Applied Soft Computing 30 (2015) 604–613

Contents lists available at ScienceDirect

Applied Soft Computing

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

An improved iterated greedy algorithm with a Tabu-based

reconstruction strategy for the no-wait flowshop scheduling problem
Jian-Ya Ding a , Shiji Song a,∗ , Jatinder N.D. Gupta b , Rui Zhang c , Raymond Chiong d ,
Cheng Wu a
a
Department of Automation, Tsinghua University, Beijing 100084, PR China
b
College of Business Administration, University of Alabama in Huntsville, Huntsville, AL 35899, USA
c
School of Economics and Management, Nanchang University, Nanchang 330031, PR China
d
School of Design, Communication and Information Technology, The University of Newcastle, Callaghan, NSW 2308, Australia

a r t i c l e i n f o a b s t r a c t

Article history: This paper proposes a Tabu-mechanism improved iterated greedy (TMIIG) algorithm to solve the no-
Received 12 July 2014 wait flowshop scheduling problem with a makespan criterion. The idea of seeking further improvement
Received in revised form 28 January 2015 in the iterated greedy (IG) algorithm framework is based on the observation that the construction phase
Accepted 1 February 2015
of the original IG algorithm may not achieve good performance in escaping from local minima when
Available online 12 February 2015
incorporating the insertion neighborhood search. To overcome this limitation, we have modified the IG
algorithm by utilizing a Tabu-based reconstruction strategy to enhance its exploration ability. A powerful
Keywords:
neighborhood search method that involves insert, swap, and double-insert moves is then applied to
No-wait flow shop
Iterated greedy algorithm
obtain better solutions from the reconstructed solution in the previous step. Empirical results on several
Tabu search benchmark problem instances and those generated randomly confirm the advantages of utilizing the new
Makespan reconstruction scheme. In addition, our results also show that the proposed TMIIG algorithm is relatively
more effective in minimizing the makespan than other existing well-performing heuristic algorithms.
© 2015 Elsevier B.V. All rights reserved.

1. Introduction implies a high utilization of the machines and thus results in low
processing costs.
The no-wait flowshop scheduling problem (NWFSP) is an impor- The NWFSP is a typical example of combinatorial optimization
tant branch of ‘zero buffer’ scheduling problems. It arises due to the problems [8]. When the problem size increases, these problems
processing characteristics of certain products, where each process become complicated and are difficult to solve by exact methods
follows the previous one immediately [1]. Application domains of like branch-and-bound or mixed integer programming techniques.
the NWFSP include chemical processing [2], plastic molding [3], To tackle this challenge, quite a number of metaheuristic meth-
food processing [3] and steel rolling [4]. Apart from these conven- ods have been designed, since they are usually able to find
tional industries, the NWFSP is also important for some advanced a high quality solution for middle- and large-scale problems
manufacturing systems such as just-in-time production systems within reasonable computational time. Representative algorithms
[5], flexible manufacturing systems [6], and robotic cells [7] where include the genetic algorithm (GA) [9], estimation of distribu-
jobs are continuously processed with no in-process waiting time. tion algorithm (EDA) [10], discrete particle swarm optimization
In this study, we consider the NWFSP with the objective of mini- (DPSO) [11], block-based evolutionary algorithm (BBEA) [12], dis-
mizing the makespan. The makespan of a schedule is defined as the crete artificial bee colony algorithm (DABC) [13], and ant colony
completion time of the last job leaving the system. This criterion is optimization (ACO) [14], among others. These algorithms have pro-
important and widely studied, since a minimum makespan usually duced competitive results for many combinatorial optimization
problems.
Although population-based metaheuristic algorithms can
provide excellent results, they are in general fairly sophisticated
∗ Corresponding author. Tel.: +86 010 62796805.
and very sensitive to algorithm parameters such as the crossover
E-mail addresses: [email protected] (J.-Y. Ding),
probability, mutation probability and population size [15]. This
[email protected] (S. Song), [email protected] (J.N.D. Gupta),
[email protected] (R. Zhang), [email protected] (R. Chiong), makes it difficult to obtain the same results when re-implementing
[email protected] (C. Wu). a reported algorithm. As Pan and Ruiz [16] pointed out, simple

http://dx.doi.org/10.1016/j.asoc.2015.02.006
1568-4946/© 2015 Elsevier B.V. All rights reserved.
 J.-Y. Ding et al. / Applied Soft Computing 30 (2015) 604–613 605

and easily adaptable algorithms are highly desirable. The iterated The rest of the paper is organized as follows. The next section
greedy (IG) algorithm [17] is an example of such algorithms. It provides the problem formulation and a literature review on the
can provide comparable or even better computational results with NWFSP with a makespan criterion. In Section 3, we present the
simple and general search mechanisms. TMIIG algorithm for the problem at hand. The effectiveness of the
The IG algorithm consists of two main phases: destruction proposed algorithm is verified through numerical experiments and
and construction. In the destruction phase, some elements are results are discussed in Section 4. Finally, Section 5 summarizes the
removed from the current solution. In the construction phase, the paper and highlights some possible future research directions.
removed elements are reinserted into the solution by applying a
greedy constructive heuristic. A local search phase to improve the
2. Background
reconstructed solution may also be integrated into the algorithm’s
framework. Although the local search is optional, we have observed
In Section 2.1, we formally define the NWFSP with a makespan
that it contributes significantly to the effectiveness of the IG algo-
criterion. Then, we review the relevant literature in Section 2.2.
rithm. As reported by some related work in the literature (e.g., see
[17,18]), the IG algorithm with a local search step performs better
than when no local search is used. In addition, most of the rel- 2.1. Problem formulation
evant studies have incorporated local search into the algorithm
framework, e.g., see [17–20,16,21,22]. These results indicate that The NWFSP is described as follows. There are n jobs to be
the local search step is an important component of the IG algorithm. processed sequentially through m machines in the same order.
In this regard, the IG algorithm can also be viewed as a method that Every job j (j = 1, 2, . . ., n) requires a predetermined processing time
iteratively applies a neighborhood search scheme to obtain a local pi,j on every machine i (i = 1, 2, . . ., m). Setup times are indepen-
optimal solution and a greedy perturbation scheme to escape from dent of the processing sequence of the jobs and are included in
local minima. the processing times. Each machine can execute at most one job
Despite its simplicity and effectiveness, the IG algorithm has at a time and a job can only be processed on one machine at a
obvious limitations when solving large-scale problems. On the time. To satisfy the no-wait constraint, each job must be processed
one hand, the algorithm adopts a “single solution” search strat- without any waiting time between consecutive operations. This
egy with a greedy insertion reconstruction technique. It may not implies that the start of a job must be delayed on the first machine
achieve good solution diversity compared to population-based when necessary. In this work, the target of scheduling is to min-
search methods. On the other hand, a neighborhood search method imize the makespan denoted as Cmax (or Cm,n ), which equals the
is usually applied once the reconstruction phase is completed, finishing time of the last job on the last machine. In addition, most
which over emphasizes on the exploration of discovered local min- assumptions related to the permutation flowshop scheduling prob-
ima regions. These facts indicate that it may be beneficial if the lem described in [23] are applicable to this problem.
exploitation ability of the IG algorithm is improved. The no-wait characteristic of the problem ensures that the com-
The above-mentioned limitations of the IG algorithm motivate pletion time difference between adjacent jobs is determined by the
us to investigate for further improvement of the algorithm’s struc- processing times of the two jobs, regardless of the other jobs in
ture. In the original perturbation scheme, randomly removed jobs the permutation. Thus, a completion time distance can be defined
are reinserted greedily into the current partial solution one by one between each pair of jobs. The completion time distance from job i
to build a complete solution. However, we notice that this insertion to job j is calculated as follows [24]:
reconstruction strategy is similar to the commonly used neighbor-  m 
hood search method based on insert moves. It may lead to repeated 
Di,j = max (ph,j − ph,i ) + pk,i . (1)
and excessive search as well as poor exploitation in the search pro- k=1,...,m
h=k
cess, since the algorithm may get stuck at local minima solutions
with respect to insertion operations. To overcome this, we propose It is worth noting that all the completion time distances can be
a Tabu-based construction method in the perturbation scheme to calculated in advance and the time complexity is O(mn2 ).
avoid repeated search and introduce solution diversities. More pre- A schedule for this problem can be expressed as a job per-
cisely, a succession of Tabu lists are built and maintained in the mutation  = [(1), (2), . . ., (n)], where (k) ∈ {1, 2, . . ., n} and
construction phase to prevent each removed object from being (k) =/ (k ), ∀ k =
/ k . It is clear that the makespan of a feasible
inserted at previously visited positions. After the reconstruction, a schedule  = [(1), . . ., (n)] is given by
more powerful neighborhood structure based on insert, swap and
double-insert moves is adopted to further enhance the exploration Cmax () = Cm,n ()
ability.
The contributions of this work can be summarized as follows: 
n

m

= D[j−1],[j] + p(1),k ,
j=2 k=1

• We have modified the structure of the IG algorithm by introduc-

where D[j−1],[j] represents the completion time distance between
ing a Tabu-based insertion strategy in the reconstruction phase.
the (j − 1)th and jth jobs in schedule , i.e. D(j−1),(j) .
Advantages of utilizing the new reconstruction scheme are con-
To simplify the expression of the makespan, a dummy job (0),
firmed via empirical results on benchmark problem instances of
which has zero processing time on all machines, is introduced at the
the NWFSP.
beginning of permutation . Thus, the schedule  is redefined as
• We have analyzed the performance of the proposed Tabu-
 = [(0), (1), . . ., (n)]. With the above developments, we obtain
mechanism improved iterated greedy (TMIIG) algorithm and
the objective function (makespan) of a permutation  as follows:
reported computational results comparing it against other well-
performing algorithms for the NWFSP. Our results show that the 
n
proposed algorithm is the most effective one. In addition, we Cmax () = Cmax ( ) = D[j−1],[j] , (2)
have updated 43 new upper bound solutions for the benchmark i=1
instances of the NWFSP, which can be a good reference for other m
studies in the future. where D[0],[1] is defined as D[0],[1] = p
k=1 (1),k
.
606 J.-Y. Ding et al. / Applied Soft Computing 30 (2015) 604–613

2009, Qian et al. [37] proposed a hybrid differential evolution (HDE)

approach based on parallel evolution and problem-dependent local
exploitation. Their computational results show that the HDE algo-
rithm is superior to the HPSO algorithm. In 2010, Tseng and Lin
[38] proposed a hybrid GA (HGA), which hybridizes a GA and an
insertion-based local search method with cut-and-repair. In 2011,
Jarboui et al. [39] proposed another hybrid algorithm based on the
GA and VNS (GA-VNS). They showed that both hybrid algorithms
can outperform the five methods proposed by Grabowski and Pem-
Fig. 1. A Gantt chart of a no-wait flowshop. pera. In 2012, Samarghandi and ElMekkawy [40] proposed a hybrid
TS and PSO algorithm and utilized a new coding and decoding tech-
Let  denote the set of all n ! possible permutation schedules nique to improve its efficiency. Their numerical results indicate
in the solution space for the NWFSP. The problem is then to find a that the TS-PSO method performs better than other heuristics and
permutation schedule * ∈  such that: metaheuristics proposed before 2008.
From this concise literature review, we have obtained the fol-
Cmax (∗ ) = minCmax (). lowing insights:
∈

To provide a better understanding of the problem formulation,

we present in Fig. 1 the Gantt chart of a no-wait flowshop with • Integer programming is seldom utilized to solve the NWFSP. Most
three machines and three jobs. It is clear that the makespan can be of the previous studies on this topic focus on designing heuristics
easily obtained by summing up the three completion time distances and metaheuristics. As a result, the effectiveness of newly pro-
(D[0],[1] , D[1],[2] and D[2],[3] ) marked in the figure. posed algorithms is generally verified by comparing the upper
bound solutions provided by existing well-performing methods.
2.2. Literature review Thus, it is meaningful to make a fair comparison among the
existing promising searching approaches and present a set of up-
The NWFSP is known to be NP-hard [25,26]. Due to its NP-hard to-date upper bounds as a reference for future research on the
nature, exact algorithms require unacceptably large computational problem.
time for solving even moderate-size problems. Therefore, in recent • According to computational results reported in the studies we
years, a great deal of research attention has been given to the have reviewed, the most promising solution methods appear
development and testing of various heuristics and metaheuristics. to be those metaheuristics proposed after the year of 2008. It
Heuristic methods start from an initial (seed) schedule of a few jobs. seems that the DPSOVND , IIGA, HDE, HGA, GA-VNS and TS/PSO
By appending and/or inserting new jobs to a known partial schedule methods are the state-of-the-art algorithms for the NWFSP
at each step, they finally arrive at a complete schedule. Important with a makespan criterion. Therefore, the proposed solution
heuristics for the no-wait flowshop to minimize the makespan have approach will be compared to these effective algorithms from the
been developed by Bonney and Gundry [27], King and Spachis [28], literature.
Gangadharan and Rajendran [29], Rajendran [2], Bertolissi [30] and
Laha and Chakraborty [31]. It is noticeable that some constructive
heuristics for the regular flowshop problem, like the one proposed 3. Tabu-mechanism improved iterated greedy algorithm
by Nawaz et al. [32], can also be adapted to minimize the makespan
in the NWFSP. This section presents the details of our TMIIG algorithm for
Compared to constructive heuristics, metaheuristic methods solving the NWFSP with a makespan criterion. The proposed algo-
can usually obtain better quality solutions for this problem, rithm consists of the following three phases: a modified NEH
and thus have gained increasing research emphasis. In 2003, algorithm to generate an initial schedule, a Tabu-based reconstruc-
Aldowaisan and Allahverdi [33] proposed two methods based on tion technique to help avoid local minima, and a neighborhood
the GA and simulated annealing (SA). In the same year, Schuster and search method to improve the current solution. The reconstruc-
Framinan [34] developed the variable neighborhood search (VNS) tion phase and the neighborhood search phase continue until a
and hybrid GA and SA (GASA) methods. Both approaches outper- pre-defined termination condition is met. In this section, we first
form the best existing heuristics for the NWFSP with a makespan describe each phase of the proposed algorithm and the stopping
criterion. Later in 2005, Grabowski and Pempera [35] presented two criterion used. Then, we present the steps of the proposed TMIIG
algorithms based on descending search and three other methods algorithm.
based on Tabu search (TS). These five methods can perform better
than the VNS and GASA methods, and are widely cited for experi-
mental comparison in the subsequent literature. In 2007, Liu et al. 3.1. Initial solution
[36] proposed several hybrid algorithms based on particle swarm
optimization (PSO) and Nawaz-Enscore-Ham (NEH) heuristic. The The NEH algorithm is an effective constructive heuristic algo-
hybrid PSO (HPSO) algorithm is superior to VNS and GASA as well. rithm for minimizing the makespan in a flowshop with an infinite
More recently, quite a number of metaheuristics have been storage buffer at each stage. Starting from an initial sort of jobs in
proposed for the NWFSP. In 2008, Pan et al. [11] presented a non-increasing order of their total processing times, it constructs
DPSO algorithm for the NWFSP. They also hybridized the variable a complete solution by inserting these jobs one by one into the
neighborhood descent (VND) algorithm based on swap and insert current partial schedule.
neighborhood structures with DPSO to further enhance its search- The original NEH algorithm is based on the hypothesis that
ing ability. The DPSOVND algorithm is able to outperform some jobs with larger total processing times should have higher pri-
existing heuristics and metaheuristics based on the makespan and orities. However, inspired by the work of Framinan et al. [41],
total flowtime criteria. In the same year, Pan et al. [19] designed we found that the standard deviation of processing times of a
an improved IG algorithm (IIGA). They showed that the perfor- job on different machines can significantly affect the makespan
mance of the IIGA is comparable with the DPSOVND algorithm. In of a schedule. Therefore, we deploy a new priority rule based on
 J.-Y. Ding et al. / Applied Soft Computing 30 (2015) 604–613 607

standard deviations to obtain the initial job sequence. The standard

deviation STDj for each job j is defined as follows:


 1 
m
STDj =  (pi,j − AVGj )2 , (3)
m−1
i=1

1
m
where AVGj = m p . Algorithm 1 describes the procedures
i=1 ij
of the modified NEH (MNEH) algorithm based on the above
developments.

Algorithm 1. The MNEH algorithm

1: Generate an initial sequence by arranging the jobs in non-increasing order
of the new priority index, STDj given in Eq. (3). Denote this sequence as
0 = [0 (1), 0 (2), . . ., 0 (n)].
2: Select the first two jobs in 0 and evaluate the two possible partial
sequences of the two jobs. The sequence with the smaller makespan will
become the current partial sequence 2 = [2 (1), 2 (2)].
3: for k = 3 to n do
4: Find the best position j* of inserting 0 (k) into the current partial sequence
k−1 by the following equation:
Cj = Dk−1 (j−1),0 (k) + D0 (k),k−1 (j) − Dk−1 (j−1),k−1 (j) , j = 1, . . ., k,
Fig. 2. An illustration of the Tabu-based reconstruction method.
j∗ = arg min Cj ,
j=1,...,k
where k−1 (0) and k−1 (k) are dummy jobs, which have zero processing time
on all machines. Then, insert 0 (k) into sequence k−1 at position j* and An illustration is provided in Fig. 2 to help achieve a better
denote the generated new sequence as k .
understanding of the proposed Tabu based reconstruction method.
5: end for
Output: sequence n and Cmax (n ) Assume that the removed jobs are R = [6, 2] and that the ini-
tial sequence is  = [3, 6, 1, 4, 2, 7, 5] as shown in Fig. 2(a). Also
assume that the maximum length of the Tabu list is ML = 3 and
Since the completion time distance matrix (Di,j )n×n can be cal- that the original Tabu lists for the removed jobs are TL6 = {5, 4, 0}
culated in advance, the main computational burden of the NEH and TL2 = {6}. Job 0 in the figure means that the first position in
algorithm lies in the inserting process of Steps 3–5. By considering the sequence is prohibited. According to the Tabu mechanism, TL6
the acceleration method for job insertion in [42], the NEH algorithm and TL2 are updated as TL6 = {4, 0, 3} and TL2 = {6, 4}. Restricted
can be completed in O(n2 ) computational effort. by the updated Tabu lists, job 6 and job 2 are reinserted into the
sequence as depicted in Fig. 2(b–c). Crosses in the figure indicate
3.2. Tabu-based reconstruction that the corresponding insertion position is prohibited. Note that
we require ML  n − d, which implies that there are always some
The original IG algorithm generates a sequence of solutions by possible slots for the insertion of a job. As we can see from this
iteratively applying a greedy constructive heuristic that contains example, previously visited insertion positions are excluded by the
destruction and construction phases. In these two phases, some easily implemented Tabu mechanism. Despite its simplicity, this
jobs are removed from a complete sequence and then reinserted method is able to improve the performance of the IG algorithm sig-
into the obtained partial sequence using the NEH-insertion method. nificantly as revealed by the computational experiments reported
However, this reconstruction mechanism may fail to overcome in Section 5.
local minima since it utilizes a similar mechanism as in the neigh-
borhood search phase, which may lead to repeated search. Due Algorithm 2. Tabu based reconstruction
to this, a new reconstruction technique that forbids some insert Input: Tabu list TLj (j = 1, . . ., n), sequence , number of removed jobs d.
1: Set R =∅ and D = .
positions is proposed as follows.
2: for k = 1 to d do
Once the initial solution is obtained, we create a Tabu list TLj 3: Select a job (without repetition) at a random position j in . Add (j − 1)
for each job j. When inserted into the partial sequence, job j is not to the end of Tabu list TL(j) . Insert job (j) at the end of sequence R and
allowed to be placed at a position immediately after any job in the remove it from sequence D .
Tabu list TLj . In each iteration, the Tabu list for each removed job is 4: if Length(TL(j) ) > ML then
5: Delete the first element in TL(j)
updated.
6: end if
To start the reconstruction, d jobs (d ∈ {1, . . ., n − 1}) are ran- 7: end for
domly selected and removed from the current complete candidate 8: for k = 1 to d
solution  containing all n jobs. For each removed job (say (j)), the 9: Evaluate each of the possible inserting positions of job R (k) that is not
job that immediately precedes it (say (j − 1)) is added to its Tabu prohibited by the Tabu list TLR (k) . Insert job R (k) into D at the position with
the smallest resulting makespan among the evaluated solutions.
list (say TL(j) ). Each Tabu list has a maximum length ML (ML < n), 10: end for
which is a parameter to be tuned. Once the number of elements in Output: sequence D .
a Tabu list TLj exceeds ML, the first element in TLj will be removed
from this list. The procedure of the Tabu-based reconstruction strategy is
The removed d jobs form a partial sequence R in the same order given in Algorithm 2. Note that only the Tabu lists for the removed
that they were selected. The partial solution of  with n − d jobs jobs are updated, and thus lines 2–7 only take O(d) time. The main
left also forms another partial sequence denoted as D in its orig- computational burden in this reconstruction procedure lies in lines
inal order of jobs. Next, the jobs in R are re-inserted into D one 8–10, where each of the possible inserting positions is evaluated.
after another. These jobs will be inserted at the positions such that By considering a similar acceleration method as in the implemen-
the makespan of the partial sequence is minimized while the Tabu tation of the MNEH algorithm, the reconstruction phase in each
constraint is not violated. iteration can be completed in O(dn) time.
608 J.-Y. Ding et al. / Applied Soft Computing 30 (2015) 604–613

3.3. Neighborhood search Algorithm 3. Neighborhood search

Input: sequence .
1: Set the swap neighborhood structure as N1 , the insert neighborhood
The schedule obtained in the reconstruction phase may not be
structure as N2 and the double-insert as N3 .
a good enough solution, since there can be better sequences in its 2: while l  3 do
neighborhood. Therefore, a local search method is adopted to fur- 3: Find the best neighbor solution  of  in Nl ().
ther improve the current sequence. In particular, the VNS method 4: if Cmax ( ) < Cmax () then
is applied to find more promising solutions in the solution space 5: Set  =  and l = 1.
6: else
through changing neighborhood structures during the search pro-
7: Set l = l + 1.
cess. 8: end if
The selection of neighborhood structures can greatly influence 9: end while
the performance of the VNS method [43]. Since a neighborhood Output: sequence .
is usually defined based on moves of jobs, the search process can
benefit much from suitably selected moves. Among various types 3.4. Acceptance and stopping criteria
of moves considered in the literature, insert and pair-wise swap
moves are most commonly used for the NWFSP. The neighborhood After a new sequence  is generated by the local search proce-
based on insert moves is defined by enumerating all possible pairs dure, a decision of whether to accept it as the incumbent solution
of positions j, k ∈ {1, . . ., n} in sequence  (j =
/ k), where job (j) is for the next iteration needs to be made. Inspired by the SA method
removed and then reinserted at position k. The neighborhood based [45], an acceptance criterion, which allows the acceptance of a
on swap moves is defined in a similar manner, which considers slightly worse solution, is adopted in this work. In this acceptance
exchanging the positions of two jobs in the sequence. criterion,  is accepted with a probability of
In this work, the selected VNS moves contain not only the insert C )  
max () − Cmax (
and swap moves [42], but also the rarely tested double-insert p = min exp ,1 ,
T
moves [31,44]. The double-insert move considers removing two
consecutive jobs from positions j and j + 1 (j = 1, . . ., n − 1) and re- where  is the solution of the current iteration and T is a constant
inserting them together into positions k and k + 1 (k = 1, . . ., n − 1, temperature. Following the suggestions of Osman and Potts [46],
and k > j + 1 or k < j) in the same order. the temperature is set as follows:
Let  denote the original sequence, and I (k, j), S (k, j) and n m
j=1
p
k=1 kj
DI (k, j) denote the sequences after the insert, swap and double- T= · T0 ,
insert moves, respectively. It can be deduced from Eq. (2) that the 10 · m · n
makespan increment caused by the above operations is given as where T0 is a parameter to be adjusted.
follows: The maximum computational time is the stopping criterion
Insert move: adopted in this work. The proposed TMIIG algorithm stops when
the maximum running time is reached. The final output solution by
Cmax = Cmax (I (k, j)) − Cmax () the algorithm is the best (i.e., one with the minimum makespan) in
all iterations along the search process.
= D[k−1][j] + D[j][k] − D[k−1][k] + D[j−1][j+1] − D[j−1][j] (4)
3.5. Overview of the TMIIG algorithm
−D[j][j+1]

Based on the above developments, the complete procedure of

Swap move:
the proposed TMIIG algorithm is described in Algorithm 4.
Cmax = Cmax (S (k, j)) − Cmax ()
⎧
⎪ D + D[j][k] + D[k][j+1] − D[k−1][k] − D[k][j] − D[j][j+1] , if j = k + 1,
⎨ [k−1][j] (5)
= D[j−1][k] + D[k][j] + D[j][k+1] − D[j−1][j] − D[j][k] − D[k][k+1] , if j = k − 1,
⎪
⎩
D[k−1][j] + D[j][k+1] + D[j−1][k] + D[k][j+1] − D[k−1][k] − D[k][k+1] − D[j−1][j] − D[j][j+1] , otherwise.

Algorithm 4. The TMIIG algorithm

Double insert moves: 1: (Initial solution): Generate an initial sequence using the MNEH
algorithm (described in Algorithm 1). Let  be the solution obtained in
this step.
Cmax = Cmax (DI (k, j)) − Cmax () 2: Initialize all the Tabu lists as empty sets: TLj : = ∅ , j = 1, . . ., n. Set the
m n
temperature T := T0 · pi,j /(10 · n · m).
= D[k−1][j] + D[j+1][k] − D[k−1][k] + D[j−1][j+2] (6) i=1 j=1
3: while the termination condition is not met do
4: (Tabu based reconstruction): With respect to the Tabu lists, apply the
−D[j−1][j] − D[j+1][j+2]
Tabu-based reconstruction method (described in Algorithm 2) to
sequence  in order to construct another complete sequence  .
Update the Tabu lists.
It is noticeable that, although the number of neighborhood solu- 5: (Neighborhood search): Use the neighborhood search method
tions is quite large, the computational time for evaluating them is (described in Algorithm 3) that hybridizes the insert, swap, and
acceptable if the above-mentioned accelerating techniques are uti- double-insert moves to improve sequence  in the previous step. The
improved solution is denoted as  .
lized. According to Eqs. (4)–(6), the makespan increments of insert,
6: (Acceptance): Set  =  with the probability of
swap and double-insert moves can be calculated in O(1) time. Thus, 1, if  < 0,
the evaluation of neighborhood solutions is accelerated and the p=
e−/T , if  ≥ 0,
computational complexity of the local search phase can be reduced where  = Cmax ( ) − Cmax ().
to O(n2 ). The procedure of the proposed neighborhood search phase 7: end while
is described in Algorithm 3. Output: The best solution found.
 J.-Y. Ding et al. / Applied Soft Computing 30 (2015) 604–613 609

4. Computational evaluation Table 1

ANOVA results for the experiment on parameter calibration of the TMIIG algorithm.

This section describes the experimental settings and results Source Df Sum of Mean F-ratio p-Value
obtained by the TMIIG algorithm for solving the NWFSP with a squares square
makespan criterion. Firstly, the experimental setup and the tested Main effects
data set are described in Section 4.1. Then, the computational A:d 5 46.0411 9.2082 100.4300 0.0000
results are reported according to the following aspects: B:T0 4 13.1675 3.2919 35.9000 0.0000
C:ML 4 213.2652 53.3163 581.5200 0.0000

• Calibration of the algorithm parameters. Interactions

• Assessment of the contribution of the proposed Tabu based AB 20 0.7311 0.0366 0.4000 0.9920
AC 20 274.1942 13.7097 149.5300 0.0000
reconstruction and neighborhood search method. BC 16 0.6905 0.0432 0.4700 0.9620
• Comparison of the effectiveness of the proposed algorithm with ABC 80 2.8585 0.0357 0.3900 1.0000
that of some existing well-performing algorithms.
Residual 14,850 1361.5033 0.0917
Total (corrected) 14,999 1912.4514
4.1. Experimental setup

The proposed algorithm and procedures were coded using C++, 3, 4}. These factors yield a total of 6 × 5 ×5 = 150 different configu-
and the computational experiments were executed on a personal rations for the proposed TMIIG algorithm. For every instance in the
computer (PC) with an Intel Core (TM) CPU running at 3.20 GHz generated data set, five replications of each algorithm were tested
in a Windows 7 Operating System environment. To compare the and the ARPD values were recorded.
performance of our proposed algorithm with other approaches, we The experimental results were analyzed by means of a multi-
conducted experiments using benchmark problems (rec01, rec03, factor analysis of variance (ANOVA) technique. A major advantage
. . ., rec41) from the OR-Library provided by Reeves [47]. This bench- of the ANOVA technique is that it calculates the magnitude of the
mark set consists of 21 problem instances with seven different F-ratio, which is a clear indicator of significance when the p-value
sizes, ranging from 20 jobs and five machines to 75 jobs and 20 is close to zero. More precisely, a large F-ratio indicates that the
machines. For each size, three instances are provided. Considering analyzed factor has a considerable effect on the response variable.
that the above problem instances are of relatively small scale, the To apply ANOVA, we checked its three main hypotheses, i.e., nor-
Taillard benchmark set was also adopted for testing the algorithm’s mality, homoscedasticity and independence of residuals. Statistical
performance. This benchmark set consists of 120 problem instances analysis showed that all of the three hypotheses could be accepted.
with 12 different sizes, ranging from 20 jobs and five machines The ANOVA results for our parameter calibration experiment are
to 500 jobs and 20 machines. For each size, 10 instances are pro- given in Table 1.
vided. We carried out 10 independent replications for the proposed According to Table 1, the length of the Tabu list, i.e., parameter
TMIIG algorithm on each of the above instances, and the termina- ML, results in the largest F-ratio. This implies that parameter ML is
tion condition was set as the maximum running time tmax = n2 /2 the most important factor that affects the average performance of
ms. the TMIIG algorithm. Therefore, a 95% confidence interval plot for
Note that a different data set has been considered for parameter the ARPD value under different selections of parameter ML is pro-
calibration and assessment of algorithm components, so as to avoid vided in Fig. 3. It is observed from this figure that the choice of ML = 0
overfitting in the results. We applied Taillard’s problem generation gives a much worse ARPD value than those obtained by non-zero
method [48] to randomly generate instances of 20 different com- values of parameter ML. This result clearly indicates the importance
binations of n and m, with m ∈ {5, 10, 15, 20}, n ∈ {40, 80, 120, 160, of incorporating the Tabu mechanism in the construction phase in
200} and processing times uniformly distributed from 1 to 99. For the original IG algorithm.
each size, five instances were generated, resulting in a total of 100 Similar comparisons were conducted for parameters d and T0 .
instances. A 95% confidence interval plot for the ARPD value under different
To compare the results obtained from the experiments, the selections of parameter d is provided in Fig. 4 and the correspond-
average relative percentage deviation (ARPD) was collected for ing figure for parameter T0 is provided in Fig. 5. It is clear from Fig. 4
measuring the average relative quality of the solutions. The ARPD that too small or too large a value of d would lead to deterioration of
is calculated as follows: the algorithm’s performance and the choice of d = 6 gives the best
result. Fig. 5 suggests that the value T0 = 0.6 gives the best aver-
1  Cr − CR∗
R

ARPD = × 100, (7) age performance. In addition, no statistical difference is observed

R CR∗
r=1

where Cr is the solution generated by a specific algorithm in the rth

(r = 1, 2, . . ., R) experiment for a given problem size, and CR∗ is the
best solution found so far. Obviously, the less the value of ARPD is,
the better the algorithm’s performance is.

4.2. Parameter calibration

Since the performance of the proposed algorithm will be

influenced by the selection of parameter values, preliminary exper-
iments were conducted under a set of potential parameter values
to find the best combinations of d, T0 and ML. Specifically, we
implemented the Design of Experiments method [49] to analyze
the effect of these factors on the ARPD. To this end, a full facto-
rial design was considered with the factors at the following levels:
d ∈ {3, 6, 9, 12, 15, 18}, T0 ∈ {0.2, 0.6, 1.0, 1.4, 1.8} and ML ∈ {0, 1, 2, Fig. 3. 95% Confidence interval for different selections of parameter ML.
610 J.-Y. Ding et al. / Applied Soft Computing 30 (2015) 604–613

Fig. 4. 95% Confidence interval for different selections of parameter d. Fig. 6. 95% Confidence interval for the TMIIG, and TMIIGNT algorithms.

follows. For each instance, the proposed TMIIG algorithm was eval-
between the selection of T0 = 0.6 and T0 = 1.0. This suggests that the
uated four times. In the first trial, the algorithm was executed
algorithm is robust for the temperature parameter in the range [0.6,
without any modification. In the following trials, however, it was
1.0].
executed with neighborhood structures selected to be based on
Based on the above results, the final parameter settings used
insert, swap, and hybrid insert-swap moves, denoted by TMIIGI ,
in our proposed TMIIG algorithm are as follows: d = 6, ML = 1, and
TMIIGS , and TMIIGIS respectively. By comparing the experimental
T0 = 0.6.
results of these four trials, we can obtain a fair comparison between
different selections of the neighborhood structures.
4.3. Effectiveness of algorithm components Fig. 7 shows a 95% confidence interval plot of the four tested
algorithms with different neighborhood structures. This figure
To test the effectiveness of the proposed Tabu based recon- clearly illustrates that it is of statistical significance that the TMIIG
struction procedure, computational experiments were designed algorithm outperforms TMIIGI , TMIIGS and TMIIGIS , which indi-
as follows. For each instance, the proposed TMIIG algorithm was cates that incorporating double-insert moves in the neighborhood
evaluated twice. In the first trial, the algorithm was executed as search step does help to improve the performance of the algorithm.
described in the previous sections without any modification. In the
second trial, however, it was executed by setting the Tabu length 4.4. Comparison with existing algorithms
to zero and thus the effect of the Tabu mechanism disappeared. By
comparing the experimental results of these two trials, the con- Firstly, the effectiveness of the proposed MNEH algorithm was
tribution of the Tabu mechanism can be clearly ascertained. For tested. Table 2 provides a detailed comparison of NEH and MNEH
notational simplicity, the algorithm without the Tabu mechanism on the Taillard’s 120 problem instances. We observe from this table
is referred to as TMIIGNT . The two algorithms were tested on the that MNEH outperforms the NEH algorithm by 0.35% on average.
aforementioned generated data set with five independent replica- It can also be seen that 10 out of 12 average makespans of MNEH
tions on each instance. are smaller than those of NEH, indicating that MNEH is superior
We present a 95% confidence interval plot in Fig. 6 to show the to NEH. It is worth noting that these improvements are obtained
performances of the TMIIG and TMIIGNT algorithms. As shown in without any additional computation efforts.
the figure, the TMIIG algorithm achieves a smaller average ARPD Then, we tested the effectiveness and efficiency of the proposed
value of 0.29, compared to the corresponding value of 0.56 for TMIIG algorithm in searching for better quality schedules. To do so,
TMIIGNT . In addition, it can also be observed from the figure that we compared its computational results with some existing algo-
the TMIIG algorithm outperforms TMIIGNT from a statistical view- rithms: the DPSOVND [11], IIGA [19], HDE [37], HGA [38], GA-VNS
point. This result suggests that the Tabu mechanism contributes [39] and TS/PSO [40]. These algorithms are the present state-of-
significantly to the effectiveness of the TMIIG algorithm. the-art algorithms for the NWFSP.
Next, we tested the effectiveness of the proposed neighborhood Since all of the six compared algorithms have previously been
search method. The computational experiments were designed as tested on the 21 benchmark instances from Reeves, numerical

Fig. 7. 95% Confidence interval for the TMIIGI , TMIIGS , TMIIGIS and TMIIG algo-
Fig. 5. 95% Confidence interval for different selections of parameter T0 . rithms.
 J.-Y. Ding et al. / Applied Soft Computing 30 (2015) 604–613 611

Table 2 can outperform these state-of-the-art algorithms for the NWFSP in

Comparison of the NEH and MNEH algorithms.
minimizing the makespan.
n×m Average makespan Average improvement Note that the instances from the Reeves benchmark set are of
ratio (%) relatively small scale and easy to solve in most cases, and the aver-
NEH MNEH age performance of different algorithms tends to be quite similar.
To further demonstrate the effectiveness of the TMIIG algorithm,
20 × 5 1540.50 1543.60 −0.20
20 × 10 2053.90 2052.90 0.05 numerical comparisons were conducted on the Taillard benchmark
20 × 20 3062.50 3089.50 −0.88 set [48], which contains a good number of large-scale instances. In
50 × 5 3520.40 3472.70 1.36 particular, the proposed algorithm was compared with the IIGA,
50 × 10 4522.00 4478.60 0.96 DPSOVND and GA-VNS algorithms. These are the best existing algo-
50 × 20 6230.40 6156.90 1.18
100 × 5 6719.00 6674.20 0.67
rithms among the six considered in the previous comparison. Since
100 × 10 8537.10 8515.00 0.26 only the performance of the GA-VNS algorithm was reported on
100 × 20 11,247.20 11,247.30 0.00 Taillard’s instances in the literature in detail, we re-implemented
200 × 10 16,354.60 16,263.80 0.56 the IIGA and DPSOVND for comparison purposes. To distinguish
200 × 20 21,089.70 21,028.80 0.29
from their original versions, the two re-implemented algorithms
500 × 20 49,710.50 49,680.40 0.06
Average NA NA 0.35 are denoted by IIGA* and DPSO∗VND respectively. The computational
results of GA-VNS were taken from [39], where the experiments
were executed on a PC with a 3.2 GB processor. As per [39], the
same maximum computational time (tmax = n2 /2 ms) was allowed
comparisons were first conducted on this data set. To this end, we for executing the GA-VNS algorithm.
collected the computational results of the compared algorithms Table 4 summarizes the computational results for the above-
from the relevant literature [11,19,37–40]. It is noticeable that mentioned algorithms, where the upper bounds are taken from
these studies reported their results with respect to different upper [39]. As revealed in the table, the total ARPD average obtained by
bound solutions: some compared their results with the makespan the proposed TMIIG algorithm is 0.14, superior to the correspond-
value provided by Rajendran [2], while the others compared the ing values of 0.42, 0.67 and 0.37 obtained by the IIGA, GA-VNS
results with their best obtained results. We carefully recorded the and DPSOVND respectively. In other words, the proposed TMIIG
actual makespan value of each compared algorithm according to algorithm is able to outperform the IIGA, GA-VNS and DPSOVND
the literature and listed the best results (including the TMIIG algo- in solution quality in the average sense. Probing further into the
rithm) for the 21 benchmark instances. The ARPD values of all statistics in the table shows that the four compared algorithms
these algorithms were recalculated based on the listed best val- achieve similar results for the small-scale instances (n = 20), while
ues according to Eq. (7). Thus, a valid comparison on ARPD values the proposed TMIIG algorithm outperforms the other algorithms at
among them is possible. a considerable margin for instances of middle and large scale. Since
Table 3 shows the experimental results for the seven algorithms all the algorithms were executed in a similar computational envi-
in comparison. As can be seen in the table, the total average of ronment with the same running time, these results again indicate
ARPD by the proposed TMIIG algorithm is 0.09, which is superior that the proposed TMIIG algorithm is superior to the algorithms
to the corresponding values of 0.16, 0.19, 0.19, 0.24, 0.47 and 0.26 compared for solving the NWFSP with a makespan criterion.
obtained by the IIGA, DPSOVND , GA-VNS, HGA, TS/PSO and HDA, Given that the proposed TMIIG algorithm is competitive against
respectively. Considering that the computational time consumed other state-of-the-art algorithms, new upper bound solutions have
by these tested algorithms generally does not exceed one second been discovered through the extensive computation tests. These
on average, time consumption is not the main issue for this prob- solutions are reported in Table 5, where the improved upper bounds
lem. Therefore, we conclude that the proposed TMIIG algorithm are highlighted in bold. The original upper bounds in the table

Table 3
Comparison of results based on Reeve’s bechmark set with a makespan criterion.

Name n×m Best IIGA DPSOVND GA-VNS HGA TS/PSO HDE TMIIG

ARPD Time ARPD Time ARPD Time ARPD Time ARPD Time ARPD Time ARPD Time

Rec1 20 × 5 1526 0.00 0.02 0.00 0.01 0.00 0.01 0.00 0.01 0.00 0.34 0.00 3.10 0.00 0.20
Rec3 20 × 5 1361 0.00 0.02 0.00 0.00 0.00 0.00 0.00 0.01 0.00 0.31 0.00 3.10 0.00 0.20
Rec5 20 × 5 1511 0.01 0.02 0.02 0.02 0.01 0.00 0.00 0.01 0.42 0.28 0.09 3.10 0.00 0.20
Rec7 20 × 10 2042 0.00 0.02 0.00 0.00 0.00 0.01 0.00 0.01 0.20 0.38 0.00 3.90 0.00 0.20
Rec9 20 × 10 2042 0.00 0.02 0.00 0.00 0.02 0.02 0.00 0.01 0.00 0.40 0.00 3.80 0.00 0.20
Rec11 20 × 10 1881 0.00 0.02 0.00 0.01 0.00 0.01 0.00 0.01 0.21 0.39 0.15 3.90 0.00 0.20
Rec13 20 × 15 2545 0.00 0.02 0.00 0.01 0.00 0.01 0.00 0.01 0.11 0.42 0.00 5.10 0.00 0.20
Rec15 20 × 15 2529 0.00 0.02 0.00 0.00 0.00 0.01 0.00 0.01 0.17 0.40 0.00 5.20 0.00 0.20
Rec17 20 × 15 2587 0.00 0.02 0.00 0.00 0.00 0.01 0.00 0.01 0.04 0.44 0.00 5.20 0.00 0.20
Rec19 30 × 10 2850 0.00 0.04 0.19 0.02 0.11 0.00 0.00 0.03 0.72 0.51 0.06 10.40 0.00 0.45
Rec21 30 × 10 2821 0.18 0.04 0.11 0.04 0.09 0.02 0.28 0.03 0.17 0.49 0.18 10.20 0.07 0.45
Rec23 30 × 10 2700 0.00 0.04 0.05 0.05 0.07 0.05 0.00 0.03 0.07 0.53 0.12 10.40 0.00 0.45
Rec25 30 × 15 3593 0.01 0.04 0.05 0.03 0.00 0.02 0.00 0.03 0.52 0.55 0.03 15.00 0.00 0.45
Rec27 30 × 15 3431 0.01 0.05 0.25 0.06 0.08 0.07 0.00 0.03 0.18 0.58 0.18 14.80 0.01 0.45
Rec29 30 × 15 3291 0.01 0.06 0.09 0.02 0.13 0.06 0.00 0.03 0.29 0.61 0.20 14.90 0.03 0.45
Rec31 50 × 10 4311 0.28 0.11 0.33 0.33 0.42 0.34 0.54 0.27 1.01 1.29 0.55 3.80 0.13 1.25
Rec33 50 × 10 4428 0.67 0.10 0.53 0.34 0.62 0.36 0.68 0.25 0.98 1.91 0.87 3.70 0.30 1.25
Rec35 50 × 10 4397 0.15 0.10 0.29 0.34 0.40 0.30 0.62 0.23 0.63 1.85 0.74 3.70 0.19 1.25
Rec37 75 × 20 8014 0.69 0.18 0.77 1.16 0.82 0.35 1.33 1.45 1.21 3.42 0.81 14.70 0.33 2.81
Rec39 75 × 20 8428 0.62 0.18 0.71 0.99 0.65 0.98 0.91 1.28 1.32 3.51 0.71 14.60 0.46 2.81
Rec41 75 × 20 8442 0.75 0.18 0.65 0.88 0.62 1.12 0.75 1.07 1.52 3.50 0.73 14.70 0.44 2.81
Average 0.16 0.06 0.19 0.21 0.19 0.18 0.24 0.23 0.47 1.05 0.26 8.21 0.09 0.79
612 J.-Y. Ding et al. / Applied Soft Computing 30 (2015) 604–613

Table 4
Comparison of results based on Taillard’s bechmark set with a makespan criterion.

n×m Upper bound (average) IIGA* DPSO∗VND GA-VNS TMIIG

ARPD Time ARPD Time ARPD Time ARPD Time

20 × 5 1480.3 0.00 0.20 0.00 0.20 0.03 0.20 0.00 0.20

20 × 10 1983 0.01 0.20 0.00 0.20 0.06 0.20 0.00 0.20
20 × 20 2971.9 0.01 0.20 0.02 0.20 0.03 0.20 0.00 0.20
50 × 5 3270.1 0.44 1.25 0.44 1.25 0.55 1.25 0.18 1.25
50 × 10 4273.6 0.52 1.25 0.26 1.25 0.47 1.25 0.06 1.25
50 × 20 5897.5 0.33 1.25 0.32 1.25 0.47 1.25 0.05 1.25
100 × 5 6223.5 0.51 5.00 0.37 5.00 0.84 5.00 0.42 5.00
100 × 10 8017.5 0.42 5.00 0.59 5.00 0.85 5.00 0.12 5.00
100 × 20 10,692.4 0.71 5.00 0.52 5.00 0.91 5.00 0.10 5.00
200 × 10 15,262.7 1.06 20.00 0.88 20.00 1.68 20.00 0.39 20.00
200 × 20 19,957.7 0.65 20.00 0.67 20.00 1.52 20.00 0.21 20.00
Average NA 0.42 20.00 0.37 20.00 0.67 20.00 0.14 20.00

Table 5 compared to other well-performing algorithms. Therefore, this

New upper bound solutions for Taillard’s benchmarks.
work not only provides a suitable metaheuristic for solving the
Instance Best Instance Best Instance Best Instance Best NWFSP, but also shows a promising direction for improving the
Ta01 1486 Ta31 3161 Ta61 6397 Ta91 15,319 performance of the IG algorithm for other problems.
Ta02 1528 Ta32 3432 Ta62 6234 Ta92 15,085 Several issues deserve further investigation. First, it would be
Ta03 1460 Ta33 3211 Ta63 6121 Ta93 15,376 interesting to implement the Tabu-based reconstruction method
Ta04 1588 Ta34 3339 Ta64 6026 Ta94 15,200 for improving the performance of other metaheuristics. Second,
Ta05 1449 Ta35 3356 Ta65 6200 Ta95 15,209
it would be meaningful to consider more powerful problem-
Ta06 1481 Ta36 3347 Ta66 6074 Ta96 15,109
Ta07 1483 Ta37 3231 Ta67 6247 Ta97 15,395 dependent neighborhood search approaches to strengthen the
Ta08 1482 Ta38 3235 Ta68 6130 Ta98 15,237 exploration ability of various algorithms for solving the NWFSP.
Ta09 1469 Ta39 3072 Ta69 6370 Ta99 15,100 Third, developing similar algorithms for other scheduling prob-
Ta10 1377 Ta40 3317 Ta70 6381 Ta100 15,340
lems involving different processing configurations and/or different
Ta11 2044 Ta41 4274 Ta71 8077 Ta101 19,681
Ta12 2166 Ta42 4177 Ta72 7880 Ta102 20,096 objective functions is a potential research direction. Finally, the
Ta13 1940 Ta43 4099 Ta73 8028 Ta103 19,913 idea of combining Tabu-based mechanisms with the IG algorithm
Ta14 1811 Ta44 4399 Ta74 8348 Ta104 19,928 for other combinatorial optimization problems could a worthwhile
Ta15 1933 Ta45 4322 Ta75 7958 Ta105 19,843 avenue for future research.
Ta16 1892 Ta46 4289 Ta76 7801 Ta106 19,942
Ta17 1963 Ta47 4420 Ta77 7866 Ta107 20,112
Ta18 2057 Ta48 4318 Ta78 7913 Ta108 20,056 Acknowledgments
Ta19 1973 Ta49 4155 Ta79 8161 Ta109 19,918
Ta20 2051 Ta50 4283 Ta80 8114 Ta110 19,935 This work is supported by the National Natural Science Founda-
Ta21 2973 Ta51 6129 Ta81 10,700 Ta111 46,689
Ta22 2852 Ta52 5725 Ta82 10,594 Ta112 47,275
tion of China under Grants 61273233, 61473141 and 61104176, as
Ta23 3013 Ta53 5862 Ta83 10,611 Ta113 46,544 well as the Research Foundation for the Doctoral Program of Higher
Ta24 3001 Ta54 5788 Ta84 10,607 Ta114 46,899 Education under Grants 20120002110035 and 20130002130010.
Ta25 3003 Ta55 5886 Ta85 10,539 Ta115 46,741
Ta26 2998 Ta56 5863 Ta86 10,690 Ta116 46,941
Ta27 3052 Ta57 5962 Ta87 10,825 Ta117 46,509
Appendix A. Supplementary Data
Ta28 2839 Ta58 5926 Ta88 10,839 Ta118 46,873
Ta29 3009 Ta59 5876 Ta89 10,723 Ta119 46,743 Supplementary data associated with this article can be found, in
Ta30 2979 Ta60 5958 Ta90 10,798 Ta120 46,847 the online version, at http://dx.doi.org/10.1016/j.asoc.2015.02.006.

are taken directly from [39]. As can be seen in the table, we have References
updated a total of 43 new best solutions for the NWFSP on Tail-
[1] M. Pinedo, Scheduling: Theory, Algorithms, and Systems, Springer, 2012.
lard’s data set. These best values can be a reference for future work [2] C. Rajendran, A no-wait flowshop scheduling heuristic to minimize makespan,
along this line of research. Interested readers are referred to the J. Oper. Res. Soc. 45 (4) (1994) 472–478.
online supplementary material of this work on the journal’s web- [3] N.G. Hall, C. Sriskandarajah, A survey of machine scheduling problems with
blocking and no-wait in process, Oper. Res. 44 (3) (1996) 510–525.
site to access the sequences of jobs relative to these updated upper [4] T. Aldowaisan, A. Allahverdi, Minimizing total tardiness in no-wait flowshops,
bounds. Found, Comput. Decis. Sci. 37 (2012) 149–162.
[5] D. Shabtay, The just-in-time scheduling problem in a flow-shop scheduling
system, Eur. J. Oper. Res. 216 (3) (2012) 521–532.
5. Conclusion and future work [6] Z. Wang, W. Xing, F. Bai, No-wait flexible flowshop scheduling with no-idle
machines, Oper. Res. Lett. 33 (6) (2005) 609–614.
In this paper, we have presented an improved iterated greedy [7] A. Agnetis, Scheduling no-wait robotic cells with two and three machines, Eur.
J. Oper. Res. 123 (2) (2000) 303–314.
algorithm with a Tabu reconstruction strategy for the NWFSP with a [8] A. Schrijver, Combinatorial Optimization: Polyhedra and Efficiency, Springer,
makespan criterion. Unlike many other reported population-based 2003.
metaheuristics in the literature, the proposed TMIIG algorithm is [9] P.-C. Chang, S.-H. Chen, Integrating dominance properties with genetic algo-
rithms for parallel machine scheduling problems with setup times, Appl. Soft
simple and can be easily replicated. Despite its simplicity, the devel- Comput. 11 (2011) 1263–1274, http://dx.doi.org/10.1016/j.asoc.2010.03.003.
oped algorithm provides promising computational results. [10] L. Wang, C. Fang, An effective estimation of distribution algorithm for the multi-
The good performance of the TMIIG algorithm largely depends mode resource-constrained project scheduling problem, Comput. Oper. Res. 39
(2) (2012) 449–460.
on the Tabu-based reconstruction strategy in the algorithm frame- [11] Q.-K. Pan, M. Fatih Tasgetiren, Y.-C. Liang, A discrete particle swarm optimiza-
work. This reconstruction technique improves the exploitation tion algorithm for the no-wait flowshop scheduling problem, Comput. Oper.
ability of the algorithm and leads to better performance when Res. 35 (9) (2008) 2807–2839.
 J.-Y. Ding et al. / Applied Soft Computing 30 (2015) 604–613 613

[12] P.-C. Chang, M.-H. Chen, M.K. Tiwari, A.S. Iquebal, A block-based evolutionary [31] D. Laha, U.K. Chakraborty, A constructive heuristic for minimizing makespan
algorithm for flow-shop scheduling problem, Appl. Soft Comput. 13 (12) (2013) in no-wait flow shop scheduling, Int. J. Adv. Manuf. Technol. 41 (1–2) (2009)
4536–4547. 97–109.
[13] Q.-K. Pan, M. Fatih Tasgetiren, P.N. Suganthan, T.J. Chua, A discrete artificial bee [32] M. Nawaz, E. Enscore, I. Ham, A heuristic algorithm for the m-machine, n-job
colony algorithm for the lot-streaming flow shop scheduling problem, Inf. Sci. flow-shop sequencing problem, Omega 11 (1) (1983) 91–95.
181 (12) (2011) 2455–2468. [33] T. Aldowaisan, A. Allahverdi, New heuristics for no-wait flowshops to minimize
[14] F. Ahmadizar, A new ant colony algorithm for makespan minimization in per- makespan, Comput. Oper. Res. 30 (8) (2003) 1219–1231.
mutation flow shops, Comput. Ind. Eng. 63 (2) (2012) 355–361. [34] C.J. Schuster, J.M. Framinan, Approximative procedures for no-wait job shop
[15] R. Chiong, S. Dhakal (Eds.), Natural Intelligence for Scheduling, Planning and scheduling, Oper. Res. Lett. 31 (4) (2003) 308–318.
Packing Problems, vol. 250, Springer, 2009. [35] J. Grabowski, J. Pempera, Some local search algorithms for no-wait flow-
[16] Q.-K. Pan, R. Ruiz, Local search methods for the flowshop scheduling problem shop problem with makespan criterion, Comput. Oper. Res. 32 (8) (2005)
with flowtime minimization, Eur. J. Oper. Res. 222 (1) (2012) 31–43. 2197–2212.
[17] R. Ruiz, T. Stützle, A simple and effective iterated greedy algorithm for the [36] B. Liu, L. Wang, Y.-H. Jin, An effective PSO-based memetic algorithm for flow
permutation flowshop scheduling problem, Eur. J. Oper. Res. 177 (3) (2007) shop scheduling, IEEE Trans. Syst. Man Cybern. B: Cybern. 37 (1) (2007)
2033–2049. 18–27.
[18] R. Ruiz, T. Stützle, An iterated greedy heuristic for the sequence dependent [37] B. Qian, L. Wang, R. Hu, D. Huang, X. Wang, A DE-based approach to no-wait
setup times flowshop problem with makespan and weighted tardiness objec- flow-shop scheduling, Comput. Ind. Eng. 57 (3) (2009) 787–805.
tives, Eur. J. Oper. Res. 187 (3) (2008) 1143–1159. [38] L.-Y. Tseng, Y.-T. Lin, A hybrid genetic algorithm for no-wait flowshop sched-
[19] Q.-K. Pan, L. Wang, B.-H. Zhao, An improved iterated greedy algorithm for the uling problem, Int. J. Prod. Econ. 128 (1) (2010) 144–152.
no-wait flow shop scheduling problem with makespan criterion, Int. J. Adv. [39] B. Jarboui, M. Eddaly, P. Siarry, A hybrid genetic algorithm for solving no-wait
Manuf. Technol. 38 (7–8) (2008) 778–786. flowshop scheduling problems, Int. J. Adv. Manuf. Technol. 54 (9–12) (2011)
[20] I. Ribas, R. Companys, X. Tort-Martorell, An iterated greedy algorithm for the 1129–1143.
flowshop scheduling problem with blocking, Omega 39 (3) (2011) 293–301. [40] H. Samarghandi, T.Y. ElMekkawy, A meta-heuristic approach for solving the
[21] J. Ding, S. Song, R. Zhang, J.N. Gupta, C. Wu, Accelerated methods for total no-wait flow-shop problem, Int. J. Prod. Res. 50 (24) (2012) 7313–7326.
tardiness minimisation in no-wait flowshops, Int. J. Prod. Res. 53 (4) (2015) [41] J.M. Framinan, R. Leisten, C. Rajendran, Different initial sequences for the
1002–1018. heuristic of Nawaz, Enscore and Ham to minimize makespan idletime or flow-
[22] M. Fatih Tasgetiren, Q.-K. Pan, P.N. Suganthan, O. Buyukdagli, A variable iter- time in the static permutation flowshop sequencing problem, Int. J. Prod. Res.
ated greedy algorithm with differential evolution for the no-idle permutation 41 (2003) 121–148, http://dx.doi.org/10.1080/00207540210161650.
flowshop scheduling problem, Comput. Oper. Res. 40 (7) (2013) 1729–1743. [42] X. Li, Q. Wang, C. Wu, Heuristic for no-wait flow shops with makespan mini-
[23] J.N.D. Gupta, E.F. Stafford Jr., Flowshop scheduling research after five decades, mization, Int. J. Prod. Res. 46 (9) (2008) 2519–2530.
Eur. J. Oper. Res. 169 (3) (2006) 699–711. [43] N. Mladenović, P. Hansen, Variable neighborhood search, Comput. Oper. Res.
[24] D. Wismer, Solution of the flowshop-scheduling problem with no intermediate 24 (11) (1997) 1097–1100.
queues, Oper. Res. 20 (3) (1972) 689–697. [44] K. Gao, Q. Pan, P. Suganthan, J. Li, Effective heuristics for the no-wait flow
[25] M.R. Garey, D.S. Johnson, Computer and Intractability: A Guide to the NP- shop scheduling problem with total flow time minimization, Int. J. Adv. Manuf.
Completeness, WH Freeman and Company, New York, NY, 1990. Technol. 66 (9–12) (2013) 1563–1572.
[26] H. Röck, The three-machine no-wait flow shop is NP-complete, J. Assoc. Com- [45] P.J. van Laarhoven, E.H. Aarts, Simulated Annealing, Springer, 1987.
put. Mach. 31 (2) (1984) 336–345. [46] I. Osman, C. Potts, Simulated annealing for permutation flow-shop scheduling,
[27] M. Bonney, S. Gundry, Solutions to the constrained flowshop sequencing prob- Omega 17 (6) (1989) 551–557.
lem, Oper. Res. Q. 27 (4) (1976) 869–883. [47] C.R. Reeves, Genetic algorithms and neighbourhood search, in: Evolutionary
[28] J. King, A. Spachis, Heuristics for flow-shop scheduling, Int. J. Prod. Res. 18 (3) Computing, Springer, 1994, pp. 115–130.
(1980) 345–357. [48] E. Taillard, Benchmarks for basic scheduling problems, Eur. J. Oper. Res. 64 (2)
[29] R. Gangadharan, C. Rajendran, Heuristic algorithms for scheduling in the no- (1993) 278–285.
wait flowshop, Int. J. Prod. Econ. 32 (3) (1993) 285–290. [49] D.C. Montgomery, Design and Analysis of Experiments, John Wiley & Sons,
[30] E. Bertolissi, Heuristic algorithm for scheduling in the no-wait flow-shop, J. 2008.
Mater. Process. Technol. 107 (1) (2000) 459–465.