International Journal of Computational Intelligence Systems

Volume 14, Issue 1, 2021, Pages 651 - 675

Enhancing Whale Optimization Algorithm with Chaotic Theory for Permutation Flow Shop Scheduling Problem

Authors
Jiang Li1, Lihong Guo1, *, Yan Li1, Chang Liu1, Lijuan Wang2, ORCID, Hui Hu3
1Changchun Institute of Optics, Fine Mechanics and Physics, Chinese Academy of Sciences, Changchun, 130033, China
2School of Computer, Jiangsu University of Science and Technology, Zhenjiang, 212003, China
3Department of Information Science and Technology, Huizhou University, Guangdong, 516007, China
*Corresponding author. Email: guolh@ciomp.ac.cn
Corresponding Author
Lihong Guo
Received 15 August 2020, Accepted 9 December 2020, Available Online 19 January 2021.
DOI
10.2991/ijcis.d.210112.002How to use a DOI?
Keywords
Whale optimization algorithm; Chaotic maps; Flow shop scheduling; Makespan; Local search; Cross selection
Abstract

The permutation flow shop scheduling problem (PFSSP) is a typical production scheduling problem and it has been proved to be a nondeterministic polynomial (NP-hard) problem when its scale is larger than 3. The whale optimization algorithm (WOA) is a new swarm intelligence algorithm which performs well for PFSSP. But the stability is still low, and the optimization results are not too good. On this basis, we optimize the parameters of WOA through chaos theory, and put forward a chaotic whale algorithm (CWA). Firstly, in this paper, the proposed CWA is combined with Nawaz–Ensco–Ham (NEH) and largest-rank-value (LRV) rule to initialize the population. Next, chaos theory is applied to WOA algorithm to improve its convergence speed and stability. On this basis, we also use cross operator and reversal-insertion operator to enhance the search ability of the algorithm. Finally, the improved local search algorithm is used to optimize the job sequence to find the minimum makespan. In several experiments, different benchmarks are used to investigate the performance of CWA. The experimental results show that CWA has better performance than other scheduling algorithms.

Copyright
© 2021 The Authors. Published by Atlantis Press B.V.
Open Access
This is an open access article distributed under the CC BY-NC 4.0 license (http://creativecommons.org/licenses/by-nc/4.0/).

1. INTRODUCTION

The flow shop scheduling problem (FSSP) [1] is a job set allocation for a set of available machines in time to satisfy a set of performance indicators. Traditional FSSP includes a set of jobs, and each job is processed on all machines in a certain order. The goal of scheduling is to arrange jobs to each machine reasonably, and to arrange the processing order and start time of jobs reasonably, so that the constrained conditions can be used and some performance indicators can be optimized. Permutation flow shop scheduling problem (PFSSP) is an important problem in the actual production process, and it is one of the most difficult NP-hard problem [2]. In essence, it is a classical combinatorial optimization problem, involving n jobs and m machines. It is generally described that n jobs which needs to go through m processes need to be processed on m machines. Each process requires different machines, and n jobs have the same processing sequence on m machines which is different from FSSP. The traditional method is difficult to solve these problems, so the intelligent algorithms are required to solve these problems. Inspired by nature, these strong metaheuristic algorithms are applied to solve NP-hard problems such as scheduling [37], image [810], feature selection [1113] and detection [1416], path planning [17,18], cyber-physical social system [19,20], texture discrimination [21], factor evaluation [22], saliency detection [23], classification [24,25], object extraction [26], gesture segmentation [27], economic load dispatch [28,29], shape design [30], big data and large-scale optimization [3133], signal processing [34], silencing efficacy prediction [35], multi-objective optimization [36,37], unit commitment [38,39], vehicle routing [40], knapsack problem [4143] and fault diagnosis [4446], and test-sheet composition [47].

Many methods are used to solve FSSP, such as ant colony optimization (ACO) [48,49], artificial bee colony (ABC) [50], genetic algorithms (GAs) [51,52], and some hybrid algorithms [5357]. Among hybrid methods, how to schedule resources reasonably and efficiently is an important factor affecting the efficiency of the model, which has been widely concerned for a long time. Up to now, the study of swarm has been very widely. Furthermore, the method of biotechnology was explored and proved to be effective. The above algorithms are optimized iteratively to simulate the evolution of nature, so that the result approaches to the optimal solution. These algorithms are mainly realized by simulating the biological characteristics of nature, e.g., whale optimization algorithm (WOA) [58] simulates predation behavior of humpback whales, and it is used to solve PFSSP [59]. There are many methods to solve PFSSP before, but they have some problems such as slow convergence and easy to fall into local optimum. WOA is a new type of algorithm to solve this problem, and it is widely used in various fields, such as wind speed forecasting system [60], image segmentation [61], economic dispatch problem [62], and neural network [63], which imitates the behavior characteristics of animal neural network and carries out distributed parallel information processing. According to the different industrial environment, the heuristic algorithm mainly optimize the following two parameters in the quality of service (QoS):

  1. Makespan—optimize the maximum makespan of the job scheduling so as to ensure the highest efficiency of the job processing.

  2. Running time—the algorithm requires lower running time.

However, the stability of WOA is poor, and the experimental results also have a lot of room for optimization, so we proposed an algorithm—chaotic whale optimization algorithm (CWA) which achieves a better stability and a smaller makespan than WOA. This algorithm uses the WOA model and improves it on this basis. Firstly, in initialization stage, WOA is combined with Nawaz–Ensco–Ham (NEH) algorithm, and largest-rank-value (LRV) rule is used to transform continuous whale population into discrete job sequence. In this way, a more suitable initial population can be generated. Secondly, CWA adopts cross selection strategy, including cross operator and reversal-insertion operator, which makes the generated search agent better. Thirdly, we improve the local search strategy so that the algorithm can escape from local optimal solutions. Finally, and most importantly, we adopt the chaotic mapping strategy, which makes the generation of the next search agent affected by the last search agent. We incorporated the above strategies into WOA algorithm, and a new variant of WOA namely CWA was proposed for solving PFSSP problem. The experimental results indicated that CWA has better performance than WOA.

The remainder of the paper is organized as follows: Some methods reported in the literatures for FSSP are provided in Section 2. The PFSSP is represented in Section 3. The standard WOA is described in Section 4 and CWA algorithm is discussed in Section 5. The experimental results of CWA are shown in Section 6. This work ends in Section 7 with the conclusions and future work.

2. RELATED WORK

In this paper, we introduced chaos theory into WOA and proposed CWA method to solve PFSSP. Therefore, the latest literatures on PFSSP, search algorithms, and chaos theory are summarized and reviewed as follows.

2.1. Permutation Flow Shop Scheduling Problem

In the field of PFSSP, bio-inspired algorithms have achieved a lot of excellent results and PFSSP was proposed by Johnson in 1954 [64]. Over the next few years, there are also many mature algorithms that have been applied. Rossit et al. [65] reviewed the related work of PFSSP past decades and provided the future work orientation of PFSSP. Han et al. [66] proposed a bi-objective model consisting of robustness and stability, which improved the efficiency of flow shop scheduling and optimized the problem. At the same time, in order to reduce the makespan, Ruiz et al. [67] improved the initialization and processing of the algorithm, and enhanced the ability of local search. Liu et al. [68] proposed a different priority rule using skewness used in NEH [69] algorithm to reduce the makespan of PFSSP. Santucci et al. [70] made a special treatment of the search operator and changed the selection strategy, which made the algorithm reach an excellent level. ABC algorithm [71] is swarm intelligence algorithm that mimic natural creatures. Li et al. [72] studied bionic principles and applied ABC to solve PFSSP to reduce the makespan. Sioud and Gagné [73] chosen a suitable neighborhood and improved the migrating bird optimization algorithm by Tabu list and restart method. Finally, it is applied to PFSSP to complete the optimization of this problem. Pagnozzi and Stützle [74] redefined the framework of random local search algorithm (RLSA), and automatically matched the search algorithm to improve the performance of RLSA. In this way, the minimum makespan can be obtained efficiently. All the above algorithms have improved the flow shop scheduling, which provided a good reference for us to reduce makespan of PFSSP.

2.2. Search Algorithms

In terms of the implementation of search algorithm, there have been many fruitful research results. Wang and Tan [75] investigated a new idea of search algorithm, which is to reuse the previous individual search information. This method improved the quality of the subsequent population and had a good application in PFSSP. Benavides and Ritt [76] observed that the optimal scheduling usually includes a few local job inversion permutation structures, so they proposed an iterative local search algorithm. Gao et al. [77] proposed a new hybrid search algorithm, which introduced the balance strategy into the improved particle swarm optimization algorithm, and then combined with tabu search to achieve efficient search for FSSPs. Wang and Wang [78] improved the search algorithm according to the characteristics of other search algorithms. It included the improved initialization scheme based on NEH and other properties based on knowledge-based search algorithm. In addition, it also used critical path to improve the performance of the search algorithm. Kwong et al. [79] optimized the algorithm by changing the search direction of the search algorithm, and combined learning strategies to optimize the performance of the algorithm. In addition, we can combine different strategies to study new search algorithms. Fu et al. [80] combined different strategies such as explosive spark and selective solution, proposed a firework algorithm, and has a good optimization for flow shop scheduling.

Through the study on the search algorithms of flow shop scheduling, we proposed a strategy to change the search agent, and introduced a probability function to make the search results escape from the local optimal solution. These methods can achieve good results for PFSSP problems.

2.3. Chaos Theory

Chaos theory is the focus of this paper, and many achievements have been achieved in this field. Wang et al. [81] proposed a chaotic krill herd (KH) algorithm by studied the behavioral specialty of krill swarm and combined chaos theory. They used 12 chaotic mapping methods to improve the algorithm. Mokhtari and Noroozi [82] studied the chaotic map attraction characteristics in particle swarm optimization (PSO) and applied it to flow shop scheduling. Alatas [83] used chaotic mapping to improve the convergence of the algorithm and enabled ABC algorithm to escape from local optimal solution. At the same time, chaos theory can also be used in hybrid algorithm. Wang et al. [84] combines chaos theory with cuckoo algorithm to improve the performance of the algorithm.

In addition, chaos theory has a very important application in cryptography and image field, through which, the useful knowledge has been investigated. Xie et al. [85] combined cryptography with chaos and proposed a new encryption scheme, which is not inferior to the previous one. Abbasinezhad-Mood and Nikooghadam [86] used chaotic mapping to solve the key exchange problem in the process of cryptographic authentication. Chaotic mapping not only improved the security of the problem, but also made the method anonymous. Xu et al. [87] mainly proposed a bit-level image encryption algorithm based on chaotic mapping, which converted one image into two, and then replaced each other by chaotic control, thus it can complete image encryption. This encryption method has good encryption. The application of chaos theory is very extensive. We have applied the above ideas to PFSSP and achieved outstanding results.

Although there are many researches on flow shop scheduling, PFSSP still does not achieve the minimum makespan. By further improving the search algorithm and introducing chaos theory, the stability of the algorithm is greatly increased and the makespan is reduced.

3. PERMUTATION FLOW SHOP SCHEDULING PROBLEM

The mathematical description of the PFSSP with minimizing makespan is as follows.

The machine M={M1,M2,M3,,Mm} processing jobs N={J1,J2,J3,,Jn}, in which each processing job contains an operator set O=Oi1,Oi2,Oi3,,Ojm, and each processing job passes through m machines in turn. The processing time of the processing job Jj on the machine Mi is given, and it is recorded as Piji=1,2,,m;j=1,2,,n.

PFSSP has the following requirements:

  1. Each job j=1,2,,n has the same machining sequence on the machine;

  2. Each machine processes each job in the same processing sequence;

  3. Each machine can only process one job at a time;

  4. Each job can only be processed by one machine at a time;

  5. The job must not be interrupted during the processing;

  6. The buffer is infinite.

The processing sequence of jobs is S=S1,S2,S3,,Sn, and n jobs will follow m machines in turn. The completion time of job Sj on machine m is represented by C(sj, m). According to the processing sequence S=S1,S2,S3,,Sn, the completion time of the PFSSP can be given as follows:

C(s1,1)=Ps1,1(1)
C(sj,1)=C(sj1,1)+Psj,1,j=2,3,,n,(2)
Csj,i=Cs1,i1+Ps1,i,i=2,3,,m,(3)
Csj,i=maxCsj1,i,Csj,i1+Psj,i,j=2,3,,n;i=2,3,,m.(4)

Therefore, the maximum completion time (makespan) can be defined as: Cmax=Csn,m. The goal of PFSSP is to find the most appropriate permutation S* to minimize the maximum completion time of the sequence, i.e., to minimize makespan.

4. WOA AND CHAOTIC MAPS

4.1. Whale Optimization Algorithm

WOA [58] is a new swarm intelligence algorithm, which simulates the process of humpback whale predation, identifies prey, and circles around it. Algorithm 1 shows the implementation process of WOA. The WOA mainly consists of the three following processes: encircling prey, bubble-net attacking, and search for prey.

Algorithm 1: The standard WOA

Initialize a population of n random search agents

Evaluate all n search agents

w* = the best search agent

t = 0

While (t < iterations)

for each search agent

  Update WOA parameters (a, A, C, l, and k)

  if (k < 0.5)

    if (|A| < l)

      xt+1=wA·D

    else if (|A| ≥ l)

      xt+1=xrandA·D

    end if

  else if (k ≥ 0.5)

    xt+1=E·ebl·cos2πl+w

  end if

end for

 Evaluate the search agent xt+1

 Update w* if xt+1 is better

t = t + 1

end while

return w*

4.1.1. Encircling prey

Humpback whales swirl around prey in the course of predation and weave foam nets, and prey is gathered by bubbles. The voice makers of humpback whales make specific sounds at the bottom to push their prey to the surface. The best search location is determined by the location of the prey or the location close to the prey, and then updated according to the best search location. The strategy can be shown below:

D=|Cxxt|(5)
xt+1=xAD(6)
A=2ara(7)
C=2r(8)
where t is number of iterations, x* is the best location of search agent, xt is the present location of search agent at iteration t. The parameter a decreases linearly from 2 to 0, and r is a random value between 0 and 1, so the range of A is −a to a.

4.1.2. Bubble-net attacking

Bubble-net attacking includes the following two behaviors: shrink ring of encirclement and continue to encircling prey. These two behaviors are modeled as the following equation:

xt+1=Eeblcos(2πl)+x(9)
E=|xxt|(10)
where b describes the shape of the logarithmic spiral, l is a random number between 0 and 1, and E is the distance between x* and xt. Humpback whales surround their prey and shrink their range, so the values of both behaviors are 1/2.
xt+1=xADk>0.5Eeblcos(2πl)+xk0.5(11)

4.1.3. Search for prey

The exploration of WOA can be achieved by random search agent xrand, which is different from exploitation using x*. This process can be given as follows:

D=|Cxrandxt|(12)
xt+1=xrandAD(13)

4.2. Chaotic Maps

In this paper, the concept of chaotic optimization algorithm (COA) [88] is introduced, chaotic variables are used. Random variables make the results of functions randomly, while the use of chaotic variables avoids randomness, which makes the results more stable [81]. The nonrepetition and ergodicity of chaos makes the method using chaotic variables have faster convergence speed than the method using random variables [88]. In this paper, a one-dimensional irreversible chaotic mapping function is investigated. Table 1 lists twelve different chaotic maps [81].

No. Name Definition
1 Chebyshev map xk+1=coskcos1(xk)
2 Circle map* xk+1=xk+0.2(1/4π)sin(2πk)mod(1)
3 Gaussian map xk+1=0xk=01/xkmod(1)otherwise,1/xkmod(1)=1xk1xk
4 Intermittency map xk+1=ε+xk+cxkn0<xkPxkP1PP<xk<1
5 Iterative map xk+1=sinmπxk,m(0,1)
6 Liebovitch map xk+1=αxk0<xkPPxkP2P1P1<xkP21β(1xk)P2<xk1
7 Logistic map xk+1=0.5xk(1xk)
8 Piecewise map xk+1=xkP0xk<PxkP0.5PPxk<0.51Pxk0.5P0.5xk<1P1xkP1Pxk<1
9 Sine map xk+1=a4sinπxk,0<a4
10 Singer map xk+1=μ(7.86xk23.31xk2+28.75xk313.302875xk4)
11 Sinusoidal map xk+1=0.5xk2sin(πxk)
12 Tent map xk+1=xk0.7xk<0.7103(1xk)xk0.7
Table 1

Chaotic maps.

5. THE PROPOSED ALGORITHM

As a new intelligent algorithm, chaos optimization algorithm (COA) [81] has been widely studied and has been applied in many fields. In this paper, COA is firstly applied to the WOA of PFSSP, and a new algorithm called CWA is then proposed. CWA maps chaotic variables linearly to the value range of optimization variables, and optimizes the search process by using the randomness, ergodicity, and regularity of chaotic variables. For random variables, chaotic mapping changes the stability of variables and optimizes the convergence speed and performance of WOA. There are four processes that will be discussed below. The first three processes of CWA and WOA are similar, but CWA has more optimization of chaotic mapping process than WOA, so CWA has better performance than WOA. The time complexity of the four processes are O(N), O(N2), O(N), O(N), therefore, the time complexity of CWA is O(N2).

5.1. Initialization

For the initialization of PFSSP, firstly, we need to solve the problem of transforming continuous search agents into discrete job sequences. To solve this problem, we use the LRV rule [89] in CWA, which can effectively transform continuous values into discrete sequences. Secondly, we use the NEH algorithm [90] to initialize population. It includes the following steps:

  1. These n jobs are arranged according to the decreasing order of machining time on the machine;

  2. Take the first two jobs in the operation sequence and select the job sequence that minimized the makespan;

  3. From the beginning of the third job, each job is inserted into each location sequence that already exists, and the sequence having smallest makespan is selected.

Afterwards, we need to calculate the maximum makespan of the generated sequence, i.e., to have a direct research of search agent performance, so as to get the best search agent. For example, we design a work sequence as shown in Table 2 and show the optimal scheduling sequence of this sequence, as shown in Table 3. Among them, the calculation of optimal scheduling sequence 1-4-3-2 is obtained by Eqs. (14). Some other sequences of this job scheduling are also shown in Table 4. The optimal job scheduling sequence is 1-4-3-2, corresponding to the completion time of this scheduling sequence, i.e., the optimal completion time, 64. Figure 1 shows the Gantt chart of the sequence 1-4-3-2.

M1 M2 M3
J1 5 6 11
J2 8 4 7
J3 11 9 3
J4 14 15 20
Table 2

Job scheduling example.

M1 M2 M3
J1 5 11 22
J4 19 34 54
J3 30 43 57
J2 38 47 64
Table 3

Evaluation of optimal scheduling sequence.

Number Job Sequence Makespan
1 2-3-4-1 79
2 4-1-3-2 70
3 3-1-2-4 73
4 4-3-2-1 70
5 1-4-3-2 64
Table 4

Evaluation of different job scheduling,

Figure 1

The Gantt chart for the job sequence 1-4-3-2.

5.2. Cross Selection

In this strategy, two operators are adopted in the experiment, namely cross operator and reversal-insertion operator.

5.2.1. Cross operator

This operator mainly exchanges the sequence of any two jobs and searches again. The main steps are as follows:

  1. Two locations are randomly selected in the search agent, which are two jobs in the job sequence;

  2. Exchange the order of these two jobs;

  3. Evaluate makespan for new sequences.

The details of cross operator can be shown in Figure 2. The number of cross operators is related to the number of jobs. In this experiment, we set the number of cross operators as one percent of the number of jobs.

Figure 2

Cross operator.

5.2.2. Reversal-insertion operator

This operator is mainly to exchange the sequence of two adjacent jobs and insert them into other locations. The main steps are as follows:

  1. Two adjacent locations are selected in the search agent, which are two adjacent jobs in the job sequence;

  2. Swap the sequences of these two jobs and remove them from the current sequence which includes k1kn jobs;

  3. Insert the exchanged jobs into the k2 possible locations in the sequence;

  4. Evaluate makespan for new sequences.

The details of cross operator can be shown in Figure 3.

Figure 3

Reversal-insertion operator.

5.3. Local Search

The improved local search strategy is used to find better results than those obtained by WOA. For each job at generation t, move the job to each position in the sequence and find the minimum makespan. After the above strategies, an optimal search agent has been generated. The operator process can be provided as follows:

  1. All search agents are orderly selected, and each selected search agent is removed from the job sequence which includes n jobs;

  2. Insert the exchanged jobs into the n1 possible locations in the sequence and the makespan of the new sequence obtained by each insertion operator is calculated, while the minimum makespan is selected;

  3. In order to avoid redundant computations, we make this strategy with small probability P.

5.4. Chaotic Mapping

According to Eqs. (1113), the variable r affects the generation of search agents at the next iteration, which changes the performance and convergence of global search. Therefore, it is of great significance for the convergence of WOA and the generation of the next population.

Although the performance of standard WOA is satisfactory, there are still some deviations from the optimal results for PFSSP. In WOA, the variable r is randomly chosen between 0 and 1, which makes the next search agent more volatile. After introducing chaotic theory, we used chaotic mapping to change these parameters, which reduced the randomness of the algorithm and improved the performance and stability of the algorithm. Therefore, WOA using chaotic mapping is called chaotic WOA.

Algorithm 2 gives the basic process of CWA, and Figure 4 gives the basic framework of CWA.

Figure 4

Framework of chaotic whale algorithm (CWA).

From Tables 5 and 6, we compare twelve different chaotic maps for WOA optimization through Reeves and Taillard benchmarks. M2 (Circle map) shows the best performance, so we get the Eq. (14), which is the best chaotic mapping of the variable r.

rt=rt1+b(a/2π)sin(2πt)mod(1)(14)
where rt is the variable r at iteration t, and in CWA, a = 0.5 and b = 0.2.

No. REC07 REC13 REC19 REC27 REC31 REC39
M1 1566.9 1964.2 2120.7 2418.4 3072.3 5170.3
M2 1566.9 1938.8 2104.3 2396.8 3070.2 5167.0
M3 1568.1 1966.1 2121.8 2416.2 3078.3 5177.2
M4 1568.7 1966.5 2118.9 2417.6 3079.8 5170.1
M5 1568.9 1950.6 2110.3 2402.0 3080.7 5167.8
M6 1568.2 1956.7 2120.3 2416.6 3077.1 5173.6
M7 1567.8 1966.9 2120.6 2417.0 3085.8 5169.9
M8 1568.7 1965.3 2120.7 2398.7 3075.2 5170.1
M9 1571.4 1961.0 2125.4 2414.7 3079.8 5168.2
M10 1566.9 1962.3 2118.4 2416.8 3071.8 5168.6
M11 1568.8 1969.7 2122.7 2411.4 3071.1 5169.0
M12 1569.2 1940.4 2106.2 2410.8 3070.9 5183.7

The best experimental results are shown in bold.

Table 5

The comparison of twelve chaotic maps on the Reeves benchmark.

No. Ta001 Ta011 Ta021 Ta031 Ta041 Ta051
M1 1281.8 1610.6 2346.4 2730.3 3054.3 3967.8
M2 1278.2 1604.0 2346.0 2720.6 3043.6 3966.7
M3 1279.9 1612.1 2355.9 2729.7 3045.8 3966.8
M4 1278.2 1604.1 2348.2 2732.4 3051.0 3974.7
M5 1278.2 1610.7 2356.6 2734.0 3045.8 3966.9
M6 1281.8 1610.7 2359.9 2730.0 3046.0 3966.8
M7 1279.9 1607.3 2354.8 2732.1 3047.9 3972.2
M8 1279.9 1608.5 2355.3 2730.3 3046.8 3971.0
M9 1281.9 1608.5 2354.8 2731.4 3043.6 3975.7
M10 1280.9 1609.0 2348.5 2731.0 3047.1 3968.9
M11 1279.9 1609.6 2349.5 2730.4 3049.1 3967.5
M12 1279.9 1610.5 2353.5 2729.4 3049.8 3971.3

No. Ta061 Ta071 Ta081 Ta091 Ta101 Ta111

M1 5493.1 5804.0 6382.4 10939.0 11417.6 26572.2
M2 5493.0 5798.4 6370.0 10914.0 11401.0 26550.1
M3 5493.0 5826.0 6374.7 10927.9 11404.3 26554.3
M4 5493.0 5809.1 6380.1 10931.2 11411.2 26557.0
M5 5493.1 5799.3 6374.6 10927.2 11401.4 26540.2
M6 5493.0 5803.0 6384.4 10933.0 11415.8 26566.3
M7 5493.0 5801.8 6370.6 10934.2 11407.6 26557.4
M8 5493.0 5800.6 6387.8 10921.1 11403.7 26551.7
M9 5493.0 5800.3 6382.0 10938.6 11406.8 26554.2
M10 5493.1 5801.2 6383.0 10922.6 11408.4 26572.7
M11 5493.0 5799.6 6385.2 10927.9 11408.6 26550.1
M12 5493.0 5803.0 6371.5 10926.2 11412.0 26576.5

The best experimental results are shown in bold.

Table 6

The comparison of twelve chaotic maps on the Taillard benchmark.

6. EXPERIMENTAL RESULTS

In the field of flow shop scheduling, there are some well-known test functions studied in this paper to investigate the performance of CWA. Carlier consists of eight instances [91] and there is Reeves datasets [92] with twenty-one instances. In addition, there is only two instances, Heller data [93]. All of them can be found at http://people.brunel.ac.uk/~mastjjb/jeb/orlib/files/flowshop1.txt. The last one can be found at http://mistic.heig-vd.ch/taillard/problemes.dir/ordonnancement.dir/ordonnancement.html by Taillard [94] which has the largest number of instances (120). The proposed algorithm was written in Java and was run in the following conditions: Intel(R) Core(TM) i3-8100 CPU 3.60GHz which possesses 8G of RAM on Windows 10. The parameters of the proposed algorithm are presented in Table 7.

Parameter Value
Population size 50
Number of iterations 3000
Small probability P 0.01
Number of runs 20
Table 7

The parameters of the proposed algorithm.

6.1. The Comparison of WOA and CWA

Table 8 and Figure 5 show the standard deviation (STD) of WOA and CWA, which is defined as follows:

STD=k=1numxkxk+12num112(15)
where num is the number of runs per problem, xkxk1 represents the difference between two adjacent experimental results.

Problem WOA CWA
REC07 9.44 0.92
REC13 11.57 4.36
REC19 11.43 2.91
REC27 13.21 4.26
REC31 15.56 3.36
REC39 13.06 6.09
Ta001 6.39 0.10
Ta011 7.26 3.32
Ta021 12.69 6.93
Ta031 13.05 1.98
Ta041 12.75 2.14
Ta051 13.48 6.92
Ta061 18.99 0.05
Ta071 18.99 2.85
Ta081 10.79 6.48
Ta091 26.94 9.99
Ta101 18.37 7.38
Ta111 33.48 26.38
Table 8

The standard deviation of WOA and CWA.

Figure 5

The standard deviation of whale optimization algorithm (WOA) and chaotic whale algorithm (CWA).

Table 8 shows the STD of the Reeves and Taillard benchmarks. The lower the numerical value, the higher the stability of the algorithm. In Figure 5, the numerical value of CWA is always lower than that of WOA, which proves that the addition of chaotic factor increases the stability of the algorithm.

6.2. The Comparison with Carlier, Reeves, and Heller Benchmarks

In this section, we compare the following algorithm with our CWA algorithm, as shown in Table 9. The comparative algorithms are as follows:

  • Hybrid genetic algorithm (HGA) [95]. HGA is a hybrid GA, which is used to solve the no wait FSSP with completion period as the goal.

  • Hybrid backtracking search algorithm (HBSA) [96]. Based on backtracking search algorithm (BSA), a HBSA is proposed for PFSP with the objective to minimize the makespan.

  • Discrete bat algorithm (DBA) [97]. DBA is constructed based on the idea of basic bat algorithm (BA), which divide whole scheduling problem into many subscheduling problems and then NEH heuristic be introduced to solve subscheduling problem.

  • Particle swarm optimization-based memetic algorithm (PSOMA) [98]. In the proposed PSOMA, both PSO-based searching operators and some special local searching operators are designed to balance the exploration and exploitation abilities. And the PSOMA utilizes several adaptive local searches to perform exploitation.

  • Particle swarm optimization based on variable neighborhood search (PSOVNS) [99]. A very efficient local search, called variable neighborhood search (VNS), was embedded in the PSO algorithm to solve the well-known benchmark suites in the literature.

Problem n * m c* HGA HBSA DBA PSOMA PSOVNS CWA
Car01 11*5 7038 BRE 0 0 0 0 0 0
ARE 0 0 0 0 0 0
WRE 0 0 0 0 0
Car02 13*4 7166 BRE 0 0 0 0 0 0
ARE 0 0 0 0 0 0
WRE 0 0 0 0 0
Car03 12*5 7312 BRE 0 0 0 0 0 0
ARE 0 0.06 0.397 0 0.42 0
WRE 1.19 1.109 0 1.189 0
Car04 14*4 8003 BRE 0 0 0 0 0 0
ARE 0 0 0 0 0 0
WRE 0 0 0 0 0
Car05 10*6 7720 BRE 0 0 0 0 0 0
ARE 0 0 0.018 0.018 0.039 0
WRE 0 0.375 0.375 0.389 0
Car06 8*9 8505 BRE 0 0 0 0 0 0
ARE 0.04 0 0 0.076 0.076 0
WRE 0 0 0.764 0.764 0
Car07 7*7 6590 BRE 0 0 0 0 0 0
ARE 0 0 0 0 0 0
WRE 0 0 0 0 0
Car08 8*8 8366 BRE 0 0 0 0 0 0
ARE 0 0 0 0 0 0
WRE 0 0 0 0 0
REC01 20*5 1247 BRE 0 0 0 0 0.16 0
ARE 0.14 0.14 0.08 0.144 0.168 0
WRE 0.16 0.16 0.16 0.321 0
REC03 20*5 1109 BRE 0 0 0 0 0 0
ARE 0.09 0.08 0.081 0.189 0.158 0
WRE 0.18 0.18 0.721 0.18 0
REC05 20*5 1242 BRE 0 0.24 0.242 0.242 0.242 0
ARE 0.29 0.24 0.242 0.249 0.249 0
WRE 0.24 0.242 0.402 0.42 0
REC07 20*10 1566 BRE 0 0 0 0 0.702 0
ARE 0.69 0.46 0.575 0.986 1.095 0.169
WRE 1.15 1.149 1.149 1.405 0.83
REC09 20*10 1537 BRE 0 0 0 0 0 0
ARE 0.64 0.07 0.638 0.621 0.651 0.046
WRE 0.65 2.407 1.691 1.366 0.325
REC11 20*10 1431 BRE 0 0 0 0 0.071 0
ARE 1.1 0 1.167 0.129 1.153 0
WRE 0 2.655 0.978 2.656 0
REC13 20*15 1930 BRE 0.36 0.1 0.415 0.259 1.036 0
ARE 1.68 0.53 1.461 0.893 1.79 0.458
WRE 1.14 3.782 1.502 2.643 0.829
REC15 20*15 1950 BRE 0.56 0.05 0.154 0.051 0.769 0
ARE 1.12 0.64 1.226 0.628 1.487 0.574
WRE 1.18 2.103 1.076 2.256 1.026
REC17 20*15 1902 BRE 0.95 0 0.368 0 0.999 0
ARE 2.32 1 1.277 1.33 2.453 0.672
WRE 2.16 2.154 2.155 3.365 1.419
REC19 30*10 2093 BRE 0.62 0.29 0.573 0.43 1.529 0.287
ARE 1.32 0.81 0.929 1.313 2.099 0.538
WRE 1.29 2.023 2.102 2.532 1.051
REC21 30*10 2017 BRE 1.44 0.69 1.438 1.437 1.487 0.642
ARE 1.57 1.5 1.671 1.596 1.671 1.472
WRE 1.83 2.231 1.636 2.033 1.636
REC23 30*10 2011 BRE 0.4 0.45 0.796 0.596 1.343 0.348
ARE 0.87 1.28 1.173 1.31 2.106 0.85
WRE 3.08 2.381 2.038 2.884 1.939
REC27 30*15 2373 BRE 1.1 0.25 1.011 1.348 1.728 0.011
ARE 1.83 1.27 1.419 1.605 2.463 1.004
WRE 2.57 2.298 2.402 3.203 1.728
REC29 30*15 2287 BRE 1.4 0.57 1.049 1.442 1.968 0.499
ARE 2.7 1.42 2.58 1.888 3.109 1.243
WRE 2.97 3.935 2.492 4.067 2.317
REC31 50*10 3045 BRE 0.43 0.43 2.299 1.51 2.594 0.427
ARE 1.34 1.91 3.392 2.254 3.232 0.99
WRE 2.66 4.532 2.692 4.237 1.478
REC33 50*10 3114 BRE 0 0 0.61 0 0.835 0
ARE 0.78 0.59 0.728 0.645 1.007 0
WRE 1.28 1.734 0.834 1.477 0
REC35 50*10 3277 BRE 0 0 0 0 0 0
ARE 0 0 0.037 0 0.038 0
WRE 0 0.092 0 0.092 0
REC39 75*20 5087 BRE 2.2 0.9 2.28 1.553 2.85 0.081
ARE 2.79 1.88 3.851 2.426 3.371 1.573
WRE 3.38 5.347 2.83 3.951 1.985
REC41 75*20 4960 BRE 3.64 1.69 3.81 2.641 4.173 1.457
ARE 4.92 2.72 5.095 3.684 4.867 2.347
WRE 3.55 6.532 4.052 5.585 3.044
Table 9

Comparison on car and Rec benchmarks.

In Table 9, we use the best relative error (BRE), the average relative error (ARE), and the worst relative error (WRE) to measure the algorithm [59]. They are provided as follows:

BRE=vvbestv(16)
ARE=vvavgv(17)
WRE=vvworstv(18)
where v* is the best scheduling result up to date, vbest is the best experimental result in this experiment, vavg is the average value of the experimental results, vworst is the worst experimental result in this experiment.

Algorithm 2: The proposed CWA

Initialize a population of n random whales search agents

Apply LRV on each search to be mapped into job permutation

Solve the PFSSP using NEH

Choose about 10% random search agents from the population

and replace them with NEH

Evaluate all n search agents

w* = the best search agent

find w*

t = 0

While (t < iterations)

for each search agent

  Update WOA parameters (a, A, C, l, and k)

  if (k < 0.5)

    if (|A| < l)

      xt+1=wA·D

    else if (|A| ≥ l)

      xt+1=xrandA·D

    end if

  else if (k ≥ 0.5)

    xt+1=E·ebl·cos2πl+w

  end if

end for

 Apply LRV on each search agent xt+1

 Perform cross selection strategy on the search agent xt+1

if fxt+1<fw

  w=xt+1

end if

 Perform Chaotic mapping strategy on the search agent xt+1

if f(xr) < f(w*)

  w* = xr

end if

w* = Perform local search strategy on the best search w*

 Evaluate the search agent xt+1

 Update w* if xt+1 is better

t = t + 1

end while

return w*

Figures 68 shows the average value of BRE, ARE, and WRE on Table 9 about PSOVNS, PSOMA, DBA, HBSA, and CWA. The average values of CWA of BRE, ARE, and WRE are 0.14, 0.44, 0.73, respectively.

Figure 6

Average of best relative error (BRE).

Figure 7

Average of average relative error (ARE).

Figure 8

Average of worst relative error (WRE).

Table 9 shows the results of comparisons with other functions on Car and Rec benchmarks. The value in bold represents the best result. Looking at the Table 8, the scheduling results for each instance are different, but CWA's scheduling results are the best in both Car and Rec benchmarks.

Figure 6 reveals the average value of BRE. From Figure 6, CWA shows a shorter maximum completion time than the other five algorithms, and the scheduling results of HBSA are second only to CWA in terms of scheduling results.

Figure 7 displays the average of ARE. Compared with the best and worst values, the average value is the most suitable value to represent the performance of the algorithm. From Figure 7, for this case, CWA scheduling results are still better than other algorithms.

Figure 8 reveals the average value of WRE. It can be seen that even in the worst case, CWA is still the best method.

CWA is superior to the other five algorithms. It can be seen from the experimental results, CWA is very effective in the field of PFSSP.

Table 10 shows the experiments on Heller benchmark instances and compares them with NEH, SG [59], and DSOMA [100]. The CWA algorithm is not only superior to the three methods, but also better than the current known best values. Therefore, the 516 of Hel1 and 136 of Hel2 can be used as better-known results for later experimenters.

Instance n*m Known Best NEH SG DSOMA CWA
Hel1 20*10 516 518 515 631 515
Hel2 100*10 136 141 137 150 135
Table 10

Comparison on Heller benchmark.

6.3. Comparison on Benchmark Taillard

In this section, we measure the performance of CWA on the Taillard benchmark instances and compare it with other algorithms. The best known value comes from [101]. This benchmark contains 120 examples, from the smallest 20*5 to the largest 500*20. We compare the Taillard function from the three aspects of the best scheduling, the average scheduling and the worst scheduling. In Table 11, we listed the experimental results of CWA and the following methods for best scheduling:

  • Discrete self-organizing migrating algorithm (DSOMA) [100]. New sampling routines have been developed that propagate the space between solutions in order to drive the algorithm.

  • Iterated greedy with referenced insertion scheme (IG_RIS) [102]. IG_RIS propose a constructive heuristic to generate an initial solution. In addition, an iteration jumping probability is proposed to change the neighborhood structure from insertion neighborhood to swap neighborhood.

  • Improved iterated greedy algorithm (IIGA) [103]. In the proposed IIGA, a speed-up method for the insert neighborhood is developed to evaluate the whole insert neighborhood of a single solution with n12 neighbors in time O(n2), where n is the number of jobs.

  • Hybrid genetic algorithm (HGA) [95]. HGA is a hybrid GA, which is used to solve the no wait FSSP with completion period as the goal.

  • Tabu-mechanism improved iterated greedy (TMIIG) [104]. The authors have modified the IG algorithm by utilizing a Tabu-based reconstruction strategy to enhance its exploration ability. A powerful neighborhood search method that involves insert, swap, and double-insert moves is then applied to obtain better solutions from the reconstructed solution in the previous step.

  • Discrete water wave optimization algorithm (DWWO) [105]. In refraction operator, a crossover strategy is employed by DWWO to avoid the algorithm falling into local optima.

Instance n*m Known Best DSOMA IG_RIS IIGA HGA TMIIG DOOW CWA
Ta001 20*5 1278 1374 1486 1449 1486 1486 1278
Ta002 1359 1408 1528 1460 1528 1528 1359
Ta003 1081 1280 1460 1386 1460 1460 1081
Ta004 1293 1448 1588 1521 1588 1588 1293
Ta005 1235 1341 1449 1403 1449 1449 1235
Ta006 1195 1363 1481 1430 1481 1481 1195
Ta007 1239 1381 1483 1461 1483 1483 1239
Ta008 1206 1379 1482 1433 1482 1482 1206
Ta009 1230 1373 1469 1398 1469 1469 1230
Ta010 1108 1283 1377 1324 1377 1377 1108
Ta011 20*10 1582 1698 2044 1955 2044 2044 1586
Ta012 1659 1833 2166 2123 2166 2166 1671
Ta013 1496 1676 1940 1912 1940 1940 1512
Ta014 1377 1546 1811 1782 1811 1811 1386
Ta015 1419 1617 1933 1933 1933 1933 1424
Ta016 1397 1590 1892 1827 1892 1892 1397
Ta017 1484 1622 1963 1944 1963 1963 1484
Ta018 1538 1731 2057 2006 2057 2057 1560
Ta019 1593 1747 1973 1908 1973 1973 1611
Ta020 1591 1782 2051 2001 2051 2051 1608
Ta021 20*20 2297 2436 2973 2912 2973 2973 2318
Ta022 2099 2234 2852 2780 2852 2852 2122
Ta023 2326 2479 3013 2922 3013 3013 2345
Ta024 2223 2348 3001 2967 3001 3001 2237
Ta025 2291 2435 3003 2953 3003 3003 2296
Ta026 2226 2383 2998 2908 2998 2998 2235
Ta027 2273 2390 3052 2970 3052 3052 2293
Ta028 2200 2328 2839 2763 2839 2839 2216
Ta029 2237 2363 3009 2972 3009 3009 2245
Ta030 2178 2323 2979 2919 2979 2979 2195
Ta031 50*5 2724 3033 2974 3161 3127 3161 3170 2724
Ta032 2834 3045 3171 3432 3438 3440 3441 2836
Ta033 2621 3036 2988 3211 3182 3213 3218 2622
Ta034 2751 3011 3113 3339 3289 3343 3349 2751
Ta035 2863 3128 3140 3356 3315 3361 3376 2863
Ta036 2829 3166 3158 3347 3324 3346 3352 2829
Ta037 2725 3021 3005 3231 3183 3234 3243 2725
Ta038 2683 3063 3040 3235 3243 3241 3239 2700
Ta039 2552 2908 2889 3072 3059 3075 3078 2554
Ta040 2782 3120 3094 3317 3301 3322 3330 2782
Ta041 50*10 2991 3638 3605 4274 4251 4274 4274 3025
Ta042 2867 3511 3470 4177 4139 4179 4180 2905
Ta043 2839 3492 3465 4099 4083 4099 4099 2888
Ta044 3063 3672 3649 4399 4480 4399 4407 3063
Ta045 2976 3633 3614 4322 4316 4324 4324 3006
Ta046 3006 3621 3574 4289 4282 4290 4294 3020
Ta047 3093 3704 3667 4420 4376 4420 4420 3115
Ta048 3037 3572 3549 4318 4304 4321 4323 3042
Ta049 2897 3541 3510 4155 4162 4158 4155 2904
Ta050 3065 3624 3603 4283 4232 4286 4286 3099
Ta051 50*20 3850 4511 4484 6129 6138 6129 6129 3934
Ta052 3704 4288 4262 5725 5721 5725 5725 3797
Ta053 3640 4289 4261 5862 5847 5873 5862 3717
Ta054 3720 4378 4338 5788 5781 5789 5789 3790
Ta055 3610 4271 4249 5886 5891 5886 5886 3715
Ta056 3681 4202 4271 5863 5875 5874 5871 3785
Ta057 3704 4315 4291 5962 5937 5968 5969 3781
Ta058 3691 4326 4298 5926 5919 5940 5926 3797
Ta059 3743 4329 4304 5876 5839 5876 5876 3852
Ta060 3756 4422 4398 5958 5935 5959 5958 3812
Ta061 100*5 5493 6151 6038 6397 6492 6397 6433 5493
Ta062 5268 6064 5933 6234 6353 6246 6268 5268
Ta063 5175 6003 5837 6121 6148 6133 6162 5175
Ta064 5014 5786 5661 6026 6080 6028 6055 5023
Ta065 5250 6021 5873 6200 6254 6206 6221 5250
Ta066 5135 5869 5732 6074 6177 6088 6121 5135
Ta067 5246 6004 5890 6247 6257 6254 6311 5247
Ta068 5094 5924 5785 6130 6225 6150 6197 5099
Ta069 5448 6154 6029 6370 6443 6391 6418 5453
Ta070 5322 6186 6049 6381 6441 6396 6404 5328
Ta071 100*10 5770 7042 6896 8077 8115 8080 8093 5785
Ta072 5349 6813 6622 7880 7986 7888 7891 5391
Ta073 5676 6943 6766 8028 8057 8042 8047 5691
Ta074 5781 7198 7037 8348 8327 8350 8364 5830
Ta075 5467 6815 6690 7958 7991 7967 7966 5497
Ta076 5303 6685 6517 7801 7823 7808 7791 5308
Ta077 5595 6827 6684 7866 7915 7880 7881 5638
Ta078 5617 6874 6729 7913 7939 7912 7924 5660
Ta079 5871 6092 6908 8161 8226 8164 8152 5840
Ta080 5845 6990 6832 8114 8186 8130 8126 5860
Ta081 100*20 6202 7854 7683 10700 10745 10722 10727 6357
Ta082 6183 7910 7739 10594 10655 10611 10604 6314
Ta083 6271 7825 7697 10611 10672 10629 10624 6356
Ta084 6269 7902 7730 10607 10630 10615 10615 6359
Ta085 6314 7901 7694 10539 10548 10563 10551 6421
Ta086 6364 7921 7745 10690 10700 10684 10680 6468
Ta087 6268 8051 7848 10825 10827 10832 10824 6406
Ta088 6401 8025 7879 10839 10863 10846 10839 6548
Ta089 6275 7969 7771 10723 10751 10763 10745 6393
Ta090 6434 8036 7818 10798 10794 10797 10787 6538
Ta091 200*10 10862 13507 13100 15319 15739 15377 15418 10885
Ta092 10480 13458 13048 15085 15534 15167 15252 10551
Ta093 10922 13521 13135 15376 15755 15416 15412 10948
Ta094 10889 13686 13112 15200 15842 15250 15304 10893
Ta095 10524 13547 13097 15209 15692 15268 15277 10537
Ta096 10326 13247 12869 15109 15622 15163 15222 10375
Ta097 10854 13910 13351 15395 15877 15441 15459 10868
Ta098 10730 13830 13225 15237 15733 15295 15307 10792
Ta099 10438 13410 13036 15100 15573 15155 15186 10465
Ta100 10675 13744 13119 15340 15803 15382 15378 10690
Ta101 200*20 11195 15027 14484 19681 20148 19723 19724 11355
Ta102 11203 15211 14690 20096 20539 20127 20091 11438
Ta103 11281 15247 14776 19913 20511 19973 19989 11517
Ta104 11275 15174 14694 19928 20461 19997 19940 11504
Ta105 11259 15047 14547 19843 20339 19900 19875 11397
Ta106 11176 15212 14734 19942 20501 20003 19980 11366
Ta107 11360 15168 14744 20112 20680 20145 20119 11544
Ta108 11334 15247 14763 20056 20614 20053 20053 11537
Ta109 11192 15136 14643 19918 20300 19998 19932 11406
Ta110 11288 15243 14703 19935 20437 19932 19916 11551
Ta111 500*20 26059 37064 35372 46689 49095 47046 46871 26483
Ta112 26520 37419 35743 47275 49461 47630 47294 26890
Ta113 26371 37059 35452 46544 48777 46977 46846 26717
Ta114 26456 37014 35687 46899 49283 47328 47185 26768
Ta115 26334 36894 35417 46741 48950 47238 47037 26611
Ta116 26477 37372 35747 46941 49533 47553 47166 26767
Ta117 26389 36998 35395 46509 48943 46944 46794 26639
Ta118 26560 36944 35568 46873 49277 47346 47127 26873
Ta119 26005 36862 35304 46743 49207 47205 47025 26282
Ta120 26457 37098 35643 46847 49092 47374 47105 26735
Table 11

Comparison on Taillard benchmark based on the best scheduling.

The bold in Table 11 represents the best value of each instance. It is obvious that the best value of each instance is obtained by CWA, so the performance of CWA is better than the other six algorithms.

Table 11 shows the results of comparisons with other six methods on Taillard benchmark based on the best scheduling and this table gives the value of scheduling time directly. Because the scale of Taillard benchmark is relatively large, so we can see that when the size of the instance is relatively small, CWA can achieve the optimal scheduling, but with the increment of the size of the instance, it is difficult to achieve the optimal scheduling. Even if the optimal scheduling value is not reached, it is close to this value, and among the seven algorithms, the maximum completion time of CWA is the smallest.

Figure 9 shows the average error (AE) [59] of best scheduling, which is defined as follows:

AE=1di=1kzziz100(19)
where z* indicates the best scheduling time for the known instances. zi is the scheduling time of CWA in the ith instance of best, mean, and worst scheduling. In this benchmark, d is equal to 120, representing 120 instances. CWA obtained the minimum value of all algorithms, which is 0.85.

Figure 9

Average error (AE) of best scheduling.

From Figure 9, the performance of DSOMA is second only to CWA, but the AE of DSOMA is 24 times that of CWA. The performance of CWA is much better than other algorithms.

Figure 10 shows the degree of improvement of CWA of best scheduling relative to other algorithms, called improvement percentage (IP) [59], which is defined as follows:

IP=mCWAmxmCWA(20)
where mcwa and mx respectively indicate the completion time of CWA and the completion time of other algorithms in these three schedulers.

Figure 10

Improvement percentage (IP) of chaotic whale algorithm (CWA) based on the best scheduling.

In Figure 10, the IP of HGA, DOOW, TMIIG, and IIGA are comparatively similar, which indicates that their performance on Taillard benchmark is comparatively close. What's more, CWA outperforms HGA and DOOW to the greatest extent and IG_RIS to the smallest extent, but it outperforms all these algorithms.

Table 12 compares the average scheduling time of the Taillard benchmark, as compared with the following algorithm:

  • Nawaz-Enscore-Ham (NEH) [90]

  • Tabu-mechanism improved iterated greedy (TMIIG) [104]

  • Improved iterated greedy algorithm (IIGA) [103]

  • Hybrid genetic algorithm (HGA) [95]

  • Discrete water wave optimization algorithm (DWWO) [105]

Instance n*m Known Best NEH TMIIG IIGA HGA DWWO CWA
Ta001 20*5 1278 1299 1486 1486 1472.7 1486 1278
Ta002 1359 1365 1528 1528 1477.6 1528 1359
Ta003 1081 1132 1460 1460 1406.8 1460 1081
Ta004 1293 1329 1588 1588 1546.4 1588 1293
Ta005 1235 1305 1449 1449 1426.6 1449 1235
Ta006 1195 1251 1481 1481 1446.6 1481 1195
Ta007 1239 1251 1483 1483 1479.4 1483 1239
Ta008 1206 1215 1482 1482 1459.2 1482 1206
Ta009 1230 1284 1469 1469 1409.2 1469 1230
Ta010 1108 1127 1377 1377 1345.5 1377 1108
Ta011 20*10 1582 1681 2044 2044 1972.6 2044 1604
Ta012 1659 1766 2166 2166 2154.6 2166 1691
Ta013 1496 1562 1940.4 1940 1931.1 1940 1530.5
Ta014 1377 1416 1811 1811 1794.3 1811 1409.35
Ta015 1419 1502 1933 1933 1934.4 1933 1450.75
Ta016 1397 1456 1892 1892 1850 1892 1423.05
Ta017 1484 1531 1963 1963 1951.5 1963 1500.35
Ta018 1538 1626 2057 2058.6 2038.5 2057 1577.4
Ta019 1593 1639 1973 1973 1953.9 1973 1620.5
Ta020 1591 1656 2051 2051 2019.3 2051 1633.05
Ta021 20*20 2297 2443 2973 2973 2938.5 2973 2346
Ta022 2099 2134 2852 2852 2814.2 2852 2146.9
Ta023 2326 2414 3019.4 3019.4 2962.4 3014.3 2381.15
Ta024 2223 2257 3001 3001 2982.5 3001 2273.5
Ta025 2291 2370 3003 3003 2995.1 3003 2340.6
Ta026 2226 2349 2998 2998 2932.8 2998 2243.9
Ta027 2273 2383 3052 3052 3004.8 3052 2316.75
Ta028 2200 2249 2839 2839 2789.5 2839 2242.05
Ta029 2237 2306 3009 3009 3005.3 3009 2282.55
Ta030 2178 2257 2979 2979 2956.3 2979 2242.9
Ta031 50*5 2724 2729 3162.4 3222 3198.3 3171.8 2720.6
Ta032 2834 2882 3441 3474.4 3453.1 3444.5 2837.9
Ta033 2621 2650 3216 3262.8 3242.5 3231.8 2624.5
Ta034 2751 2782 3346.6 3374.8 3349.8 3350.8 2759.35
Ta035 2863 2868 3364.6 3406 3369.1 3376.7 2863.65
Ta036 2829 2835 3347.6 3382.2 3356.4 3356.2 2833.75
Ta037 2725 2806 3235.8 3267.6 3246.4 3245.3 2725
Ta038 2683 2700 3242.4 3275.2 3264.9 3240.3 2694.2
Ta039 2552 2606 3078.8 3123.2 3088.2 3086.8 2570.25
Ta040 2782 2801 3327.2 3377.2 3341.5 3336.5 2785.3
Ta041 50*10 2991 3175 4276.8 4311.2 4289.1 4281.5 3043.55
Ta042 2867 3073 4185 4201 4193.5 4184.5 2918.55
Ta043 2839 2994 4107.6 4124 4108.6 4105.5 2916.7
Ta044 3063 3218 4405 4439.2 4517.4 4405.7 3072.45
Ta045 2976 3186 4330.4 4347 4333.2 4324.7 3025.95
Ta046 3006 3148 4297.2 4330 4301.6 4295.2 3050.1
Ta047 3093 3277 4429.6 4441.4 4412.6 4420 3145.65
Ta048 3037 3170 4327 4357.6 4331.9 4323.3 3053.55
Ta049 2897 3025 4164.2 4194.8 4173 4161.7 2931.55
Ta050 3065 3267 4286.2 4301 4279 4286.2 3115
Ta051 50*20 3850 4006 6139.6 6154.6 6149.7 6138.5 3966.67
Ta052 3704 3958 5741 5762.2 5751.5 5733.5 3829
Ta053 3640 3866 5882.4 5907 5884.4 5865.5 3760
Ta054 3720 3953 5791.4 5802.6 5804.7 5790.7 3837.15
Ta055 3610 3872 5899.4 5930.6 5909.7 5893.5 3759.82
Ta056 3681 3861 5883.4 5912.6 5890.6 5874.3 3820.91
Ta057 3704 3927 5974 6012 5974.2 5974 3814
Ta058 3691 3914 5945.4 5970.2 5951 5930.5 3842.58
Ta059 3743 3970 5883.2 5900.6 5873.5 5876 3882.91
Ta060 3756 4036 5959 5982 5963.5 5958.8 3872
Ta061 100*5 5493 5514 6413.4 6586.4 6557.2 6438.3 5493
Ta062 5268 5284 6252.2 6428.2 6409.4 6285.5 5279.27
Ta063 5175 5222 6135.8 6292.8 6260.4 6164.2 5186.27
Ta064 5014 5023 6031.4 6184.8 6159 6050.7 5023
Ta065 5250 5261 6217.8 6358.4 6325.6 6223.3 5254.4
Ta066 5135 5154 6096 6269.2 6225.7 6122.7 5137.8
Ta067 5246 5282 6263.4 6442.2 6409 6308.8 5258.6
Ta068 5094 5140 6156.6 6338.6 6308.6 6210.8 5124.9
Ta069 5448 5489 6395 6557.4 6516.3 6422 5465.4
Ta070 5322 5336 6401.6 6561 6542.5 6427.2 5327.3
Ta071 100*10 5770 5897 8094.2 8224.6 8173.5 8100.8 5798.4
Ta072 5349 5466 7909 8061.4 8048.1 7938.3 5403.3
Ta073 5676 5747 8054.4 8152 8142.2 8051 5699.2
Ta074 5781 5924 8362.4 8484.2 8437.5 8378.2 5860
Ta075 5467 7991 7981.2 8087.6 8046.4 7980 5539.8
Ta076 5303 7823 7821.6 7930.4 7883.7 7809 5325.6
Ta077 5595 7915 7887.2 7983.8 8007 7889 5676.3
Ta078 5617 7939 7924.8 8051.4 8049.2 7931.7 5688
Ta079 5871 8226 8172 8312.6 8290.4 8183.8 5902.3
Ta080 5845 8186 8148.8 8263.6 8255.3 8141 5898.7
Ta081 100*20 6202 10745 10782.4 10887 10826.6 10744 6370
Ta082 6183 10655 10623 10755.6 10762.5 10608.7 6350
Ta083 6271 10672 10650 10774.6 10740.4 10637 6390.9
Ta084 6269 10630 10647.2 10765.8 10679.7 10629.8 6379.5
Ta085 6314 10548 10579.8 10690.6 10658.2 10568.3 6449.9
Ta086 6364 10700 10696.2 10819.4 10753 10690 6503.2
Ta087 6268 10827 10849.2 10997.2 10913.6 10843.2 6489.3
Ta088 6401 10863 10862 11011.6 10905.2 10850.2 6571.9
Ta089 6275 10751 10780.4 10875 10859 10758 6433
Ta090 6434 10794 10817.2 10944.6 10928.8 10808.5 6594.5
Ta091 200*10 10862 15739 15397.4 15751.4 15843.5 15449.3 10913.95
Ta092 10480 15534 15203 15572.2 15645.3 15281.3 10894.8
Ta093 10922 15755 15433.2 15762.4 15882.4 15454.5 11002
Ta094 10889 15842 15279.8 15625 15927 15304.7 13054.8
Ta095 10524 15692 15280.8 15626.2 15763.8 15295.8 10651.2
Ta096 10326 15622 15189.6 15557 15669.9 15235.5 10447.91
Ta097 10854 15877 15476.6 15842 15962.3 15508.3 10889.4
Ta098 10730 15733 15319.2 15685.6 15833.2 15340 10805.7
Ta099 10438 15573 15185.2 15525.8 15626.2 15240.2 10485.4
Ta100 10675 15803 15410.8 15754.4 15869.1 15412.2 10723.3
Ta101 200*20 11195 20148 19770.8 20127.8 20331.3 19736.7 11401
Ta102 11203 20539 20148.8 20536.6 20763.5 20118.8 11500.5
Ta103 11281 20511 19995.2 20355.6 20583.8 19999 11585.6
Ta104 11275 20461 20032 20349.6 20594.6 19967.2 11552.1
Ta105 11259 20339 19936 20331.4 20544.4 19923.8 11450.3
Ta106 11176 20501 20050.6 20407 20661.9 20011.3 11436.4
Ta107 11360 20680 20167.4 20539.2 20804.3 20147.5 11586.2
Ta108 11334 20614 20095.6 20449.8 20665.7 20084 11585.7
Ta109 11192 20300 20022.2 20358.2 20574.9 19975.5 11457.3
Ta110 11288 20437 19968.6 20284 20587.6 19946.3 11603.45
Ta111 500*20 26059 49095 47147.2 48253.4 49289.4 46965.5 26550.1
Ta112 26520 49461 47692.2 48699 49948.2 47429.8 27020.9
Ta113 26371 48777 47049.6 48116.4 49139 46878.7 26802.2
Ta114 26456 49283 47375.8 48513.2 49657.8 47232.8 26845.3
Ta115 26334 48950 47280 48345.8 49423.1 47099.2 26659.55
Ta116 26477 49533 47591.2 48701.6 49848.4 47331.5 26833.8
Ta117 26389 48943 47087.2 48186 49366.5 46866.3 26722.3
Ta118 26560 49277 47421.4 48468.2 49675.8 47248.7 26990.3
Ta119 26005 49207 47248.6 48297.2 49429.3 47083.5 26436.7
Ta120 26457 49092 47402.4 48483.4 49545.6 47183.2 26803
Table 12

Comparison on Taillard benchmark based on the mean scheduling.

Tables 8 and 11 represent the average of the algorithm. This shows that the CWA's scheduling results can achieve a greater degree of optimization in different instances. Similarly, the proposed CWA method is superior to all other algorithms in mean scheduling. It can achieve the minimum scheduling time for each instance.

Figure 11 is AE of all algorithms in mean scheduling. The maximum value of IIGA is 46.84, and the minimum value of CWA is 1.65, that means they are 28 times different. It can be seen from the diagram that the optimization range of CWA is very large, also shown in Figure 12. At the same time, it can be seen from Figure 12, the performance of NEH algorithm is the worst in Taillard instance, while the performance of the other four algorithms is comparatively similar. It shows that CWA improves other algorithms to a similar degree, and the degree of improvement is very large.

Figure 11

Average error (AE) of mean scheduling.

Figure 12

Improvement percentage (IP) of chaotic whale algorithm (CWA) based on the mean scheduling.

Table 13 shows the worst scheduling of Taillard benchmark instances. Comparing CWA with HGA [95], the results show that CWA is always better than HGA, including smaller or larger instances.

Instance n*m Known Best HGA CWA Instance n*m Known Best HGA CWA
Ta001 20*5 1278 1485 1278 Ta061 100*5 5493 6594 5493
Ta002 1359 1505 1359 Ta062 5268 6469 5284
Ta003 1081 1431 1081 Ta063 5175 6320 5193
Ta004 1293 1573 1293 Ta064 5014 6198 5023
Ta005 1235 1445 1235 Ta065 5250 6397 5265
Ta006 1195 1471 1195 Ta066 5135 6296 5139
Ta007 1239 1496 1239 Ta067 5246 6460 5266
Ta008 1206 1475 1206 Ta068 5094 6359 5123
Ta009 1230 1429 1230 Ta069 5448 6561 5465
Ta010 1108 1368 1108 Ta070 5322 6592 5335
Ta011 20*10 1582 1998 1622 Ta071 100*10 5770 8230 5800
Ta012 1659 2166 1692 Ta072 5349 8101 5424
Ta013 1496 1942 1534 Ta073 5676 8215 5740
Ta014 1377 1811 1434 Ta074 5781 8505 5860
Ta015 1419 1947 1467 Ta075 5467 8096 5525
Ta016 1397 1879 1539 Ta076 5303 7947 5328
Ta017 1484 1971 1504 Ta077 5595 8081 5692
Ta018 1538 2066 1571 Ta078 5617 8105 5694
Ta019 1593 1973 1646 Ta079 5871 8349 5940
Ta020 1591 2032 1674 Ta080 5845 8340 5903
Ta021 20*20 2297 2972 2362 Ta081 100*20 6202 10932 6435
Ta022 2099 2835 2160 Ta082 6183 10847 6373
Ta023 2326 2984 2372 Ta083 6271 10821 6374
Ta024 2223 2994 2316 Ta084 6269 10797 6408
Ta025 2291 3017 2338 Ta085 6314 10777 6460
Ta026 2226 2964 2263 Ta086 6364 10853 6535
Ta027 2273 3028 2353 Ta087 6268 11031 6568
Ta028 2200 2826 2267 Ta088 6401 10992 6593
Ta029 2237 3009 2299 Ta089 6275 10963 6482
Ta030 2178 2979 2246 Ta090 6434 11067 6620
Ta031 50*5 2724 3229 2735 Ta091 200*10 10862 15916 10950
Ta032 2834 3475 2838 Ta092 10480 15764 10570
Ta033 2621 3277 2624 Ta093 10922 16026 11045
Ta034 2751 3384 2762 Ta094 10889 16111 10911
Ta035 2863 3404 2864 Ta095 10524 15829 10553
Ta036 2829 3377 2835 Ta096 10326 15731 10575
Ta037 2725 3280 2725 Ta097 10854 16029 10908
Ta038 2683 3288 2705 Ta098 10730 15933 10803
Ta039 2552 3121 2577 Ta099 10438 15759 10492
Ta040 2782 3383 2814 Ta100 10675 15934 10727
Ta041 50*10 2991 4306 3054 Ta101 200*20 11195 20458 11461
Ta042 2867 4235 2923 Ta102 11203 20889 11496
Ta043 2839 4124 2953 Ta103 11281 20636 11609
Ta044 3063 4549 3071 Ta104 11275 20753 11638
Ta045 2976 4367 3051 Ta105 11259 20601 11466
Ta046 3006 4332 3083 Ta106 11176 20780 11442
Ta047 3093 4444 3165 Ta107 11360 20915 11627
Ta048 3037 4347 3076 Ta108 11334 20814 11601
Ta049 2897 4188 2965 Ta109 11192 20757 11501
Ta050 3065 4304 3120 Ta110 11288 20712 11626
Ta051 50*20 3850 6172 3984 Ta111 500*20 26059 49580 26582
Ta052 3704 5790 3859 Ta112 26520 50354 27161
Ta053 3640 5929 3831 Ta113 26371 49399 26832
Ta054 3720 5827 3864 Ta114 26456 50004 26838
Ta055 3610 5950 3794 Ta115 26334 49847 26716
Ta056 3681 5911 3803 Ta116 26477 50046 27048
Ta057 3704 6001 3798 Ta117 26389 49591 26832
Ta058 3691 5971 3880 Ta118 26560 49942 27020
Ta059 3743 5899 3826 Ta119 26005 49697 26513
Ta060 3756 5979 3922 Ta120 26457 50002 26886
Table 13

Comparison on Taillard benchmark based on the worst scheduling.

Figure 13 shows the improvement degree of CWA relative to HGA, and CWA is significantly improved. With the increment of instance size, the optimization degree of CWA increases, and reached 85.7% when the instance is 500*20. The experimental results show that the optimization of CWA is very successful.

Figure 13

Improvement percentage (IP) achieved by chaotic whale algorithm (CWA) for hybrid genetic algorithm (HGA) on the worst scheduling.

7. CONCLUSIONS AND FUTURE WORK

This paper proposed a new CWA which combined chaotic mapping, inserted local search, and cross selection operator. The chaotic map stabilizes the random factors in the algorithm and increases the stability and optimization of the algorithm. What's more, these strategies also make the algorithm avoid falling into the local optimum through local search with small probability and circular cross selection operator. NEH is used to initialize the population so that a better result can be obtained in the initialization stage. In the process of initializing population, LRV rule is used to transform the continuous search agent into discrete job sequence, which optimizes the rationality of job generation. The experimental results show that it is feasible to use chaos strategy to optimize whale algorithm.

Compared with other scheduling algorithms, CWA has other advantages, such as fewer variables to be adjusted, faster convergence speed, and more stability. Because the parameters have a great influence on the performance of the algorithm, its stability is improved when chaotic mapping is used to make the parameters tend to be controllable [106]. However, the algorithm still has uncertainty, which needs further optimization and improvement. In future, except the methods used in the current paper, some of the most presentative metaheuristic algorithms [107], such as BA [108110], biogeography-based optimization (BBO) [111,112], ACO [113], cuckoo search (CS) [114,115], earthworm optimization algorithm (EWA) [116], elephant herding optimization (EHO) [117,118], moth search (MS) algorithm [119], firefly algorithm (FA) [120], ABC [121123], harmony search (HS) [124], monarch butterfly optimization (MBO) [125,126], PSO [127,128], genetic programming [129], CS [84,130,131] and more recently, the KH algorithm [81,132135].

In this paper, CWA is tested on Carlier, Reeves, Heller, and Taillard benchmarks. Since only these four groups of benchmark functions have been tested, another limitation is the lack of research on broader dimensions. But from the results of the experiment, the proposed CWA algorithm has better stability and optimization results. In the current work, we only apply CWA to PFSSP. In the future, the algorithm will be used to solve other engineering problems. We are ready to apply this algorithm to the differential distributed job shop scheduling system, and use it to study big data of ocean.

CONFLICTS OF INTEREST

The authors declare that they have no conflicts of interest.

AUTHORS' CONTRIBUTIONS

Conceptualization, J.L. and L.W.; methodology, L.G. and H.H.; software, Y.L. and L.W.; validation, J.L.; writing-original draft preparation, J.L. and C.L.; writing-review and editing, L.G.; All authors have read and agreed to the published version of the manuscript.

Funding Statement

This work was supported by the National Natural Science Foundation of China (No. 61977059).

ACKNOWLEDGMENTS

The authors would like to thank the anonymous reviewers and the editor for their careful reviews and constructive suggestions to help us improve the quality of this paper.

REFERENCES

6.D. Gao, G.-G. Wang, and W. Pedrycz, Solving fuzzy job-shop scheduling problem using DE algorithm improved by a selection mechanism, IEEE Trans. Fuzzy Syst., Vol. 28, 2020, pp. 3265-3275.
51.D.E. Goldberg, Genetic algorithms in search, Optim. Mach. Learn., Vol. 1, 1989, pp. 1.
Journal
International Journal of Computational Intelligence Systems
Volume-Issue
14 - 1
Pages
651 - 675
Publication Date
2021/01/19
ISSN (Online)
1875-6883
ISSN (Print)
1875-6891
DOI
10.2991/ijcis.d.210112.002How to use a DOI?
Copyright
© 2021 The Authors. Published by Atlantis Press B.V.
Open Access
This is an open access article distributed under the CC BY-NC 4.0 license (http://creativecommons.org/licenses/by-nc/4.0/).

Cite this article

TY  - JOUR
AU  - Jiang Li
AU  - Lihong Guo
AU  - Yan Li
AU  - Chang Liu
AU  - Lijuan Wang
AU  - Hui Hu
PY  - 2021
DA  - 2021/01/19
TI  - Enhancing Whale Optimization Algorithm with Chaotic Theory for Permutation Flow Shop Scheduling Problem
JO  - International Journal of Computational Intelligence Systems
SP  - 651
EP  - 675
VL  - 14
IS  - 1
SN  - 1875-6883
UR  - https://doi.org/10.2991/ijcis.d.210112.002
DO  - 10.2991/ijcis.d.210112.002
ID  - Li2021
ER  -