An alternative hybrid evolutionary technique focused on allocating machines and sequencing operations

Article history: Received November 4 2015 Received in Revised Format April 6 2016 Accepted April 7 2016 Available online April 7 2016 We present here a hybrid algorithm for the Flexible Job-Shop Scheduling Problem (FJSSP). This problem involves the optimal use of resources in a flexible production environment in which each operation can be carried out by more than a single machine. Our algorithm allocates, in a first step, the machines to operations and in a second stage it sequences them by integrating a Multi-Objective Evolutionary Algorithm (MOEA) and a path-dependent search algorithm (Multi-Objective Simulated Annealing), which is enacted at the genetic phase of the procedure. The joint interaction of those two components yields a very efficient procedure for solving the FJSSP. An important step in the development of the algorithm was the selection of the right MOEA. Candidates were tested on problems of low, medium and high complexity. Further analyses showed the relevance of the search algorithm in the hybrid structure. Finally, comparisons with other algorithms in the literature indicate that the performance of our alternative is good. © 2016 Growing Science Ltd. All rights reserved


Introduction
The design of production plans involves decisions on the allocation of limited resources in order to optimize efficiency-related short-term objectives (Bihlmaier et al., 2009;Nowicki & Smutnicki, 2005;Armentano & Scrich, 2000).The framework of analysis of this kind of problems is the Job-Shop Scheduling Problem (JSSP) (Agnetis et al., 2001;Lin et al., 2011;Heckman & Beck, 2011;Nazarathy & Weiss, 2010), which assumes a class of jobs, consisting of ordered sequences of operations that have to be distributed over several machines.One of the goals is to minimize the makespan, i.e. the total processing time of the jobs (Heinonen & Pettersson, 2007;Chao-Hsien & Han-Chiang, 2009;Della Croce et al., 2014).Flexible JSSP (FJSSP) generalizes this problem.It assumes that operations can be performed on different machines.Thus, it involves the decision of the allocation of operations on machines, a NP-Hard problem (Ullman, 1975;Papadimitriou, 1994).While most of the literature on this problem focuses on its single objective versions, some authors state that several objectives have to be optimized as to achieve an efficient production process (Chinyao & Yuling, 2009).Based on the latter motivation we present here an algorithm for a multi-objective version of FJSSP.On the grounds of a preliminary analysis of the problem we decided to base our alternative on an evolutionary approach (Goldberg, 1989;Pezzella et al., 2008).One of the main advantages of this strategy is that evolutionary algorithms can be easily adapted to the problem at hand.They are quite efficient in handling single-objective problems, but their high rate of convergence hampers their usefulness on multi-objective versions.This is because their fast convergence leads to a loss of diversity, indicated by poorly distributed Pareto frontiers.A multi-objective algorithm based on an underlying evolutionary component should be, therefore, complemented by an efficient search procedure in order to diversify solutions with a few rounds of evaluation of the fitness functions.This is the approach we followed in this paper, providing a methodological ground for the design of such a hybrid algorithm, as introduced in Frutos et al. (2010) and Frutos and Tohmé (2015).We present here a Multi-Objective Evolutionary Algorithm (MOEA) (Coello et al., 2006) joined by a local search procedure (MOSA, Multi-Objective Simulated Annealing) to solve a FJSSP (Hansmann et al., 2014;Tsai & Lin, 2003;Wu et al., 2004;Nidhiry & Saravanan, 2012).From now on, we will call this hybrid structure a Multi-Objective Hybrid Evolutionary Algorithm or MOHEA.The rest of the paper is organized as follows.Section 1.1 discusses some of the literature on the FJSSP while section 1.2 introduces the formal description of multiple-objective optimization problems.Section 2 presents our formal characterization of FJSSP while in section 3 the MOHEA for this framework is introduced.Results of running the algorithm are shown in section 4, and in section 5 we present the conclusions.

Approaches to the FJSSP
The FJSSP is particularly hard to solve.It has been analyzed for 1, 2 and 3 machines and some arbitrary number of jobs.Very few developments have been devoted to the case of 4 or more machines for at least 3 jobs, due to the combinatorial explosion of feasible sequences.A brief survey of the literature on the problem shows that, (Brandimarte, 1993) proposed a hierarchical approach, distinguishing the allocation from the sequencing subproblem, where the former is handled as a routing problem while the latter is seen as a Job-Shop one.(Mesghouni et al., 1997), instead, attacked the problem with genetic algorithms.On the other hand, (Kacem et al., 2002) proposed a localization approach to the control of the genetic algorithm, yielding good solutions and minimizing the makespan and the workload of the machines.Tay and Wibowo (2004) and Ho and Tay (2005) introduced dispatch rules to generate populations and a scheme structure under which each generation explores the search space.They also studied the representation of solutions in order to achieve a more efficient makespan.This approach was generalized in (Ho et al., 2007), in which the evolutionary algorithm is complemented by learning under schemes and composite dispatch rules.(Fattahi et al., 2007) proposed a hierarchical approach to the FJSSP in which the allocation sub-problem is solved by a Taboo Search treatment, while the sequencing one is handled by Simulated Annealing.This approach was tested on twenty instances of the FJSSP, although only the simplest ones got solved.(Zhang & Gen, 2005) presents a genetic algorithm working on multiple scenarios, in which each one corresponds to an operation and each feasible machine to a state.(Pezzella et al., 2008), also used a genetic algorithm aided by Kacem's et al. (2002) localization approach.This allows for intelligent mutations that reassign operations from heavy loaded to less loaded machines.(Yazdani et al., 2010) proposed the minimization of makespan by handling, in parallel, variable neighbourhoods while (Yang et al., 2010) solved the FJSSP with an improved constraint satisfaction adaptive neural network.Recent approaches involve hybridizations with an artificial bee colony algorithm (Li et al., 2011) or a shuffled frog-leaping algorithm (Li et al., 2012).On the other hand, the local search procedure included has been defined on the critical path (Xiong et al., 2012) or performed hierarchically over the different objectives (Yuan & Xu, 2015).

Multi-Objective Optimization: Basic Concepts
Let us assume that several goals (objectives) have to be minimized.Thus, a vector . The main goal of Multi-Objective Optimization is to find the corresponding * FP .A good approximation should yield a few feasible candidates; close enough to the frontier (Frutos & Tohmé, 2009).

The flexible job-shop scheduling problem
The FJSSP is defined in terms of m machines, , and a class of n independent jobs, n j j 1 J {J }   .
Each job Jj amounts to a set of sequenced operations, operation will be allocated only once.From the many possible objectives that can be pursued in this setting we choose the minimization of the total processing time, (makespan) Eq. ( 1), and the minimization of the total operational cost given by Eq. ( 2). S .As we will see the two objectives are in conflict, which makes this problem interesting.

A multi-objective hybrid evolutionary algorithm
Evolutionary algorithms imitate genetic processes, improving solutions by breeding new solutions up from older ones.The solutions are represented as a fixed number of chromosomes, composed by smaller units called genes.They codify the hereditary features of an individual (solution).In the case of sequencing problems the chromosomes indicate the programming of jobs.It is assumed that among all possible chromosomes one codifies the optimal sequence.To show how this works we will consider the case of instance MF01 analyzed in (Frutos et al., 2010), with three jobs and four machines (3×4).The first and second jobs require three operations each while the third one requires only two.This amounts to eight operations with processing times and operational costs shown in Table 1.Solutions have to be codified in terms of the characteristics of the problem, respecting its constraints.We will use two chromosomes for each solution.The first one determines the solution of the allocation sub-problem while the second the solution of the sequencing of operations sub-problem.The size of the allocation chromosomes is the number of total operations in the problem.The size of the sequencing chromosomes is the number of machines in M. In the allocation chromosome each gene is: 0→M1, 1→M2, 2→M3, 3→M4 (Third column, rows 3 to 10 in Table 2).For the sequencing chromosome each gene is: 0→1│2│3, 1→1│3│2, 2→2│1│3, 3→2│3│1, 4→3│1│2 and 5→3│2│1 (Third row, columns 4 to 7 in Table 2).This means that, in the first chromosome, gene 0 corresponds to an allocation to machine M1; if it is 1 to machine M2 and so on.In the second chromosome, if the gene is 0 the sequence of jobs is 1, 2 and 3; if it is 1 the sequence is 1, 3 y 2, and so on.A crucial role in the algorithm is played by the chromosome decoding procedure, since it is in charge of the interpretation and representation of the solutions of the FJSSP.This procedure is the most computationally costly in the algorithm.It runs by fetching the information in the genes and looking up in the times and costs table.After this, the objectives are evaluated on the resulting solution candidate (see Figure 1).The initial population is randomly generated: the allocation chromosome is obtained by assigning at random values between 0 and m-1 (between 0 and 3 in our example) to the genes.On the other hand, the sequencing chromosome is generated by the random assignation of values between 0 and n!-1 (between 0 and 5 in the example) to its genes.Then, the crossover and mutation genetic operators are applied over the candidates.Crossover exchanges segments of chromosomes between pairs of candidates, making the new candidates carry a mixture of the information of its 'parents'.Mutation introduces random feasible changes in the chromosomes, as to yield new solution candidates and allowing the exploration of different areas of the search space.The crossover operator is applied on chromosomes of the same type, exchanging segments of the allocation chromosome and segments of the sequencing chromosome among candidates.In our algorithm we use uniform crossover, which exchanges the genes in the same position in the chromosomes of two candidates, breeding two new solution candidates.The version of the mutation operator used here swaps two genes in the same chromosome (Two-swap).On the allocation chromosome it operates on a 20 % of the genes, while on the sequencing chromosome it acts only over the 10 %.After a careful examination of local search algorithms we selected Simulated Annealing as our 'improves' operator.The main components of this algorithm are the way in which the neighborhood of solutions is generated and the probability function ( T )  e   , in particular the value of δ.The general structure of our Multi-Objective Simulated Annealing is the same as in (Frutos et al., 2010).The neighborhood is generated by taking both chromosomes in a candidate and altering at random the value of some genes (with the same percentages as the mutation operator).This induces a change of the machines on which operations are run and at the same time a change in their ordering.The determination of δ, somehow lacking in the multi-objective literature is as follows: et al., 2010) have shown that in appropriate settings this specification yields very good solutions.As for those settings, they are given by the following parameters: Ti (initial temperature), Tf (final temperature), the cooling function ( k 1 k T T    , where α is the rate of cooling and k the iteration step), and the number of iterations M ( , where  is a control parameter).The hybrid structure was designed in the PISA (A Platform and Programming Language Independent Interface for Search Algorithms) environment (Frutos & Tohmé, 2009) (see Fig. 2).

Implementation and design of experiments
Any instance of the FJSSP is specified by the number of jobs, the number of machines and the total number of operations to be performed.For our experiments we use the problems reported in (Frutos et al., 2010): MF01 (3 jobs × 4 machines with 8 operations), MF02 (4 × 5 with 12 operations), MF03 (10 × 7 with 29 operations), MF04 (10 × 10 with 30 operations) and MF05 (15 × 10 with 56 operations).We compared the global performance of the evolutionary stage by exchanging three selectors (see Fig. 2): Nondominated Sorting Genetic Algorithm II (NSGAII) (Deb et al., 2002) (with a O(g×N 2 ) complexity, for a population of size 2N), Strength Pareto Evolutionary Algorithm II (SPEAII) (Zitzler et al., 2002) (its complexity is O((N+N')2*log(N+N')), where N is the size of the population and N' the size of the file that stores the results) and Indicator-Based Evolutionary Algorithm (IBEA) (Zitzler & Künzli, 2004) (of O(N 2 ) complexity).A preliminary analysis of the improvement process showed that it tended to become stable at the 200th generation.We chose then the limit of 250 generations, just to leave room for any later improvement.The parameters and characteristics of the computing equipment used during these experiments were as follows: size of the population: 200, type of cross-over: uniform; probability of crossover: 0.90, type of mutation: two-swap, probability of mutation: 0.01, type of local search: simulated annealing (Ti: 850, Tf: 0.01, α: 0.95, ω: 10), probability of local search: 0.01, CPU: 3.00 GHZ and RAM: 1.00 GB.In the case of the IBEA selector we chose the additive epsilon index.We run each algorithm 30 times and the undominated solutions were picked up.

Selection of MOEA
As a first instance, we compare the performance of the MOEAs using the following metrics: the multiplicative version of the Unary Epsilon Index (Ie), the Hypervolume Index (IH) and the R2 index (IR2).These indexes allow distinguishing differences among the approximations when the dominance rankings yield very similar results (Knowles et al., 2005).For the parameters of the indexes we kept the initial specifications of PISA.Unary indexes were obtained using normalized approximation sets (Zitzler & Künzli, 2004) and the reference class generated by PISA.A non-parametric test was run on the approximation sets in order to obtain valid conclusions on the quality of the optimization methods.In our analysis we used Fisher' test (Knowles et al., 2005) with a confidence level α = 0.05.Tables 3, 4 and 6 indicate that problems MF01, MF02 and MF04 yield no significantly different results.Tables 5 and 7, instead, show significant differences in the results of MF03 and MF05 between IBEA and the other two algorithms, SPEAII and NSGAII.Given these results, we have to note that an assessment based on a relatively small number of runs requires further analysis in order to ensure its robustness.We have to see that the results reported in the previous subsections are statistically significant.We proceed as follows.For each algorithm we take the final outcome on each problem.This is written as a vector.Then we take a component by component distance to the vector of solutions (of the same dimensionality).Then we postulate different hypotheses, one the null hypothesis (that the algorithms do not yield differences) and alternative ones, indicating differences among the algorithms.To see this, we start introducing a distance to the frontier variable, which is basically a variant of a taxicab metric.Let us remark that the results we present below are robust under changes in the underlying metric: the same analysis based on the Euclidean and the supremum metric yield analogous results.The distance of the frontier is obtained as the addition of the distances to their corresponding values in the frontier of the actually obtained values f1 and f2: , where x,i d is the distance yield by algorithm x on observation i.The values of i refer to the observations (i=1,…,n), 1,i f and 2,i f to values on the frontier and * 1,i f , * 2,i f to the actual output of the algorithm on i.Notice that this distance (unlike the taxicab one) is negative.In case an algorithm does not reach a solution, we assign the maximal distance found for the other algorithms on that observation.Table 8 shows the P-values of differences in means test for the different algorithms.As indicated, the null hypothesis is the means are equal.All the results are significant, meaning, in particular, that the number of cases considered were enough to make the assessment.
It can be seen that IBEA is more efficient (in the sense that the distance to the frontier is much shorter, indicated by Pr(T>t)).On the other hand, there are no significant differences between SPEAII and NSGAII.

MOHEA vs. MOEA: Why Hybrid?
Now we report the results of experiments comparing the MOEA (IBEA in our case) and the MOEA complemented with a search process (IBEA + Simulated Annealing).In Figures 3,4,5,6 and 7 (Left) we show the Pareto frontiers for both algorithms.The MOEA yields an incomplete frontier, which, moreover is sometimes dominated by the frontier obtained by the MOHEA.Furthermore, the latter exhibits a better distribution of solutions.In Figs. 3,4,5,6 and Fig. 7 (Right) show the mean number of undominated solutions (S) found by the MOHEA and the MOEA, for different generation numbers (G).
It can be seen that for the 250 generations run by the two algorithms, there exists a clear difference between them.Other experiments, not reported here, indicated that the MOEA reached the undominated solutions for these problems around generation 500.Putting this together with the number of evaluations in the search process we conclude that for similar results, the MOHEA on average required 35,2 % less evaluations than MOEA.We compared also the results under our MOHEA with those obtained by the Hybrid Artificial Bee Colony Algorithm (HABC) introduced by Li et al., (2011) and by the Multi-Population Interactive Coevolutionary Algorithm (MPICA) presented in (Xing et al., 2011).These two algorithms were implemented in C++.The parameters for them were taken from the publications in which they have been presented.They were also run 30 times each and the outcomes were evaluated according to the same metrics used in the choice of the selector.Fisher's test was used again with a confidence level α = 0.05.Tables 9, 10 and 11 show no significant differences for MF01, MF02 and MF03 under the different algorithms and indexes.Tables 12 and 13 show significant differences in problems MF04 and MF05, between MOHEA and MPICA over HABC.As in the previous section we run a robustness analysis, again by the same procedure.The results are reported in Table 14, which shows that MOHEA yields better results than MPICA and HABC.In turn, these last two algorithms do not exhibit significant differences.Finally, we also compare the mean running times of the three algorithms (see Table 15).We see that HABC runs faster that MOHEA and MPICA, but as seen above, its solutions are not as good as those of MOHEA.On the other hand MOHEA ran faster than MPICA in four out of five cases.

Conclusion
We presented a Multi-Objective Hybrid Evolutionary Algorithm (MOHEA) for the Flexible Job-Shop Scheduling Problem (FJSSP).Our algorithm integrates two meta-heuristic procedures: a Multi-Objective Evolutionary Algorithm (MOEA) and a Multi-Objective Simulated Annealing (MOSA) algorithm.Individuals are coded in a way that facilitates the application of two basic genetic operators.Different MOEAs were tested for the first component of the MOHEA.IBEA showed to perform better than NSGAII and SPEAII.We also compared MOHEA with the Hybrid Artificial Bee Colony Algorithm (HABC) and the Multi-Population Interactive Coevolutionary Algorithm (MPICA).The running time performances of MOHEA and MPICA were better than that of HABC.Only in one case MPICA improved over MOHEA.We can conclude that our hybrid structure intended to provide solutions to the FJSSP obtains good solutions at a reasonable time.
Técnicas (CONICET) and the Universidad Nacional del Sur (UNS) for Grant PGI 24/ZJ34.We want also thank Dr. Ana C. Olivera for her constant support and help during this research.

Table 1 A
Flexible Job-Shop Scheduling Problem

Table 8 P
-values in the difference of means test between IBEA, NSGAII and SPEAII.

Table 14 P
-values in the difference of means test between MOHEA, HABC and MPICA

Table 15
Mean Running Time for MOHEA, HABC and MPICA