International Journal of Engineering Research in Africa ISSN: 1663-4144, Vol. 36, pp 124-136 doi:10.4028/www.scientific.net/JERA.36.124 © 2018 Trans Tech Publications, Switzerland
Submitted: 2017-12-27 Revised: 2018-01-23 Accepted: 2018-01-24 Online: 2018-06-01
Hybrid Metaheuristics for Solving the Blocking Flowshop Scheduling Problem Ghita Lebbar1,a*, Abdellah El Barkany2,b, Abdelouahhab Jabri3,c and Ikram El Abbassi4,d 1,2,3
Mechanical Engineering Laboratory, Faculty of Sciences and Techniques, Sidi Mohammed Ben Abdellah University, Fez, Morocco
4
ECAM-EPMI, Research Laboratory in Industrial Eco-innovation and Energetics Quartz-Lab, Cergy-Pontoise, France
a*
[email protected],
[email protected],
[email protected] and
[email protected]
Keywords: Blocking flowshop scheduling, makespan, SAGA, simulated annealing, genetic algorithm.
Abstract: This paper suggests two evolutionary optimization approaches for solving the blocking flow shop scheduling problem with the maximum completion time (makespan) criterion, namely the genetic algorithm (GA) and the simulated annealing genetic algorithms (SAGA) that combines the simulated annealing (SA) with the (GA), respectively. The considered problem and the proposed algorithms have some parameters to be adjusted through a design of experiments with exorbitant runs. In fact, a Taguchi method is presented to study the parameterization problem empirically. The performance of the proposed algorithms is evaluated by applying it to Taillard’s well-known benchmark problem, the experiment results show that the SA combined with GA method is advanced to the GA and to the compared algorithms proposed in the literature in minimizing makespan criterion. Ultimately, new known upper bounds for Taillard’s instances are reported for this problem, which can be used thereafter as a basis of benchmark in eventual investigations. Introduction In the manufacturing world, production scheduling is a decision-making and a branch of convoluted movement of information that design the operations planning and control processes. A sturdy production sequence impedes bottlenecks, alleviates idle time and curtails lead-time, thus, minimizing operating cost in one hand and achieving a considerable level of efficacy and productiveness in the other. The permutation flowshop scheduling problems (PFSP) is one of the most extensively researched combinatorial optimization scheduling problems in the literature and the practical applications. In the classical PFSP, a set of N jobs have to be performed sequentially, by a set of M serial machines in the same permutation and a work in process inventory is permitted since intermediate buffers have infinite capacity between consecutive machines. However, in real-life situations, there are many production factories, where the buffers capacity is zero or limited owing to the system features [1]. In similar situations, the basic PFSP becomes blocking flowshop scheduling problem (BFSP), for which, a job cannot leave a machine unless the next immediate downstream machine is available for processing. This problem is denoted as 𝐹𝐹𝑚𝑚 |𝑏𝑏𝑏𝑏𝑏𝑏𝑏𝑏𝑏𝑏|𝐶𝐶𝑚𝑚𝑚𝑚𝑚𝑚 conforming to the standard notation introduced by [2]. With respect to the computational complexity of the problem, (Hall and Sriskandarajah, 1996) [3] proved, relying on the results obtained by [4], that the 𝐹𝐹𝑚𝑚 |𝑏𝑏𝑏𝑏𝑏𝑏𝑏𝑏𝑏𝑏|𝐶𝐶𝑚𝑚𝑚𝑚𝑚𝑚 problem with m ≥ 3 is strongly NP-hard. Thereafter, (Allahverdi et al. 2008) [5] showed that the BFSP is NP-hard even for BFSP with m ≥ 2. As a result, exact methods such as branch and bound and dynamic programing are nots to solve the BFSP optimally, especially when it is concerning wide size problem. All rights reserved. No part of contents of this paper may be reproduced or transmitted in any form or by any means without the written permission of Trans Tech Publications, www.scientific.net. (#108616811, Sidi Mohammed Ben Abdellah University, Morocco-01/06/18,16:44:40)
International Journal of Engineering Research in Africa Vol. 36
125
In view of the complexity of the considered problem, literature on the resolution of the BFSP with the makespan criterion shows a variety of approaches, namely constructive heuristics and improvement heuristics. Among constructive heuristics developed to solve the BFSP. (Mccornick et al.1989) [6], proposed an innovating heuristic, recognized as Profile Fitting (PF), so as to minimize cycle time in a blocking assembly line with m serial stations. (Leisten, 1990) [7] suggested another constructive technique to tackle the BFSP with permutation and non-permutation, in order to increase the exploitation of buffers and decrease the machine blocking times. He concluded that the developed heuristic was not apt to provide fitter solutions than results introduced by [8] heuristic produced originally for the traditional PFSP. On the basis of the afore mentioned heuristics, (Ronconi, 2004) [9] presented tree constructive heuristics, Known as a MinMax (MM) heuristic, (MME) heuristic that represents the combination of MM with NEH, and (PFE) heuristic which combines the PF with NEH heuristic, respectively. Their results indicated that the tree algorithms outperform the NEH heuristic in problems with up to 500 jobs and 20 machines. (Pan et al. 2010) [10] proposed recently two constructive heuristics namely, the weighted profile fitting (wPF) and Pan-Wang (Pw) algorithm relying on the PF method. Thereafter, they combined the NEH procedure with PF, wPF and Pw, and called heuristics, PF-NEH (x), wPF-NEH(x) and Pw-NEH(x), where 𝑥𝑥 ∈ [1, n] is the number of sequences to be generated. The computational results showed that the Pw-NEH provided promising solutions for the considered problem, compared to the existing algorithms. Regarding the meta-heuristics [11] proposed a genetic algorithm (GA) using the idea of deliberately slowing down processing of certain operations. Computational results indicated that the genetic algorithm performs notably better than the improvement heuristics presented by [12] . (Grawborski and pempra, 2000) [1] developed two tabu search algorithms, called (TS) and (TS+M) algorithms and compared them with the uppers bound of Taillard benchmark suite reported by [13]. They noted that tabu search algorithms outperform the branch and bound approach and Carrafa’s genetic algorithm. (Liu and reeves, 2001) [14] proposed a hybrid particle swarm algorithm (HPSO) to solve the BFSP. A hybrid discrete differential evolution algorithm (HDDE) was suggested by [15]. (Ribas et al. 2011) [16] introduced an iterated greedy (IG) algorithm. The author declared that the IG algorithm outperformed the HDDE and produced novel best solution for Taillard benchmark suite. (Han et al. 2012) [17] proposed an improved discrete artificial bee colony (IABC) algorithm for solving effectively the BFSP. A discrete artificial bee colony algorithm combining differential evolution (DE-ABC) was presented by [18]. (Sadaqa et al. 2015) [19] suggested a Meta-heuristic Randomized Priority Search (Meta-RaPS) on the basis on the NEH algorithm. Recently, (Han et al. 2016) [20] proposed a modified fruit fly optimization algorithm (MFFO) basing on the smell-based and vision-based search. The computational results of the MFFO indicated that the proposed algorithm is superior comparing to the algorithms suggested in the literature. The combination of the constituent from various metaheuristics is presently one of the most fruitful tendencies in optimization field. In fact, considerable algorithms that do not preserve the design of single classical metaheuristic have been developed ([21], [22], [23]). The principal incentive for the hybridization of those algorithms is to combine the qualities and the advantages of the individual combined approaches, minimize the effect of their corresponding disadvantages and subsequently enhance the systems performance. In this study, by combining the exploration ability of GA and The power to escape the local optimum of (SA), an effective genetic-simulated annealing algorithm (SAGA) algorithm is suggested for solving the blocking flow shop scheduling problem with makespan criterion. The rest of this paper is organized as follows. In Section 2, the description of Fm|blocking|Cmax is presented. Section 3 outlines the proposed SAGA algorithm. In section 4, algorithms parameters selection is discussed. Section 6 provides the experimental results. Finally, section 7 reports conclusion and future research areas.
126
International Journal of Engineering Research in Africa Vol. 36
1. Formulation of the Blocking Flowshop Scheduling Problem In the 𝐹𝐹𝑚𝑚 |𝑏𝑏𝑏𝑏𝑏𝑏𝑏𝑏𝑏𝑏|𝐶𝐶𝑚𝑚𝑚𝑚𝑚𝑚 problem, a set (𝑗𝑗1 , … , 𝑗𝑗𝑛𝑛 ) of independent n jobs have to be performed sequentially on a set (𝑀𝑀1 , … , 𝑀𝑀𝑚𝑚 ) of m machines in the same order without any intermediate buffers. Each job is composed of (𝑂𝑂𝑗𝑗1 , … , 𝑂𝑂𝑗𝑗𝑗𝑗 ) operations 𝑂𝑂𝑗𝑗𝑗𝑗 , where, each operation has a positive integer processing time. It is assumed that, the setup times are included in the processing times of operations. All jobs are ready for processing at the time zero. Each machine can process at most one job and each job can be processed on at most one machine. Further, Owing to the lack of the intermediate buffers between machines, jobs remains blocked until the next downstream machine is free. That is, the current job on the machine must be blocked after finishing its operation until the next machine is available for processing. The target is to search a sequence of jobs under the aforementioned restrictions for which the makespan is minimized. Let Π= (π1,π2,π3,…,πn) be a set of job permutations, where, π(j) represents the index of the job arranged at the j-th position of π. The Dπ(1),k represents the departure time of job π(1) in machine k. In accordance with [24], the departure time Dπ(j),k of each job in each machine can be calculated using the following recursive expressions: 𝑆𝑆π(𝑗𝑗),0 = 0 (1) Dπ(1),k = Dπ(1),k−1 + pπ(1),k
Sπ(j),0 = Dπ(j)−1,1
k = 1,2, . +. . , m − 1
j = 2,3, … , n
Dπ(j),k = max�Dπ(j),k−1 + pπ(j),k , Dπ(j)−1,k+1 � 𝑘𝑘 = 1, … , m − 1 j = 2, … , n Dπ(j),m = Dπ(j),m−1 + pπ(j),m
j = 1,2, . . . , n
(2) (3)
(4) (5)
In the recursive on top, pπ(j),k represents the processing time of the job π(j) in the sequence. 𝑆𝑆π(𝑗𝑗),0 denotes the start time Sπ(j),0 of the job π(j) on the first machine. Equation (2) calculates the departure time of the first job on machine k. Equation (4) calculates the departure time of the job π(j) on machine k, and Equation (5) computes the departure time of the job π(j) on machine m. According to [24], the makespan value of π can be then calculated as: Cmax (π) = Dπ(n),m
(6)
Hence, the aim of the 𝐹𝐹𝑚𝑚 |𝑏𝑏𝑏𝑏𝑏𝑏𝑏𝑏𝑏𝑏|𝐶𝐶𝑚𝑚𝑚𝑚𝑚𝑚 problem is to find the best sequence π∗ in the set of all permutations Π, so that the completion time of the last job in the permutation is minimized, thus: Cmax (π∗ ) ≤ Cmax (π) π∗ , π 𝜖𝜖 Π
(7)
2. The Proposed SAGA for Solving the BFSP 2.1 The basic genetic algorithm
Genetic algorithms are stochastic search method that are operative to a large gamut of combinatorial optimization problems. GA have been inspired from the biological mechanism of reproduction and natural selection. They are founded on the durability of the fittest insight, yielding to find efficiently a new generation with superior fitness function. In scheduling context, GA consider sequences of job as candidate solution or chromosomes. Figure below depicts an example of a chromosome with permutation encoding. The set of chromosomes generated at the start of the genetic search are referred to as an initial population, within each individual is evaluated by its corresponding value of makespan. Further, before reaching a termination criterion, in every new generation, a new population that retains the same size PS is formed using the fundamental elements of the genetic search procedure namely, selection, crossover and mutation operators.
International Journal of Engineering Research in Africa Vol. 36
127
Figure 1. Example of chromosome with permutation encoding Selection operator is arguably one of the most important operator since it directs the evolution, guides the search and eliminates the feeble chromosomes. GA uses a selection to choice a pair of candidate solutions from the population based on their fitness. Selected individuals are used to create new offsprings with which a new generation is formed. Various selection operators exist to select the parents, that is, roulette wheel selection, tournament, fitness proportionate Selection, rank selection, and so one. In this study, the deterministic tournament selection is adopted for which the best individuals that possess the best fitness value are randomly selected for crossover. The crossover is a process of producing one or more offsprings from one-selected parents. Various tools exist to apply crossover operation. In this study, the two-points crossover is used, in which two-points are randomly selected for sharing the parents. The genes outside the selected points are inherited from the parents to the child. The residual genes will be moved to the second parent in the same instruction from left to right. As a result, two offsprings are generated as shown in Fig.2. The ratio of how many individuals will be chosen for mating is called the crossover probability Pc. However, the higher this probability, the more the population undergoes change.
Figure 2. Crossover operation Mutation is the third operator used in GA process, it is applied to preserve genetic diversity from one generation to the next, to prevent premature convergence and provide a process to escape from local optima. In this work, shift change mutation is used, in this type of mutation, a job at one position is removed and implanted at another random position with a mutation probability Pm as described in Fig.3.
Figure 3. Mutation operation The termination criterion is when the program reaches an absolute number of generations without any improvement in the optimal solution. Algorithm 1 represents the more detailed pseudo-code of the genetic algorithms.
128
International Journal of Engineering Research in Africa Vol. 36
Step 1: Initialize PS, Pc, Pm, max_iter, 𝛼𝛼 Step 2: Generate the initial population P (n) consisting of a random sequences of jobs Step 3: Calculate fitness value ‘makespan’ of each sequence in current population Step 4: select the best fitness g_oldbest.from the current population with the tournament selection approach Step 5: While number of maximum iterations has not been attained: For each sequence: 1 to PS: Select randomly two sequences from P (n) If random ≤ 𝑃𝑃𝑃𝑃 : Apply the two point’s crossover operator and produce two children If random ≤ 𝑃𝑃𝑃𝑃 : Apply the shift change mutation operator Copy the obtained offspring sequences to the new_population P (n+1) Evaluation of the New_population gi (P (n+1)) Step 6: If the stop criterion is satisfied, terminate de SAGA procedure and return the optimal value found, otherwise, go to the step 4
Algorithm 1. The pseudo-code of the GA approach 2.2 The basic simulated annealing Simulated annealing (SA) is a stochastic computational method for finding global optimum to a large combinatorial optimization problem. It is a Monte Carlo approach that simulates the process of slow cooling after heating studied in statistical mechanics systems. The concept of slow cooling is explained as slow diminishing in the probability of accepting worse solutions as the solution space is explored. An appropriate annealing schedule is defined to decrease regularly the control parameter of the optimization known as the temperature. Relying on the iterative progress, SA procedure allows escaping local minima, which are not global and explore globally the search space within a reasonable computing time. In the implementation of the process, a primary solution S is randomly generated, the move from the current solution S to a neighborhood solution 𝑆𝑆 ∗ depends on the probability P of the BoltzmanGibbs distribution of the objective function difference. Besides, this transition is accepted if the following condition is satisfied: 𝑃𝑃 = 𝑚𝑚𝑚𝑚𝑚𝑚 �1, 𝑒𝑒
−∆𝑓𝑓 𝑇𝑇
�≥Ω
(8)
Where, ∆𝑓𝑓 = 𝑓𝑓(𝑆𝑆) − 𝑓𝑓(𝑆𝑆 ∗ ) is the difference between the objective function of the two states, Ω is a randomly number generated in the interval [0, 1] and T is the current control parameter in the process. Starting from a high value of the temperature 𝑇𝑇0 , a search space is performed at each temperature for a certain number of iterance appointed as the markov chain length that depicts the number of moves performed for a fixed value of the temperature. In order to ensure Boltzmann annealing for finding the global minimum, the control parameter T have to decrease logarithmically with the time. Given that, 𝟎𝟎 < 𝛼𝛼 < 𝟏𝟏 is the cooling rate parameter, the annealing schedule is consequently defined as: 𝑇𝑇𝑛𝑛+1 = 𝛼𝛼 𝑇𝑇𝑛𝑛 (9) 2.3 The proposed SAGA algorithm
Genetic algorithms has substantiated to be an efficient and performing optimization approach in several combinatorial optimization problems. In spite of its strength to discover promptly the search space through the genetic operations of reproduction, mutation and crossover. The situation for which it could be stuck in a local optimum is still dissatisfying, in particular, concerning the amount of time absorbed to find an optimal solution. Thus, a more suitable strategy is the use of the integrated method of GA with other approach in which this fatal disadvantage will be avoided. The
International Journal of Engineering Research in Africa Vol. 36
129
simulated annealing is also stochastic method that also search space by means of a series of iterations as mentioned above. Through the cooling process adopted, SA has a strong ability to avoid to be trapped in a local optimum by seeking the optimum around its initial solution, which tends to find local improvements efficiently. Aiming at combining the strength of the both strategies in one hand, and compensate the drawback of the genetic algorithm method in the other, this section introduces the suggested SAGA algorithm for solving the BFSP under study. Following the production of the initial population, a new set of children sequences are evolved through genetic operators of reproduction, crossover and mutation. Each individual is selected by the metropolis criterion of SA formulated in eq (8), given that, ∆𝑓𝑓 in the case of the BFSP represents the difference of makespan between the new population P(n+1) and the old one P(n), denoted as g_newbest and g_oldbest respectively. In other words, the proposed SAGA proceeds in two stages: in the first stage, a set of offsprings is developed by executing genetic implementation. Thereafter, the offsprings obtained are accepted to go into the next generation with use of the local selection of SA relying on the aforementioned probability that depends on the system Temperature T, such that the next generation contains solely the elements whose makespan value equals the highest value in terms of quality. The proposed in this paper admits the replacement of the old population if g_newbest is better than g_oldbest and in the contrary case with a Boltzmann probability superior to a random number generated between [0, 1]. This process is repeated iteratively until the maximum number of iterations is attained and the global optimal value of makespan is obtained. The specific steps of SAGA are illustrated in algorithm 2. Step 1: Initialize PS, Pc, Pm, max_iter, 𝛼𝛼, T=T0 Step 2: Generate the initial population P(n) consisting of a random sequences of jobs Step 3: Calculate fitness value ‘makespan’ of each sequence in current population Step 4: select the best fitness g_oldbest.from the current population Step 5: While number of maximum iterations has not been attained : For each sequence: 1 to PS: Select randomly two sequences from P (n) If random ≤ 𝑃𝑃𝑃𝑃 : Apply the two point’s crossover operator and produce two children If random ≤ 𝑃𝑃𝑃𝑃 : Apply the shift change mutation operator Copy the obtained offspring sequences to the new_population P (n+1) Evaluation of the New_population gi (P (n+1)) Shoot the best and the worst finesses: g _newbest = min [g] g _newworst= max [g] If g _newbest < g _oldbest : Provisory-population = New-population [g_newbest,: ] Else If g _newbest == g _oldbest : Go to the next iteration Else If g_newbest > g _oldbest : Calculate X= exp [-(g _newbest - g _oldbest ) / T] If X > Ω : Provisory_population=New_population [g_newbest , : ] New_population [: , g_ newworst ] = Provisory_population P (n+1) = New_population Reduce the temperature by setting T = 𝛼𝛼 *T Step 6: If the stop criterion is satisfied, terminate de SAGA procedure and return the optimal value found, otherwise, go to the step 5
Algorithm 2. The pseudo-code of the SAGA approach
130
International Journal of Engineering Research in Africa Vol. 36
3. Taguchi Experimental Design The forcefulness and the efficacy of optimization approaches substantially depends on the adequate selection of parameters. In fact, concerning the genetic algorithms method, four parameters including population size, maximum number of iterations, crossover probability and mutation probability have to be neatened as design factors. Furthermore, to handle correctly the proposed SAGA strategy, two other factors are added, namely, the cool rate and the temperature parameter. As a result, to select the best combination of the afore-mentioned parameters, the Taguchi method is used in this section. The overall aim of this method is to explore the factors that influence the mean and the variance of the process performance. Based on the design of experiment, Taguchi method minimizes the impact of noise factor that cannot be controlled by designers and find the best level of the powerful controllable factors. To measure the amount of variation existing in the response variable, an indicator of variation is used (signal-to-noise ratio (S/N)). Usually, there are three classes of the performance characteristics to investigate the S/N ratio, nominal-the-best, larger-the-better, and smaller-the-better. As the present study revolves around a minimization problem, the smaller the fitness value the better. Thus, the S/N ration have to be measured using the smaller-the-better, as stated below: 𝑺𝑺
𝑵𝑵
1
= −10 log10 � ∑𝑛𝑛𝑖𝑖=1 𝑛𝑛
1
𝑦𝑦𝑖𝑖2
(10)
�
Where, 𝑛𝑛 and y represent the number of observations and the response variable, respectively. To direct parameter design, three levels have been attributed to each factor as illustrated in table (1-2). Moreover, relying on the orthogonal array distribution, 27 sets of experiment have been executed. Figures (4-5) show the response graphs obtained for the SAGA and the GA approaches. With respect to the SAGA method, the corresponding response graph indicates that the best combination parameter, which determines a robust parameter design for minimizing the makespan, requires a population size of 400, an iteration number of 600, a mutation probability of 0.1, a crossover probability of 0.9, a cool rate of 0.5 and a temperature of 700. Whereas, regarding the calibration of the classical genetic algorithm GA optimization approach, the four above-mentioned considered factors have to be adjusted such as : population size = 400, iteration number =400, mutation probability= 0.1 and the crossover probability =0.4. Table 3 gathers the selected level of each factor for the two-optimization tools. Table 1. Level range design of each factor of genetic algorithms Factors Level 1 Level 2 Level 3
Population size 100 200 400
Iteration number 200 400 600
Crossover probability 40 % 60 % 90 %
Mutation probability 10 % 40 % 70 %
Table 2. Level range design of each factor of simulated annealing genetic algorithm Factors
Population size
Level 1 Level 2 Level 3
100 200 400
Iteration number 200 400 600
Crossover probability 40 % 60 % 90 %
Mutation probability 10 % 40 % 70 %
cool rate
Temperature
0.3 0.5 0.7
500 700 1000
International Journal of Engineering Research in Africa Vol. 36
131
Figure 4. Mean effect plot for SN ratio of GA parameters
Figure 5. Mean effect plot for SN ratio of SAGA parameters Table 3. Optimal parameters of the proposed optimization approaches Method SAGA GA
Population size 400 400
Iteration number 600 400
Crossover probability 90 % 40 %
Mutation probability 10 % 10 %
cool rate 0.5 -
Temperature 500 -
4. Computational Experiments 4.1 Test problem To evaluate the relative performance of the proposed algorithms, computational test were performed on the commonly known of Taillard’s benchmark suite [25]. The benchmark consists of 120 flowshop problems of different sizes, ranging from 20 jobs and 5 machines to 500 jobs and 20 machines, and each problem includes 10 instances. The proposed SAGA and GA approaches are written in python 3.5 and implemented in a computer microprocessor core i3, with 4,0 GB of RAM memory. The proposed algorithms are compared to other forcful existing metaheuristics in the literature, namely, IABC( han et al. 2012) [17]), DE-ABC( han et al. 2015) [18], IG( Ribas et al.2011) (16], and MFFO ( han et al. 2016) [20]. For each instance, every algorithm is run five independent replications, to assess the performance of each algorithm, a measure of quality denoted as ARDP is adopted. This measure computes the average percentage relative difference between the best makespan retained after the run of each method with the referenced one obtained by the branch and bound of (Ronconi, 2005). The ARDP of each algorithm is formulated as follows: 𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝑖𝑖 =
1
50
𝟓𝟓 ∑10 𝑙𝑙=1 ∑𝒌𝒌=𝟏𝟏
𝒊𝒊 𝑪𝑪𝑹𝑹 𝒍𝒍 −𝑪𝑪𝒍𝒍,𝒌𝒌
𝑪𝑪𝑹𝑹 𝒍𝒍
∗ 100%
(11)
𝑖𝑖 Where, 𝐶𝐶𝑙𝑙,𝑘𝑘 and 𝐶𝐶𝑙𝑙𝑅𝑅 denote the makespan of the lth instance provided by the ith algorithm in the kth run and the referenced one got by Ronconi, respectively. Evidently, the greater the value of 𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝑖𝑖 , the best the result provided by the algorithm.
132
International Journal of Engineering Research in Africa Vol. 36
4.2 Comparison of stat-of-the art algorithms Table 4 and figure 5 depict the comparisons results of ARDP obtained by IABC, DE-ABC, IG, MFFO, GA and SAGA algorithms taking into account the same experimental design. Concerning genetic algorithms method, it can be seen from Table 4 and figure 5 that for instances of 20*5, 20*20, 50*5 and 50*10, the GA is lightly superior to the compared algorithms, for the instance,20*10, the total average of ARDP obtained by GA is equal to that obtained by IABC algorithm. Conversely, for the remaining instances it is inferior to the five compared instances. As follows from the table and the figure shown below (4,6) the total average of ARDP, 4, 45 obtained by the proposed SAGA algorithm is larger than the respective values of 4.10, 4.24, 4.23, 4.38 and 4.26 yielded by IABC, DE-ABC, IG, MMFO and GA, respectively. It is clear that the proposed SAGA algorithm exceed the IABC, DE-ABC, IG, MFFP and GA algorithms through most of the 12 blocking flowshop scheduling problem sizes. To profoundly asses the effectiveness of the proposed SAGA algorithm, the convergence curves of different algorithms are analyzed as shown in figure 7. To outlines the convergence curves of the best makespan values produced by IG, MMFO, GA and SAGA algorithms for the selected instance Ta109 of Taillard’s benchmarks suite. It can be observed from the figure that, as the computation time increases, the convergence curve of the proposed SAGA achieves the undermost level compared to the state-of-the-art algorithms. In the final analysis, the computational results demonstrate the efficiency and the performance of the proposed SAGA approach for solving the 𝐹𝐹𝑚𝑚 |𝑏𝑏𝑏𝑏𝑏𝑏𝑏𝑏𝑏𝑏|𝐶𝐶𝑚𝑚𝑚𝑚𝑚𝑚 problem. The principal reason for which the SAGA can be more performant than the traditional GA and the other powerful metaheuristic may be owing to the combination of the benefits of GA in the exploration ability and the cooling process of SA algorithm in escaping the local optimum in which, the GA may be trapped. 4.3 New upper bounds To get more perspicacity concerning the performance of the proposed SAGA algorithm, its conduct was investigated on the 120 instances introduced by Taillard for solving the blocking flowshop scheduling problem. Table (5-6) summarize the makespan values yielded by B&B developed by Ronconi, MMFO presented by Han2016 and those obtained by the proposed GA and SAGA algorithms. It can be clearly noted that in most of the Taillard instances presented above, the suggested SAGA provides better upper bounds (highlighted in bold) than those of all the other compared algorithms. 4.5
ARDP
4.4 4.3 4.2 4.1 4 3.9
IABC
IG
DE-ABC
MFFO
GA
SAGA
Methods
Figure 6. Benchmark results of the total avearge ARPD indicator
International Journal of Engineering Research in Africa Vol. 36
133
Table 4. Comparison results of the ARPD measure
Makespan
Problems Instance 20*5 20*10 20*20 50*5 50*10 50*20 100*5 100*10 100*20 200*10 200*20 500*20 Average 16000 15800 15600 15400 15200 15000 14800 14600
IABC 0.41 2.38 3.30 4.68 6.15 6.29 2.31 5.97 5.39 4.14 4.46 3.7 4.10
DE-ABC 0.34 2.34 3.29 4.89 6.25 6.45 2.67 6.27 5.53 4.37 4.7 3.72 4.24
IG 0.43 2.37 3.30 4.91 6.33 6.38 2.89 6.10 5.48 4.25 4.7 3.7 4.23
MFFO 0.36 2.38 3.28 4.94 6.23 6.33 3.27 6.63 5.79 4.68 4.88 3.77 4.38
GA 0.4 2.39 3.30 4.96 6.26 6.3 3.01 6.07 5.61 4.57 4.7 3.6 4.26
SAGA 0.48 2.41 3.31 4.94 6.42 6.54 3.26 6.71 5.83 4.81 4.97 3.72 4.45
MMFO IG SAGA GA
0
20
40
60
Generation
80
100
Figure 7. Convergence curves of instance Ta109 Table 5. Comparison of best solutions for Taillard’s benchmark (new solutions are in bold) Instance Ta001 Ta002 Ta003 Ta004 Ta005 Ta006 Ta007 Ta008 Ta009 Ta010 Ta011 Ta012 Ta013 Ta014 Ta015 Ta016 Ta017 Ta018 Ta019 Ta020 Ta021 Ta022 Ta023 Ta024 Ta025 Ta026 Ta027 Ta028 Ta029 Ta030
B&B 1384 1411 1294 1448 1366 1363 1381 1384 1378 1283 1736 1897 1677 1622 1658 1640 1634 1741 1777 1847 2530 2297 2560 2399 2538 2467 2502 2411 2421 2407
MMFO 1374 1408 1280 1448 1341 1363 1381 1379 1373 1283 1698 1833 1659 1535 1617 1590 1622 1731 1749 1782 2436 2234 2479 2348 2435 2383 2390 2328 2363 2323
SAGA 1374 1408 1280 1448 1341 1363 1381 1379 1373 1283 1698 1833 1659 1535 1617 1590 1622 1731 1749 1782 2436 2234 2465 2348 2435 2372 2390 2328 2363 2323
GA 1374 1408 1280 1448 1341 1363 1381 1379 1373 1283 1698 1833 1659 1535 1617 1590 1622 1731 1749 1782 2436 2234 2479 2348 2435 2383 2390 2328 2363 2323
Instance Ta031 Ta032 Ta033 Ta034 Ta035 Ta036 Ta037 Ta038 Ta039 Ta040 Ta041 Ta042 Ta043 Ta044 Ta045 Ta046 Ta047 Ta048 Ta049 Ta050 Ta051 Ta052 Ta053 Ta054 Ta055 Ta056 Ta057 Ta058 Ta059 Ta060
B&B 3151 3395 3184 3303 3272 3400 3228 2360 3104 3264 3913 3798 3723 3885 3934 3831 3957 3774 3784 3928 4886 4668 4666 4650 4475 4521 4576 4688 4532 4846
MMFO 2997 3189 3007 3120 3159 3161 3013 3054 2908 3116 3638 3487 3481 6665 3628 3611 3681 3563 3532 3624 4500 4276 4266 4344 4268 4280 4308 4310 4310 4415
SAGA 2989 3192 2998 3117 3158 3156 3013 3050 2897 3094 3624 3475 3480 3666 3620 3607 3671 3562 3529 3619 4496 4276 4257 4338 4261 4275 4305 4310 4308 4415
GA 3005 3207 3014 3127 3169 3178 3115 3052 2908 3117 3642 3497 3491 3670 6332 3615 3685 3569 3531 3623 4503 4278 4274 4365 4269 4282 4308 4321 4322 4417
134
International Journal of Engineering Research in Africa Vol. 36
Table 6. Comparison of best solutions for Taillard’s benchmark (new solutions are in bold) Instance
B&B
MMFO
SAGA
GA
Instance
B&B
MMFO
SAGA
GA
Ta061
6455
6130
6135
6141
Ta091
14113
13384
13357
13424
Ta062
6214
5992
5985
6008
Ta092
14127
13294
13280
13311
Ta063
6124
5913
5910
5921
Ta093
14416
13416
13416
13463
Ta064
5976
5710
5702
5723
Ta094
14435
13337
13335
13344
Ta065
6173
5938
5931
5949
Ta095
14113
13340
13311
13354
Ta066
6094
5808
5799
5811
Ta096
13309
13081
13026
13085
Ta067
6262
5959
5941
6002
Ta097
14563
13561
13571
13607
Ta068
6061
5865
5862
5881
Ta098
14329
13504
13502
13644
Ta069
6474
6106
6093
6102
Ta099
13923
13217
13201
13308
Ta070
6366
6112
6105
6141
Ta100
14435
13400
13389
13439
Ta071
7496
6988
6971
6998
Ta101
15579
14192
14889
14908
Ta072
7281
6714
6709
6729
Ta102
15728
14963
14952
14962
Ta073
7400
6834
6820
6892
Ta103
15915
15043
15026
15107
Ta074
7670
7102
7989
7125
Ta104
16039
14949
14930
14987
Ta075
7317
6804
6799
6935
Ta105
15938
14857
14857
14899
Ta076
7301
6615
6607
6639
Ta106
15911
14909
14925
14929
Ta077
7247
6770
6770
6811
Ta107
15898
14980
14972
15009
Ta078
7315
6806
6801
6859
Ta108
16022
15005
15000
15021
Ta079
7631
6997
6964
7055
Ta109
15817
14834
14782
15013
Ta080
7411
6910
6895
6928
Ta110
15969
14954
14921
14954
Ta081
8347
7774
7751
7799
Ta111
38334
35838
36504
36609
Ta082
8372
7838
7883
7892
Ta112
38642
35970
35946
36108
Ta083
8265
7777
7769
7802
Ta113
38163
35757
35615
36778
Ta084
8365
7818
7803
7836
Ta114
38625
36070
36045
36628
Ta085
304
7799
7769
7836
Ta115
38492
35869
35762
35869
Ta086
8450
7841
7870
7851
Ta116
38551
36099
36099
37006
Ta087
8507
7929
7973
7979
Ta117
38179
35760
35760
36805
Ta088
8584
7976
7975
7993
Ta118
38664
35921
35865
36792
Ta089
8341
7894
7832
7889
Ta119
38339
35739
35695
36885
Ta090
8489
7929
7906
7918
Ta120
38540
36120
35971
36184
5. Conclusions This paper suggests SAGA approach for solving the blocking flowshop scheduling problem (BFSP) with the maximum completion time criterion. The proposed algorithm combines the features of GA and the annealing mechanism of SA algorithm. The efficiency of the algorithm is evaluated on the commonly know 120 instances presented by Taillard. Compared to the results found by IABC, IG, DE-ABC, and GA, the computational results show the high-performance of the proposed SAGA approach for solving the𝐹𝐹𝑚𝑚 |𝑏𝑏𝑏𝑏𝑏𝑏𝑏𝑏𝑏𝑏|𝐶𝐶𝑚𝑚𝑚𝑚𝑚𝑚 problem. Moreover, the results illustrates that for 82 out of 120 instances, SAGA has provided best upper bounds than MFFO and B&B methods.
International Journal of Engineering Research in Africa Vol. 36
135
From the outcome of our investigation, it is possible to conclude that, due to the reasonable incorporation, the developed SAGA has a great potential for solving the BFSP, since it avails from the exploration ability of GA and The power to escape the local optimum of SA. As future direction, we intend to extend our research to solve the blocking flowshop scheduling problem with several criteria, as a single criterion is regarded as deficient for real-life and practical applications. Acknowledgements The authors would like to thank the editors and anonymous reviewers for their useful comments and their constructive suggestions, which helped to improve this paper in order to make it appropriate for publication in International Journal of Engineering Research in Africa. References [1]. Grabowski J, Skubalska E, Smutnicki C. On flowshop scheduling with release and due dates to minimize maximum lateness. Journal of Operational Research Society, 34 (1983) 615–20. [2]. Graham RL, Lawler EL, Lenstra JK, Rinnooy Kan AHG. Optimization and approximation in deterministic sequencing and scheduling: a survey. Annals of Discrete Mathematics, 5 (1979) 287–362. [3]. Hall NG, Sriskandarajah CA survey of machine scheduling problems with blocking and nowait in process. Operations Research, 44 (1996) 510–25. [4]. Papadimitriou CH, Kanellakis PC. Flowshop scheduling with limited temporary storage”. Journal of the Association for Computing Machinery, 27 (1980) 533–49. [5]. Allahverdi, A., C. T. Ng, T. E. Cheng, and M. Y. Kovalyov. A Survey of Scheduling Problems with Setup Times or Costs”. European Journal of Operational Research, 187(3) (2008) 985–1032. [6]. McCormick, S. T., M. L. Pinedo, S. Shenker, and B. Wolf. Sequencing in an Assembly Line with Blocking to Minimize Cycle Time. Operations Research, 37 (6) (1989) 925–935. [7]. Leisten, R. Flowshop sequencing problems with limited buffer storage, International Journal of Production Research, 28 (1990)2085–100. [8]. Nawaz, M., Enscore, E.E.J. and Ham, I. A heuristic algorithm for the m-machine, n-job flowshop sequencing problem, OMEGA – International Journal of Management Science, 11(1983) 91–95. [9]. Ronconi, D. P. A note on constructive heuristics for the flowshop problem with blocking. International Journal of Production Economics, 87 (1) (2004) 39-48. [10]. Pan, Q. K., and L. Wang. Effective Heuristics for the Blocking Flowshop Scheduling Problem with Makespan Minimization. Omega, 40 (2) (2012)218–229. [11]. Caraffa, V., S. Ianes, T. P. Bagchi, and C. Sriskandarajah. Minimizing Makespan in a Blocking Flowshop using Genetic Algorithms. International Journal of Production Economics, 70 (2) (2001)101–115. [12]. Abadi INK, Hall NG. Sriskandarajh C. Minimizing cycle time in a blocking flowshop. Operations Research, 48(2000)177–80. [13]. Ronconi, D.P. A branch-and-bound algorithm to minimize the makespan in a flowshop problem with blocking, Annals of Operations Research, 138 (1) (2005) 53–56. [14]. Liu JY, Reeves CR. Constructive and composite heuristic solutions to the P//Sci scheduling problem. European Journal of Operational Research, 132 (2) (2001) 439–52.
136
International Journal of Engineering Research in Africa Vol. 36
[15]. Wang L, Pan QK, Suganthan PN, Wang WH, Wang YM. A novel hybrid discrete differential evolution algorithm for blocking flowshop scheduling problems. Computers and Operations Research, 37 (3) (2010) 509–20. [16]. Ribas I, Companys R, Tort-Martorell X. An iterated greedy algorithm for the flowshop scheduling with blocking. OMEGA-The international Journal of Management Science, 39 (2011) 293–301. [17]. Han, Y. Y., Q. K. Pan, J. Q. Li, and H. Y. Sang. An Improved Artificial Bee Colony Algorithm for the Blocking Flowshop Scheduling Problem. International Journal of Advanced Manufacturing Technology. 60 (2012) 1149–1159. [18]. Han, Y. Y., Gong, D., & Sun, X. A discrete artificial bee colony algorithm incorporating differential evolution for the flow-shop scheduling problem with blocking. Engineering Optimization. 47(7) (2015) 927-946. [19]. Sadaqa, M., & Moraga, R. J. Scheduling Blocking Flowshops Using Meta-RaPS. In Procedia Computer Science. 61 (2015) 533-538. [20]. Han, Y., Gong, D., Li, J., & Zhang, Y. Solving the blocking flowshop scheduling problem with makespan using a modified fruit fly optimisation algorithm. International Journal of Production Research, 54 (22) (2016) 6782-6797. [21]. Mezmaz, M., Melab, N., & Talbi, E. G. Combining metaheuristics and exact methods for solving exactly multi-objective problems on the grid. Journal of Mathematical Modelling and Algorithms in Operations Research, 6(3) (2007). [22]. Bianchi, L., Birattari, M., Chiarandini, M., Manfrin, M., Mastrolilli, M., Paquete, L. & Schiavinotto, THybrid metaheuristics for the vehicle routing problem with stochastic demands. Journal of Mathematical Modelling and Algorithms, 5 (1) (2006)91-110. [23]. Blum, C., Puchinger, J., Raidl, G. R., & Roli, A. Hybrid metaheuristics in combinatorial optimization: A survey. Applied Soft Computing, 11 (6) (2011) 4135-4151. [24]. Ronconi, D. P., and V. A. Armentano. Lower Bounding Schemes for Flowshops with Blocking in-Process. Journal of the Operational Research Society, 52 (2001) 1289–1297. [25]. Taillard, E. Benchmarks for Basic Scheduling Problems. European Journal of Operational Research, 64 (1993) 278–285.