.0161E−024.6617E−021.2686E−014.3715E−02
F6Mean6.0654E+026.3646E+026.3019E+026.5423E+026.3760E+026.6740E+026.5698E+026.7686E+02
Std2.1157E+003.7348E+001.7134E+001.6261E+001.7642E+001.9326E+004.6951E+002.7664E+00
F7Mean7.0000E+027.0002E+027.0000E+027.0000E+027.0072E+021.5189E+038.7646E+022.2393E+03
Std2.2570E−032.9020E−021.3482E−038.7157E−057.3273E−028.9320E+014.7746E+011.2078E+02
F8Mean8.5432E+028.6053E+028.0000E+028.8905E+028.4584E+021.3779E+031.1792E+031.4521E+03
Std1.0967E+011.6686E+013.6566E−148.6441E+003.5131E+001.8239E+013.2398E+012.9142E+01
F9Mean9.7625E+021.1167E+031.0293E+031.2139E+031.1217E+031.5389E+031.3524E+031.6603E+03
Std1.8384E+015.2566E+011.4174E+011.5608E+012.3528E+012.4766E+015.7128E+014.4152E+01
F10Mean4.5061E+032.8314E+031.0005E+032.5130E+031.7142E+031.4090E+049.7217E+031.5638E+04
Std7.3326E+024.9080E+023.6802E−013.0465E+021.1210E+024.2865E+027.3561E+026.6454E+02
F11Mean5.4622E+037.0212E+036.0840E+031.2356E+046.0156E+031.4763E+041.0780E+041.6078E+04
Std8.2249E+028.2361E+023.4059E+024.2241E+022.9035E+023.8872E+021.2156E+035.6881E+02
F12Mean1.2000E+031.2015E+031.2003E+031.2019E+031.2003E+031.2035E+031.2017E+031.2039E+03
Std1.6879E−027.4723E−013.8324E−021.3953E−013.1793E−022.7726E−015.4827E−016.3396E−01
F13Mean1.3003E+031.3006E+031.3004E+031.3004E+031.3004E+031.3059E+031.3021E+031.3090E+03
Std7.0850E−021.0539E−014.6607E−025.7507E−024.0805E−023.3150E−011.0097E+004.5530E−01
F14Mean1.4003E+031.4007E+031.4003E+031.4003E+031.4002E+031.6117E+031.4390E+031.7458E+03
Std3.4372E−023.3983E−012.4591E−021.4305E−011.9778E−021.9335E+011.2122E+012.2414E+01
F15Mean1.5064E+031.5261E+031.5184E+031.5306E+031.5380E+038.5442E+052.2421E+048.1612E+06
Std1.1733E+008.1862E+002.4146E+001.9423E+004.0054E+003.0121E+052.0084E+043.5799E+06
F16Mean1.6207E+031.6212E+031.6186E+031.6217E+031.6196E+031.6225E+031.6220E+031.6229E+03
Std7.6452E−016.4079E−015.3736E−012.3124E−012.7325E−011.9779E−015.4200E−012.8123E−01
F17Mean4.9877E+051.5315E+062.9183E+061.1501E+073.7307E+047.9017E+072.3543E+071.0507E+09
Std3.3574E+051.1971E+069.9324E+052.9081E+061.0219E+043.3931E+071.5184E+072.9493E+08
F18Mean2.5937E+034.3262E+031.9498E+031.4977E+042.0278E+032.7039E+092.9459E+052.2565E+10
Std3.2450E+021.7094E+033.3471E+011.0067E+042.1925E+016.7210E+089.5604E+054.0705E+09
F19Mean1.9255E+031.9408E+031.9178E+031.9432E+031.9192E+032.4003E+032.0586E+035.6575E+03
Std1.7111E+012.2768E+012.5989E+004.1544E+002.0956E+009.2196E+014.3932E+011.2839E+03
F20Mean2.7843E+035.3820E+037.6201E+031.2000E+042.2434E+034.8444E+045.7962E+045.5550E+05
Std7.1792E+022.5939E+031.9478E+033.9044E+034.2596E+011.3875E+041.8990E+044.0634E+05
F21Mean3.0989E+057.4692E+051.6829E+064.4611E+067.2525E+031.3752E+075.2585E+062.0742E+08
Std1.3879E+051.4451E+066.7850E+051.6576E+068.6239E+024.3410E+063.6101E+069.4565E+07
F22Mean2.9087E+033.3087E+032.8861E+032.8979E+032.7960E+035.0267E+033.7516E+038.5416E+05
Std3.0349E+023.4025E+021.3306E+021.7525E+021.2731E+022.7959E+023.3540E+027.4423E+05
F23Mean2.5000E+032.6440E+032.6440E+032.5000E+032.6441E+032.5000E+032.5000E+032.5015E+03
Std2.2452E−082.9114E−028.6295E−056.6937E−041.2726E−023.8664E−040.0000E+005.7444E−01
F24Mean2.6000E+032.6786E+032.6603E+032.6000E+032.6769E+032.6000E+032.6000E+032.6004E+03
Std5.1366E−134.2888E+003.7369E+004.5493E−032.0714E+002.0774E−030.0000E+001.3536E−01
F25Mean2.7000E+032.7232E+032.7156E+032.7000E+032.7165E+032.7000E+032.7000E+032.7000E+03
Std5.4726E−137.1652E+001.4428E+001.6050E−051.4787E+008.9987E−060.0000E+009.4600E−03
F26Mean2.7368E+032.7805E+032.7039E+032.7005E+032.7004E+032.7050E+032.7040E+032.8000E+03
Std4.8878E+017.4227E+011.8368E+016.9352E−022.8754E−024.1763E−011.8142E+011.0257E−04
F27Mean3.0010E+034.1326E+033.5633E+033.3253E+033.5511E+032.9000E+032.9000E+036.2548E+03
Std2.1319E+021.3759E+023.0854E+026.1412E+024.2828E+027.0698E−051.3876E−127.1969E+02
F28Mean3.0000E+035.9252E+034.2260E+033.0345E+034.5327E+033.0000E+033.0000E+031.8710E+04
Std3.1447E−079.3591E+027.1001E+011.8876E+021.0657E+028.6107E−041.3876E−121.8366E+03
F29Mean3.1000E+031.4373E+074.5595E+036.2146E+038.5455E+034.0677E+033.1000E+034.6542E+07
Std2.6081E−033.5718E+073.0829E+022.4175E+031.5276E+031.1930E+030.0000E+002.4654E+08
F30Mean3.2000E+032.9902E+041.3395E+041.0882E+041.5651E+043.2403E+033.2000E+036.5958E+07
Std9.6308E−051.0889E+041.0396E+035.1662E+031.0830E+033.2663E+010.0000E+001.9283E+07
Table 21

The comparison of the statistical results obtained by all the algorithms for IEEE CEC2014 benchmark set at D = 100

FunctionCDESSAALCPSOCLPSODECLSWDEESSACMSSACSSA
F1Mean3.0373E+071.6890E+081.0391E+081.5800E+094.5335E+073.8793E+094.9043E+081.2539E+10
Std6.2481E+068.2566E+071.8017E+072.7095E+087.9694E+065.7769E+081.3112E+082.1241E+09
F2Mean1.0115E+043.4788E+045.9123E+022.6788E+046.5006E+072.2886E+114.3842E+103.1731E+11
Std1.0061E+043.3344E+045.5123E+022.8121E+049.9092E+061.2638E+109.1261E+091.4322E+10
F3Mean3.8941E+034.1767E+031.9645E+035.0625E+032.7773E+032.6526E+052.4072E+053.3420E+05
Std1.4962E+036.2394E+035.9028E+021.7422E+034.3237E+021.0278E+042.4386E+041.5722E+04
F4Mean6.7495E+028.5885E+026.2133E+025.9610E+027.6936E+025.0318E+044.8556E+031.1088E+05
Std3.6040E+011.2037E+022.8402E+012.8656E+012.8265E+017.0766E+031.1494E+031.2470E+04
F5Mean5.2000E+025.2119E+025.2067E+025.2126E+025.2033E+025.2132E+025.2073E+025.2135E+02
Std9.6536E−053.9274E−024.0364E−022.7287E−022.3212E−022.5498E−021.1313E−012.9799E−02
F6Mean6.2646E+026.9565E+026.7992E+027.3881E+026.9428E+027.4993E+027.3219E+027.6331E+02
Std4.4928E+006.1348E+003.8667E+002.2096E+002.0372E+003.1247E+009.2966E+002.5280E+00
F7Mean7.0000E+027.0004E+027.0000E+027.0000E+027.0118E+022.9078E+031.0831E+033.9215E+03
Std3.0756E−031.5114E−011.0194E−041.5296E−032.3155E−021.3311E+025.6687E+011.4485E+02
F8Mean9.4185E+021.0147E+038.0000E+021.1964E+039.3585E+022.1008E+031.6015E+032.2069E+03
Std2.1854E+014.0285E+011.4162E−138.7542E+009.1889E+001.5963E+013.3608E+014.3962E+01
F9Mean1.1138E+031.5713E+031.3290E+031.7844E+031.6009E+032.3387E+031.8800E+032.3926E+03
Std3.0262E+011.5456E+023.3715E+013.0506E+014.6289E+013.2350E+014.4718E+014.2799E+01
F10Mean9.9232E+036.6889E+031.0015E+031.1868E+043.6278E+033.1069E+042.0832E+043.2403E+04
Std1.3765E+039.8903E+028.2675E−014.9386E+022.3908E+023.8020E+027.7268E+027.8972E+02
F11Mean1.0804E+041.5441E+041.3692E+042.9800E+041.4038E+043.0720E+041.9658E+043.3206E+04
Std1.1231E+031.1278E+038.0169E+028.4141E+026.7302E+025.0284E+021.1201E+038.3069E+02
F12Mean1.2000E+031.2022E+031.2005E+031.2032E+031.2005E+031.2042E+031.2023E+031.2045E+03
Std1.1272E−023.5883E−014.5458E−022.0633E−015.1288E−021.6855E−016.8393E−014.2190E−01
F13Mean1.3004E+031.3007E+031.3005E+031.3006E+031.3004E+031.3080E+031.3030E+031.3097E+03
Std3.6105E−028.9022E−023.5586E−027.1605E−023.6641E−022.4736E−017.2045E−012.6337E−01
F14Mean1.4003E+031.4006E+031.4003E+031.4004E+031.4002E+032.0542E+031.5150E+032.3653E+03
Std1.9093E−022.8319E−011.7061E−021.8109E−011.5161E−024.2125E+012.0347E+013.8085E+01
F15Mean1.5162E+031.6103E+031.5539E+031.5813E+031.6432E+037.1870E+067.7084E+043.1953E+07
Std2.3681E+002.4699E+013.3440E+002.4362E+001.0666E+011.4393E+064.4472E+048.1341E+06
F16Mean1.6431E+031.6452E+031.6408E+031.6465E+031.6424E+031.6466E+031.6451E+031.6473E+03
Std9.6941E−013.6811E−015.7743E−012.1849E−015.1940E−012.7699E−018.9190E−013.3166E−01
F17Mean1.7872E+061.5241E+071.5337E+071.0572E+086.1941E+054.5688E+086.9203E+072.2487E+09
Std6.1682E+051.0119E+073.2867E+062.3680E+071.9965E+059.7137E+072.9614E+074.2971E+08
F18Mean3.0950E+031.6532E+052.1420E+037.5992E+032.9321E+031.5597E+101.2178E+064.4608E+10
Std3.7091E+026.7432E+057.2154E+011.4856E+041.2325E+022.3392E+096.2212E+065.2934E+09
F19Mean1.9752E+032.0297E+031.9860E+032.0107E+031.9914E+034.4807E+032.3930E+031.3461E+04
Std2.3966E+012.1378E+011.4933E+011.5642E+001.7086E+014.0873E+028.9316E+011.7666E+03
F20Mean3.3312E+031.7176E+041.2120E+043.4767E+043.8046E+032.7952E+051.5113E+051.8854E+06
Std3.1286E+024.0600E+033.6516E+037.2559E+035.0775E+025.9664E+043.0685E+041.1003E+06
F21Mean1.6299E+065.6538E+067.0572E+064.1190E+076.7877E+041.5503E+081.7784E+075.2156E+08
Std6.9097E+053.7075E+062.0203E+068.1843E+061.9718E+043.9951E+078.6440E+061.0203E+08
F22Mean3.5471E+034.9940E+034.0170E+035.1262E+033.8985E+031.1893E+045.3235E+034.0948E+05
Std3.6910E+024.2864E+022.7537E+022.2846E+022.8120E+021.7024E+035.7711E+022.8636E+05
F23Mean2.5000E+032.6523E+032.6482E+032.5000E+032.6488E+032.5000E+032.5000E+032.5009E+03
Std9.7777E−112.7861E+009.7546E−031.4844E−047.4011E−023.2680E−040.0000E+002.5340E−01
F24Mean2.6000E+032.7912E+032.7610E+032.6000E+032.8005E+032.6000E+032.6000E+032.6004E+03
Std2.4628E−075.4624E+002.0714E+004.0190E−033.9007E+002.7675E−030.0000E+001.3023E−01
F25Mean2.7000E+032.7908E+032.7572E+032.7000E+032.7668E+032.7000E+032.7000E+032.7000E+03
Std1.4700E−091.6993E+014.4215E+007.6659E−063.9150E+005.9650E−060.0000E+007.3391E−03
F26Mean2.8000E+032.8135E+032.7959E+032.7373E+032.7398E+032.7579E+032.7967E+032.8000E+03
Std2.5333E−136.2806E+013.2300E+014.8494E+014.9461E+014.5940E+011.8157E+014.8990E−05
F27Mean2.9000E+035.6893E+034.9494E+033.4481E+035.1651E+032.9000E+032.9000E+031.0792E+04
Std2.7241E−051.9037E+023.1812E+021.2474E+036.8814E+023.8811E−052.3126E−121.1237E+03
F28Mean3.0000E+039.7895E+036.7116E+033.0000E+038.0358E+033.0000E+033.0000E+034.1848E+04
Std2.1634E−081.4031E+037.5170E+021.7129E−015.6013E+024.4692E−042.3126E−123.3839E+03
F29Mean3.1000E+038.1801E+076.4705E+036.6327E+031.3394E+043.9823E+033.1000E+031.5244E+06
Std6.2494E−041.4183E+088.1234E+026.6160E+021.6613E+038.0442E+020.0000E+001.0125E+06
F30Mean3.2000E+032.8557E+053.9946E+042.1366E+041.8428E+043.3531E+033.2000E+034.8488E+08
Std1.1902E−042.0854E+059.8120E+032.6105E+041.4322E+031.6098E+020.0000E+002.0860E+08
Table 7

The overall ranking of all the methods based on average fitness results on IEEE CEC2014 benchmark set at D = 50

FunctionCDESSAALCPSOCLPSODECLSWDEESSACMSSACSSA
F113452768
F224135768
F334251678
F435214768
F516352748
F613254768
F724315768
F834152768
F913254768
F1054132768
F1114362758
F1214362758
F1315342768
F1425341768
F1513245768
F1634152768
F1723451768
F1834152768
F1934152768
F2023451678
F2123451768
F2245231768
F2327648315
F2418647315
F2518647315
F2667321548
F2737645218
F2827546318
F2927456318
F3027546318
Mean2.20004.83333.10004.20003.30005.90004.70007.7000
Rank16243758
w/l/tNA26/1/320/7/325/3/216/9/528/0/222/3/530/0/0
Table 8

The overall ranking of all the methods based on average fitness results on IEEE CEC2014 benchmark set at D = 100

FunctionCDESSAALCPSOCLPSODECLSWDEESSACMSSACSSA
F114362758
F224135768
F334152768
F435214768
F515362748
F614263758
F734125768
F834152768
F913254768
F1043152768
F1114263758
F1214362758
F1315342768
F1425341768
F1514235768
F1635162748
F1723461758
F1835142768
F1915243768
F2014352768
F2123461758
F2214352768
F2328637415
F2427648315
F2528647315
F2668412357
F2727546318
F2827546318
F2928456317
F3027654318
Mean2.03335.03333.00004.43333.43335.96674.46677.6333
Rank16243758
w/l/tNA27/1/220/9/127/3/021/7/229/1/022/4/430/0/0
Table 9

The average ranking of the involved algorithms achieved by the Friedman test at D = 50 and 100

AlgorithmD = 50D = 100
Average rankingRankAverage rankingRank
CDESSA2.143311.92111
ALCPSO4.648954.78896
CLPSO3.172223.19562
DECLS4.237844.44334
WDE3.432233.59893
ESSA6.051176.01007
CMSSA4.687864.46335
CSSA7.626787.57898
The overall ranking of all the methods based on average fitness results on IEEE CEC2014 benchmark set at D = 50 The overall ranking of all the methods based on average fitness results on IEEE CEC2014 benchmark set at D = 100 The average ranking of the involved algorithms achieved by the Friedman test at D = 50 and 100 Comparing the average Friedman ranking value for D = 50 and 100 with D = 30, it can be found that the average ranking value of the improved CDESSA drops from 2.3622 to 2.1433 and 1.9211 with the increasing dimensionality. It indicates that the enhanced CDESSA holds more advantages in higher dimensionality problems contrary to the competitors; it also means that the introduced mechanisms play a greater role in the high-dimensional problems.

Experiment 2: results for FS problems

In essence, FS is a kind of discrete optimization problem. This problem aims to choose as few features as possible and get rid of irrespective features to acquire the best classification precision. In this section, the presented CDESSA is adopted to deal with 12 diverse datasets from the UCI machine learning repository [66] comparing with four advanced FS algorithms such as BMFO, BSSA, BWOA, BFOA, SSAPSO [32]. The parameters of all the competitors are set as their original versions in the literature [6, 67–69], respectively. The details of the data sets are described in Table 10, including feature numbers and sample numbers of each dataset.
Table 10

The details of the selected datasets from UCI

DatasetFeature numbersSample numbers
Breastcancer10699
BreastEW31569
Exactly141000
primary-tumor18399
heart14270
M-of-n141000
SpectEW23267
CongressEW17435
CTG3222126
Cleveland_heart14303
Brain_Tumor210,36850
Tumors_1415,010308
The details of the selected datasets from UCI As we all know, in the beginning, CDESSA is designed for continuous real optimization problems, so it cannot be used to solve FS problems directly. It must be carried out binary transformation as follows. In the initial stage, every dimension of each agent is assigned 0 or 1 randomly to acquire the binary value. This operation can be denoted as below.where indicates the jth dimension of the value of ith agent. During the optimization process, each dimension of the continuous solution of each agent will be transformed into a binary value by the following transfer functions called the S-shaped transfer function,where indicates each dimension of the position of each agent, means the output binary value of the S-shaped transfer function. These functions realize the aim of transforming continuous real value to binary value. During the process, each algorithm executes N (N means the population size of each method) times on each dataset and carries out K-fold crossover at each time. During each crossover, the data of each dataset will be divided into three sections: training set, validation set, and test set. First of all, the K nearest neighbor classifier (KNN) is applied to train and classify the data of the training set; then the training model will be verified on the validation set; finally, the selected features are performed on the test set with the objective of gaining the accuracy value. In this work, whether each feature will be selected depends on the binary value of the corresponding dimension of the solution. If the value is 0, it indicates this feature is chosen; if the value is 1, it means this feature is abandoned. Each FS scheme that corresponds to the position of the agent is estimated based on the fitness acquired by the method. The fitness can be calculated by Eq. (15) as follows:where means the accuracy acquired by the algorithm on the validation set. indicates the total feature number in the corresponding dataset. is the feature number which is selected by the algorithm. The parameters and are weight coefficients of the accuracy and feature number, respectively. In this experiment, the maximum iteration is 50. The population size is 10. Each algorithm runs ten times randomly, Wilcoxon singed-rank test [62] was also utilized with a 0.05 significance level to measure whether there is a statistical significance between CDESSA and other competitors. The comparison results of fitness obtained by competitors are shown in Appendix Table 22. From Appendix Table 22, it can be found that CDESSA acquires the best results on Breastcancer, primary-tumor, heart, M-of-n, SpectEW; at the same time, CDESSA obtains second best results on BreastEW, Exactly, Cleveland_heart, Tumors_14. The boxplot of fitness obtained by the competitors on datasets is demonstrated in Fig. 7 also supports this conclusion, and in Fig. 7, it can be seen that the fitness fluctuation range of CDESSA is smaller than the competitors on most datasets. The ranking values of peers illustrated in Table 11 indicate that our proposed algorithm CDESSA gets the best ranking value 2.5667 and first rank No.1. The results of the fitness state clearly that the presented CDESSA acquires the best FS set on the datasets. The convergence pictures of the compared algorithms on fitness are shown in Fig. 8. The convergence graphs indicate our proposed CDESSA holds a faster convergence speed against the competitors.
Table 22

The statistics results of all the methods on the fitness value

DatasetBMFOBSSABWOABFOASSAPSOCDESSA
MeanStdMeanStdMeanStdMeanStdMeanStdMeanStd
Breastcancer0.03090.00100.03220.00170.03190.00200.03160.00110.03550.00260.02950.0023
BreastEW0.03950.00260.04180.00130.03900.00220.03840.00210.05150.00260.03870.0020
Exactly0.03360.00660.03420.00500.03190.00390.04140.00560.04540.00680.03240.0046
Primary-tumor0.55750.00790.55630.00810.55770.00650.55820.00570.54880.00820.54490.0041
Heart0.09830.00570.09810.00540.10040.00780.10000.00700.11460.00790.09220.0062
M-of-n0.03060.00140.03160.00240.02850.00170.03300.00320.03860.00200.02910.0013
SpectEW0.10430.00760.10500.00520.09910.00590.10640.00680.12120.00410.09750.0062
CongressEW0.03010.00320.03150.00220.03080.00280.03280.00310.04170.00360.03170.0033
CTG30.07660.00150.07660.00260.07600.00280.07740.00160.08490.00150.07680.0017
Cleveland_heart0.09870.00750.10440.00710.10580.00810.10860.00370.11790.00690.10440.0064
Brain_Tumor20.11340.01480.11830.02550.09980.01740.12310.02190.15540.02490.12150.0181
Tumors_140.30960.00740.31140.00940.31480.01200.31600.01060.33650.00980.31080.0056
Fig. 7

The boxplot of the fitness among competitors on the datasets

Table 11

The ranking value of the compared algorithms achieved by the Friedman test on fitness value

DatasetBMFOBSSABWOABFOASSAPSOCDESSA
Breastcancer2.33.63.53.45.62.6
BreastEW34.42.42.762.5
Exactly3.22.52.14.855.452.9
Primary-tumor4.13.84.24.72.41.8
Heart3.33.13.835.62.2
M-of-n3.53.31.654.45.92.25
SpectEW3.13.32.853.6562.1
CongressEW2.73.12.853.562.85
CTG32.92.82.63.862.9
Cleveland_heart2.62.72.94.15.63.1
Brain_Tumor22.932.93.85.23.2
Tumors_1433.12.93.75.92.4
Mean3.053.2252.88753.85.47082.5667
Rank342561
Fig. 8

The convergence graphs of the competitors on the datasets

The boxplot of the fitness among competitors on the datasets The ranking value of the compared algorithms achieved by the Friedman test on fitness value The convergence graphs of the competitors on the datasets The comparison results of errors acquired by peers are demonstrated in Appendix Table 23. It can be found that CDESSA obtains the lowest error value on Exactly, primary-tumor, heart, SpectEW, CongressEW, Tumors_14; meanwhile, CDESSA obtains second best results on Breastcancer, M-of-n. The boxplot of error value obtained by the peers on datasets is illustrated in Fig. 9 also affirms this conclusion; at the same time, it can be found that the error volatility of CDESSA is lower than the compared algorithms on most of the datasets. The ranking values of all the compared algorithms shown in Table 12 make clear that our presented algorithm CDESSA obtains the best ranking value 2.5708 and first rank No.1. The results of the error value state clearly that the FS scheme found by CDESSA holds the highest accuracy rate on the classifier at the datasets.
Table 23

The statistics results of all the methods on the error value

DatasetBMFOBSSABWOABFOASSAPSOCDESSA
MeanStdMeanStdMeanStdMeanStdMeanStdMeanStd
Breastcancer0.01140.00140.01150.00190.01150.00240.01280.00210.00930.00250.01070.0018
BreastEW0.01670.00330.01930.00180.01580.00200.01580.00290.02200.00250.01590.0029
Exactly0.00500.00600.00550.00440.00300.00340.01200.00520.01500.00660.00450.0047
Primary-tumor0.55540.00890.55200.00860.55470.00770.55430.00700.53910.00910.53690.0047
Heart0.07410.00580.07960.00630.07780.00870.07780.00780.08890.01020.06850.0072
M-of-n0.00200.00140.00200.00200.00000.00140.00300.00270.00700.00190.00100.0008
SpectEW0.08620.00810.08630.00600.08230.00690.08770.00710.09740.00410.08020.0062
CongressEW0.01270.00350.01150.00200.01160.00230.01370.00260.01610.00340.01150.0031
CTG30.05740.00130.05740.00230.05760.00210.05710.00170.06000.00170.05760.0021
Cleveland_heart0.07790.00700.08400.00800.08220.00830.08920.00470.09210.00820.08420.0080
Brain_Tumor20.09250.01570.10080.02760.08080.01820.10270.02320.13420.02680.10130.0193
Tumors_140.29750.00780.30060.01020.30340.01290.30440.01120.31830.01070.29750.0061
Fig. 9

The boxplot of the error among competitors on the datasets

Table 12

The ranking value obtained by the pees based on the Friedman test on error value

DatasetBMFOBSSABWOABFOASSAPSOCDESSA
Breastcancer3.64.43.83.952.652.6
BreastEW3.454.22.552.855.62.35
Exactly3.252.552.054.755.353.05
Primary-tumor4.63.844.82.41.4
Heart3.253.353.83.35.12.2
M-of-n3.73.22.14.255.72.05
SpectEW3.33.22.953.855.72
CongressEW3.453.453.23.354.82.75
CTG333.52.92.95.33.4
Cleveland_heart2.43.13.43.74.93.5
Brain_Tumor22.93.333.65.053.15
Tumors_143.13.534.14.92.4
Mean3.33333.46253.06253.78334.78752.5708
Rank342561
The boxplot of the error among competitors on the datasets The ranking value obtained by the pees based on the Friedman test on error value The comparison results of the feature number gained by compared algorithms are illustrated in Appendix Table 24. From Appendix Table 24, it can be concluded that CDESSA obtains the smallest feature number on Exactly, primary-tumor, heart, CongressEW, Cleveland_heart; at the same time, CDESSA obtains second best results on BreastEW, M-of-n, CTG3. The boxplot of feature number acquired by the peers on datasets is demonstrated in Fig. 10 also indicates this, meanwhile, it can be discovered that CDESSA has a lower fluctuation on the feature number than rivals on most datasets in Fig. 10. The ranking values of competitors shown in Table 13 indicate that the proposed algorithm CDESSA obtains the best ranking value 2.7042 and No.1 rank value. The results of feature numbers indicate that CDESSA acquires the shortest FS set on the datasets. All the feature number results indicate that the FS set gained by CDESSA contains the fewest uncorrelated feature information.
Table 24

The statistics results of all the methods on the feature number

DatasetBMFOBSSABWOABFOASSAPSOCDESSA
MeanStdMeanStdMeanStdMeanStdMeanStdMeanStd
Breastcancer3.60000.18893.70000.27183.70000.25303.60000.30624.65000.24703.64890.3665
BreastEW13.75000.570114.25000.771414.15000.612514.40000.704318.25000.760814.07030.9146
Exactly7.45000.27577.45000.29147.40000.29167.80000.18898.10000.17927.30000.2091
Primary-tumor10.10000.377910.75000.537910.65000.535010.60000.452212.30000.65749.92670.5356
Heart6.75000.51256.55000.43586.85000.37066.70000.18297.80000.53176.42460.5628
M-of-n7.45000.13507.60000.17927.25000.23007.65000.26338.25000.13387.40000.2714
SpectEW9.35000.83869.45000.65799.20000.69649.50000.599012.80000.56589.78970.4318
CongressEW6.10000.75436.35000.75546.10000.50126.25000.34968.60000.53346.07500.5534
CTG39.10000.47618.85000.38089.00000.57939.80000.499011.90000.57248.90000.4977
Cleveland_heart6.50000.46246.35000.46256.80000.81556.55000.38027.85000.61476.23620.4794
Brain_Tumor25305.300040.73464543.5500178.72604795.700099.15155271.350063.67755482.7500244.83785216.650071.2003
Tumors_148114.250056.94657792.2000152.96637858.0000178.59368101.950058.088410,524.3000376.63807960.450092.1092
Fig. 10

The boxplot of the feature number among competitors on the datasets

Table 13

The ranking value obtained by the pees based on the Friedman test on feature number

DatasetBMFOBSSABWOABFOASSAPSOCDESSA
Breastcancer3.053.053.252.962.75
BreastEW2.153.233.762.95
Exactly3.152.82.64.855.651.95
Primary-tumor1.953.83.853.262.2
Heart3.152.43.53.15.92.95
M-of-n33.851.7462.45
SpectEW32.952.63.163.35
CongressEW2.93.152.753.3562.85
CTG33.42.552.24.1562.7
Cleveland_heart3.152.853.73.162.2
Brain_Tumor24.7124.15.73.5
Tumors_144.41.91.94.262.6
Mean3.16672.79172.75423.64585.93752.7042
Rank423561
The boxplot of the feature number among competitors on the datasets The ranking value obtained by the pees based on the Friedman test on feature number

Experiment 3: results for constrained engineering optimization problems

In this section, our proposed CDESSA is adopted to deal with five constrained engineering optimization problems. When solving these problems, for CDESSA the population size is set to 50. The maximum iteration is 2000, CDESSA runs 100 times randomly. The statistical results of all the compared algorithms are extracted from the corresponding references due to the best results reported by researchers.

Tension/compression spring design problem

Finding solutions using a developed approach needs a logical modeling process [70-73]. The spring design problem is a widespread case in the engineering design field. The objective of the spring design problem is to devise a minimum weight () compression spring holding three variables: wire diameter (), mean coil diameter (), and the number of active coils () [74-76]. This problem holds the mathematical model as below: consider . object . subject to:variable scope . The best solution obtained by CDESSA for the tension/compression spring design problem is shown in Table 14. From Table 14, it can be found that CDESSA is superior to both original SSA and other competitors in solving tension/compression spring design problems. All the results indicate that CDESSA maintains a more competitive performance for tension/compression spring design problems, in which GA3 is the worst performer, ESSA performs the second best.
Table 14

The optimal schemes acquired by the peers for the tension/compression spring design problem

AlgorithmOptimum variablesOptimum weight
dDP
GA3 [75]0.0519890.36396510.8905220.0126810
CPSO [77]0.0517280.35764411.2445460.0126740
PSO [52]0.0157280.35764411.2445430.0126747
NM-PSO [78]0.0516200.35549811.3332720.0126300
DE [79]0.0516090.35471411.4108310.0126702
SSA [6]0.0512070.34521512.0040320.0126763
ESSA [80]0.0517190.35743411.2471230.0126653
CDESSA0.0516910.35677611.2855580.0126652
The optimal schemes acquired by the peers for the tension/compression spring design problem

Welded beam design problem

Design a welded beam with minimum fabrication cost is another kind of typical engineering design problem. This problem holds four design variables, which are reported in detail in related refs [81]. To solve this problem, the corresponding mathematical model is built as below: consider object subject to:variable scope where The best solution gained by CDESSA is shown in Table 15. From Table 17, it can be concluded that CDESSA further outperforms all the compared methods. The results got by other methods are all bigger than 1.72, but the proposed CDESSA is less than 1.7, the difference between CDESSA and other peers is obvious. It means the proposed CDESSA can find a more promising solution within the variable boundaries and restraints.
Table 15

The optimal results acquired by the peers for the welded beam design problem

AlgorithmOptimum variablesOptimum cost
hltb
GA3 [75]0.2059863.4713289.0202240.2064801.728226
CPSO [77]0.2023693.5442149.0482100.2057231.728024
CAEP [82]0.2057003.4705009.0366000.2057001.724852
NM-PSO [78]0.2058303.4683389.0366240.2057301.724717
WCA [83]0.2057283.4705229.0366200.2057291.724856
SSA [6]0.2057003.4714009.0366000.2057001.724910
ESSA [80]0.1971983.4852138.9809460.2082881.723317
CDESSA0.2057193.2532649.0368250.2057291.695277
Table 17

The optimal schemes acquired by the peers for the three-bar design problem

AlgorithmOptimum variablesOptimum weight
A1A2
Ray and Sain [91]0.7950000.395000264.3
DEDS [92]0.788675130.40824828263.8958434
PSO-DE [93]0.78867510.4082482263.8958434
MBA [90]0.78856500.4085597263.8958522
CS [94]0.788670.40902263.9716
SSA [6]0.7886654142580650.408275784444547263.8958434
CDESSA0.7886750.408248263.8958434
The optimal results acquired by the peers for the welded beam design problem

Pressure vessel design problem

This design problem aims to devise a pressure vessel using a minimum total cost under the constraints of material cost, forming cost, and welding cost. This problem contains four variables: the shell thickness (Ts), the head thickness (Th), inner radius (R), the cylindrical section length (L); in which, both Ts and Th are the integral multiples of 0.625 in [84]. The corresponding mathematical model of this problem is built as follows: consider . object . subject to:variable scope . The best solution gained by the proposed CDESSA is in Table 16. As shown in Table 16, it can be observed that the best result gained by CDESSA is much smaller than other peers, quality of the solution acquired by the competitors is more than 6000, our method CDESSA is only 5453.2428, the difference between CDESSA and other compared approaches is more than 600, so it means CDESSA is much better than other methods. All these results indicate that our method has a much more competitive performance in solving this problem.
Table 16

The best results acquired by the peers for the pressure vessel design problem

AlgorithmOptimum variablesOptimum cost
TsThRL
Branch-bound [85]1.1250000.62500058.29100043.6900007198.0428
Lagrangian multiplier [84]1.1250000.62500058.29100043.6900007198.0428
CDE [86]0.8125000.43750042.098400176.6376006059.7340
GA3 [75]0.8125000.43750042.097400176.6540006059.9463
CPSO [77]0.8125000.43750042.091300176.7465006061.0777
HPSO [87]0.8125000.43750042.098400176.6366006059.7143
G-QPSO [88]0.8125000.43750042.098400176.6372006059.7208
CDESSA0.7500000.37500041.966408178.3066735453.2428
The best results acquired by the peers for the pressure vessel design problem

Three-bar design problem

The goal of this problem is to design a truss using three bars with the minimum weight, it is a classical constraining optimization problem, it includes two variables: the cross-sectional areas of the bars (A1, A2) [89, 90]. To solve this problem its corresponding mathematical model can be constructed as follows: consider . object . subject to: variable scope . where , , . The optimal results got by CDESSA and other competitors are demonstrated in Table 17. Based on the values in Table 17 it can be observed that DEDS, PSO-DE, SSA, and CDESSA obtain the same best result 263.8958434, Ray and Sain acquire the worst result 264.3. CDESSA shows better competitiveness in comparison to other evolutionary algorithms. So it can be used to solve the three-bar design problem effectively. The optimal schemes acquired by the peers for the three-bar design problem

Multiple disk clutch brake design problem

This design problem is a kind of classical discrete optimization problem, its goal is to design a multiple disk clutch brake with both the minimum mass of the multiple disk clutch brake systems. This problem involves five variables: internal diameter (ri), external diameter (ro), the thickness of the disc (t), activating force (F), the number of frictional-force(Z). At the same time, the value range of ri and ro is [54, 73] and [83], respectively; their step size is both one. The variable t changes in the region [1], and its step size is 0.5. The value range of F is [600, 1000] and its step size is 10. The variable Z changes in the region, and its step size is one. It means the type of all the variables value is discrete [83, 95]. The corresponding mathematical model of this problem can be built as below: consider . object . subject to: where , , , , , , , , , , , . , , , , , . , , , , . , . This problem is a discrete combinatorial optimization problem, it needs to satisfy more constraints, so it is harder to solve. The best solutions acquired by CDESSA and other compared methods are shown in Table 18. It can be observed that CDESSA gets the best solution with the lowest function value 0.235242; CBA ranks second; WCA and TLBO acquire the same result 0.313656, PVS obtains the worst results 0.313660. It can be found that CDESSA is superior to other competitors and CDESSA holds more advantages.
Table 18

The optimal results acquired by the peers for the multiple disk clutch brake problem

AlgorithmOptimum variablesOptimum cost
rirotFZ
CBA [96]80.00000090.0000003.0000001000.0000002.0000000.263684
WCA[83]70.00000090.0000001.000000910.0000003.0000000.313656
PVS [95]70.00000090.0000001.000000980.0000003.0000000.313660
TLBO [97]70.00000090.0000001.000000810.0000003.0000000.313656
CDESSA70.00000090.0000001.000000970.0000002.0000000.235242
The optimal results acquired by the peers for the multiple disk clutch brake problem

Conclusions and future directions

In this study, the performance of SSA is enhanced by chaotic initialization and DE mechanism. The presented CDESSA was firstly estimated on IEEE CEC2014 benchmark problems to test the efficacy in solving high-dimensional and multimodal problems. Then, it is utilized to deal with feature selection problems and constrained engineering optimization problems. During the experiments, the proposed CDESSA is compared with the state-of-the-art algorithms. The results and evaluations verified the improved performance of CDESSA in terms of quality of results and convergence trends, and the Friedman test also shows that CDESSA is significantly superior to those competitors. The results verify that the chaotic initialization and DE mechanism enhance the exploration and exploitation capability of the original SSA effectively. The proposed mechanism in CDESSA can improve the equilibrium between the diversification and intensification cores of SSA and mitigate its convergence and stagnation shortcomings. In future works, our research plans to focus on the following themes: firstly, introduce the surrogate model to CDESSA to reduce the computation cost; secondly, CDESSA will be applied to handle other challenging optimization problems, such as multi-objective problems, and dynamic optimization problems; then, chaotic initialization and DE strategy will be adopted to enhance the performance of other representative computational intelligence algorithms for engineering optimization problems, such as monarch butterfly optimization (MBO) [102], earthworm optimization algorithm (EWA) [103], elephant herding optimization (EHO) [104], moth search (MS) algorithm [105]; finally, the advanced adaptive DE [106, 107] and adaptive distributed DE [108-110] a lgorithms will be arranged to further enhance the performance of the proposed method.
  7 in total

1.  Coevolutionary particle swarm optimization using Gaussian distribution for solving constrained optimization problems.

Authors:  Renato A Krohling; Leandro dos Santos Coelho
Journal:  IEEE Trans Syst Man Cybern B Cybern       Date:  2006-12

2.  Local Binary Pattern-Based Adaptive Differential Evolution for Multimodal Optimization Problems.

Authors:  Hong Zhao; Zhi-Hui Zhan; Ying Lin; Xiaofeng Chen; Xiao-Nan Luo; Jie Zhang; Sam Kwong; Jun Zhang
Journal:  IEEE Trans Cybern       Date:  2019-08-08       Impact factor: 11.448

3.  Resource-Aware Distributed Differential Evolution for Training Expensive Neural-Network-Based Controller in Power Electronic Circuit.

Authors:  Xiao-Fang Liu; Zhi-Hui Zhan; Jun Zhang
Journal:  IEEE Trans Neural Netw Learn Syst       Date:  2021-05-07       Impact factor: 10.451

4.  hPSD: A Hybrid PU-Learning-Based Spammer Detection Model for Product Reviews.

Authors:  Zhiang Wu; Jie Cao; Yaqiong Wang; Youquan Wang; Lu Zhang; Junjie Wu
Journal:  IEEE Trans Cybern       Date:  2018-11-02       Impact factor: 11.448

5.  Adaptive Distributed Differential Evolution.

Authors:  Zhi-Hui Zhan; Zi-Jia Wang; Hu Jin; Jun Zhang
Journal:  IEEE Trans Cybern       Date:  2019-10-21       Impact factor: 11.448

6.  Study of biological networks using graph theory.

Authors:  Wei Gao; Hualong Wu; Muhammad Kamran Siddiqui; Abdul Qudair Baig
Journal:  Saudi J Biol Sci       Date:  2017-11-14       Impact factor: 4.219

7.  HSSAGA: Designation and scheduling of nurses for taking care of COVID-19 patients using novel method of Hybrid Salp Swarm Algorithm and Genetic Algorithm.

Authors:  Moein Qaisari Hasan Abadi; Sara Rahmati; Abbas Sharifi; Mohsen Ahmadi
Journal:  Appl Soft Comput       Date:  2021-04-30       Impact factor: 6.725
  7 in total
  1 in total

1.  Multilevel threshold image segmentation for COVID-19 chest radiography: A framework using horizontal and vertical multiverse optimization.

Authors:  Hang Su; Dong Zhao; Hela Elmannai; Ali Asghar Heidari; Sami Bourouis; Zongda Wu; Zhennao Cai; Wenyong Gui; Mayun Chen
Journal:  Comput Biol Med       Date:  2022-05-18       Impact factor: 6.698
  1 in total

北京卡尤迪生物科技股份有限公司 © 2022-2023.