Int. J.

Production Economics 145 (2013) 387–396

Contents lists available at SciVerse ScienceDirect

Int. J. Production Economics

journal homepage: www.elsevier.com/locate/ijpe

An effective estimation of distribution algorithm for solving

the distributed permutation flow-shop scheduling problem
Sheng-yao Wang n, Ling Wang, Min Liu, Ye Xu
Department of Automation, Tsinghua National Laboratory for Information Science and Technology (TNList), Tsinghua University, Beijing 10084, China

art ic l e i nf o a b s t r a c t

Article history: In this paper, an effective estimation of distribution algorithm (EDA) is proposed to solve the distributed
Received 13 June 2012 permutation flow-shop scheduling problem (DPFSP). First, the earliest completion factory rule is
Accepted 3 May 2013 employed for the permutation based encoding to generate feasible schedules and calculate the schedule
Available online 14 May 2013
objective value. Then, a probability model is built for describing the probability distribution of the
Keywords: solution space, and a mechanism is provided to update the probability model with superior individuals.
Distributed permutation flow-shop By sampling the probability model, new individuals can be generated among the promising search
scheduling region. Moreover, to enhance the local exploitation, some local search operators are designed based on
Estimation of distribution algorithm the problem characteristics and utilized for the promising individuals. In addition, the influence of
Probability model
parameter setting of the EDA is investigated based on the Taguchi method of design of experiments, and
Design of experiment
a suitable parameter setting is suggested. Finally, numerical simulations based on 420 small-sized
instances and 720 large-sized instances are carried out. The comparative results with some existing
algorithms demonstrate the effectiveness of the proposed EDA in solving the DPFSP. In addition, the new
best-known solutions for 17 out of 420 small instances and 589 out of 720 large instances are found.
& 2013 Elsevier B.V. All rights reserved.

1. Introduction achieve higher product quality, lower production costs and lower
management risks (Kahn et al., 2004).
The permutation flow-shop scheduling problem (PFSP) has Scheduling in distributed systems is more difficult than the
been concentrated on by many researchers due to its wide classical shop scheduling, because it should determine the assign-
applications in economics and industrial engineering (Hejazi and ment of jobs to factories as well as the processing sequence in each
Saghafian, 2005). The PFSP has been proved to be NP-complete factory. Obviously, both sub-problems are related to each other
when the number of machines is more than three (Garey et al., and cannot be solved sequentially if high performance is desired
1976). After the pioneering work of Johnson (1954), much research (Naderi and Ruiz, 2010). Compared to the classical shop schedul-
work has been carried out on the PFSP (Cheng and Janiak, 2000; ing, the literature on the distributed scheduling is relatively
Suliman, 2000; Chung et al., 2002; Cheng and Kovalyov, 2003; limited and the study on this topic is in its infancy. Jia et al.
Cheng et al., 2004, 2013; Ruiz and Maroto, 2005; Lin et al., 2008; (2002, 2003) studied the distributed job shop problem under
Tseng and Lin, 2010a, 2010b; Sun et al., 2012; Shabtay et al., 2013; different criteria and employed a standard genetic algorithm (GA)
Wang et al., 2013c, 2013d). In most research of the PFSP, a common to solve the problem. Later, Jia et al. (2007) refined the previous GA
assumption is that there is only one production center or factory, to solve the small-sized and medium-sized distributed scheduling
which means that all jobs are assumed to be processed in the same problems. Chan et al. (2005, 2006) proposed an adaptive GA to
factory. Nevertheless, with the development of the business solve the distributed job shops with makespan criterion for larger
concept, coproduction between companies is more and more problems. De Giovanni and Pezzella (2010) proposed an improved
common nowadays (Wang and Shen, 2007). Besides, multi-plant GA to solve the distributed and flexible job-shop scheduling
companies and supply chains are taking a more important role in problem. As for the distributed permutation flow-shop scheduling
practice (Moon et al., 2002). Therefore, the distributed manufac- problem (DPFSP), Naderi and Ruiz (2010) presented six mixed
turing strategy comes into being, which enables companies to integer linear programming models and developed two factory
assignment rules and 14 heuristics based on dispatching rules,
effective constructive heuristics and variable neighborhood des-
n
Correspondence to: Room 427A, Main Building, Tsinghua University,
cent methods. Besides, to evaluate the proposed models and
Beijing 100084, China. Tel.: +86 10 62783125; fax: +86 10 62786911. algorithms, the authors generated 420 small-sized instances and
E-mail address: [email protected] (S.-y. Wang). 720 large-sized instances which are available at http://soa.iti.es,

0925-5273/$ - see front matter & 2013 Elsevier B.V. All rights reserved.
http://dx.doi.org/10.1016/j.ijpe.2013.05.004
388 S.-y. Wang et al. / Int. J. Production Economics 145 (2013) 387–396

along with the best known solution for each instance. Based on factory contains the same set of m machines M ¼ fM 1 ; M 2 ; …; M m g.
these instances, Gao et al. (2012) proposed a tabu search algorithm A job J i is formed by a sequence of m operations fOi;1 ; Oi;2 ; …; Oi;m g
for solving the DPFSP and tested the performance of the proposed to be performed one after another, where the execution of Oi;j
algorithm. However, the authors centred their study only on a part requires machine Mj and processing time t i;j 4 0. When a job is
of the instances and no results were listed for direct comparisons. assigned to a certain factory, it cannot be transferred to another
Estimation of distribution algorithm (EDA) is a relatively novel factory and all its operations can only be processed in the factory.
population-based optimization algorithm, which has led to increas- Besides, the following assumptions for the classical flow-shop
ing studies and wide applications during recent years (Larranaga and scheduling are adopted. All jobs are independent and available for
Lozano, 2002). Considering different kinds of the relationships processing at time 0. Each machine can process only one job at a
between variables, the EDA has different complexity of the model. time and each job can be processed on only one machine at a time.
Accordingly, the EDA can be classified as a univariate model, bivariate Preemption is not allowed, i.e., each operation must be completed
model or multivariate model. Univariate models assume that the without interruption once it is started. Setup times of machines
variables are independent of each other, e.g., the population-based and move times between operations are negligible. The DPFSP is
incremental learning (Baluja, 1994), the univariate marginal distribu- to determine both the assignment of jobs to the factories and
tion algorithm (Mühlenbein and Paass, 1996) and the compact GA the sequences of jobs in all the factories to minimize a certain
(Harik et al., 1998). Bivariate models assume that each variable is scheduling objective function. In this paper, we consider the
associated with another one, e.g., the mutual information maximiza- maximum completion time (makespan) as the criterion.
tion for input clustering (De Bonet et al., 1997), the combining Let λk ¼ ½λk ð1Þ; λk ð2Þ; ⋯; λk ðnk Þ be the sequence of the jobs in
optimizers with mutual information trees (Baluja and Davies, 1997) factory k, where nk is the total number of the jobs assigned to
and the bivariate marginal distribution algorithm (Pelikan and factory k. C i;j is denoted as the completion time of Oi;j . For a
Mühlenbein, 1999). Multivariate models consider the relationship schedule Λ of the DPFSP, i.e., a set of sequences fλ1 ; λ2 ; ⋯; λF g, we
between all the variables, e.g., the factorized distribution algorithms can calculate the makespan C max as follows:
(Mühlenbein and Mahnig, 1999), the extended compact GA (Harik,
1999) and the Bayesian optimization algorithm (Pelikan et al., 1999). C λk ð1Þ;1 ¼ t λk ð1Þ;1 ; k ¼ 1; 2; ⋯; F ð1Þ
For more details about the EDA, please refer to Larranaga and Lozano
(2002). C λk ðiÞ;1 ¼ C λk ði−1Þ;1 þ t λk ðiÞ;1 ; k ¼ 1; 2; ⋯; F; i ¼ 2; 3; ⋯; nk ð2Þ
So far, the EDA-based algorithms have been applied to a variety
of academic and application problems, such as feature selection
(Saeys et al., 2003), inexact graph matching (Cesar et al., 2005), C λk ð1Þ;j ¼ C λk ð1Þ;j−1 þ t λk ð1Þ;j ; k ¼ 1; 2; ⋯; F; j ¼ 2; 3; ⋯; m ð3Þ
software testing (Sagarna and Lozano, 2005), single machine
scheduling (Chen and Chen, 2013) flow-shop scheduling (Jarboui C λk ðiÞ;j ¼ maxfC λk ði−1Þ;j ; C λk ðiÞ;j−1 g þ t λk ðiÞ;j ;
et al., 2009), resource-constrained project scheduling (Wang and
Fang, 2012), multi-dimensional knapsack problem (Wang et al., k ¼ 1; 2; ⋯; F; i ¼ 2; 3; ⋯; nk ; j ¼ 2; 3; ⋯; m ð4Þ
2012a), flexible job-shop scheduling (Wang et al., 2012b, 2013a,
2013b), and so on. However, to the best of our knowledge, there is
C max ¼ maxC nk ;m ; k ¼ 1; 2; ⋯; F ð5Þ
no research work about the EDA for solving DPFSP. In this paper,
we will propose an effective EDA to solve the DPFSP with the The objective of solving the DPFSP is to find a schedule with the
criterion to minimize the maximum completion time. Specifically, minimum makespan.
the earliest completion factory rule is employed for the permuta-
tion based encoding to generate feasible schedules and calculate
the schedule objective value. Meanwhile, a probability model is
built with the superior individuals for generating new individuals, 3. Estimation of distribution algorithm
and a mechanism is provided to update the probability model.
Besides, some local search operators are designed based on the As a relatively new paradigm in the field of evolutionary
problem characteristics and utilized to enhance the exploitation computation, estimation of distribution algorithm employs expli-
capability. In addition, the influence of parameters is investigated cit probability distributions in optimization (Larranaga and
based on Taguchi method of design of experiment, and a suitable Lozano, 2002). Compared with the GA, the EDA reproduces new
parameter setting is suggested. Finally, we use the benchmark population implicitly instead of the crossover and mutation
instances generated by Naderi and Ruiz (2010) to test the perfor- operators. In the EDA, a probability model of the most promising
mances of the EDA and to compare it with some existing methods area is built by statistical information based on the search
to solve the DPFSP. experience, and then the probability model is used for sampling
The remainder of the paper is organized as follows: In Section to generate the new individuals. Meanwhile, the probability model
2, the DPFSP is described. In Section 3, the basic EDA is introduced is updated in each generation with the potential individuals of the
briefly. Then, the framework of the EDA for solving the DPFSP is new population. In such an iterative way, the population evolves,
proposed in Section 4. The influence of parameter setting is and finally satisfactory solutions can be obtained.
investigated based on design of experiment testing in Section 5, The general framework of the EDA is illustrated in Fig. 1.
and computational results and comparisons are provided as well. The critical step of the above procedure is to estimate the
Finally we end the paper with some conclusions and future work probability distribution. The EDA makes use of the probability
in Section 6. model to describe the distribution of the solution space. The
updating process reflects the evolutionary trend of the population.
Due to the difference of problem types, a proper probability model
2. Distributed permutation flow-shop scheduling problem and a suitable updating mechanism should be well developed to
estimate the underlying probability distribution. Nevertheless, the
The distributed permutation flow-shop scheduling problem EDA pays more attention to global exploration while its exploita-
(Naderi and Ruiz, 2010) can be described as follows. There are n tion capability is relatively limited. So, an effective EDA should
jobs J ¼ fJ 1 ; J 2 ; …; J n g to be processed in F factories, where each balance the exploration and the exploitation abilities.
 S.-y. Wang et al. / Int. J. Production Economics 145 (2013) 387–396 389

Obviously, the ECF rule aims at obtaining a schedule with small

Initialize the population
makespan and balancing the workload of factories. With the ECF
rule, an individual of the EDA can be decoded to a feasible
schedule Λ. Then the makespan of the schedule can be calculated
Select superior sub-population
as mentioned in Section 2.

Estimate the probability distribution 4.2. Probability model and updating mechanism
of the superior sub-population

Different from the GA that produces offspring through cross-

over and mutation operators, the EDA does it by sampling
Sample the probability model to
generate new population according to a probability model. So, the probability model has a
great effect on the performance of the EDA. In this paper, the
probability model is designed as a probability matrix P.
No The element pij(l) of the probability matrix P represents the
Stopping condition
is met? probability that job j appears before or in position i of the solution
sequence at generation l. The value of pij refers to the importance
Yes
of a job when deciding the scheduling order. For all values of i and
End
j, pij is initialized to pij(0)¼ 1/n, which ensures that the whole
solution space can be sampled uniformly.
Fig. 1. General framework of the EDA. In each generation of the EDA, the new individuals are
generated via sampling the solution space according to the
probability matrix P. For every position i, job j is selected with a
Procedure ECF rule. probability pij. If job j has already appeared, it means that job j has
For k =1 to F been scheduled. Then, the whole jth column of probability matrix
λ k (1) = π ( k); P will be set as zero and all the elements of P will be normalized to
maintain that each row sums up to 1. In such a way, an individual
nk = 1 ;
is constructed until all the jobs appear in the sequence, and then
For k = F + 1 to n its makespan can be calculated. In the EDA, a population with
Find the factory f that can process job π (k ) with the P_Size individuals are generated.
earliest completion time; Next, it determines the superior sub-population that consists of
nf = nf + 1 ; the best SP_Size solutions, where SP_Size ¼ η%  P_Size. And then
the probability matrix P is updated according to the following
f
λ ( nf ) = π ( k ) ; equation:
Fig. 2. Pseudo code of the ECF rule. α SP_Size
pij ðl þ 1Þ ¼ ð1−αÞpij ðlÞ þ ∑ I k ; ∀i; j ð7Þ
i  SP_Size k ¼ 1 ij
4. EDA for DPFSP
where α∈ð0; 1Þ is the learning rate of P, and I kij is the following
In this section, an estimation of distribution algorithm (EDA) is indicator function of the kth individual in the superior sub-
presented to solve the DPFSP. The framework of the EDA is population.
illustrated, following the encoding and decoding schemes, prob- (
1; if job j appears before or in position i
ability model and its updating mechanism, and local search I kij ¼ ð8Þ
0; else
scheme.
The updating process can be regarded as a kind of increased
learning, where the second term on the right hand side of the
4.1. Encoding and decoding schemes
equation represents learning information from the superior sub-
population. Note that ∑SP_Size
k ¼ 1 ij
I k indicates that the total appearance
Every individual of the population denotes a solution of the
number of all the jobs before or in position i is i  SP_Size.
DPFSP, which is represented by a sequence of all the job numbers
as Eq. (6) to determine the schedule order of all the jobs. For
example, a solution π ¼ ½2; 3; 1; 4 implies that job 2 is scheduled 4.3. Local search scheme
first, and next are job 3 and job 1, in sequence. Job 4 is the last job
to be scheduled. It is widely accepted that a local search procedure is efficient in
π ¼ ½πð1Þ; πð2Þ; ⋯; πðnÞ ð6Þ improving the solutions generated by the EDA (Wang et al.,
2012a). In this paper, some local search operators are designed
To decode a sequence is to arrange the factories for all the jobs based on the problem characteristics to enhance the local exploi-
and determine the processing order in each factory so as to tation around the best solution found by the EDA. Since the
generate a feasible schedule. Considering the characteristics of makespan of a solution can be reduced by improving the schedule
the DPFSP, we employ an effective decoding rule called earliest in the factory with the latest completion time, we design the local
completion factory (ECF) rule. For each job in the sequence to be search operators as follows.
scheduled, the ECF rule assigns it to the factory that can complete Job-swap: In the factory with the latest completion time,
the job with the earliest completion time. The factory assignment randomly select two different jobs from the processing sequence
is the same method as that implemented by Naderi and Ruiz and then swap them.
(2010). Furthermore, the jobs in the same factory are processed in Job-insert: In the factory with the latest completion time,
the order as they appear in the sequence π. The pseudo code of the randomly choose two different jobs from the sequence and then
ECF rule is illustrated in Fig. 2. insert the back one before the front one.
390 S.-y. Wang et al. / Int. J. Production Economics 145 (2013) 387–396

Job-inverse: In the factory with the latest completion time, and {20, 50, 100, 200, 500}  20 and the number of factories F is
invert the subsequence between two different random positions of from {2, 3, 4, 5, 6, 7}.
a job sequence. The EDA is coded in C language and run on a 3.2 GHz Intel Core
Factory-swap: Randomly select two different jobs, one from i5 processor. To evaluate the performance of the EDA, same as in
the factory with the latest completion time and the other from literature (Naderi and Ruiz, 2010), we evaluate the experimental
another randomly selected factory; then swap the factories results by relative percentage deviation (RPD) as follows:
assigned to them.
alg−opt
In each step of the local search, these operators are performed RPD ¼  100 ð9Þ
opt
sequentially in the above order (one time for one operator) to
generate another solution, and then the new solution replaces the where opt is the makespan of best-known solutions from
old one if it has a smaller makespan. The above procedure is http://soa.iti.es and alg corresponds to the makespan of the
applied 200 times on the best individual of the current population solution obtained by a certain algorithm. If the obtained RPD is
in every generation. less than 0, it implies that a new best solution is found.

5.1. Parameters setting

4.4. Procedure of the EDA

The proposed EDA contains several key parameters: P_Size (the

With the above design, the procedure of the EDA for solving the
population size), α (the learning rate of P), η (the parameter
DPFSP is illustrated in Fig. 3.
associated with the superior sub-population), and Gen (the max-
It can be seen that the EDA contains two main phases in every
imum number of generations). To investigate the influence of
generation. At the global exploration phase, a probability model is
these parameters on the performance of the EDA, we implement
built with the superior individuals of the entire population to
the Taguchi method of design of experiment (DOE) (Montgomery,
generate the new individuals. At the local exploitation phase, the
2005) by using a moderate-sized instance (Naderi and Ruiz, 2010),
best solution adopts multiple local search operators based on the
i.e., I_4_16_5_4, where 4_16_5 denotes the size (F¼ 4, n ¼16 and
problem characteristics for further exploitation. The algorithm
m ¼5) of the instance and the last 4 denotes that it is the fourth
stops when the maximum number of generations Gen is reached.
instance of this size. Combinations of different values of these
parameters are listed in Table 1.
For each parameter combination, the EDA is run 10 times
5. Computational results and comparisons independently and the average makespan value obtained by the
EDA is calculated as the average response variable (ARV) value.
To test the performance of the EDA, numerical tests are carried According to the number of parameters and the number of factor
out with two sets of benchmarks (Naderi and Ruiz, 2010), which levels, we choose the orthogonal array L16 ð44 Þ. That is, the total
are available at http://soa.iti.es. The first set consists of 420 small- number of treatments is 16, the number of parameters is 4,
sized instances, where n ¼{4, 6, 8, 10, 12, 14, 16}, m ¼{2, 3, 4, 5} and
F¼{2, 3, 4}. The second set consists of 720 large-sized instances, Table 1
which is extended from the benchmark of Taillard (1993). The Combinations of parameter values.
combinations of n  m are {20, 50, 100}  5, {20, 50, 100, 200}  10,
Parameters Factor level

1 2 3 4

P_Size 50 100 150 200

η 10 20 30 40
α 0.1 0.2 0.3 0.4
Initialize the probability matrix Gen 500 1000 1500 2000

Generate P_Size individuals by

Table 2
sampling the probability matrix Orthogonal array and ARV values.

Experiment Number Factor ARV

Local search on the best individual P_Size η α Gen

1 1 1 1 1 443.4
2 1 2 2 2 445.1
Select the best SP_Size individuals 3 1 3 3 3 447.5
4 1 4 4 4 446.3
and update the probability matrix 5 2 1 2 3 443.1
6 2 2 1 4 444.5
7 2 3 4 1 446.6
No 8 2 4 3 2 446.2
Stopping condition 9 3 1 3 4 444.5
10 3 2 4 3 445.6
is met? 11 3 3 1 2 443.2
12 3 4 2 1 444.8
Yes 13 4 1 4 2 444.7
14 4 2 3 1 445.8
End 15 4 3 2 4 443.9
16 4 4 1 3 443.6
Fig. 3. Framework of the EDA for the DPFSP.
 S.-y. Wang et al. / Int. J. Production Economics 145 (2013) 387–396 391

and the number of factor levels is 4. The orthogonal array and the Fig. 5 illustrates the Gantt chart of the best solution obtained by
obtained ARV values are listed in Table 2. the EDA for instance I_4_16_4_1. As for the other 403 small-sized
According to the orthogonal table, we illustrate the trend of instances, the best makespan values by the EDA are equal to the
each factor level in Fig. 4. Then, we figure out the response value of best known ones.
each parameter to analyze its significance rank. The results are For the small-sized instances, the CPU times employed by both
listed in Table 3. the EDA and the heuristic algorithms are extremely short. Similar
From Table 3 it can be seen that the learning rate α of P is the to Naderi and Ruiz (2010), we will comment on the CPU times
most significant one among the four parameters. That is, the based on the large-sized instances in the next sub-section.
learning rate of the matrix for machine assignment is crucial to
the EDA. A large value of α could lead to premature convergence. 5.3. Results and comparison for large-sized instances
In addition, η ranks the second, which implies that the number of
the superior sub-population to update the probability model is Next, we carry out tests with the large-sized instances. Table 5
also important. A small value of η can help the algorithm build an presents the results of the experiments, averaged for each value of
accurate model. Besides, the significant rank of the population size F (120 data per average).
is the third. A large value of P_Size makes the algorithm sample the Form Table 5, it can be seen that the EDA outperforms other
solution space sufficiently. However, a large population size will algorithms in solving all the large-sized instances. On average, the
cause a large amount of computational budget. It can be seen from EDA yields −1.63% RPD to the best known values. In particular, the
Fig. 4 that it makes no improvement when the size is too large. EDA obtains new best makespan values for 589 out of 720 large-
Similar conclusion can be drawn for the maximum number of the sized instances, which are listed in Tables B.1–B.6 in Appendix B
generations. According to the above analysis, a good choice of (grouped by each value of F).
parameter combination is suggested as P_Size ¼ 150, η ¼10, α ¼0.1 Besides, the CPU times employed by the EDA, VND(a), VND(b),
and Gen ¼1000. NEH1 and NEH2 for the instances grouped by F are listed in
Table 6.
From Table 6, it can be seen that all the heuristic algorithms
5.2. Results and comparison for small-sized instances spend the average CPU time below 0.15 s, while the EDA spends
much more. The reason is that, the heuristic algorithm constructs
Considering the 420 small-sized instances, we compare the a solution based on some heuristic rules while the EDA performs
EDA with several heuristic algorithms (Naderi and Ruiz, 2010). For search procedure among the whole solution space. Fortunately,
each instance, we run the EDA 10 times independently and obtain the average running time of the EDA is acceptable and does not
the best makespan and the RPD. Table 4 summarizes the results
grouped by each combination of n and F (20 data per average) as Table 3
Naderi and Ruiz (2010), where the results of the comparative Response value and rank of each parameter.
algorithms are directly from literature.
Level P_Size η α Gen
From Table 4, it can be seen that the EDA is the best one among
all the algorithms for solving the small-sized instances. The 1 445.575 443.925 443.675 445.15
corresponding RPD values of the best solutions by the EDA are 2 445.1 445.25 444.225 444.8
negative for the instances {2, 3, 4}  16 and 4  14, which implies 3 444.525 445.3 446 444.95
that some of the best know solutions are updated by the EDA. In 4 444.5 445.225 445.8 444.8
Delta 1.075 1.375 2.325 0.35
particular, our EDA obtains new best makespan values for 17 Rank 3 2 1 4
instances, which are listed in Table A1 in Appendix A. In addition,

Fig. 4. Factor level trend of the EDA.

392 S.-y. Wang et al. / Int. J. Production Economics 145 (2013) 387–396

Table 4
RPD of the algorithms for the small-sized instances.

Fn SPT1 LPT1 Johnson1 CDS1 Palmer1 NEH1 VND(a) SPT2 LPT2 Johnson2 CDS2 Palmer2 NEH2 VND(b) EDA

24 12.95 19.31 6.13 2.69 7.75 2.61 0.00 10.02 17.80 3.76 0.72 5.22 0.15 0.15 0.00
26 13.30 29.53 10.34 7.56 7.57 4.42 1.26 10.38 28.37 8.08 4.60 4.83 1.44 1.44 0.00
28 17.31 35.43 9.75 12.33 9.97 4.43 2.10 16.16 33.09 7.98 10.33 8.64 2.53 2.52 0.00
2  10 21.03 36.93 8.38 9.72 10.18 4.28 3.22 17.49 37.12 7.52 8.38 9.25 3.27 2.89 0.00
2  12 20.40 39.36 10.14 11.81 10.77 6.95 3.38 17.23 37.42 8.29 10.49 10.32 4.13 3.90 0.00
2  14 17.73 40.00 10.27 6.59 10.57 6.05 2.70 17.34 38.36 9.21 5.24 9.24 3.16 2.82 0.00
2  16 17.11 42.42 12.33 10.53 10.97 6.66 2.92 16.95 41.63 10.99 8.26 9.00 3.84 3.30 −0.04
34 8.87 6.42 5.31 5.01 4.82 0.43 0.00 5.36 5.41 1.99 2.55 1.50 0.43 0.00 0.00
36 21.24 27.63 6.91 8.50 9.40 4.04 0.68 11.93 25.26 5.28 6.99 3.40 1.39 1.39 0.00
38 17.71 26.87 10.14 9.17 11.84 5.08 1.86 17.93 25.14 8.30 8.67 6.98 2.43 2.43 0.00
3  10 24.48 37.98 12.67 13.56 13.97 8.42 2.59 16.23 35.43 10.07 10.81 11.67 3.79 3.63 0.00
3  12 25.92 42.12 14.90 14.66 14.39 7.66 3.66 19.55 40.14 10.12 11.00 11.36 5.08 5.08 0.00
3  14 23.44 41.00 14.62 16.45 17.97 10.54 4.48 19.95 38.56 14.04 13.47 14.05 4.90 4.46 0.00
3  16 25.31 41.71 14.39 15.55 15.68 8.18 3.50 22.83 39.39 10.41 11.60 11.06 3.98 3.81 −0.18
44 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00
46 13.34 14.52 7.65 4.99 6.02 3.10 0.00 11.31 12.52 5.62 3.39 2.98 0.47 0.00 0.00
48 15.21 16.86 9.20 9.23 10.17 4.19 0.77 13.89 16.44 5.56 8.61 7.04 0.77 0.77 0.00
4  10 22.65 34.47 13.19 13.90 16.21 7.07 1.57 16.84 30.64 8.93 9.30 7.23 2.30 2.22 0.00
4  12 25.51 38.72 13.61 18.13 15.49 8.58 4.23 20.54 34.06 8.68 14.49 12.39 4.97 4.68 0.00
4  14 26.61 42.89 13.50 15.47 16.53 10.94 4.25 22.31 39.79 8.51 11.23 12.25 4.54 4.46 −0.02
4  16 28.70 44.01 19.12 17.62 19.34 9.79 5.08 24.09 40.67 13.96 13.69 15.84 5.59 5.55 −0.19
Average 18.99 34.34 10.60 10.64 11.41 5.88 2.30 15.63 29.39 7.97 8.28 8.30 2.82 2.64 −0.02

Note: The bold values mean better results.

Fig. 5. Best solution of instance I_4_16_4_1 found by the EDA.

Table 5
RPD of the algorithms for the large-sized instances.

Instance (F) SPT1 LPT1 Johnson1 CDS1 Palmer1 NEH1 VND(a) SPT2 LPT2 Johnson2 CDS2 Palmer2 NEH2 VND(b) EDA

2 18.71 33.70 12.68 9.29 10.44 2.92 0.10 17.71 32.36 11.58 8.52 9.34 1.21 0.32 −1.55
3 19.95 33.05 13.40 10.83 11.72 3.42 0.10 17.58 31.57 11.18 8.92 9.43 1.15 0.35 −1.79
4 20.07 33.30 13.48 11.19 11.95 4.21 0.06 17.23 30.42 10.67 8.54 8.90 1.11 0.46 −1.76
5 20.04 32.89 13.18 11.29 12.24 4.28 0.11 16.73 29.57 10.24 8.07 8.76 0.92 0.46 −1.77
6 20.32 32.58 13.57 11.50 12.70 4.73 0.11 16.07 28.55 9.94 7.83 8.45 0.95 0.51 −1.52
7 21.04 32.02 13.61 11.49 12.53 4.86 0.10 15.42 27.14 9.71 7.31 8.21 0.81 0.45 −1.37
Average 20.02 32.92 13.35 10.93 11.93 4.07 0.10 16.79 29.93 10.55 8.20 8.85 1.03 0.43 −1.63

Note: The bold values mean better results.

 S.-y. Wang et al. / Int. J. Production Economics 145 (2013) 387–396 393

Table B2
Table 6
New best makespan obtained by the EDA for large-sized instances (F¼ 3).
CPU time spent by the algorithms (s).

Instance Best EDA Instance Best EDA Instance Best EDA

Instance (F) EDA VND(a) VND(b) NEH1 NEH2
known known known
2 90.390 0.246 0.120 0.009 0.017
Ta001_3 598 576 Ta036_3 1064 1045 Ta072_3 2157 2109
3 101.015 0.182 0.105 0.007 0.020
Ta002_3 594 582 Ta037_3 1072 1030 Ta073_3 2223 2190
4 111.375 0.129 0.104 0.007 0.024
Ta003_3 533 505 Ta038_3 1016 1003 Ta074_3 2327 2312
5 120.625 0.114 0.086 0.006 0.028
Ta004_3 630 602 Ta039_3 990 955 Ta075_3 2216 2178
6 130.609 0.120 0.083 0.005 0.031
Ta005_3 579 565 Ta040_3 1041 1020 Ta076_3 2111 2078
7 140.734 0.091 0.076 0.006 0.035
Ta006_3 574 554 Ta041_3 1417 1365 Ta077_3 2186 2170
Average 115.791 0.147 0.096 0.007 0.026
Ta007_3 566 546 Ta042_3 1347 1320 Ta078_3 2283 2230
Ta008_3 590 560 Ta043_3 1335 1316 Ta079_3 2345 2302
Ta009_3 594 556 Ta044_3 1411 1368 Ta080_3 2315 2288
Table A1 Ta010_3 531 501 Ta045_3 1406 1372 Ta081_3 2985 2943
New best makespan obtained by the EDA for small-sized instances. Ta011_3 905 875 Ta046_3 1390 1355 Ta082_3 2938 2912
Ta012_3 961 927 Ta047_3 1440 1400 Ta083_3 2996 2948
Instance F n m Best known EDA RPD Ta013_3 872 843 Ta048_3 1408 1373 Ta084_3 2954 2928
Ta014_3 796 760 Ta049_3 1350 1310 Ta085_3 3000 2969
I_2_16_5_1 2 16 5 526 523 −0.57 Ta015_3 835 806 Ta050_3 1429 1402 Ta086_3 3030 3001
I_2_16_5_3 2 16 5 652 650 −0.31 Ta016_3 783 772 Ta051_3 2121 2087 Ta087_3 2994 2973
I_3_16_3_1 3 16 3 340 339 −0.29 Ta017_3 855 825 Ta052_3 2044 1999 Ta088_3 3064 3048
I_3_16_3_3 3 16 3 350 349 −0.29 Ta018_3 894 856 Ta053_3 2009 1981 Ta089_3 3014 2972
I_3_16_3_5 3 16 3 362 361 −0.28 Ta019_3 899 843 Ta054_3 2079 2027 Ta090_3 3041 2996
I_3_16_4_2 3 16 4 458 457 −0.22 Ta020_3 946 893 Ta055_3 2027 1990 Ta091_3 3995 3950
I_3_16_4_4 3 16 4 430 429 −0.23 Ta021_3 1509 1472 Ta056_3 2037 1987 Ta092_3 3913 3867
I_3_16_4_5 3 16 4 419 414 −1.19 Ta022_3 1441 1387 Ta057_3 2030 1999 Ta093_3 4008 4004
I_3_16_5_1 3 16 5 453 452 −0.22 Ta023_3 1576 1523 Ta058_3 2034 2021 Ta094_3 3941 3932
I_3_16_5_3 3 16 5 476 475 −0.21 Ta024_3 1479 1446 Ta059_3 2081 2061 Ta096_3 3853 3817
I_3_16_5_4 3 16 5 524 522 −0.38 Ta025_3 1512 1474 Ta060_3 2096 2030 Ta097_3 4007 3960
I_4_14_5_4 4 14 5 425 423 −0.47 Ta026_3 1511 1462 Ta061_3 1972 1925 Ta099_3 3871 3820
I_4_16_3_3 4 16 3 312 311 −0.32 Ta027_3 1505 1470 Ta062_3 1874 1854 Ta100_3 3918 3914
I_4_16_4_1 4 16 4 323 319 −1.24 Ta028_3 1487 1422 Ta063_3 1850 1826 Ta101_3 4727 4690
I_4_16_4_2 4 16 4 359 358 −0.28 Ta029_3 1509 1471 Ta064_3 1771 1760 Ta102_3 4767 4756
I_4_16_5_3 4 16 5 365 363 −0.55 Ta030_3 1414 1407 Ta065_3 1863 1842 Ta103_3 4742 4733
I_4_16_5_4 4 16 5 447 441 −1.34 Ta031_3 1007 979 Ta066_3 1800 1797 Ta104_3 4728 4722
Ta032_3 1082 1056 Ta067_3 1854 1836 Ta105_3 4695 4689
Ta033_3 1000 968 Ta069_3 1935 1917 Ta109_3 4767 4722
Ta034_3 1074 1030 Ta070_3 1874 1866
Table B1
Ta035_3 1077 1049 Ta071_3 2294 2267
New best makespan obtained by the EDA for large-sized instances (F¼ 2).

Instance Best EDA Instance Best EDA Instance Best EDA

known known known
increase greatly as the problem size increases. Meanwhile, the
Ta001_2 770 751 Ta035_2 1508 1488 Ta069_2 2823 2797 solutions with much better quality can be found by the EDA.
Ta002_2 783 768 Ta036_2 1480 1477 Ta070_2 2730 2728 Besides, the distributed scheduling in a real life scenario can be
Ta003_2 676 645 Ta037_2 1482 1439 Ta071_2 3182 3133 solved offline, so solution quality is more important than the
Ta004_2 803 765 Ta038_2 1428 1412 Ta072_2 2969 2903
efficiency of the algorithm.
Ta005_2 751 731 Ta039_2 1389 1339 Ta073_2 3093 3035
Ta006_2 722 709 Ta040_2 1465 1444 Ta074_2 3235 3199 So, it is concluded that the EDA is more effective than the
Ta007_2 731 708 Ta041_2 1823 1776 Ta075_2 3060 3006 existing methods in solving the DPFSP with the criterion to
Ta008_2 745 711 Ta042_2 1743 1722 Ta076_2 2902 2871 minimize the makespan, especially for the large-sized problems.
Ta009_2 742 720 Ta043_2 1731 1714 Ta077_2 3050 3025
The superiority of the EDA owes to the following aspects. (1) With
Ta010_2 672 645 Ta044_2 1828 1776 Ta078_2 3124 3081
Ta011_2 1071 1050 Ta045_2 1811 1779 Ta079_2 3247 3205
the permutation based encoding and the ECF rule based decoding,
Ta012_2 1145 1117 Ta046_2 1845 1767 Ta080_2 3235 3137 it is helpful to obtain a schedule with small makespan. (2) With
Ta013_2 1051 1001 Ta047_2 1902 1821 Ta081_2 3892 3833 the well-designed probability model and the suitable updating
Ta014_2 934 913 Ta048_2 1822 1787 Ta082_2 3842 3792 mechanism, it is helpful to explore search procedure effectively,
Ta015_2 1013 966 Ta049_2 1739 1710 Ta083_2 3866 3836
especially within the promising area of the solution space. (3) With
Ta016_2 959 928 Ta050_2 1854 1824 Ta084_2 3844 3800
Ta017_2 1029 991 Ta051_2 2585 2560 Ta086_2 3972 3899 multiple local search operators based on the problem character-
Ta018_2 1083 1032 Ta052_2 2511 2467 Ta087_2 3887 3875 istics, it is helpful to improve the good schedule by enhancing the
Ta019_2 1063 1028 Ta053_2 2513 2423 Ta088_2 4040 3967 exploitation capability.
Ta020_2 1117 1073 Ta054_2 2539 2485 Ta089_2 3929 3867
Ta021_2 1723 1686 Ta055_2 2470 2437 Ta090_2 3950 3899
Ta022_2 1589 1571 Ta056_2 2508 2453 Ta091_2 5721 5700
Ta023_2 1773 1736 Ta057_2 2496 2445 Ta092_2 5587 5553 6. Conclusions
Ta024_2 1676 1642 Ta058_2 2538 2479 Ta094_2 5711 5669
Ta025_2 1781 1697 Ta059_2 2551 2515 Ta097_2 5720 5708 In this paper, an effective estimation of distribution algorithm
Ta026_2 1691 1658 Ta060_2 2537 2494 Ta098_2 5659 5646
was proposed for solving the distributed permutation flow-shop
Ta027_2 1733 1684 Ta061_2 2846 2810 Ta099_2 5495 5470
Ta028_2 1661 1621 Ta062_2 2725 2705 Ta100_2 5663 5612 scheduling problem with the criterion to minimize the makespan.
Ta029_2 1694 1668 Ta063_2 2665 2658 Ta102_2 6498 6490 To the best of our knowledge, this is the first reported work of the
Ta030_2 1645 1609 Ta064_2 2582 2571 Ta103_2 6554 6489 EDA for solving the DPFSP. To be specific, the earliest completion
Ta031_2 1436 1407 Ta065_2 2685 2681 Ta104_2 6450 6445 factory rule was employed for the permutation based encoding to
Ta032_2 1508 1490 Ta066_2 2656 2626 Ta105_2 6404 6402
Ta033_2 1393 1368 Ta067_2 2719 2678 Ta106_2 6479 6411
generate feasible schedules. A probability model was designed to
Ta034_2 1484 1449 Ta068_2 2637 2614 Ta109_2 6503 6466 generate the new individuals and a mechanism was provided to
update the probability model suitably. Some local search operators
394 S.-y. Wang et al. / Int. J. Production Economics 145 (2013) 387–396

Table B3 Table B4
New best makespan obtained by the EDA for large-sized instances (F¼ 4). New best makespan obtained by the EDA for large-sized instances (F¼5).

Instance Best EDA Instance Best EDA Instance Best EDA Instance Best EDA Instance Best EDA Instance Best EDA
known known known known known known

Ta001_4 528 492 Ta036_4 875 840 Ta071_4 1862 1847 Ta001_5 469 442 Ta035_5 719 708 Ta069_5 1240 1228
Ta002_4 513 491 Ta037_4 841 832 Ta072_4 1722 1717 Ta002_5 460 437 Ta036_5 741 723 Ta070_5 1202 1192
Ta003_4 465 442 Ta038_4 827 806 Ta073_4 1807 1777 Ta003_5 432 393 Ta037_5 729 712 Ta071_5 1614 1586
Ta004_4 537 519 Ta039_4 792 768 Ta074_4 1911 1882 Ta004_5 487 469 Ta038_5 704 689 Ta072_5 1493 1480
Ta005_4 510 488 Ta040_4 832 819 Ta075_4 1800 1771 Ta005_5 459 434 Ta039_5 677 661 Ta073_5 1546 1531
Ta006_4 490 479 Ta041_4 1182 1149 Ta076_4 1716 1688 Ta006_5 450 436 Ta040_5 716 696 Ta074_5 1634 1620
Ta007_4 497 471 Ta042_4 1127 1123 Ta077_4 1772 1757 Ta007_5 445 435 Ta041_5 1048 1029 Ta075_5 1551 1522
Ta008_4 499 483 Ta043_4 1150 1127 Ta078_4 1820 1810 Ta008_5 468 441 Ta042_5 1028 995 Ta076_5 1472 1449
Ta009_4 500 475 Ta044_4 1190 1160 Ta079_4 1878 1856 Ta009_5 443 427 Ta043_5 1035 1012 Ta077_5 1530 1508
Ta010_4 463 434 Ta045_4 1210 1164 Ta080_4 1885 1855 Ta010_5 401 391 Ta044_5 1060 1030 Ta078_5 1570 1549
Ta011_4 809 783 Ta046_4 1178 1153 Ta082_4 2499 2465 Ta011_5 747 730 Ta045_5 1048 1035 Ta079_5 1615 1596
Ta012_4 881 832 Ta047_4 1227 1188 Ta083_4 2524 2511 Ta012_5 808 770 Ta046_5 1048 1027 Ta080_5 1638 1594
Ta013_4 792 756 Ta048_4 1215 1166 Ta084_4 2498 2482 Ta013_5 731 703 Ta047_5 1096 1056 Ta081_5 2242 2227
Ta014_4 720 682 Ta049_4 1130 1111 Ta085_4 2523 2514 Ta014_5 657 634 Ta048_5 1061 1031 Ta082_5 2228 2204
Ta015_4 754 720 Ta050_4 1230 1188 Ta086_4 2560 2545 Ta015_5 706 671 Ta049_5 1012 990 Ta083_5 2241 2231
Ta016_4 710 690 Ta051_4 1883 1850 Ta087_4 2545 2528 Ta016_5 657 640 Ta050_5 1087 1058 Ta084_5 2224 2208
Ta017_4 789 744 Ta052_4 1785 1768 Ta088_4 2608 2590 Ta017_5 732 693 Ta051_5 1730 1706 Ta085_5 2252 2239
Ta018_4 811 773 Ta053_4 1806 1754 Ta089_4 2561 2518 Ta018_5 749 720 Ta052_5 1650 1626 Ta086_5 2278 2264
Ta019_4 784 760 Ta054_4 1831 1786 Ta090_4 2560 2537 Ta019_5 737 712 Ta053_5 1647 1607 Ta087_5 2257 2249
Ta020_4 834 803 Ta055_4 1807 1766 Ta091_4 3098 3095 Ta020_5 771 755 Ta054_5 1687 1642 Ta088_5 2316 2303
Ta021_4 1406 1367 Ta056_4 1779 1759 Ta092_4 3063 3043 Ta021_5 1348 1303 Ta055_5 1636 1630 Ta089_5 2280 2249
Ta022_4 1321 1293 Ta057_4 1768 1753 Ta094_4 3064 3063 Ta022_5 1261 1234 Ta056_5 1635 1619 Ta090_5 2267 2258
Ta023_4 1443 1405 Ta058_4 1819 1780 Ta096_4 3016 2992 Ta023_5 1390 1347 Ta057_5 1638 1605 Ta092_5 2559 2543
Ta024_4 1399 1357 Ta059_4 1846 1820 Ta097_4 3140 3123 Ta024_5 1321 1301 Ta058_5 1668 1626 Ta094_5 2553 2541
Ta025_4 1420 1368 Ta060_4 1844 1795 Ta098_4 3086 3076 Ta025_5 1360 1306 Ta059_5 1690 1668 Ta095_5 2554 2538
Ta026_4 1391 1356 Ta061_4 1505 1494 Ta099_4 3005 3002 Ta026_5 1339 1288 Ta060_5 1655 1647 Ta096_5 2509 2503
Ta027_4 1389 1362 Ta062_4 1449 1431 Ta101_4 3828 3813 Ta027_5 1339 1298 Ta061_5 1256 1233 Ta097_5 2622 2604
Ta028_4 1348 1319 Ta063_4 1426 1407 Ta103_4 3867 3861 Ta028_5 1299 1258 Ta062_5 1198 1183 Ta098_5 2582 2564
Ta029_4 1394 1363 Ta064_4 1373 1354 Ta104_4 3839 3827 Ta029_5 1339 1301 Ta063_5 1176 1164 Ta099_5 2536 2509
Ta030_4 1345 1301 Ta065_4 1436 1426 Ta105_4 3830 3826 Ta030_5 1289 1240 Ta064_5 1142 1123 Ta100_5 2570 2565
Ta031_4 794 782 Ta066_4 1422 1391 Ta106_4 3826 3813 Ta031_5 680 664 Ta065_5 1189 1178 Ta102_5 3357 3337
Ta032_4 859 848 Ta067_4 1437 1419 Ta107_4 3938 3909 Ta032_5 740 723 Ta066_5 1167 1142 Ta104_5 3347 3312
Ta033_4 807 778 Ta068_4 1402 1383 Ta108_4 3900 3887 Ta033_5 691 667 Ta067_5 1192 1173 Ta106_5 3310 3290
Ta034_4 865 829 Ta069_4 1494 1485 Ta109_4 3873 3852 Ta034_5 740 707 Ta068_5 1161 1142 Ta107_5 3384 3376
Ta035_4 848 831 Ta070_4 1452 1444 Ta110_4 3899 3886

Table B5
New best makespan obtained by the EDA for large-sized instances (F ¼6).
Table B6
Instance Best EDA Instance Best EDA Instance Best EDA New best makespan obtained by the EDA for large-sized instances (F¼ 7).
known known known
Instance Best EDA Instance Best EDA Instance Best EDA
Ta001_6 431 410 Ta033_6 630 595 Ta064_6 984 969 known known known
Ta002_6 433 404 Ta034_6 657 630 Ta066_6 997 988
Ta003_6 390 369 Ta035_6 656 629 Ta067_6 1015 1011 Ta001_7 415 386 Ta036_7 610 587 Ta065_7 922 910
Ta004_6 460 432 Ta036_6 664 646 Ta068_6 993 985 Ta002_7 403 382 Ta037_7 584 576 Ta066_7 892 878
Ta005_6 436 404 Ta037_6 663 630 Ta069_6 1075 1059 Ta003_7 375 360 Ta038_7 576 561 Ta067_7 907 898
Ta006_6 431 402 Ta038_6 632 614 Ta070_6 1035 1026 Ta004_7 432 413 Ta039_7 556 534 Ta068_7 881 876
Ta008_6 424 414 Ta039_6 607 586 Ta071_6 1432 1421 Ta005_7 410 385 Ta040_7 588 560 Ta069_7 953 933
Ta009_6 413 396 Ta040_6 643 617 Ta072_6 1332 1323 Ta006_7 396 383 Ta041_7 881 879 Ta070_7 926 916
Ta010_6 379 365 Ta041_6 959 941 Ta073_6 1378 1366 Ta008_7 402 386 Ta042_7 869 848 Ta071_7 1306 1298
Ta011_6 710 695 Ta042_6 924 914 Ta074_6 1466 1446 Ta009_7 401 377 Ta043_7 893 866 Ta072_7 1213 1211
Ta012_6 754 730 Ta043_6 949 928 Ta075_6 1372 1360 Ta010_7 368 347 Ta044_7 905 890 Ta073_7 1245 1243
Ta013_6 703 673 Ta044_6 979 957 Ta077_6 1360 1350 Ta011_7 699 670 Ta045_7 908 890 Ta074_7 1343 1322
Ta014_6 619 605 Ta045_6 967 950 Ta078_6 1400 1382 Ta012_7 724 706 Ta046_7 895 880 Ta075_7 1256 1244
Ta015_6 670 652 Ta046_6 959 941 Ta079_6 1427 1416 Ta013_7 688 650 Ta047_7 944 907 Ta076_7 1192 1190
Ta016_6 638 613 Ta047_6 1014 969 Ta080_6 1444 1422 Ta014_7 601 586 Ta048_7 900 883 Ta077_7 1253 1231
Ta017_6 681 671 Ta048_6 972 943 Ta082_6 2030 2011 Ta015_7 637 628 Ta049_7 869 848 Ta078_7 1272 1261
Ta018_6 720 692 Ta049_6 928 905 Ta083_6 2061 2047 Ta016_7 615 591 Ta050_7 943 910 Ta079_7 1303 1299
Ta019_6 704 702 Ta050_6 1006 972 Ta084_6 2042 2019 Ta017_7 674 671 Ta051_7 1553 1523 Ta080_7 1319 1302
Ta020_6 746 723 Ta051_6 1636 1599 Ta085_6 2068 2054 Ta018_7 695 692 Ta052_7 1476 1449 Ta081_7 1904 1901
Ta021_6 1296 1264 Ta052_6 1559 1524 Ta088_6 2121 2109 Ta020_7 715 707 Ta053_7 1470 1438 Ta082_7 1896 1884
Ta022_6 1213 1191 Ta053_6 1526 1512 Ta089_6 2085 2063 Ta021_7 1263 1243 Ta054_7 1514 1469 Ta083_7 1919 1907
Ta023_6 1328 1320 Ta054_6 1580 1546 Ta090_6 2087 2071 Ta024_7 1247 1241 Ta055_7 1479 1463 Ta085_7 1933 1912
Ta024_6 1283 1267 Ta055_6 1550 1528 Ta091_6 2244 2237 Ta025_7 1271 1253 Ta056_7 1469 1447 Ta086_7 1941 1938
Ta025_6 1313 1281 Ta056_6 1556 1515 Ta095_6 2214 2208 Ta027_7 1237 1232 Ta057_7 1465 1438 Ta088_7 2007 1972
Ta026_6 1272 1256 Ta057_6 1530 1509 Ta097_6 2267 2259 Ta029_7 1245 1240 Ta058_7 1481 1455 Ta089_7 1947 1933
Ta027_6 1294 1251 Ta058_6 1549 1527 Ta098_6 2236 2227 Ta030_7 1201 1168 Ta059_7 1514 1502 Ta091_7 2004 1997
Ta028_6 1252 1240 Ta059_6 1598 1569 Ta099_6 2188 2183 Ta031_7 565 539 Ta060_7 1509 1477 Ta096_7 1954 1948
Ta029_6 1289 1257 Ta060_6 1575 1545 Ta100_6 2228 2220 Ta032_7 610 589 Ta061_7 956 944 Ta097_7 2028 2024
Ta030_6 1235 1205 Ta061_6 1070 1060 Ta105_6 2958 2949 Ta033_7 562 539 Ta062_7 913 905 Ta098_7 1989 1983
Ta031_6 596 592 Ta062_6 1035 1022 Ta034_7 604 573 Ta063_7 909 891 Ta102_7 2733 2732
Ta032_6 666 644 Ta063_6 1019 1003 Ta035_7 586 570 Ta064_7 869 860
 S.-y. Wang et al. / Int. J. Production Economics 145 (2013) 387–396 395

were also designed based on the problem characteristics to Harik, G., 1999. Linkage learning via probabilistic modeling in the ECGA. Illigal
enhance the exploitation. The influence of parameter setting was Report no. 99010, Illinois Genetic Algorithms Lboratory, University of Illinois at
Urbana-Champaign, Illinois.
investigated by using DOE test. Extensive testing results and Harik, G.R., Lobo, F.G., Goldberg, D.E., 1998. The compact genetic algorithm. In:
comparisons demonstrated the effectiveness of the proposed Proceedings of the IEEE Conference on Evolutionary Computation, Indianapolis,
EDA in solving the DPFSP. The new best-known solutions for 17 523–528.
Hejazi, S.R., Saghafian, S., 2005. Flowshop-scheduling problems with makespan
out of 420 small-sized instances and 589 out of 720 large-sized criterion: a review. International Journal of Production Research 43,
instances were presented as well. The future work is to design 2895–2929.
EDA-based algorithms for distributed job-shop scheduling pro- Jarboui, B., Eddaly, M., Siarry, P., 2009. An estimation of distribution algorithm for
minimizing the total flowtime in permutation flowshop scheduling problems.
blem and multi-objective distributed schedule problem. Computer & Operations Research 36, 2638–2646.
Jia, H.Z., Fuh, J.Y.H., Nee, A.Y.C., Zhang, Y.F., 2002. Web-based multi-functional
scheduling system for a distributed manufacturing environment. Concurrent
Engineering-Research and Applications 10, 27–39.
Acknowledgments
Jia, H.Z., Fuh, J.Y.H., Nee, A.Y.C., Zhang, Y.F., 2007. Integration of genetic algorithm
and Gantt chart for job shop scheduling in distributed manufacturing systems.
This research is partially supported by the National Key Basic Computers & Industrial Engineering 53, 313–320.
Jia, H.Z., Nee, A.Y.C., Fuh, J.Y.H., Zhang, Y.F., 2003. A modified genetic algorithm for
Research and Development Program of China (No. 2013CB329503), distributed scheduling problems. Journal of Intelligent Manufacturing 14,
the National Science Foundation of China (Nos. 61174189 and 351–362.
61025018), the Doctoral Program Foundation of Institutions of Johnson, S.M., 1954. Optimal two- and three-stage production schedules with
setuptimes included. Naval Research Logistics Quarterly 1, 61–68.
Higher Education of China (No. 20100002110014), and the
Kahn, K.B., Castellion, G.A., Griffin, A., 2004. The PDMA Handbook of New Product
National Science and Technology Major Project of China (No. Development, 2nd ed.. Wiley, New York.
2011ZX02504-008). Larranaga, P., Lozano, J.A., 2002. Estimation of Distribution Algorithms: A New Tool
for Evolutionary Computation. Springer, Netherlands.
Lin, B.M.T., Lu, C.Y., Shyu, S.J., Tsai, C.Y., 2008. Development of new features of ant
colony optimization for flowshop scheduling. International Journal of Produc-
Appendix A tion Economics 112, 742–755.
Mühlenbein, H., Paass, G., 1996. From recombination of genes to the estimation of
distributions I: binary parameters. Lecture Notes in Computer Science 1141,
See Table A1. 178–187.
Montgomery, D.C., 2005. Design and Analysis of Experiments. John Wiley & Sons,
Arizona.
Appendix B Moon, C., Kim, J., Hur, S., 2002. Integrated process planning and scheduling with
minimizing total tardiness in multi-plants supply chain. Computers & Industrial
Engineering 43, 331–349.
See Tables B1–B6.
Mühlenbein, H., Mahnig, T., 1999. Convergence theory and applications of the
factorized distribution algorithm. Jounal of Computing and Information Tech-
nology 7, 19–32.
References Naderi, B., Ruiz, R., 2010. The distributed permutation flowshop scheduling
problem. Computers & Operations Research 37, 754–768.
Baluja, S., 1994. Population-based incremental learning: a method for integrating Pelikan, M., Goldberg, D.E., Cantú-Paz, E., 1999. BOA: the bayesian optimization
genetic search based function optimization and competitive learning. Technical algorithm. In: Proceedings of the Genetic and Evolutionary Computation, San
Report CMU-CS-94-163, Carnegie Mellon University Pittsburgh, PA. Francisco, 525–532.
Baluja, S., Davies, S., 1997. Using optimal dependency-trees for combinatorial Pelikan, M., Mühlenbein, H., 1999. The bivariate marginal distribution algorithm. In:
optimization: learning the structure of the search space. In: Proceedings of Benítez, J.M. (Ed.), Advances in Soft Computing: Engineering Design and
the 14th International Conference on Machine Learning, San Francisco, 30–38. Manufacturing. Springer-Verlag, London.
Cesar, R.M., Bengoetxea, E., Bloch, I., Larranaga, P., 2005. Inexact graph matching for Ruiz, R., Maroto, C., 2005. A comprehensive review and evaluation of
model-based recognition: evaluation and comparison of optimization algo- permutation flowshop heuristics. European Journal of Operational Research
rithms. Pattern Recognition 38, 2099–2113. 165, 479–494.
Chan, F.T.S., Chung, S.H., Chan, L.Y., Finke, G., Tiwari, M.K., 2006. Solving distributed Saeys, Y., Degroeve, S, Aeyels, D., Van de Peer, Y., Rouze, P., 2003. Fast feature
FMS scheduling problems subject to maintenance: genetic algorithms selection using a simple estimation of distribution algorithm: a case study on
approach. Robotics and Computer-Integrated Manufacturing 22, 493–504. splice site prediction. Bioinformatics 19, 179–188.
Chan, F.T.S., Chung, S.H., Chan, P.L.Y., 2005. An adaptive genetic algorithm with Sagarna, R., Lozano, J., 2005. On the performance of estimation of distribution
dominated genes for distributed scheduling problems. Expert Systems with algorithms applied to software testing. Applied Artificial Intelligence 19,
Applications 29, 364–371. 457–489.
Chen, S.H., Chen, M.C., 2013. Addressing the advantages of using ensemble Shabtay, D., Bensoussan, Y., Kaspi, M., 2013. A bicriteria approach to maximize the
probabilistic models in estimation of distribution algorithms for scheduling weighted number of just-in-time jobs and to minimize the total resource
problems. International Journal of Production Economics 141, 24–33. consumption cost in a two-machine flow-shop scheduling system. Interna-
Cheng, T.C.E., Ding, Q., Lin, B.M.T., 2004. A concise survey of scheduling with time- tional Journal of Production Economics 136, 67–74.
dependent processing times. European Journal of Operational Research 152, Suliman, S.M.A., 2000. A two-phase heuristic approach to the permutation flow-
1–13. shop scheduling problem. International Journal of Production Economics 64,
Cheng, T.C.E., Janiak, A., 2000. A permutation flow-shop scheduling problem with 143–152.
convex models of operation processing times. Annals of Operations Research Sun, L.H., Sun, L.Y., Wang, M.Z., Wang, J.B., 2012. Flow shop makespan minimization
96, 39–60. scheduling with deteriorating jobs under dominating machines. International
Cheng, T.C.E., Kovalyov, M.Y., 2003. Scheduling a single server in a two-machine Journal of Production Economics 138, 195–200.
flow shop. Computing 70, 167–180. Taillard, E., 1993. Benchmarks for basic scheduling problems. European Journal of
Cheng, T.C.E., Wu, C.C., Chen, J.C., Wu, W.H., Cheng, S.R., 2013. Two-machine Operational Research 64, 278–285.
flowshop scheduling with a truncated learning function to minimize the Tseng, L.Y., Lin, Y.T., 2010a. A genetic local search algorithm for minimizing total
makespan. International Journal of Production Economics 141, 79–86. flowtime in the permutation flowshop scheduling problem. International
Chung, C.S., Flynn, J., Kirca, O., 2002. A branch and bound algorithm to minimize the Journal of Production Economics 127, 121–128.
total flow time for m-machine permutation flowshop problems. International Tseng, L.Y., Lin, Y.T., 2010b. A hybrid genetic algorithm for no-wait flowshop
Journal of Production Economics 79, 185–196. scheduling problem. International Journal of Production Economics 128,
De Bonet, J.S., Isbell Jr., C.L., Viola, P., 1997. MIMIC: finding optima by estimating 144–152.
probability densities, Advances in Neural Information Processing Systems. MIT Wang, L., Fang, C., 2012. An effective estimation of distribution algorithm for the
Press, Cambridge pp. 424–430. multi-mode resource-constrained project scheduling problem. Computer &
De Giovanni, L., Pezzella, F., 2010. An improved genetic algorithm for the distributed Operations Research 39, 449–460.
and flexible job-shop scheduling problem. European Journal of Operational Wang, L., Shen, W., 2007. Process Planning and Scheduling for Distributed
Research 200, 395–408. Manufacturing. Springer, London.
Gao, J., Chen, R., Deng, W., 2012. An efficient tabu search algorithm for the Wang, L., Wang, S.Y., Fang, C., 2012a. A hybrid estimation of distribution algorithm
distributed permutation flowshop scheduling problem. International Journal for solving multidimensional knapsack problem. Expert Systems with Applica-
of Production Research. http://dx.doi.org/10.1080/00207543.2011.644819. tions 39, 5593–5599.
Garey, M.R., Johnson, D.S., Sethi, R., 1976. The complexity of flowshop and jobshop Wang, L., Wang, S.Y., Liu, M., 2013a. A Pareto-based estimation of distribution
scheduling. Mathematics of Operations Research 1, 117–129. algorithm for the multi-objective flexible job-shop scheduling problem.
396 S.-y. Wang et al. / Int. J. Production Economics 145 (2013) 387–396

International Journal of Production Research. http://dx.doi.org/10.1080/ Wang, S.Y., Wang, L., Liu, M., Xu, Y., 2013c. An enhanced estimation of distribution
00207543.2012.752588. algorithm for solving hybrid flow-shop scheduling problem with identical
Wang, L., Wang, S.Y., Xu, Y., Zhou, G., Liu, M., 2012b. A bi-population based parallel machines. The International Journal of Advanced Manufacturing
estimation of distribution algorithm for the flexible job-shop scheduling Technology. http://dx.doi.org/10.1007/s00170-013-4819-y.
problem. Computers & Industrial Engineering 62, 917–926. Wang, X.L., Xie, X.Z., Cheng, T.C.E., 2013d. A modified artificial bee colony algorithm
Wang, S.Y., Wang, L., Liu, M., Xu, Y., 2013b. An effective estimation of distribution for order acceptance in two-machine flow shops. International Journal of
algorithm for the flexible job-shop scheduling problem with fuzzy processing Production Economics 141, 14–23.
time. International Journal of Production Research. http://dx.doi.org/10.1080/
00207543.2013.765077.