Anda di halaman 1dari 22

This article was downloaded by: [National University of Sciences & Technology]

On: 30 October 2014, At: 23:22


Publisher: Taylor & Francis
Informa Ltd Registered in England and Wales Registered Number: 1072954 Registered
office: Mortimer House, 37-41 Mortimer Street, London W1T 3JH, UK

Engineering Optimization
Publication details, including instructions for authors and
subscription information:
http://www.tandfonline.com/loi/geno20

A discrete artificial bee colony


algorithm incorporating differential
evolution for the flow-shop scheduling
problem with blocking
a

Yu-Yan Han , Dunwei Gong & Xiaoyan Sun

School of Information and Electrical Engineering, China


University of Mining and Technology, Xuzhou, PR China
Published online: 18 Jun 2014.

Click for updates


To cite this article: Yu-Yan Han, Dunwei Gong & Xiaoyan Sun (2014): A discrete artificial bee colony
algorithm incorporating differential evolution for the flow-shop scheduling problem with blocking,
Engineering Optimization, DOI: 10.1080/0305215X.2014.928817
To link to this article: http://dx.doi.org/10.1080/0305215X.2014.928817

PLEASE SCROLL DOWN FOR ARTICLE


Taylor & Francis makes every effort to ensure the accuracy of all the information (the
Content) contained in the publications on our platform. However, Taylor & Francis,
our agents, and our licensors make no representations or warranties whatsoever as to
the accuracy, completeness, or suitability for any purpose of the Content. Any opinions
and views expressed in this publication are the opinions and views of the authors,
and are not the views of or endorsed by Taylor & Francis. The accuracy of the Content
should not be relied upon and should be independently verified with primary sources
of information. Taylor and Francis shall not be liable for any losses, actions, claims,
proceedings, demands, costs, expenses, damages, and other liabilities whatsoever or
howsoever caused arising directly or indirectly in connection with, in relation to or arising
out of the use of the Content.
This article may be used for research, teaching, and private study purposes. Any
substantial or systematic reproduction, redistribution, reselling, loan, sub-licensing,
systematic supply, or distribution in any form to anyone is expressly forbidden. Terms &

Downloaded by [National University of Sciences & Technology] at 23:22 30 October 2014

Conditions of access and use can be found at http://www.tandfonline.com/page/termsand-conditions

Downloaded by [National University of Sciences & Technology] at 23:22 30 October 2014

Engineering Optimization, 2014


http://dx.doi.org/10.1080/0305215X.2014.928817

A discrete artificial bee colony algorithm incorporating


differential evolution for the flow-shop scheduling problem
with blocking
Yu-Yan Han, Dunwei Gong and Xiaoyan Sun
School of Information and Electrical Engineering, China University of Mining and Technology,
Xuzhou, PR China
(Received 26 December 2013; accepted 19 May 2014)
A flow-shop scheduling problem with blocking has important applications in a variety of industrial systems
but is underrepresented in the research literature. In this study, a novel discrete artificial bee colony (ABC)
algorithm is presented to solve the above scheduling problem with a makespan criterion by incorporating
the ABC with differential evolution (DE). The proposed algorithm (DE-ABC) contains three key operators.
One is related to the employed bee operator (i.e. adopting mutation and crossover operators of discrete DE
to generate solutions with good quality); the second is concerned with the onlooker bee operator, which
modifies the selected solutions using insert or swap operators based on the self-adaptive strategy; and the
last is for the local search, that is, the insert-neighbourhood-based local search with a small probability is
adopted to improve the algorithms capability in exploitation. The performance of the proposed DE-ABC
algorithm is empirically evaluated by applying it to well-known benchmark problems. The experimental
results show that the proposed algorithm is superior to the compared algorithms in minimizing the makespan
criterion.
Keywords: artificial bee colony; flow-shop scheduling; blocking; differential evolution; makespan

1.

Introduction

Flow-shop scheduling problems can be generally classified into the following two categories
according to whether there are buffers or not: one with infinite buffers, and the other with finite
buffers. The former does not result in blocking any job since there are enough intermediate
buffers to store those completed jobs. The latter only maintains a limited capacity of in-process
inventories, which means that there are either no buffers or buffers with a limited capacity owing
to the finite storage facilities. The flow-shop scheduling problem with no intermediate buffers
considered here, named the blocking flow-shop (BFS) scheduling problem, is a special case of
the latter. For this case, a job must remain in the current machine until the next machine is available
for processing, which increases the waiting time or the productive cycle, and thus decreases the
production efficiency.
Owing to the above process characteristics, the BFS scheduling problem has been a typical
problem with a strong engineering background. In the chemical industry, partially processed jobs
(i.e. physical or chemical materials) are held in machines because there is no intermediate storage
Corresponding

author. Email: dwgong@vip.163.com

2014 Taylor & Francis

Downloaded by [National University of Sciences & Technology] at 23:22 30 October 2014

Y.-Y. Han et al.

(Suhami and Mah 1981). For various processes in a manufacturing enterprise (Grabowski and
Pempera 2000), the intermediate product cannot be stored in some stages. In the case of the iron
and steel industry (Gong, Tang, and Duin 2010), the blocking of ingots in the soaking pit will result
in extra energy consumption since the blocked ingot requires a high temperature. With respect to
the computational complexity, it has been proved that the BFS scheduling problem with more than
two machines is NP hard (Storn and Price 1997, Allahverdi, Ng, and Cheng 2008). Therefore, it
is of great importance to seek appropriate methods to tackle the BFS problem.
With regard to the algorithms for solving the BFS scheduling problem with a makespan criterion, the existing studies can be broadly classified into constructive heuristics and improved
metaheuristics. The former type uses specific rules to assign each job with a priority index to construct a sequence, such as NawazEnscoreHam (NEH), profile fitting (PF) and MinMax (MM).
Nawaz, Enscore, and Ham (1983) first designed an NEH heuristic for the traditional flow-shop
problem with a makespan criterion, and the experimental results demonstrated the effectiveness
and efficiency of the proposed heuristic in comparison with existing algorithms. Since then, the
NEH heuristic has gained much attention and has been successfully applied to the BFS scheduling problem (Ribas, Companys, and Tort-Martorell 2011). McCormich et al. (1989) developed
a constructive heuristic, known as PF, for solving permutation problems in an assembly line.
In this heuristic, PF tries to generate a partial sequence by adding an unscheduled job so as to
minimize the sum of idle and blocking time on machines. Ronconi (2004) presented the MM
heuristic based on the makespan properties. In addition, they proposed two composite constructive heuristics, called MME (the combination of MM and NEH) and PFE (the combination of PF
and NEH). Their empirical results showed that MME and PFE heuristics are superior to the NEH
heuristic. Recently, Pan and Wang (2012) designed two simple constructive heuristics for the BFS
scheduling problem, called the weighted profile fitting (wPF) and PanWang (PW) heuristics,
based on the PF approach. They also developed three improved constructive heuristics, namely,
PF-NEH, wPF-NEH and PW-NEH, by combining the procedure of the NEH heuristic with the
PF, wPF and PW, respectively. A series of BFS scheduling problem instances demonstrated that
the presented constructive heuristics perform significantly better than the existing ones.
The above constructive heuristics can rapidly yield feasible solutions. However, the quality
of these solutions is somewhat worse than that obtained by the metaheuristic algorithms (Li,
Wang, and Wu 2009). Caraffa et al. (2001) proposed a genetic algorithm (GA) to minimize the
makespan criterion for the BFS scheduling problem. In this work, each job sequence of the problem is viewed as a chromosome. Grabowski and Pempera (2007) presented tabu search algorithms
with and without multimoves (TS and TS + M, respectively) for the flow-shop scheduling problem with blocking, where a dynamic tabu list is used to overcome the local optima. Qian et al.
(2009) developed an effective hybrid differential evolution (HDE) to solve the flow-shop scheduling problem with limited buffers between consecutive machines. Thereafter, Wang et al. (2010)
designed a novel hybrid discrete differential evolution (HDDE) to optimize the makespan criterion
for the BFS scheduling problem. To avoid the regions with local optima, Davendra et al. (2012)
adopted an enhanced differential evolution (EDEc) to solve the flow-shop scheduling problem
with blocking based on the segregation bias rules. The proposed EDEc algorithm obtained 49 new
upper bounds for the Taillard problems. Since then, DE has been successfully applied to different
problems owing to its powerful performance (Ramesh, Kannan, and Baskar 2012; Zhu, Yan, and
Zhao 2013). In addition, for the same problem, Wang and Tang (2012) presented a discrete particle
swarm optimization (DPSO) to minimize the makespan objective of the BFS scheduling problem
with m machines. In this work, they adopted a self-adaptive strategy to control the diversity of
population. In addition, they provided a stochastic variable neighbourhood search approach to
improve the exploration.
The artificial bee colony (ABC) algorithm, one of the most recent swarm intelligence
approaches, was presented by Karaboga (2005). As the name implies, this algorithm simulates

Downloaded by [National University of Sciences & Technology] at 23:22 30 October 2014

Engineering Optimization

the foraging behaviour of honey bee colonies. ABC was commonly used for optimization problems with continuous variables, and showed promise in terms of accuracy and efficiency (Kang,
Li, and Ma 2013). However, owing to the continuous nature of the basic ABC algorithm, it
cannot be directly used to generate a feasible job permutation for the flow-shop scheduling
problem. To overcome such a drawback, Pan et al. (2011) first proposed a discrete artificial
bee colony (DABC) algorithm for the lot-streaming flow-shop scheduling problem. Thereafter, Han, Duan, and Zhang (2012) applied an improved DABC algorithm to tackle the BFS
scheduling problem, and the experimental results demonstrated that the proposed algorithm
outperforms the HDDE, DABC, GA, TS and TS + M algorithms in minimizing the makespan
criterion.
Among the aforementioned metaheuristic algorithms, the DDE and DABC algorithms are the
most powerful and have been successfully applied to the BFS scheduling problem. However, both
the employed and onlooker bees in the existing DABC algorithms adopted the insert or swap
operator to produce new neighbouring solutions, which may be local optima. Besides, the DABC
algorithm spends a lot of computation time on repeated search, which greatly reduces the convergence speed and the efficiency of the algorithm. To tackle these problems, a hybrid algorithm,
integrating DDE and DABC algorithms, was proposed. In the literature, hybrid algorithms often
obtain results of high quality. Fan, Liang, and Zahara (2004) employed a hybrid simplex search and
particle swarm optimization (PSO) to solve multimodal optimization problems. Xu et al. (2014)
embedded a local search into the artificial immune algorithm for solving a distributed permutation
flow-shop scheduling problem. Huang and Shiau (2008) presented a hybrid algorithm based on the
column generation and the constructive heuristic for a proportionate flexible flow-shop scheduling
problem. Following that, for the same problem, Shiau and Huang (2012) proposed a hybrid twophase encoding particle swarm optimization (TPEPSO) algorithm and the experimental results
demonstrated its robustness. Han et al. (2014) embedded the estimation of distribution algorithm
(EDA) into NSGA-II to replace traditional crossover and mutation operators and the experimental
results demonstrated the superiority of the hybrid algorithm in terms of quality. In this study, the
proposed hybrid algorithm, DE-ABC, was also empirically demonstrated to perform better in
exploration and exploitation. According to the notation introduced by Graham et al. (1979), the
BFS scheduling problem with makespan criterion under study is denoted as Fm|blocking|Cmax
throughout this article.
To sum up, in this study, the proposed algorithm makes three main contributions: (1) the
mutation and crossover operators are adopted to generate good solutions, instead of the insert
or swap operator in the employed bee stage; (2) the insert or swap operator based on the selfadaptive strategy is employed to modify the given solutions in the onlooker bee stage; and (3) the
insert-neighbourhood-based local search with a small probability is performed so as to improve
the algorithms capability in exploitation.
The remainder of this article is organized as follows. In Section 2, the description of
Fm|blocking|Cmax is given. Section 3 addresses the basic ABC algorithm. The proposed algorithm
is presented in detail in Section 4. Section 5 provides the experimental results. Finally, the article
ends with some conclusions in Section 6.

2.

Formulation of the blocking flow-shop scheduling problem

Table 1 lists the symbols and notations that will be used throughout this article.
The BFS scheduling problem with no buffers is formulated as follows. There are n jobs and m
machines. Each job from sequence has to be processed on m machines without intermediate
buffers in the same series. Job (j) has a sequence of m operations Oj,k . Operation Oj,k corresponds

Y.-Y. Han et al.

Downloaded by [National University of Sciences & Technology] at 23:22 30 October 2014

Table 1.

Symbols and notation.

n
m
PS
= {(1), (2), . . . , (n)}
(j)

i
Oj,k
Pj,k
Dj,k
Sj,0

Total number of jobs


Total number of machines
Population size
Job permutation
jth job of sequence
Set of all sequences,  = {1 , 2 , ..., PS }
ith sequence i {1, 2, ..., PS}
Operation of job (j = 1, 2, ..., n) on machine k, k = 1, 2, ...m
Processing time of job (j = 1, 2, ..., n) on machine k, k = 1, 2, ...m
Departure time of job (j = 1, 2, ..., n) on machine k, k = 1, 2, ...m
Start time of job (j = 1, 2, ..., n) on the first machine

to the processing of job (j) on machine k during an uninterrupted processing time, Pj,k . Moreover,
the following constraints are considered for the Fm|blocking|Cmax problem in this article:
At any time, each machine can process at most one job and each job can be processed on at
most one machine.
Since the flow shop has no intermediate buffers, a job cannot leave a machine until the next
machine downstream is available for processing.
Both the set-up time and transportation time of each job are included in the processing time.
The purpose is to seek a schedule for the processing sequence of jobs on all machines under the
above constraints so that the maximal completion time (i.e. makespan) is minimized. According
to Ronconi (2004), the departure time of each job on each machine can be calculated using the
following equations:
S1,0 = 0

(1)

D1,k = D1,k1 + P1,k


Sj,0 = Dj1,1

k = 1, 2, . . . , m 1

(2)

j = 2, 3, . . . . n

(3)

Dj,k = max{Dj,k1 + Pj,k , Dj1,k+1 }


Dj,m = Dj,m1 + Pj,m

k = 1, 2, . . . , m 1

j = 1, 2, . . . , n

j = 2, 3, . . . , n

(4)
(5)

Since the start time of the first job on the first machine, S1,0 , is 0, the makespan of the job
permutation, = {(1), (2), . . . , (n)}, is equal to the time when the last job in the processing
sequence is finished at machine m, and its value can be represented by Equation (6):
Cmax ( ) = Dn,m

(6)

Therefore, the objective of the formulated optimization problem can be formulated as follows:
min
 Cmax ( )

3.

(7)

Basic artificial bee colony algorithm

As the name implies, the basic ABC algorithm simulates the foraging behaviour of three kinds
of honey bee, i.e. employed bees, onlooker bees and scout bees. The ABC algorithm is an iterative process like other swarm intelligence-based algorithms (Karaboga and Akay 2009). First,
a population representing a number of food sources is randomly initialized; each food source is

Downloaded by [National University of Sciences & Technology] at 23:22 30 October 2014

Engineering Optimization

Figure 1.

Pseudo-code of the basic artificial bee colony (ABC) algorithm.

a candidate solution of an optimization problem. Then, these food sources are updated by three
kinds of honey bee. For the first kind, a new food source is generated based on the current one
using an employed bee, and its fitness value is calculated. With respect to the second kind, the
food source sharing with the employed bee is further modified using an onlooker bee. Then, the
fitness value of the modified food source is calculated. For the last kind, a new food source is
randomly produced to update the unchanged one in the entire population using the scout bee.
After a number of iterations, the best food source obtained so far is saved and taken as the optimal
solution of the optimization problem. Figure 1 depicts the pseudo-code of the above process.

4. The proposed differential evolutionartificial bee colony algorithm


In this study, a DE-ABC algorithm incorporating DABC and DDE algorithms is designed to
enhance the ability of DABC in exploration and exploitation. The proposed algorithm adopts
both the MME heuristic and the random method to initialize the population. Then, the employed
bees explore new and unknown areas in the search space using the DDE approach. Following
that, the onlooker bees share these solutions and adopt the insert or swap operator to modify them
according to a self-adaptive strategy. Finally, the scout bees perform several insert operators on
the unchanged solutions in the population and replace them.
4.1. Initializing population
The MME heuristic, first proposed by Ronconi and Armentano (2001), has aroused much interest
and been successfully applied to the BFS scheduling problem with the makespan criterion. The
MME heuristic contains two main components, i.e. the MM and NEH heuristics, and employs
the shortest critical path to reduce the blocking time of a job on machines. Because of the good
performance of the MME heuristic, an initial solution with high quality is generated by adopting
the MME heuristic in this study. In addition, to maintain the diversity of the initial population, the
other solutions are randomly generated in the entire search space. Figure 2 states the pseudo-code
of the MME heuristic.
With the above heuristic, the procedure of initializing a population is summarized as follows.
Step 1: Perform the MME heuristic given in Figure 2 to yield a solution.

Downloaded by [National University of Sciences & Technology] at 23:22 30 October 2014

Y.-Y. Han et al.

Figure 2.

Pseudo-code of the MME heuristic.

Step 2: Randomly generate a solution in the search space. If it is distinct from all the existing
solutions in the population, then put it into the population; otherwise, discard it.
Step 3: Repeat Step 2 until the population has PS individuals.
4.2. Exploring new and unknown areas using the discrete differential evolution approach
The insert and swap operators, through which the neighbouring solutions are generated, show the
superiority of the DABC algorithm. However, in the existing DABC algorithm, both the employed
and onlooker bees adopt the insert or swap operator to yield new solutions. Because of such a
repeated search, the neighbouring solutions obtained by them may be local optima. Thus, the
previous DABC algorithm finds it difficult to explore the entire solution space, which indicates
that it is weak in exploration. The merit of DDE is that it takes advantage of the differences among
individuals in the population to seek the global optimal solution. That is, DDE has the capability
to escape from the local optima. Therefore, the DDE algorithm is incorporated into DABC to
yield new solutions instead of the insert or swap operator in the employed bee stage. Since DDE
contains two critical operators, i.e. the mutation and crossover operators, an individual or job
permutation is first generated by performing the mutation operator. The job permutation, ,
itself may not include all jobs since some jobs may exist repeatedly, whereas others may be lost.
Therefore, the crossover operator is applied to generate a complete sequence of jobs. Figure 3
gives the pseudo-code of the mutation and crossover operators.
To simply illustrate the aforementioned steps, suppose that there are five jobs, and the values
of the parameters pmu and pc are equal to 0.9 and 0.1, respectively. Let ref = {1, 2, 3, 4, 5}. An
example of constructing a new solution using the mutation and crossover operators is shown in
the following.
(1) Randomly select three solutions, a = {2, 1, 4, 3, 5}, b = {4, 2, 5, 1, 3} and c =
{2, 5, 4, 3, 1}, from the population.
(2) Perform the mutation operator.
Set i = 1 and rand() = 0.8 < 0.9
(1) = (b (1) c (1)) 1 a (1) = ((4 2) + 2 + 5)%5 + 1 = 5

Downloaded by [National University of Sciences & Technology] at 23:22 30 October 2014

Engineering Optimization

Figure 3.

Pseudo-code of the mutation and crossover operators.

i = 2 and rand() = 0.6 < 0.9


(2) = (b (2) c (2)) 1 a (2) = ((2 5) + 1 + 5)%5 + 1 = 4
i = 3 and rand() = 0.95 > 0.9
(3) = (b (3) c (3)) 0 a (3) = (0 + 4 + 5)%5 + 1 = 5
i = 4 and rand() = 0.1 < 0.9
(4) = (b (4) c (4)) 1 a (4) = ((1 3) + 3 + 5)%5 + 1 = 2
i = 5 and rand() = 0.9
(5) = (b (5) c (5)) 0 a (5) = (0 + 5 + 5)%5 + 1 = 1
So the job permutation, = {5, 4, 5, 2, 1}, is generated using the mutation operator.
(3) Execute the crossover operator.
Set j = 1 and rand() = 0.3 > 0.1, so put the job, (1), into temp , that is, temp = {5}
Set j = 2 and rand() = 0.05 < 0.1, so no job is selected.
Set j = 3 and rand() = 0.6 > 0.1; as (3) already exists in temp , this job is discarded.
Set j = 4 and rand() = 0.08 < 0.1, so no job is selected.
Set j = 5 and rand() = 0.75 > 0.1, so put job (5) into temp , that is, temp = {5, 1}.
In the following, generate a sub-sequence, new , by removing the jobs included in temp from
ref , so new = {2, 3, 4}. Take job 5 from temp and insert it into the best position of new to obtain a
sub-sequence, new = {2, 3, 4, 5}. Similarly, insert job 1 into the current sub-sequence, and obtain
the complete sequence, new = {2, 3, 1, 5, 4}, with the smallest value of makespan.

Downloaded by [National University of Sciences & Technology] at 23:22 30 October 2014

Y.-Y. Han et al.

Figure 4.

Pseudo-code of the self-adaptive strategy.

4.3. Modifying the selected solutions using a self-adaptive strategy


The onlooker bees use the tournament selection to share some solutions generated by the employed
bees, and adopt the insert or swap operator to modify them based on a self-adaptive strategy so as to
search for outstanding neighbouring solutions. The insert operator randomly selects two positions,
p1 and p2 (p1 < p2), from a sequence and moves all jobs between the positions p1 + 1 and p2
forward a position in turn, whereas the swap operator just interchanges the corresponding jobs
between positions p1 and p2 (p1 < p2). Figure 4 describes the pseudo-code of the self-adaptive
strategy based on the insert and swap operators.
4.4.

Updating unchanged solutions

As addressed in the basic ABC algorithm, after the employed bees and onlooker bees have finished
their search, the algorithm checks whether or not there are any exhausted or unchanged solutions.
Next, these unchanged solutions are replaced with the new ones discovered by the scout bees.
The scout bees play a role in realizing the negative feedback mechanism and the fluctuation
property in the self-organization of the ABC algorithm (Akay and Karaboga 2012). In the basic
ABC algorithm, the scout bees randomly generate solutions to replace those unchanged ones.
This process enhances the diversity of the population, but it also reduces the search efficiency.
Therefore, each scout bee performs several insert operators on the unchanged solutions in this
study.
4.5. Local search
As a simple neighbourhood search approach, the local search seeks a better solution in the neighbourhood of a given solution. It has been shown that the insert-neighbourhood-based local search
(Ruben and Stutzle 2008) is superior to TS and the simulated annealing (SA) algorithm (Ruben and
Concepcion 2005). Therefore, this local search is adopted to enhance the ability of the proposed
algorithm in exploitation. In this study, the local search is performed on the solutions selected by

Downloaded by [National University of Sciences & Technology] at 23:22 30 October 2014

Engineering Optimization

Figure 5.

Pseudo-code of the local search.

the onlooker bees stage with a small probability of pls, which controls whether or not a solution
performs the local search. Specifically, a random value, r, is generated in the range of [0,1]. If r
is less than pls, then the solution performs the local search. Figure 5 states the pseudo-code of the
local search.
4.6.

Summary of the proposed algorithm

In summary, the steps of the proposed algorithm are as follows:


Step 1: Set the values of the parameters used in the proposed algorithm and initialize the
population by using MME and the random method stated in Subsection 4.1.
Step 2: Perform the mutation and crossover operators presented in Subsection 4.2 to produce
outstanding solutions.
Step 3: Modify the selected solutions using the self-adaptive strategy given in Subsection 4.3.
Step 4: Carry out the local search to seek a better solution from the neighbourhood of a given
solution with the probability of pls.
Step 5: Obtain new solutions by applying several insert operators to the unchanged solutions
in the whole population and replace the exhausted ones according to the approach described
in Subsection 4.4.
Step 6: Judge whether the termination criterion of the algorithm is met or not. If it is, stop the
evolution of the population and output the optimal solutions; otherwise, go to Step 2.

5.

Experiments

In this section, the proposed algorithm is applied to some typical instances of the BFS scheduling
problem and compared with six existing available algorithms to evaluate the performance of the
proposed algorithm. The proposed algorithm is written with C++, and implemented on a personal
computer with Pentium Dual 2.79 GHz and 1.96 GB of memory.
5.1. Experimental setting
There are 120 instances from the well-known benchmark set provided by Taillard (1993). These
instances have also been used by the existing algorithms by treating them as Fm|blocking|Cmax .
They are divided into 12 groups of different sizes, each consisting of 10 instances of the same

10

Y.-Y. Han et al.

size. The values of parameters, the size of instances and the maximal computation time adopted
by the proposed algorithm are set as follows:

Downloaded by [National University of Sciences & Technology] at 23:22 30 October 2014

The population size, PS, the mutation probability, pmu, the crossover probability, pc, and the
local search probability, pls, are set to 20, 0.9, 0.1 and 0.2, respectively.
The number of jobs for each instance is chosen from the set {20,50,100,200,500} and that of
machines is chosen from {5,10,20}.
For each instance, the maximal computation time is set to 5 m n milliseconds.
The proposed DE-ABC algorithm is compared with the existing metaheuristics, i.e. TS + M
(Grabowski and Pempera 2007), HDDE (Wang et al. 2010), DABC (Han et al. 2011), IABC (Han,
Duan, and Zhang 2012), DPSOsvns (Wang and Tang 2012) and EDEc (Davendra et al. 2012). For
each instance, each method is independently run five times, the minimal makespan is recorded and
compared with the referenced makespan taken from the branch-and-bound method of Ronconi
(2005), and the average relative percentage difference of the five runs is obtained. For all instances
in a group, such average relative percentage differences are computed and denoted as the ARPD.
The makespan of the jth instance provided by the ith algorithm in the tth run is denoted as
i
Cj,t
, the referenced value of which, provided by Ronconi (2005), is CjR . In addition, the average
relative percentage difference obtained by the ith algorithm, denoted as ARPDi , can be expressed
as follows:
5
10
i
1   CjR Cj,t
ARPDi =
100%
(8)
50 j=1 t=1
CjR
It is clear that the larger the value of ARPDi , the better the result produced by the algorithm.
During the solving of the above instances, if a better solution of an instance is obtained using
the proposed algorithm than that of the comparative algorithms, its upper bound is updated. The
computational results related to the following aspects are reported in the experiments:

comparison results between the proposed algorithm and the six comparative ones
update of upper bounds of some benchmark problems
convergence of different algorithms for six instances
non-parametric test on ARPD indicator
sensitivity analysis on parameters pmu, pc and pls used in the proposed algorithm.

5.2. Results and analysis


5.2.1.

Comparison of DABC, IABC and DE-ABC

In this section, the proposed DE-ABC algorithm is compared with two existing ABC algorithms,
DABC and IABC. The essential difference between DABC and DE-ABC is that the former uses
the insert or swap operator in the employed bee stage, whereas the latter adopts the mutation
and crossover operators of DDE. In addition to the above difference, for the IABC and DEABC algorithms, the other difference is that the latter employs only one of the four initialization
strategies provided by the former. Although the four initialization strategies of IABC have been
recognized as superior to the single one, DDE is beneficial in generating good solutions if the
proposed algorithm is better than IABC. Thus, this section provides the following experimental
results to demonstrate the performance of DDE in exploration.
Table 2 reports the corresponding comparison results on the premise of the same computational
time and experimental environment. From Table 2: (1) for instances with a small scale, such as
20 5, 20 10 and 20 20, the proposed algorithm is slightly inferior to DABC and IABC
in the value of ARPD; (2) except for the above cases, the proposed algorithm is superior to

Engineering Optimization

11

Table 2. Average relative percentage difference (ARPD) values of DABCa , IABCa and DE-ABCa .

Downloaded by [National University of Sciences & Technology] at 23:22 30 October 2014

DABC

IABC

DE-ABC

nm

ARPD

MinAD

MaxAD

ARPD

MinAD

MaxAD

ARPD

MinAD

MaxAD

Time (s)

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

0.44
2.40
3.30
4.60
6.09
6.27
2.08
5.63
5.11
3.49
3.91
3.25
3.88

0.40
2.38
3.30
4.29
5.71
5.99
1.74
5.40
4.73
3.22
3.62
3.10
3.65

0.46
2.40
3.30
4.90
6.44
6.54
2.43
5.88
5.41
3.81
4.23
3.46
4.11

0.41
2.38
3.30
4.68
6.15
6.29
2.31
5.97
5.39
4.14
4.46
3.70
4.10

0.32
2.35
3.29
4.32
5.80
6.03
2.07
5.68
5.13
3.86
4.24
3.56
3.89

0.46
2.39
3.30
5.03
6.46
6.57
2.54
6.34
5.72
4.44
4.67
3.89
4.32

0.34
2.34
3.29
4.89
6.25
6.45
2.67
6.27
5.53
4.37
4.70
3.72
4.24

0.21
2.25
3.26
4.52
5.93
6.15
2.29
5.93
5.15
3.91
4.45
3.51
3.96

0.46
2.40
3.30
5.21
6.55
6.72
3.09
6.62
5.89
4.74
5.02
3.94
4.50

0.50
1.00
2.00
1.25
2.50
5.00
2.50
5.00
10.00
10.00
20.00
50.00
9.15

Note: Data in italic indicate the best value of ARPD for each group of instances among these algorithms.
a
Pentium Dual 2.79 GHz and 1.96 GB memory.

the comparative ones; and (3) on average, the ARPD value of DE-ABC is equal to 4.22, much
better than the values of 3.88 and 4.10 gained by DABC and IABC, respectively. The minimal
and maximal values of ARPD (MinAD and MaxAD, respectively) produced by DE-ABC are also
much larger than those generated by the DABC and IABC algorithms. Furthermore, the superiority
of the DE-ABC algorithm over the comparative ones increases along with the increment of the
problem size.
The following conclusion can be drawn from the above experimental results: the proposed
algorithm can take full advantage of the differentiation information among the population in the
employed bee stage to enhance the capability in exploration.
5.2.2.

Comparison of EDEc, HDDE and DE-ABC

The HDDE and EDEc algorithms proposed by Wang et al. (2010) and Davendra et al. (2012),
respectively, can obtain better solutions than some existing algorithms, i.e. TS + M, GA and HDE,
suggesting that the algorithms based on DDE are more powerful. Although DDE has the capability
in exploration, it lacks the capability in exploitation. However, DABC can successfully search
for local optima in the neighbourhood of a solution, so the DE-ABC algorithm, by incorporating
DDE with DABC, can balance the capability in exploration and exploitation, which can be verified
from Table 3.
As illustrated in Table 3: (1) for instances of 20 5, 20 10 and 20 20, the proposed
algorithm is slightly inferior to EDEc and HDDE; (2) except for the above cases and the 100 5
instance, the proposed algorithm outperforms the comparative ones; (3) the ARPD value, 4.22,
generated by DE-ABC is much larger than the values of 3.79 and 3.75 obtained by EDEc and
HDDE, respectively, indicating that the hybrid algorithm based on DDE and DABC does better
in exploration and exploitation than the above comparative ones.
From the above experimental results, it can be concluded that DE-ABC overcomes the shortage
of DDE in exploitation by incorporating DABC and obtains solutions with high quality.
5.2.3.

Comparison of TS + M, DPSOsvns and DE-ABC

For blocking scheduling problems, there are two typical algorithms, i.e. TS + M and DPSOsvns .
The former is a benchmark and the latter has been recently developed. In this study, the proposed

12

Y.-Y. Han et al.

Table 3. Average relative percentage difference (ARPD) values of HDDEa , DE-ABCa and EDEcb .
EDEc

Downloaded by [National University of Sciences & Technology] at 23:22 30 October 2014

nm

HDDE

DE-ABC

ARPD

MinAD

MaxAD

ARPD

MinAD

MaxAD

ARPD

MinAD

MaxAD

0.46
2.39
3.30
4.80
5.78
5.82
3.23
5.92
4.74
3.23
3.32
2.56
3.79

0.00
0.57
2.13
3.03
4.50
3.68
1.46
3.62
3.28
2.58
3.09
2.12
2.51

1.83
5.36
4.48
6.06
6.95
8.13
3.92
6.24
5.25
4.84
3.73
3.05
4.98

0.45
2.38
3.30
4.58
5.94
6.19
1.78
5.44
4.86
3.41
3.83
2.91
3.75

0.42
2.37
3.29
4.37
5.66
6.02
1.47
5.22
4.66
3.21
3.65
2.72
3.59

0.46
2.39
3.30
4.85
6.19
6.44
2.09
5.72
5.14
3.68
4.08
3.13
3.97

0.34
2.34
3.29
4.89
6.25
6.45
2.67
6.27
5.53
4.37
4.70
3.72
4.24

0.21
2.25
3.26
4.52
5.93
6.15
2.29
5.93
5.15
3.91
4.45
3.51
3.96

0.46
2.40
3.30
5.21
6.55
6.72
3.09
6.62
5.89
4.74
5.02
3.94
4.50

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

Note: Data in italic indicate the best value of ARPD for each group of instances among these algorithms.
a
Pentium Dual 2.79 GHz and 1.96 GB memory.
b
MacBook Pro, 2.4 GHz Intel Core 2 Duo, 2 GB RAM.

Table 4. Average relative percentage difference (ARPD) values of TS + Ma , DPSObsvns and DE-ABC.
TS + M

DE-ABC

DPSOsvns

nm

ARPD

Time (s)

ARPD

Time (s)

ARPD

MinAD

MaxAD

Time (s)

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

0.24
1.77
2.94
0.55
3.52
4.26
2.62
2.66
3.02
0.58
2.31
1.47
1.68

2.70
4.60
7.60
6.20
10.80
19.30
12.40
22.10
39.40
44.30
79.40
209.00
38.15

0.07
1.94
3.12
3.36
4.25
4.60
1.48
4.42
3.38
2.53
2.82
1.63
2.79

1.00
2.00
3.00
2.50
5.00
10.00
5.00
10.00
20.00
20.00
40.00
100.00
18.20

0.34
2.34
3.29
4.89
6.25
6.45
2.67
6.27
5.53
4.37
4.70
3.72
4.24

0.21
2.25
3.26
4.52
5.93
6.15
2.29
5.93
5.15
3.91
4.45
3.51
3.96

0.46
2.40
3.30
5.21
6.55
6.72
3.09
6.62
5.89
4.74
5.02
3.94
4.50

0.50
1.00
2.00
1.25
2.50
5.00
2.50
5.00
10.00
10.00
20.00
50.00
9.15

Note: Data in italic indicate the best value of ARPD for each group of instances among these algorithms.
a
Pentium P-IV, 1000 MHz and 30,000 iterations.
b
Intel 2.33 GHz CPU and 2 GB memory.

algorithm is compared with these, and associated comparison results are reported in Table 4. It
should be noted that the TS + M algorithm employed 30,000 iterations on an Intel 1 GHz CPU,
whereas both DPSOsvns and DE-ABC performed fewer than 2000 iterations on a Pentium 2.33
and 2.79 GHz CPU, respectively. Although the implementation environment in this study is better
than that of TS + M, fewer iterations are adopted here. Therefore, it can be seen that the value of
ARPD produced by DE-ABC is comparable with that of TS + M.
From Table 4: (1) for all instances, DE-ABC clearly prevails over TS + M and DPSOsvns in the
value of ARPD; and (2) even the MinAD value of DE-ABC is much larger than that generated
by the TS + Mand DPSOsvns algorithms. These values demonstrate that the proposed algorithm
performs better in the value of ARPD than the comparative algorithms.

Engineering Optimization

Downloaded by [National University of Sciences & Technology] at 23:22 30 October 2014

5.2.4.

13

New upper bounds

The excellent performance of the proposed algorithm encouraged the authors to seek new upper
bounds of Taillards benchmarks for the Fm|blocking|Cmax scheduling problem. Ronconi (2005)
first reported the upper bounds of 120 instances; following that, Grabowski and Pempera (2007)
obtained 94 out of 120 upper bounds using the TS+M algorithm. Later, more algorithms, i.e.
HDDE, DABC, DPSOsvns , EDEc and IABC, provided newer upper bounds. Thus, in Tables 5 and
6, for each instance, the best solution yielded by the proposed algorithm is reported to update
some upper bounds.
As stated in Tables 5 and 6: (1) for the instances with a small scale, i.e. 20 5, 20 10 and
20 20, 30 best values of makespan obtained by DE-ABC are equal to the corresponding best
ones yielded by HDDE, DABC, DPSOsvns , EDEc and IABC, respectively; (2) for the remaining
90 instances, zero, six (6/90 = 7%), one (1/90 = 1%), zero, 11 (11/90 = 12%) and 10 (10/90 =
11%) new solutions produced by TS + M, DABC, HDDE, DPSOsvns , EDEc and IABC are found,
respectively, while 62 out of 90 (62/90 = 69%) new upper bounds provided by the proposed
algorithm are further improved. These experimental results clearly demonstrate that the proposed
algorithm can update more upper bounds of Taillards benchmarks than the comparative ones.
5.2.5.

Convergence curves of different algorithms

To further evaluate the makespan values of different algorithms, in this section, the convergence
curves of the makespan values are investigated. For the Fm|blocking|Cmax scheduling problem,
Ta53, Ta72, Ta86, Ta96, Ta109 and Ta115 of Taillards benchmarks are selected, and their best values of makespan obtained by HDDE, DABC, DPSOsvns , EDEc, IABC and DE-ABC are calculated,
respectively, with the increment of computation time, shown as Figures 68.
From Figures 68: (1) for each instance, each algorithm can obtain better and better makespan
values as the computation time increases; and (2) for the same computation time, the makespan
values of the proposed algorithm are the smallest among these algorithms.
In summary, compared with the other algorithms for the Fm|blocking|Cmax scheduling problem,
the algorithm proposed in this study does better in both exploration and exploitation. Therefore,
it can achieve more outperforming solutions.
5.2.6. Non-parametric testing in the value of ARPD
To evaluate whether the differences in the value of ARPD between the algorithm proposed in this
study and the comparative ones are significant, the MannWhitney U distribution test, a nonparametric testing method, was conducted. Table 7 lists the experimental results; the symbols +
and in the table denote that the proposed algorithm is significantly superior and inferior to
the comparative method, respectively, whereas the symbol 0 indicates that there is no significant
difference between them.
Table 7 reports that: (1) the proposed algorithm is significantly inferior to HDDE, DABC, EDEc
and IABC for the instance of 20 5; (2) for the 20 10 and 20 20 instances, the proposed
algorithm is not significantly different from the comparative algorithms except for TS + M; (3)
for the 100 5 instance, the proposed algorithm is inferior to EDEc; (4) there is no significant
difference between DE-ABC and IABC for the 500 20 instance; and (5) except for the above
cases, the proposed algorithm is significantly superior to TS + M, HDDE, DPSOsvns , DABC,
EDEc and IABC.
From the above experimental results, it can be concluded that the proposed algorithm is
significantly superior to the comparative ones for most instances.

14

Y.-Y. Han et al.

Downloaded by [National University of Sciences & Technology] at 23:22 30 October 2014

Table 5. Values of makespan for the instances with n = 20 and 50.


Data set

Instance

TS + M

DABC

HDDE

DPSOsvns

EDEc

IABC

DE-ABC

20 5

Ta01
Ta02
Ta03
Ta04
Ta05
Ta06
Ta07
Ta08
Ta09
Ta10

1387
1408
1280
1448
1341
1363
1381
1379
1373
1283

1374
1408
1280
1448
1341
1363
1381
1379
1373
1283

1374
1408
1280
1448
1341
1363
1381
1379
1373
1283

1374
1408
1280
1448
1341
1363
1381
1379
1373
1283

1374
1408
1280
1448
1341
1363
1381
1379
1373
1283

1374
1408
1280
1448
1341
1363
1381
1379
1373
1283

1374
1408
1280
1448
1341
1363
1381
1379
1373
1283

20 10

Ta11
Ta12
Ta13
Ta14
Ta15
Ta16
Ta17
Ta18
Ta19

1698
1836
1674
1555
1631
1603
1629
1754
1759

1698
1833
1659
1535
1617
1590
1622
1731
1749

1698
1833
1659
1535
1617
1590
1622
1731
1749

1698
1833
1659
1535
1617
1590
1622
1731
1749

1698
1833
1659
1535
1617
1590
1622
1731
1749

1698
1833
1659
1535
1617
1590
1622
1731
1749

1698
1833
1659
1535
1617
1590
1622
1731
1749

Ta20
Ta21
Ta22
Ta23
Ta24
Ta25
Ta26
Ta27
Ta28
Ta29
Ta30

1782
2449
2242
2483
2348
2450
2398
2397
2345
2363
2334

1782
2436
2234
2479
2348
2435
2383
2390
2328
2363
2323

1782
2436
2234
2479
2348
2435
2383
2390
2328
2363
2323

1782
2436
2234
2479
2348
2435
2383
2390
2328
2363
2323

1782
2436
2234
2479
2348
2435
2383
2390
2328
2363
2323

1782
2436
2234
2479
2348
2435
2383
2390
2328
2363
2323

1782
2436
2234
2479
2348
2435
2383
2390
2328
2363
2323

50 5

Ta31
Ta32
Ta33
Ta34
Ta35
Ta36
Ta37
Ta38
Ta39
Ta40

3163
3348
3173
3277
3338
3330
3168
3228
3068
3285

3009
3215
3040
3142
3180
3192
3046
3084
2923
3134

3027
3225
3032
3145
3188
3190
3049
3081
2917
3127

3038
3233
3050
3181
3196
3213
3055
3084
2935
3137

3028
3227
3031
3140
3173
3194
3053
3089
2936
3127

3021
3203
3022
3128
3186
3193
3031
3081
2921
3135

3016
3219
3025
3136
3162
3177
3024
3059
2921
3125

50 10

Ta41
Ta42
Ta43
Ta44
Ta45
Ta46
Ta47
Ta48
Ta49
Ta50

3776
3641
3588
3786
3745
3747
3778
3708
3668
3729

3654
3508
3487
3687
3643
3607
3707
3587
3527
3637

3665
3522
3497
3682
3650
3621
3716
3591
3553
3642

3686
3548
3517
3732
3669
3674
3729
3615
3573
3672

3681
3534
3524
3710
3663
3634
3731
3603
3563
3656

3661
3512
3497
3680
3636
3600
3708
3572
3541
3631

3647
3496
3495
3672
3637
3619
3688
3589
3534
3623

50 20

Ta51
Ta52
Ta53
Ta54
Ta55
Ta56
Ta57
Ta58
Ta59
Ta60

4627
4411
4388
4479
4359
4372
4402
4444
4423
4609

4515
4307
4278
4382
4282
4295
4324
4322
4318
4429

4516
4298
4282
4379
4283
4301
4325
4341
4328
4443

4547
4349
4343
4456
4283
4334
4371
4384
4343
4520

4541
4329
4319
4416
4304
4328
4369
4361
4365
4452

4521
4289
4289
4378
4278
4290
4312
4328
4320
4433

4517
4278
4266
4344
4276
4299
4311
4336
4323
4415

20 20

Note: Data in bold indicate the best upper bound for each instance among the comparative algorithms.

Engineering Optimization

15

Downloaded by [National University of Sciences & Technology] at 23:22 30 October 2014

Table 6. Values of makespan for the instances with n = 100, 200 and 500.
Data set

Instance

TS + M

DABC

HDDE

DPSOsvns

EDEc

IABC

DE-ABC

100 5

Ta61
Ta62
Ta63
Ta64
Ta65
Ta66
Ta67
Ta68
Ta69
Ta70

6639
6481
6299
6120
6340
6244
6346
6289
6559
6509

6226
6117
6017
5835
6056
5910
6074
5997
6201
6247

6274
6121
6035
5839
6060
5956
6101
6022
6234
6247

6270
6098
6018
5842
6026
5951
6094
5964
6221
6209

6155
6042
5941
5772
5997
5876
6008
5908
6153
6160

6232
6105
6007
5810
6056
5903
6063
5980
6218
6237

6179
6072
5976
5783
6001
5891
6049
5948
6187
6195

100 10

Ta71
Ta72
Ta73
Ta74
Ta75
Ta76
Ta77
Ta78
Ta79
Ta80

7320
7108
7233
7413
7168
6993
7092
7143
7327
7299

7119
6824
6959
7213
6899
6719
6874
6927
7144
7037

7140
6853
6957
7246
6895
6754
6869
6931
7133
7056

7146
6882
6997
7283
6940
6768
6924
6961
7192
7086

7053
6806
6925
7201
6853
6702
6841
6897
7059
6990

7097
6806
6951
7173
6852
6697
6835
6886
7090
6982

7069
6781
6911
7188
6840
6661
6826
6859
7039
6993

100 20

Ta81
Ta82
Ta83
Ta84
Ta85
Ta86
Ta87
Ta88
Ta89
Ta90

8101
8105
8071
8081
8074
8151
8273
8248
8116
8261

7866
7919
7918
7875
7899
7940
8009
8059
7980
8005

7878
7931
7944
7923
7900
7971
8031
8086
8018
8017

8068
8015
7985
7999
7977
8054
8122
8185
8084
8102

7894
7902
7897
7919
7905
7979
8041
8061
7984
8025

7829
7881
7873
7883
7873
7923
8002
8042
7952
7945

7841
7883
7856
7857
7859
7870
7993
8026
7928
7973

200 10

Ta91
Ta92
Ta93
Ta94
Ta95
Ta96
Ta97
Ta98
Ta99
Ta100

14,220
14,089
14,149
14,156
14,130
13,963
14,386
14,256
13,954
14,224

13,711
13,647
13,732
13,678
13,666
13,410
13,901
13,798
13,626
13,764

13,722
13,596
13,762
13,686
13,702
13,492
13,925
13,839
13,656
13,733

13,806
13,721
13,867
13,706
13,832
13,517
13,943
13,885
13,726
13,818

13,725
13,624
13,721
13,675
13,723
13,420
13,898
13,809
13,585
13,703

13,622
13,567
13,632
13,605
13,528
13,352
13,859
13,722
13,507
13,638

13,617
13,514
13,584
13,532
13,504
13,311
13,814
13,644
13,505
13,623

200 20

Ta101
Ta102
Ta103
Ta104
Ta105
Ta106
Ta107
Ta108
Ta109
Ta110

15,334
15,522
15,713
15,687
15,443
15,472
15,522
15,540
15,394
15,523

15,006
15,229
15,329
15,233
15,149
15,224
15,199
15,322
15,162
15,239

15,057
15,263
15,360
15,276
15,123
15,223
15,288
15,283
15,207
15,254

15,117
15,309
15,495
15,452
15,257
15,370
15,289
15,402
15,337
15,379

15,111
15,326
15,440
15,400
15,215
15,372
15,327
15,410
15,280
15,357

14,931
15,180
15,218
15,188
15,076
15,220
15,185
15,203
15,093
15,132

14,901
15,087
15,205
15,109
15,001
15,108
15,113
15,139
15,066
15,141

500 20

Ta111
Ta112
Ta113
Ta114
Ta115
Ta116
Ta117
Ta118
Ta119
Ta120

37,860
38,044
37,732
38,062
37,991
38,132
37,561
37,750
37,730
38,014

36,960
37,359
36,827
37,328
37,032
37,221
36,970
37,219
37,060
37,243

37,206
37,485
37,142
37,291
37,267
37,477
37,133
37,161
36,939
37,389

37,614
37,965
37,477
37,986
37,592
37,969
37,528
37,790
37,455
37,818

37,139
37,474
37,154
37,471
37,162
37,372
37,324
37,562
37,112
37,294

36,908
37,192
36,872
37,245
36,936
37,244
36,845
37,045
36,816
37,137

36,865
37,132
36,778
37,132
36,722
37,131
36,825
37,019
36,885
37,056

Note: Data in bold indicate the best upper bound for each instance among the comparative algorithms.

16

Y.-Y. Han et al.


4700

4600

7200

4500
4450
4400

7100
7000
6900

4350
4300

6800

4250
4200

10

15

6700

(s)
25

20

10

15

Figure 6.

25

30

35

40

(s)
50

45

Convergence curves of instances Ta 53 and Ta72.

x 104
1.41

8400
DE-ABC
DABC
HDDE
IABC
DPSOsvns
EDEc

1.4
1.39
1.38

DE-ABC
DABC
HDDE
IABC
DPSOsvns
EDEc

8300

8200
Makespan

Makespan

20

Ta 72 (10010 problem)

Ta 53 (5020 problem)

1.37
1.36
1.35

8100

8000
1.34
1.33

7900

1.32
1.31

10

15

20

25

30

35

40

45

(s)
50

7800

10

Figure 7.

1.59

15

20

(s)
30

25

Ta 86 (10020 problem)

Ta 96 (20010 problem)

Convergence curves of instances Ta86 and Ta96.

x 104

x 104
3.84

DE-ABC
DABC
HDDE
IABC
DPSOsvns
EDEc

1.58
1.57

DE-ABC
DABC
HDDE
IABC
DPSOsvns
EDEc

3.82
3.8

1.56
3.78
1.55

Makespan

Makespan

Downloaded by [National University of Sciences & Technology] at 23:22 30 October 2014

DE-ABC
DABC
HDDE
IABC
DPSOsvns
EDE

7300

makespan

4550

makespan

7400

DE-ABC
DABC
HDDE
IABC
DPSOsvns
EDE

4650

1.54
1.53

3.74
3.72

1.52

3.7

1.51

3.68

1.5
1.49

3.76

10

20

30

40

50

60

70

80

90

(s)
100

Ta 109 (20020 problem)

Figure 8.

Convergence curves of instances Ta109 and Ta115.

3.66

20

40

60

80

100

Ta 115 (50020 prolbem)

120

(s)
140

Engineering Optimization
Table 7.

Non-parametric testing results in the average relative percentage difference (ARPD) value.

DE-ABC vs

TS + M

HDDE

DPSOsvns

DABC

EDEc

IABC

+
+
+
+
+
+
+
+
+
+
+
+

0
0
+
+
+
+
+
+
+
+
+

+
0
0
+
+
+
+
+
+
+
+
+

0
0
+
+
+
+
+
+
+
+
+

0
0
0
+
+

+
+
+
+
+

0
0
+
+
+
+
+
+
+
+
0

20 5
20 10
20 20
50 5
50 10
50 20
100 5
100 10
100 20
200 10
200 20
500 20

4.3

4.25

ARPD

Downloaded by [National University of Sciences & Technology] at 23:22 30 October 2014

17

4.2

4.15

4.1
0.1

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

Parameter pmu
Figure 9.

5.2.7.

Influence of parameter pmu on the average relative percentage difference (ARPD).

Sensitivity analysis on parameters pmu, pls and pc

There are three important parameters, pmu, pc and pls, in the proposed algorithm. To investigate
the influences of these parameters on the performances of DE-ABC, their values were changed
from 0 to 1.0 with a step size of 0.1, and the corresponding values of ARPD were obtained and
are shown in Figures 911. In these figures, the horizontal coordinate denotes the value of a
parameter, and the vertical coordinate represents the average value of ARPD for 120 instances.
Figure 9 illustrates that: (1) the trajectory tendency is relatively stable when the value of pmu
varies in the range of [0, 0.1]; (2) the value of ARPD gradually reduces to a low level when the
value of pmu changes from 0.1 to 0.2 and from 0.3 to 0.8; and (3) when pmu is equal to 0.9, the
ARPD value reaches the largest value, 4.22. It can be concluded that the parameter pmu has no
evident influence on the value of ARPD of the proposed algorithm. Therefore, in this study, the
value of pmu is set as 0.9.

18

Y.-Y. Han et al.

4.4
4.35
4.3

ARPD

4.2
4.15
4.1
4.05
4
0.1

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

Parameter pc
Figure 10.

Influence of parameter pc on the average relative percentage difference (ARPD).

4.4

4.2

ARPD

Downloaded by [National University of Sciences & Technology] at 23:22 30 October 2014

4.25

3.8

3.6

3.4

3.2
-0.1

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

Parameter pls
Figure 11.

Influence of parameter pls on the average relative percentage difference (ARPD).

Figure 10 reports that: (1) when the value of pc is in the range of [0.3, 1], the value of ARPD
becomes smaller and smaller along with the increment in pc. The reason is that as the solution
obtained by the mutation operator can be considered with a smaller and smaller probability,

Downloaded by [National University of Sciences & Technology] at 23:22 30 October 2014

Engineering Optimization

19

equivalently, the mutation operator works with the same probability; and (2) if pc = 1, then the
value of ARPD is the smallest, indicating that only the crossover operator is used to generate a
new solution based on the current reference one, resulting in trapping into local optima of the
optimization problem. Thus, the value of pc is set as 0.1 in this study.
Figure 11 demonstrates that: (1) when the value of pls is equal to 0, suggesting that the selected
solutions do not undergo the local search at all, the proposed algorithm will have a bad performance; and (2) the larger the value of pls, the smaller the value of ARPD obtained by the proposed
algorithm. This is the reason why the selected solutions will perform the local search with a large
probability, which needs to spend a long time and loses opportunities to generate promising solutions by a number of iterations. Therefore, an appropriate value of pls leading to a good ARPD is
set as 0.2.

6.

Conclusions

Since the BFS scheduling problem plays a key role in real-world applications, it is necessary to
develop effective methods for this problem. In recent years, both ABC and DE have proven to be
effective algorithms for solving the BFS scheduling problem. Thus, a novel algorithm, DE-ABC,
is presented in this study by incorporating the merits of DDE with those of DABC to solve the
BFS problem with makespan criterion. First, the mutation and crossover operators with powerful
exploration are adopted to generate solutions in the employed bee stage, and then the insert or swap
operator based on the self-adaptive strategy is proposed to modify the selected solutions in the
onlooker bee stage. Finally, an efficient local search operator based on the insert neighbourhood
is designed to improve the algorithms capability in exploitation.
The performances of the proposed algorithm in this article are measured on a set of 120
benchmark instances proposed by Taillard. The experimental results demonstrate the superiority
of the proposed algorithm in terms of the ARPD indicator by comparison with the results yielded
by some algorithms presented in the literature, such as TS + M, DPSOsvns , DABC, HDDE, EDEc
and IABC.
It is worth mentioning that the excellent performance of the proposed algorithm in exploration
and exploitation may be attributed to the combination of DDE, the insert or swap operator based
on the self-adaptive strategy and the local search. Future work could apply the proposed algorithm
to scheduling problems with other criteria and extend the ideas proposed in this study to multiobjective scheduling problems.
Funding
This research is jointly supported by the National Natural Science Foundation of China [grant no. 61375067]; Fundamental
Research Funds for the Central Universities [grant no. 2013XK09]

References
Akay, B., and D. Karaboga. 2012. A Modified Artificial Bee Colony Algorithm for Real-Parameter Optimization.
Information Sciences 192: 120142.
Allahverdi, A., C. T. Ng, and T. C. E. Cheng. 2008. A Survey of Scheduling Problems with Setup Times or Costs.
European Journal of Operational Research 187: 9851032.
Caraffa, V., S. Ianes, T. P. Bagchi, and C. Sriskandarajah. 2001. Minimizing Makespan in a Blocking Flowshop Using
Genetic Algorithms. International Journal of Production Economics 70: 102115.
Davendra, D., I. Zelinka, M. Bialic-Davendra, R. Senkerik, and R. Jasek. 2012. Clustered Enhanced Differential Evolution
for the Blocking Flow Shop Scheduling Problem. Central European Journal of Operations Research 20: 679717.
Fan, S. K. S., Y. C. Liang, and E. Zahara. 2004. A Hybrid Simplex Search and Particle Swarm Optimization for the
Global Optimization of Multimodal Functions. Engineering Optimization 36 (4): 401418.

Downloaded by [National University of Sciences & Technology] at 23:22 30 October 2014

20

Y.-Y. Han et al.

Gong, H., L. X. Tang, and C. W. Duin. 2010. A Two-Stage Flow Shop Scheduling Problem on a Batching Machine and
a Discrete Machine with Blocking and Shared Setup Times. Computers & Operations Research 37: 960969.
Grabowski, J., and J. Pempera. 2000. Sequencing of Jobs in Some Production System. European Journal of Operational
Research 125: 535550.
Grabowski, J., and J. Pempera. 2007. The Permutation Flow Shop Problem with Blocking. A Tabu Search Approach.
Omega 35: 302311.
Graham, R., E. Lawler, J. Lenstra, and A. R. Kan. 1979. Optimization and Approximation in Deterministic Sequencing
and Scheduling: A Survey. Annals of Discrete Mathematics 5 (1): 287326.
Han, Y. Y., J. H. Duan, and M. Zhang. 2011. Apply the Discrete Artificial Bee Colony Algorithm to the Blocking Flow
Shop Problem with Makespan Criterion. In proceeding of Chinese Control and Decision Conference 21312135.
Mianyang.
Han, Y. Y., D. W. Gong, X. Y. Sun, and Q. K. Pan. 2014. An Improved NSGA-II Algorithm for Multi-Objective LotStreaming Flow Shop Scheduling Problem. International Journal of Production Research 52 (8): 22112231.
Han, Y. Y., Q. K. Pan, J. Q. Li, and H. Y. Sang. 2012. An Improved Artificial Bee Colony Algorithm for the Blocking
Flowshop Scheduling Problem. International Journal of Advanced Manufacturing Technology 60: 11491159.
Huang, Y. M., and D. F. Shiau. 2008. Combined Column Generation and Constructive Heuristic for a Proportionate
Flexible Flow Shop Scheduling. International Journal of Advanced Manufacturing Technology 38: 691704.
Kang, F., J. J. Li, and Z. Y. Ma. 2013. An Artificial Bee Colony Algorithm for Locating the Critical Slip Surface in Slope
Stability Analysis. Engineering Optimization 49 (2): 207223.
Karaboga, D. 2005. An Idea Based on Honey Bee Swarm for Numerical Optimization. Technical Report TR06. Turkey:
Computer Engineering Department, Erciyes University.
Karaboga, D., and B. Akay. 2009. A Comparative Study of Artificial Bee Colony Algorithm. Applied Mathematics and
Computation 214: 108132.
Li, X. P., Q. Wang, and C. Wu. 2009. Efficient Composite Heuristics for Total Flowtime Minimization in Permutation
Flow Shops. Omega 37: 155164.
McCormich, S. T., M. L. Pinedo, S. Shenker, and B. Wolf. 1989. Sequencing in an Assembly Line with Blocking to
Minimize Cycle Time. Operations Research 37: 925936.
Nawaz, M., E. E. J. Enscore, and I. Ham. 1983. A Heuristic Algorithm for the m-Machine, n-Job Flow Shop Sequencing
Problem. International Journal of Management Science 11: 9195.
Pan, Q. K., M. F. Tasgetiren, P. N. Suganthan, and T. J. Chua. 2011. A Discrete Artificial Bee Colony Algorithm for the
Lot-Streaming Flow Shop Scheduling Problem. Information Sciences 181: 24552468.
Pan, Q. K., and L. Wang. 2012. Effective Heuristics for the Blocking Flowshop Scheduling Problem with Makespan
Minimization. Omega 40: 218229.
Qian, B., L. Wang, D. X. Huang, W. L. Wang, and X. Wang. 2009. An Effective Hybrid DE-Based Algorithm for
Multi-Objective Flow Shop Scheduling with Limited Buffers. Computers & Operations Research 36 (1): 209233.
Ramesh, S., S. Kannan, and S. Baskar. 2012. An Improved Generalized Differential Evolution Algorithm for MultiObjective Reactive Power Dispatch. Engineering Optimization 44 (4): 391405.
Ribas, I., R. Companys, and X. Tort-Martorell. 2011. An Iterated Greedy Algorithm for the Flowshop Scheduling Problem
with Blocking. Omega 39: 293301.
Ronconi, D. P. 2004. A Note on Constructive Heuristics for the Flowshop Problem with Blocking. International Journal
of Production Economics 87: 3948.
Ronconi, D. P. 2005. A Branch-and-Bound Algorithm to Minimize the Makespan in a Flowshop Problem with Blocking.
Annals of Operations Research 138: 5365.
Ronconi, D. P., and V. A. Armentano. 2001. Lower Bounding Schemes for Flowshops with Blocking in-Process. Journal
of the Operational Research Society 52: 12891297.
Ruben, R., and M. Concepcion. 2005. A Comprehensive Review and Evaluation of Permutation Flowshop Heuristics.
European Journal of Operational Research 165: 479494.
Ruben, R., and T. Stutzle. 2008. An Iterated Greedy Heuristic for the Sequence Dependent Setup Times Flowshop Problem
with Makespan and Weighted Tardiness Objectives. European Journal of Operational Research 187: 11431159.
Shiau, D. F., and Y. M. Huang. 2012. A Hybrid Two-Phase Encoding Particle Swarm Optimization for Total Weighted
Completion Time Minimization in Proportionate Flexible Flow Shop Scheduling. International Journal of Advanced
Manufacturing Technology 58: 339357.
Storn, R., and K. Price. 1997. Differential EvolutionA Simple and Efficient Adaptive Scheme for Global Optimization
Over Continuous Spaces. Journal of Global Optimization 11: 341359.
Suhami, I., and R. S. H. Mah. 1981. An Implicit Enumeration Scheme for the Flowshop Problem with No Intermediate
Storage. Computers and Chemical Engineering 5: 8391.
Taillard, E. 1993. Benchmarks for Basic Scheduling Problems. European Journal of Operational Research 64: 278285.
Wang, L., Q. K. Pan, P. N. Suganthan, W. H. Wang, andY. M. Wang. 2010. A Novel Hybrid Discrete Differential Evolution
Algorithm for Blocking Flow Shop Scheduling Problems. Computers & Operations Research 37 (3): 509520.
Wang, X. P., and L. X. Tang. 2012. A Discrete Particle Swarm Optimization Algorithm with Self-Adaptive Diversity
Control for the Permutation Flowshop Problem with Blocking. Applied Soft Computing 12: 652662.
Xu, Y., L. Wang, S. Y. Wang, and M. Liu. 2014. An Effective Hybrid Immune Algorithm for Solving the
Distributed Permutation Flow-Shop Scheduling Problem. Engineering Optimization, 46 (9): 12691283.
doi:10.1080/0305215X.2013.82767
Zhu, J., X. F. Yan, and W. X. Zhao. 2013. Chemical Process Dynamic Optimization Based on the Differential Evolution
Algorithm with an Adaptive Scheduling Mutation Strategy. Engineering Optimization 45 (10): 12051221.

Anda mungkin juga menyukai