Documenti di Didattica
Documenti di Professioni
Documenti di Cultura
Mohamed Kurdi
PII: S2210-6502(18)30191-3
DOI: https://doi.org/10.1016/j.swevo.2018.10.012
Reference: SWEVO 461
Please cite this article as: M. Kurdi, Ant colony system with a novel Non-DaemonActions procedure for
multiprocessor task scheduling in multistage hybrid flow shop, Swarm and Evolutionary Computation
BASE DATA (2018), doi: https://doi.org/10.1016/j.swevo.2018.10.012.
This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to
our customers we are providing this early version of the manuscript. The manuscript will undergo
copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please
note that during the production process errors may be discovered which could affect the content, and all
legal disclaimers that apply to the journal pertain.
ACCEPTED MANUSCRIPT
Ant colony system with a novel Non-DaemonActions procedure for multiprocessor tasks
scheduling in multistage hybrid flow shop
ABSTRACT
This paper presents an ant colony system with a novel Non-DaemonActions procedure (ACSNDP)
PT
algorithm for multiprocessor tasks scheduling in multistage hybrid flow shop. A DaemonActions
procedure is an optional component of ant colony optimization (ACO), which integrates problem-
RI
specific actions that cannot be performed by single ants such as the actions performed by local
search routines. In many applications to hard combinatorial optimization problems, ACO performs
SC
best when integrated with local search routines because they move the ants’ solutions to their local
optimums. However, such an integration is shown as an effective approach only experimentally, and
does not have any effects on the convergence properties of ACO theoretically, since the validity of
U
convergence proofs depends only on the way solutions are constructed and not on the fact that the
AN
solutions are moved or not to their local optimums. Furthermore, it can be noticed that the traditional
DaemonActions procedure does not interfere in the way solutions are constructed because local
search routines have been always integrated with ACO in a daemon fashion i.e. they have been made
M
hidden to the ants, and the ants do not how their solutions have been relocated. Consequently, the
ants may perform limited exploitation (intensification) because they cannot exploit the problem-
D
specific knowledge brought by the local search routine in such a way that enables them to construct
TE
these local optimums by themselves in the upcoming tour. In order to overcome these limitations, a
novel Non-DaemonActions procedure that can interfere positively in the way solutions are
constructed is proposed, which is basically based on the following idea. Iteratively at the end of each
EP
tour, the local search routine tries to improve the constructed solution, and then if a local optimum
has been found, the ant learns the modifications made by the local search routine on its solution, and
performs the corresponding modifications in the pheromone concentrations and heuristic information
C
that will probably enable it to construct this local optimum by itself, in the upcoming tour, before the
AC
application of the local search over again. If the ant can do that, it will be able to construct more
accurate local optimums in the upcoming tours by repeating the whole process over and over again,
and thus enhance its exploitation capabilities. The proposed algorithm is tested on 700 well-known
benchmark instances, with the proposed Non-DaemonActions procedure, and without it using the
classical alternatives, and also compared with other 12 algorithms well-known in the literature.
Computational results verify the improvements achieved by the proposed procedure, and show the
superiority of the proposed algorithm over 7 of the compared works in terms of solution quality.
1
ACCEPTED MANUSCRIPT
Keywords: Multistage hybrid flow shop; multiprocessor tasks; Ant colony optimization;
DaemonActions procedure; local search; Metaheuristic.
1 Introduction
PT
performance goals such as a flow time, tardiness, lateness, and makespan can be optimized
[1]. Due to growing competition between manufacturing companies in terms of shorter
RI
product life cycles, custom-made products, and variable demand pattern, effective scheduling
has become a requirement for survival in the market place [2]. Flow shop problem (FSP) is a
SC
well-known type of workshops in which a set of jobs flow in the same order through multiple
production stages, where each stage consists of only one processor (machine). Hybrid flow
shop problem (HFSP), also called flexible flow shop problem (FFSP) is an extension of the
U
FSP that allows the availability of one or more identical, uniform, or unrelated parallel
AN
processors in any stage (at least one stage must have more than one processor), as illustrated
in Fig. 1, and thus provides additional flexibility and efficiency for the production system [3].
M
HFSP has been widely used in several industrial settings such as airplane engine production,
electronics manufacturing, semiconductors, and petrochemical production [4]. Given its
D
essential complexity and practical relevance, HFSP has been addressed by a significant
number of researchers and industrialists since the first HFSP research paper proposed by
TE
Arthanari and Ramamurthy [5] in 1971. Detailed discussions and classifications of the
relevant research on HFSP can be found in the literature surveys of Ruiz and Vázquez-
EP
Rodríguez [6], Ribas et al. [7], and Lee and Loong [8]. As stated in these survey papers,
current available research on HFSP with respect to complexity can be classified into three
categories: two-stage, three-stage and multistage (m-stage, where m>3) HFSPs. Meanwhile,
C
the solution approaches for solving these HFSPs can be classified into exact algorithms and
AC
approximation algorithms.
With regard to exact algorithms, Ruiz and Vázquez-Rodríguez [6] pointed out that
branch and bound (B&B) is the most favored one for finding optimal solutions for the HFSP,
but even a relatively simple HFSP such as a two-stage problem with either stage having two
identical processors is NP-hard in the strong sense [9]; therefore, the general efficiency of the
available B&B algorithms is rather bounded. For a detailed discussion and taxonomy of the
B&B algorithms that have been proposed to solve various HFSPs, the reader is referred to the
2
ACCEPTED MANUSCRIPT
literature survey of Kis and Pesch [10]. Since exact algorithms require prohibitive amounts of
computational cost for finding optimal solutions even for moderately sized HFSP, researchers
usually prefer developing approximation algorithms, which do not guarantee finding optimal
solutions, but there is a considerable probability of finding near-optimal solutions in
acceptable computational cost. These approximation algorithms include, but not limited to,
constructive heuristics and iterative (or improvement) heuristics.
PT
Constructive heuristics are procedures that construct the solutions step by step (job by
job or stage by stage). Guinet et al. [11] proposed a heuristic for the makespan minimization
RI
in the two-stage HFSP based on Johnson’s rule. Gupta and Tunc [12] applied four heuristics
to minimize makespan for the two-stage HFSP with separable setup and removal times. Lee
SC
and Vairaktarakis [13] presented a heuristic to minimize makespan for solving the two-stage
HFSP. Riane et al. [14] applied two heuristics to minimize makespan for the three-stage
HFSP. Hunsucker and Shah [15] evaluated the performance of different dispatching rules for
U
a constrained HFSP. Kyparisis and Koulamas [16] developed some heuristics to minimize the
AN
makespan for the HFSP with uniform parallel machines in each stage. [17] Moccellin et al.
proposed constructive heuristics based on priority rules for scheduling HFSP with machine
M
blocking and setup times, with the objective of makespan minimization. However,
constructive heuristics are relatively quick, but can easily stagnate in local optimums; on the
D
other hand, iterative heuristics are generally slower, but produce solutions of a top quality,
since they have the ability to escape from local optimums by operating iteratively on
TE
batch processing machines and total weighted tardiness minimization. Shao et al. [19]
presented a hybrid discrete teaching-learning based metaheuristic for solving a no-idle HFSP
AC
with total tardiness minimization. Jin et al. [20] applied a genetic algorithm (GA) to minimize
makespan for a real-life printed circuit board assembly line, which is a three-stage HFSP.
Besbes et al. [21] developed a GA for makespan minimization in HFSP with non-fixed
availability constraints. Costa et al. [22] presented a novel GA to address the makespan
minimization for a HFSP with parallel batching and eligibility constraints. Choi and Wang
[23] proposed a novel decomposition-based approach, which combines both the shortest
processing time and GA, to minimize makespan for the HFSP with stochastic processing
3
ACCEPTED MANUSCRIPT
times. Wardono and Fathi [24] applied a tabu search (TS) algorithm to minimize makespan in
HFSP with limited buffer capacities. Chen et al. [25] developed a TS algorithm to minimize
makespan for the HFSP with precedence and blocking constraints. Zhang et al. [26] handled
the HFSP with lot streaming and the objective of total flow time minimization using an
effective modified migrating birds optimization (MBO) algorithm. Meng et al. [27] proposed
an improved (MBO) algorithm to handle the same problem, but with the objective of
PT
makespan minimization. Peng et al. [28] developed an improved artificial bee colony (ABC)
algorithm for a real-world hybrid flow shop rescheduling problem in steelmaking-refining
RI
continuous casting process, where the solution quality was measured by employing a
weighted sum of six realistic objectives. The no-wait HFSP under makespan and flowtime
SC
criteria was addressed by Riahi and Kazemi [29] using a hybrid ACO with simulated
annealing (SA) algorithm. A memetic algorithm (MA) was proposed by Tavakkoli-
Moghaddam et al. [30] in HFSP with processor blocking environment. Tang et al. [31]
U
proposed a neural network model and algorithm to solve the dynamic HFSP. Zandieh et al.
AN
[32] applied an immune algorithm (IA) for the HFSP with sequence-dependent setup times.
Li et al. [33] proposed a hybrid fruit fly optimization algorithm (HFOA) to solve the hybrid
M
flowshop rescheduling problem with flexible processing time in steelmaking casting systems.
Wang et al. presented [34] novel multidisciplinary approaches to artificial swarm intelligence
D
for heterogeneous computing and cloud scheduling. Li et al. [35] developed a discrete ABC
algorithm to solve the multiple resource constrained HFSP in a steelmaking casting
TE
production system, with the objective of makespan minimization. Marichelvam and Geetha
[36] addressed a bi-objective (makespan and mean flow time) real-industrial HFSP using a
EP
cuckoo search (CS) algorithm. Bathrinath et al. [37] proposed metaheuristic approaches
(improved versions of particle swarm optimization (PSO) and harmony search (HS)
C
algorithms) for solving identical parallel processor problem with the objective of
mınimization of makespan and total tardiness. Han et al. [38] developed an improved NSGA-
AC
II algorithm for a multi-objective (the delivery time penalty and the load imbalance penalty)
HFSP. Gong et al. [39] presented a novel hybrid ABC algorithm for a multi-objective
(makespan and earliness time) HFSP with blocking lot-streaming. Li et al. [40] applied an
energy-aware multi-objective optimization algorithm (EA-MOA) for solving the HFSP with
consideration of the setup energy consumptions, where two objectives, namely, the
minimization of the makespan and the energy consumptions, were considered
simultaneously. The classical HFSP was handled by the following studies: TS [41], bat
4
ACCEPTED MANUSCRIPT
algorithm (BA) [42], ACO [43], artificial immune system (AIS) [44], novel discrete ABC
[45], hybrid variable neighbourhood search (HVNS) that combines the chemical-reaction
optimization (CRO) and the estimation of distribution (EDA) [46], hybrid PSO and
bottleneck heuristic [47], FOA [48], and improved CS [49].
In this work, the minimization of the makespan for the HFSP with multiprocessor task
(HFSPMT) is considered. HFSPMT is an extension of the HFSP that allows any task within a
PT
job to be processed by one or more identical processors simultaneously at any stage.
HFSPMT models several real-life industrial settings as for the textile industry, bio-process
RI
industry, cars manufacturing, parallel computing, distributed systems, real-time machine
vision systems, transportation problems, berth allocation of container terminal, and work-
SC
force assignment [50]. In the literature so far, 5 heuristic and 23 metaheuristic approaches
have been proposed for this problem with the objective of makespan minimization. The
heuristic approaches are priority rules-based heuristic [51], iterated greedy (IG) [52], parallel
U
greedy approach (PGA) [53], discrepancy search heuristic (DS) [54], and the heuristic of
AN
multistage hybrid flowshops (HMHF) [55]. The metaheuristic approaches are SA [56] [57],
restricted SA (RSA) [58], TS [59], GA [60] [61] [62] [63], ACO [3], PSO [64] [65] [66]
M
[67], HS [68] [69], discrete firefly (DF) [70], shuffled frog-leaping algorithm (SFLA) [71],
MA [72], two hybrid algorithms: PSO with TS and PSO with SA [73], hybrid AIS with IG
D
[74], AIS based on novel dispatching rules [75], and hybrid ABC [76].
ACO is a well-known metaheuristic approach which has been applied successfully for
TE
specific or centralized actions that cannot be performed by single ants such as a local search
process. In many applications to hard combinatorial optimization problems, ACO performs
C
best when integrated with local search routines because they move the ants’ solutions to their
local optimums [77]. However, such an integration is shown as an effective approach only
AC
experimentally, and does not have any effects on the convergence properties of ACO
theoretically, since the validity of convergence proofs (as presented in [78]) depends only on
the way solutions are constructed and not on the fact that the solutions are moved or not to
their local optimums by a local search routine [79]. This gap between the experimental and
theoretical findings has inspired us to search for a novel Non-DaemonActions procedure that
can interfere (hopefully, positively) in the way solutions are constructed. Generally, it can be
observed that the traditional DaemonActions procedure cannot interfere in the way solutions
5
ACCEPTED MANUSCRIPT
are constructed because local search routines have been always integrated with ACOs in a
daemon fashion i.e. they have been made hidden to ants, and ants do not know how their
solutions have been moved to the new locations (their local optimums); for that reason, the
ants cannot exploit the knowledge brought by the local search routine in such a way that
enables them to construct these local optimums by themselves in the upcoming tour.
Consequently, this may reduce the exploitation (or intensification) capabilities of ACO.
PT
In this work, in order to overcome the previously discussed limitations of the
RI
traditional DaemonActions procedure and solve HFSPMT with the objective of makespan
minimization more effectively, a novel Non-DaemonActions procedure is proposed. In the
SC
proposed procedure, the local search routine moves the ants’ solutions to their local
optimums in the current iteration, but in the meantime informs the ants with the
U
corresponding changes in the relative pheromone concentrations and heuristic information,
by which they will probably construct these local optimums by themselves using the state
AN
transition rule, in the upcoming tour, before the application of the local search routine over
again. If the ants can do that, they will be able to construct more accurate local optimums in
M
the upcoming tours by repeating the whole process over and over again, and thus acquire
better exploitation capabilities. To the best of our knowledge, this is the first work that
D
considers informing the ants with the improvements made on their solutions by the local
TE
search routine, and then letting them update the pheromone concentrations and heuristic
information accordingly.
The remainder of this paper is organized as follows. In the next section, HFSPMT
EP
definition and formulation are presented. Section 3 describes the encoding and decoding
method. In Section 4, the ACSNDP framework is explained. Section 5 reports the
C
computational results. The conclusions and further works are provided in Section 6.
AC
2 Problem definition
6
ACCEPTED MANUSCRIPT
processing time ( Pij ) . Processors might process, at most, one job at a time and without pre-
emption. The objective is to minimize makespan i.e. the maximum completion time of all
tasks in the last stage. The problem can be denoted by FK ( Pm1 , ..., Pm k ) | sizeij | C max using the
classification scheme of Graham et al. [80], and can be modelled as a mixed integer linear
programming (MILP) problem as follows (a revised version of the formulation proposed by
Lin et al. [76]). To simplify the formulation, the following notations are used.
PT
Parameters:
RI
n number of jobs
k number of stages
mi number of identical parallel processors at stage i
SC
pij processing time of job j at stage i
sizeij number of parallel processors required to process job j at stage i
T planning horizon for which the schedule is to be developed
Decision variables:
U
AN
Xijt a binary time index, which is equal to 1 if job j is processed at stage i in time
period t and equal to 0 otherwise
Xidjt a binary time index, which is equal to 1 if job j is processed at stage i on processor d
M
Cmax makespan
TE
Let job 0 be the dummy initial job. Then, the FK ( Pm1 ,..., Pm k | sizeij | C max can be formulated
as follows.
EP
Min Cmax
Subject to,
C
∑
n
j =1
X ijt sizeij ≤ mi , i = 1,..., k , t = 1,...T , (1)
AC
∑
n
j =1
X idjt ≤ 1, i = 1,..., k , d = 1,..., mi , t = 1,...T , (2)
∑
T
t =1
X ijt = pij , i = 1,..., k , j = 1,..., n,
(5)
7
ACCEPTED MANUSCRIPT
S ij ≤ t + T (1 − X ijt ), i = 1, ..., k , j = 1, ..., n, t = 1,..., T ,
(6)
C max ≥ C kj , j = 1,..., n ,
(8)
PT
Sij ∈ {1,..., T} , i = 1,..., k , j = 1,..., n, t = 1,..., T ,
(10)
Ci 0 = 0, i = 1,..., k ,
RI
(11)
SC
(12)
Constraint set (1) specifies that the number of processors assigned for each job at each stage
and during each time period should not exceed the number of available processors. Constraint
U
set (2) states at most one job can be assigned for each processor at each stage and during each
AN
time period. Constraint set (3) imposes the condition that no job task can be started until its
preceding task is finished. Constraint set (4) is the starting time constraint, which is taken
M
from the processing time requirement. Constraint set (5) represents the time period that each
job at each stage occupies. Constraint set (6) specifies that, at each stage, each job will
D
occupy its required number of processors from its starting time at that stage. Constraint set
(7) denotes that, at each stage, each job will occupy its required number of processors until its
TE
finishing time at that stage. Constraint set (8) gives the makespan. Finally, constraint sets (9)–
(12) define the decision variables.
EP
An example of a 5×2 HFSPMT is given in Table 1. The data include the number of
identical parallel processors at each stage. For example, m1 = 3 specifies that there are 3
C
processors at the first stage. The data also include the processing time and the number of
AC
processors required at each stage for each job. For example, job 1 must be processed by 2
processors in the first stage for 10 time units, and then by one processor in the second stage
for 15 time units. A possible solution of the 5×2 HFSPMT represented by a Gantt chart is
given in Fig. 2.
8
ACCEPTED MANUSCRIPT
let
be the set of processors on which job
is processed at stage i. If we define =
{ , . . . , } for each stage i, then the corresponding schedule is represented by a processing
order (, ), where = { , . . . , } and = { , . . . , }. For example, the solution of the
5×2 HFSPMT given in Fig. 2 can be represented by = { , } and = { , },
where = { , , , , } and = { , , , , }. The individual lists
and sets
are as follows: = (4,2,5,3,1), = (4,2,3,5,1), = {1,2,3}, =
PT
{1,2,3}, = {1}, = {2,3}, = {2,3}, = {1,2}, = {1,2}, = {1}, =
{1,2}, = {1}. For more details, please refer to [59].
RI
Due to the complex nature of HFSPMT, most of the studies suggest decomposing it
into two sequential decision problems: determining the jobs-permutation at the first stage
SC
by a metaheuristic, followed by a decoding method to generate the jobs-permutations for the
subsequent stages ( , … , ) [65]. Furthermore, several studies such as Oğuz et al. [59],
U
Sivrikaya-Şerifoğlu and Ulusoy [60], and Ying and Lin [3] recommend using two-directional
AN
(i.e. forward and backward) scheduling at the first stage; therefore, the proposed ACSNDP
uses the same approach. That is, the forward schedule is constructed by determining using
an ACS approach, and then ( , … , ) by queuing the jobs by the first-come-first-served
M
(FCFS) manner i.e. by arranging them in ascending order according to their completion times
at the previous stage. The sets of processors = { , . . . , } are obtained as follows. The
D
jobs at each stage i are handled in the order they appear in the list , by assigning them to the
TE
first available sizeij parallel processors simultaneously, while respecting the problem
constraints. The tiebreak rule gives preference to the job with the smaller order number. On
EP
the other hand, the same procedure is applied to the inverse problem to obtain the backward
schedule. A proof of the equivalence between the problem and its inverse can be found in [3].
C
ACO, which are inspired by the foraging behaviour of real ants. Ants are capable of finding
the shortest path from a food resource to their nest without using visual cues, but by using a
chemical substance called pheromone which serves as a communication media between
them. Initially, since there are no clues (pheromone trails) about the food resources, ants
search the areas surrounding their nests randomly. As soon as an ant finds a food resource, it
carries a portion of food back to the nest, and while walking, it lays pheromone on the ground
in order to remember the path and also guide other ants to the food. Since ants reaching a
9
ACCEPTED MANUSCRIPT
food resource by shorter paths will come back to the nest earlier than ants travelling through
longer paths, the shorter paths will have higher traffic volumes than those of the longer ones,
and thus the pheromone will accumulate more quickly on shorter paths than on longer ones.
Consequently, by means of the accumulated pheromone, the search process becomes more
organized and focused around the good areas because the more pheromone on a path, the
more likely that it will be selected by ants [82]. The described foraging behaviour of ants can
PT
be used to solve scheduling problems by simulation: the objective function value (e.g.
makespan) corresponds to the quality of the food source (e.g. distance), artificial ants
RI
searching in search space simulate real ants searching in their environment, and an adaptive
memory corresponds to the pheromone trail [3].
SC
The pseudocode of ACSNDP is presented in Algorithm 1. As shown in Algorithm 1,
ACSNDP is composed of two steps: the initialization step and the main loop. In the
initialization step, the parameters β , ρ1 , ρ 2 , r0 ,τ 0 , l are set; a colony of l ants are positioned
U
on a dummy start job (node); the initial pheromone concentrations are set to small values in
AN
the range [0,τ 0 ] ; and the heuristic information are set, as proposed by Ying and Lin [3], to
the jobs’ ranking indexing of the selected constructive heuristic (SCH) ȥ at the first stage. In
M
the main loop step, each ant repeatedly applies the state transition rule to select the next job
from the set of unvisited jobs, adds it to the currently constructed path (list of jobs-
D
permutation at first stage), and update the concentration of pheromone that exists on it by
TE
applying the local update rule, until a full path (jobs-permutation at first stage) is constructed;
and then, the forward and backward schedules are constructed and the one that has the
minimum makespan is selected. Once all of the ants have terminated their tours, the Non-
EP
DaemonActions procedure is applied on the iteration-best solution to improve it, and then the
concentrations of pheromone on the visited jobs of each ant are updated once more by
C
applying the global updating rule. The main loop step is repeated again until the maximum
AC
number of tours has been reached [82]. In the following sections, the constructive heuristics
that are used to assign heuristic information to jobs, the state transition rule, the local update
rule, the Non-DaemonActions procedure, and the global update rule are discussed in details.
10
ACCEPTED MANUSCRIPT
[3], five constructive heuristics: shortest processing time heuristic; shortest final stage’s
processing time heuristic; revised Campbell, Dudek, and Smith heuristic; revised Raghu and
Rajendran heuristic; and revised Palmer’s heuristic are applied on each benchmark instance,
and the one that produces the minimum makespan is selected from them. For more details,
please refer to [3].
PT
Algorithm 1 Pseudocode of the proposed ACSNDP
RI
/* Initialization */
t = 0 ; set the parameters β , ρ1 , ρ 2 , r0 ,τ 0 , l ;
Let V be the set of jobs (nodes);
SC
Place all ants on a dummy start job;
For each edge ( d , e) between the current job d and each other job e ∈V \ d do
Set the initial pheromone concentrations τ de (t ) : U (0,τ 0 ) ; Set the heuristic informationηde (t ) ;
U
End For
a
Let x1 (t ) be the path (jobs-permutations at the first stage) constructed so far for an ant a;
AN
x a (t ) = [ x1a (t ), . . . , xka (t )] the schedule (jobs-permutations at all stages) corresponding to an ant a;
Cmax ( x a (t )) its objective function value;
x g −b (t ) the global-best solution; Cmax ( x g −b (t )) = ∞ ;
M
/* Main loop */
Repeat
D
i −b i −b
Let x (t ) the iteration-best solution; Cmax ( x (t )) = ∞ ;
For each ant a = 1,..., l do
TE
x1a (t ) = ∅ ;
Repeat
Determine the next job e ∈V \ x1 (t ) using the state transition rule; x1 (t ) = x1 (t ) ∪ e ;
a a a
EP
11
ACCEPTED MANUSCRIPT
End If
i −b
For each edge (u , v) ∈ x1 (t ) do
Apply the global updating rule;
End For
For each edge (u , v ) do
τ de (t + 1) = τ de (t ) ;
End For
x g −b (t + 1) = x g −b (t ) ; C max ( x g − b ( t + 1)) = C max ( x g − b ( t )) ; t = t + 1 ;
PT
Until t reaches the maximum number of tours
Return x g −b (t ) ;
RI
While constructing its tour, an ant a currently located at job d chooses the next job e,
SC
to move to, using following rule.
{
β
arg max {τ du ( t )η du ( t )} if r ≤ r0
e= u∈ N d ( t )
a
(13)
E if r > r0
U
Where r : U (0,1) ; r0 ∈ [0,1] is a user-specified parameter; N da (t ) = V \ x1a (t ) is the current set
AN
of unvisited jobs (feasible nodes that are connected to node d); τ du is the pheromone
concentration on the corresponding edge (d,u), and represents the a posteriori effectiveness
M
of the move from job d to job u ( i.e. indicates how profitable it has been in the previous
tours to make a move from d to u); η du (t ) is the heuristic information, which is the job u’s
D
ranking index of the SCH ȥ at the first stage, and expresses the a priori effectiveness of the
TE
move (i.e. attractiveness or desirability of the move); β is a positive constant used to amplify
the influence of the heuristic information; and E ∈ N da (t ) is a job selected randomly
EP
a
d
AC
r ≤ r0 , the ants select the next job using Formula 13 and exploit the knowledge obtained
about the search space by favoring the best moves that have the highest pheromone
concentrations and the best heuristic information; otherwise, ants use Formula 14 to explore
the search space by investigating previously unseen moves [82]. Once an ant a finishes its
tour, the jobs-permutation of n jobs at the first stage (which has been already stored in x1a (t ) )
is used to determine the jobs-permutations on the rest of stages using the FCFS manner.
12
ACCEPTED MANUSCRIPT
4.3 Local update rule
During its tour, an ant changes the pheromone concentrations on its path (visited
edges) by applying the local updating rule as follows.
τ de (t ) = (1 − ρ2 )τ de (t ) + ρ2τ 0 (15)
Where τ 0 = LB − 1 is the maximum initial pheromone concentration that is set as the inverse of
the lower bound of the makespan, and ρ 2 ∈ [0,1] is the local pheromone decay coefficient.
PT
This rule is used to decrease the attractiveness of edges dynamically in order to shuffle the
tour i.e. controls the influence of ants’ search history of the previous tours on their decisions
RI
in the current one. If the pheromone concentrations are decreased, ants will forget the
previous tours and explore new ones. If ants explore different paths, then there is a higher
SC
probability that one of them will find an improving solution than there in the case that they
focus on searching in small regions of the search space (the narrow neighbourhoods of the
U
previous best path); consequently, this will provide a better exploration and delay the
AN
premature convergence [81].
makes use of the information found in the surrounding environment such as the pheromone
trails and heuristic information. Once an ant has built a solution, it uses the
EP
changed; the trails can either increase, ants excrete pheromone on the edges they use
(exploitation), or decrease due to pheromone evaporation (exploration). Finally,
AC
13
ACCEPTED MANUSCRIPT
convergence properties of ACO theoretically. This is because the validity of convergence
proofs (as presented in [78]) depends only on the way solutions are constructed and not on
the fact that the solutions are moved or not to their local optimums by a local search routine
[79]; so theoretically, it would be better if the local search routine can directly affect the way
solutions are constructed (in a positive manner). However, it can be seen that the traditional
DaemonActions procedure cannot interfere in the way solutions are constructed by ants
PT
because local search routines have been always integrated with ACOs in a daemon fashion
i.e. the modifications done by the local search routine on the visited edges of ants’ paths,
RI
which relocate them on the best found neighbourhood regions, are applied without informing
the decision holders, which are the ants themselves, and the pheromone concentrations and
SC
heuristic information, associated with the modified edges, that affect the ants’ decisions when
they come to these edges in the upcoming tour are not modified accordingly. Consequently,
the ants may follow the path found by the local search routine when applied on the best ant’s
U
solution (because of the global updating rule), but they will probably not follow the paths
AN
found by it when applied on their solutions; this in turn, may not enable the ants to fully
exploit the knowledge brought by the local search, and reduces the exploitation
M
chance for following the locally optimized paths (previously found by the local search
routine) because they know what kind of modifications have been performed on the original
TE
paths, and thus can make the corresponding modifications in pheromone concentrations and
heuristic information (associated with the modified edges) that enable them to follow these
EP
paths; for this reason, the proposed local search routine is called the Non-DaemonActions
procedure. The difference between the traditional DaemonActions procedure and the
C
proposed one is explained graphically in Fig. 3. As it can be seen from Fig. 3, the basic
assumption is as follows. If an ant cannot find a shorter path from nest to food source in the
AC
near neighbourhood unless an external local search routine is applied on its solution, this
indicates that its knowledge about the surrounding environment i.e. pheromone trails and
heuristic information are not completely accurate and should be updated according to the
knowledge provided by the local search routine, in such a way that enables it to find this path
by itself at the next tour. If an ant can find the shorter path by itself, this will probably entitle
it for finding much shorter paths in the future tours.
14
ACCEPTED MANUSCRIPT
Ying and Lin [3] proposed an effective ACO with a DaemonActions procedure that
integrates a local search routine for HFSPMT, in which, the local search routine starts from
an initial jobs list and tries to improve the current list by swapping job positions in the list
with each of the six nearest jobs. If the change produces a better solution, then it is kept;
otherwise, it is ignored and swapped repeatedly until all the remaining pairs of jobs have
been tested. For this particular DaemonActions procedure, the modifications proposed to
PT
convert it into a Non-DaemonActions are swapping the jobs along with their pheromone
concentrations and heuristic information. The pseudocode of the proposed Non-
RI
DaemonActions procedure is given in Algorithm 2.
Algorithm 2 Pseudocode of the Non-DaemonActions procedure
SC
i −b
Let x1 = [ x11 ,...x1n ] be the array of jobs-permutation at the first stage for the iteration-best solution;
τ = [τ1 ,...τ n ] the array of their corresponding pheromone concentrations;
η = [η1 ,...ηn ] the array of their corresponding heuristic ranking values;
U
s ∈ {2,...n − 1} a randomly selected position and x1 s = x1i − b [ s ] the job located at s;
For Itr = 1 to 6 do
AN
Let e = s + Itr be the neighbouring position the from right side;
If e ≤ n then
M
i −b
Let x1e = x1 [e] be the neighbouring job from the right side;
x1i −b [ s ] = x1e ; x1i −b [e] = x1s ; // swapping the positions of the two jobs
i −b
D
Let x1,se be the resulting jobs-permutation after swapping the positions of x1s and x1e ;
Construct the jobs-permutations for the other stages in a FCFS manner for the forward and
TE
i −b i −b
backward schedules; Compute their objective function values Cmax ( xse, f (t )) and Cmax ( xse,b (t )) ;
Cmax ( xsei −b (t )) = min (Cmax ( xsei −,bf (t )), Cmax ( xsei −,bb (t ))) ;
i −b i −b
If Cmax ( xse ) < Cmax ( x
EP
) then
temp = τ [ x1s ] ; τ [ x1s ] = τ [ x1e ] ; τ [ x1e ] = temp ; // swapping their corresponding pheromone values
temp = η[ x1s ] ; η[ x1s ] = η[ x1e ] ;η[ x1e ] = temp ; // swapping their corresponding heuristic values
C
break;
i −b i −b
Else x1 [ s ] = x1s ; x1 [e] = x1e // undo the swapping
AC
End If
End If
End For
For Itr = 1 to 6 do
Let e = s − Itr be the neighbouring position from the left side;
If e ≥ 1 then
i −b
Let x1e = x1 [e] be the neighbouring job from the left side;
x1i −b [ s ] = x1e ; x1i −b [e] = x1s ; // swapping the positions of the two jobs
i −b
Let x1,es be the resulting jobs-permutation after swapping the positions of x1e and x1s ;
15
ACCEPTED MANUSCRIPT
Construct the jobs-permutations for the other stages in a FCFS manner for the forward and
i −b i −b
backward schedules; Compute their objective function values Cmax ( xes , f (t )) and Cmax ( xes,b (t )) ;
Cmax ( xesi −b (t )) = min (Cmax ( xesi −,bf (t )), Cmax ( xesi −,bb (t ))) ;
i −b i −b
If Cmax ( xes ) < Cmax ( x ) then
temp = τ [ x1s ] ; τ [ x1s ] = τ [ x1e ] ; τ [ x1e ] = temp // swapping their corresponding pheromone values
temp = η[ x1s ] ; η[ x1s ] = η[ x1e ] ;η[ x1e ] = temp // swapping their corresponding heuristic values
break;
PT
i −b i −b
Else x1 [ s ] = x1s ; x1 [e] = x1e // undo the swapping
End If
End If
RI
End For
While Ying and Lin [3] proposed applying the local search routine only on the jobs-
permutation at the first stage, Lin et al. [76] recommended applying it on the jobs-
SC
permutation at the last stage as well, and Chou [65] suggested applying a cocktail of
decoding methods for assigning the jobs to processors in the rest of stages (i.e. i = 2,..., k );
U
therefor, an additional step of refinement, which is not reported in Algorithm 2, was
AN
occasionally performed by applying the local search routine proposed Ying and Lin [3] on a
randomly selected stage, that is other than the first stage, with a probability of 10%; however,
this additional step was found most effective for the smallest size instances (i.e. n = 5 ).
M
The global updating rule is applied after all ants have completed their tours, their
TE
forward and backward schedules have been constructed, and the Non-DaemonActions
procedure has been applied to improve the best schedule of current iteration. This rule makes
the ants in favor of exploitation by encouraging them to search in the neighbourhood of the
EP
iteration-best solution; this is done by increasing the pheromone concentrations of the edges
that belong to its first stage’s jobs-permutation and decreasing the rest, as follows.
C
τ de (t ) = (1 − ρ1 )τ de (t ) + ρ1∆τ de (t ) (16)
AC
1 i −b
if (d , e ) ∈ x1 (t )
∆τ de (t ) = Cmax ( xi −b (t )) (17)
0 otherwise
Where x1i −b (t ) is the jobs-permutation at the first stage of the iteration-best solution,
Cmax ( xi −b (t )) is its makespan value, and ρ1 ∈ [0,1] is the global pheromone decay coefficient
which controls the influence of the iteration-best solution; while large values of ρ1 cause
rapid evaporation and increase the influence of the iteration-best solution, so exploration is
16
ACCEPTED MANUSCRIPT
emphasized; on the other hand, small values of it cause slow evaporation and decrease the
influence of the iteration-best solution, so exploitation is emphasized [82].
5 Computational results
To illustrate the effectiveness of the proposed algorithm, 700 test instances were
considered from two well-known benchmark problem sets. The first set was produced by
PT
Oğuz et al. [59]. In this set, two classes of random instances were generated: Type-Ι and
Type-ΙΙ. For the Type-Ι test instances, the number of processors available at different stages
RI
was randomly chosen to be between one and five; for the Type-ΙΙ test instances, it was set at
five processors for all stages. Moreover, the test instances of Type-Ι and Type-ΙΙ were
SC
categorized into 30 sub-sets by using the following experimental design: the number of jobs n
was set as 5, 10, 20, 50, and 100; the number of stages k was given to be 2, 5, and 8; and the
number of processor requirements for the task of job j at stage i (sizeij) was chosen randomly
U
from the set {1,..., mi }. For each combination of n and k, 10 instances with integer job
AN
processing times uniformly distributed over [1,100] were randomly generated; so, the test bed
is composed of 2×5×3×10=300 benchmark instances. In fact, the complete set contains
M
another 300 instances produced over the range [1,1000], but it has been ignored since that
Oğuz et al. [59] did not report their results on them because they found that the performance
D
does not depend on the interval over which job processing times are uniformly distributed.
TE
The second set was proposed by Sivrikaya-Şerifoğlu and Ulusoy [60]. This set also has two
classes of random instances, namely Type-A and Type-B. For comparison purposes, the
experimental design of this benchmark problem set was equivalent to the benchmark problem
EP
set proposed by Oğuz et al. [59], except that the number of stages (k) was given to be 2, 5, 8
and 10. So, the test bed was composed of 2×5×4×10=400 benchmark instances.
C
The algorithm was implemented in C++, and the tests were run on a PC with 3.40
AC
GHz Intel(R) Core (TM) i7-3770 CPU and 8.00 GB. The parameters were set as follows: the
number of ants 20, the maximum number of iterations (tours) 1000, and the rest of
parameters were set as proposed by Ying and Lin [3]: ρ1 = 0.1 , ρ 2 = 0.1 , τ 0 = (LB) −1 , and
β = 3 ; and r0 = 0.3 for small size problems (i.e. n ≤ 20) and r0 = 0.7 for large size ones (i.e.
n > 20) . In order to study the effects of the algorithmic parameters on the solution quality, a
series of experiments were done an instance of the maximum available size (100x10) called
10010HB1, which belongs to Type-B (which is harder than Type-A) of the second benchmark
set. In all of the following experiments, only one parameter was changed while keeping the
17
ACCEPTED MANUSCRIPT
other parameters fixed. The effects of varying the size of the colony l are shown in Fig. 4; it
is clear that the algorithm finds better solutions in less number of iterations as l increases.
Fig.5 clarifies the effects of varying ρ1 ; it can be noticed that as ρ1 increases, the performance
decreases because of an imbalance between the exploitation and exploration of the search
space i.e. the ants are excessively encouraged to forget their previous search experience and
explore new regions in the neighbourhood of the iteration-best solution. The effects of
PT
varying ρ 2 are shown in Fig. 6; it can be observed that as ρ 2 grows, the solution quality
diminishes because the ants are hurriedly motivated to forget their search history and explore
RI
new areas in the search space randomly. It can be also observed that performance diminishes
slightly when ρ 2 grows from 0.1 to 0.3, and more apparently at the values of 0.5 and 0.7.
SC
Fig. 7 represents the effects of varying β ; it can been seen that as β enlarges, the premature
convergence occurs more rapidly at earlier iterations because the ants are extremely induced
U
to neglect their previous search experience about the search space (represented by the
AN
pheromone concentrations) and explore new areas in favor of the amplified heuristic
information. It can be also seen that when β values are 5 and 7, the search starts from higher
quality solutions, but converges very soon nearby these solutions. On the hand, when β
M
values are 1 and 3, the search starts from inferior solutions, but converges slowly to near-
D
optimal solutions. The effects of varying r0 are explained in Fig. 8; it can be noted that
whereas the algorithm exhibits the best convergence at r0 = 0.7 , it exhibits worse
TE
exploitation ( r0 = 0.5 ), or excessive exploitation ( r0 = 0.9 ). Finally, Fig. 9 shows the effects
EP
of varying the neighbourhood size; it can be noticed that as it gets larger, the performance
enhances. It can be also noticed that even with small neighbourhood size such as 3, the
C
quality.
18
ACCEPTED MANUSCRIPT
transition rule, accordingly (ACSNDP-SP); and an ACSNDP that that swaps only the
positions and pheromone concentrations of the jobs (ACSNDP-SJP).
The average relative deviation of the best solution (ARD-BS) and the average relative
deviation of the average solution (ARD-AvS) from lower bound (LB), for each subset, were
calculated by the following formulas, respectively:
10
BSi − LBi
ARD − BS = ∑( ) × 100 /10 (18)
PT
i =1
LBi
10
AvSi − LBi
ARD − AvS = ∑ ( ) ×100 /10 (19)
RI
i =1
LBi
Where BSi and AvSi are the best solution and average solution, respectively, obtained from 5
SC
independent runs on each instance i; LBi was calculated based on the lower bound developed
by Oğuz et al. [59] using the following formula, where the set of jobs is denoted by J and the
U
set of stages is denoted by M:
AN
i −1
1
m
LB = max
i∈M
min
j∈J k =1
∑
pkj +
mi
∑ j∈J
pij × sizeij +
min
j∈J
∑ pkj (20)
k =i +1
M
Table 2 and Table 3 summarize the computational results of ARD-BS, ARD-AvS, and
the average computational time (ACT) in seconds obtained for the first and second
D
benchmark problem sets, respectively. From Table 2 and Table 3, it can be seen that ACSNDP
TE
is capable of obtaining the best results in terms of number of lowest ARD-BS and ARD-AvS
values obtained over the 70 subsets. For the ARD-BS case, the compared algorithms can be
ordered from best to worst as follows: ACSNDP 55 (78.57%), ACSNDP-SJP 27 (38.57%),
EP
19
ACCEPTED MANUSCRIPT
RI ARD−AvS =100×(OAARD−AvS − ACSNDPARD−AvS )/ OAARD−AvS (22)
PT
computational cost 123% in ACT. Regarding ACSNDP-SJ that uses the traditional
DaemonActions procedure proposed by Ying and Lin [3], the improvements are 2.49 % in
RI
ARD-BS and 2.24% in ARD-AvS, almost in the same computational cost; these
improvements are not so big, but they still close a critical part of the gap to the optimum, and
SC
validate the advantages of the proposed Non-DaemonActions procedure over the traditional
one in improving the exploitation. It can be also noted that ACSNDP-SPJ is also better than
U
ACSNDP-SJ, but the improvement occurred are tiny, which indicates that changing the
pheromone concentrations alone without changing the corresponding heuristic information is
AN
not enough to enable the ants to follow the paths found by the local search routine.
Concerning ACSNDP-SP, its performance is even worse than ACSNDP-SJ, it seems that ants
M
cannot perform a local search process effectively by themselves and the reconstructed paths
were not better than the original ones; this may be due to the fact that the local search routine
D
can interact with the search space directly, but ants can interact with the search space only
TE
indirectly via the probabilistic model. Furthermore, a comparison of convergence curves for
the five versions on the instance 10010HB1 is shown in Fig.10, it can be seen that the order
of the compared algorithms from best to worst in terms of convergence properties is the same
EP
one as obtained in the cases of ARD-BS and ARD-AvS, which supports the results obtained
in ARD-BS and ARD-AvS and confirms the superiority of ACSNDP in terms of convergence
C
20
ACCEPTED MANUSCRIPT
(50.00%), MAJOS 9 (30.00%), SAWCW 6 (20.00%), GAOE 6 (20.00%), ACSNDP 4 (13.33%),
HIAYing 3 (10.00%), ACSYL 2 (6.67%), and GAJOS 1 (3.33%); therefore, ACSNDP is better
than HIAYing, ACSYL, and GAJOS. Table 6 shows the computational results for the second
benchmark problem set. Similarly, the order is as follows: HABCLYH 20 (50.00%), MAJOS 14
(35.00%), SAWCW 11 (27.50%), ACSNDP 6 (15.00%), HIAYing 5 (12.50%), ACSYL 3 (7.50%),
GASU 2 (5.00%), and GAJOS 2 (5.00%). Consequently, the performance of ACSNDP is better
PT
than ACSYL, HIAYing, GASU, and GAJOS. To make a more comprehensive comparison, the
overall overage of AV-BS was calculated over the available instances for each algorithm.
RI
Table 7 shows the number of instances solved (NIS), overall average of AV-BS values for
ACSNDP and the other algorithms (OA), and the improvement achieved by ACSNDP to the
SC
other algorithms. From Table 7 we can notice that ACSNDP yields relative improvements to
all other algorithms except HABCLYH, SAWCW, and HIAYing.
This study also measured the effectiveness of the available algorithms that use the
U
same lower bound approach by considering the number of lowest ARD-BS values obtained
AN
over the subsets. The computational results for the first benchmark problem set are provided
in Table 8, and the algorithms can be ordered as follows: IGYing 16 (53.33%), ACSNDP 6
M
effectiveness of ACSNDP on the first benchmark problem set, ANOVA test was carried on
the ARD-BS values. Fig. 11 reports the means and 95% confidence intervals (CIs). From Fig.
TE
11, it can be noted that ACSNDP is also the second best algorithm in terms of statistical
significance. In addition, the resulting p-value is 0.000019 which means that the compared
EP
has also got the second rank after IGYing. The results of ANOVA test on the second
benchmark problem set are shown in Fig. 12, where it can be observed that ACSNDP is again
at the second rank after IGYing, but there are no significant differences between the compared
algorithms since the resulting p-value is 0.951754. The relative improvements made by
ACSNDP in ARD-BS were also calculated, and the computational results are presented in
Table 10. As revealed from Table 10, ACSNDP yields improvements to all other algorithms
except IGYing and PSOCFA.
21
ACCEPTED MANUSCRIPT
6 Conclusion and further works
PT
was proposed. Unlike the traditional DaemonActions procedure, the proposed Non-
DaemonActions procedure informs the ants about the changes occurred in their paths that
RI
take them to their local optimums, and give them the opportunity to act accordingly i.e. do
the corresponding changes in the environment (changing the pheromone concentrations and
SC
heuristic information) that enable them to reach these local optimums by themselves in the
upcoming tour.
A brief study of the effects of the algorithmic parameters (such as β , ρ1 , ρ 2 , r0 , size of
U
the colony, and size of neighbourhood) on the solution quality was performed and the results
AN
were analysed. The proposed algorithm was compared with other four versions of it that use
different strategies during the Non-DaemonActions procedure phase: ACSNDP-SN that
M
traditional DaemonActions procedure proposed by Ying and Lin [3]); ACSNDP-SP that
TE
swaps only the pheromones concentrations of the jobs and reconstruct the paths using the
state transition rule; and ACSNDP-SJP that that swaps only the positions and pheromone
concentrations of the jobs. The tests we done on the well-known benchmarks problem sets of
EP
Oğuz et al. [59] and Sivrikaya-Şerifoğlu and Ulusoy [60]. The comparisons were based on
four performance criteria: the number of lowest ARD-BS (the average relative deviation of
C
the best solution) values, the number of lowest ARD-AvS (the average relative deviation of
AC
the average solution) values, the overall average of ARD-BS, the overall average of ARD-BS
values over the two benchmark problem sets, and convergence curves.
The computational results verified the superiority of the Non-DaemonActions
procedure over all of the considered alternatives in terms of effectiveness and robustness. The
highest relative improvements with respect to solution quality (ARD-BS and ARD-AvS
values) were in the case of ACSNDP-SN, but with a considerable increase in the
computational cost. In the comparison made with ACSNDP-SJ, the results of ACSNDP of
were better, at almost the same computational cost. Furthermore, it was observed that
22
ACCEPTED MANUSCRIPT
ACSNDP-SPJ is also better than ACSNDP-SJ, but the improvements occurred were tiny.
Concerning ACSNDP-SP, its performance was even worse than ACSNDP-SJ.
ACSNDP was also compared with other 12 works well-known in the literature:
ACSYL [3], IGYing [52], SAWCW [57], TSOZDJL [59], GASU [60], GAOE [61], PSOTL [64], GAJOS
& MAJOS [72], PSOCFA [73], HIAYing [74], and HABCLYH [76]. The comparisons were based
on five performance criteria: the number of lowest AV-BS (the average of the best solution)
PT
obtained by the compared algorithms over the subsets of the first benchmark set and subsets
of the second benchmark, the overall overage of AV-BS over all of the solved instances
RI
(either one, the other or both of benchmark sets), the number of lowest ARD-BS values
obtained over the subsets of the first benchmark set and subsets of the second benchmark, the
SC
overall overage of ARD-BS over all of the solved instances, and ANOVA test. The
computational results showed that ACSNDP can give relative improvements in ARD-BS and
ARD-AvS values to all of the compared algorithms, except HABCLYH [76], SAWCW [57], and
U
HIAYing [74] in the case of the AV-BS; and IGYing [52] and PSOCFA [73] in the case of the
AN
ARD-BS; Therefore, ACSNDP can be considered as an effective approach for solving
HFSPMT.
M
since it yields improvements in the premature convergence when applied on HFSPMT, which
is one of the most intractable combinatorial optimization problems. Furthermore, the
TE
proposed approach is effective, simple to understand, and easy to implement; and these
characteristics are sometimes welcome in real applications, especially industrial applications.
EP
Acknowledgements
The author would like to thank Kuo-Ching Ying and Shih-Wei Lin for their
C
References
[1] Singh, M. R., & Mahapatra, S. S. (2016). A quantum behaved particle swarm
optimization for flexible job shop scheduling. Computers & Industrial Engineering, 93,
36-44.
https://doi.org/10.1016/j.cie.2015.12.004
[2] Singh, M. R., Singh, M., Mahapatra, S. S., & Jagadev, N. (2016). Particle swarm
optimization algorithm embedded with maximum deviation theory for solving multi-
objective flexible job shop scheduling problem. The International Journal of Advanced
Manufacturing Technology, 85(9-12), 2353-2366.
https://doi.org/10.1007/s00170-015-8075-1
23
ACCEPTED MANUSCRIPT
[3] Ying, K. C., & Lin, S. W. (2006). Multiprocessor task scheduling in multistage hybrid
flow-shops: an ant colony system approach. International Journal of Production
Research, 44(16), 3161-3177.
https://doi.org/10.1080/00207540500536939
[4] Behnamian, J., & Zandieh, M. (2011). A discrete colonial competitive algorithm for
hybrid flowshop scheduling to minimize earliness and quadratic tardiness
penalties. Expert Systems with Applications, 38(12), 14490-14498.
https://doi.org/10.1016/j.eswa.2011.04.241
[5] T.S. Arthanary and K.G. Ramamurthy. (1971). An extension of two machines
PT
sequencing problem, OPSEARCH, 8(4),10–22.
[6] Ruiz, R., & Vázquez-Rodríguez, J. A. (2010). The hybrid flow shop scheduling
problem. European journal of operational research, 205(1), 1-18.
RI
https://doi.org/10.1016/j.ejor.2009.09.024
[7] Ribas, I., Leisten, R., & Framiñan, J. M. (2010). Review and classification of hybrid
SC
flow shop scheduling problems from a production system and a solutions procedure
perspective. Computers & Operations Research, 37(8), 1439-1454.
https://doi.org/10.1016/j.cor.2009.11.001
[8] “Lee, T., & Loong, Y. (2019). A review of scheduling problem and resolution methods
U
in flexible flow shop. International Journal of Industrial Engineering
Computations, 10(1), 67-88.
AN
https://doi.org/10.5267/j.ijiec.2018.4.001
[9] Gupta, J. N. (1988). Two-stage, hybrid flowshop scheduling problem. Journal of the
M
https://doi.org/10.1016/j.ejor.2003.12.026
[11] Guinet, A., Solomon, M. M., Kedia, P. K., & Dussauchoy, A. (1996). A computational
study of heuristics for two-stage flexible flowshops. International Journal of Production
EP
separable setup and removal times. European Journal of Operational Research, 77(3),
415-428.
AC
https://doi.org/10.1016/0377-2217(94)90407-3
[13] Lee, C. Y., & Vairaktarakis, G. L. (1994). Minimizing makespan in hybrid
flowshops. Operations Research Letters, 16(3), 149-158.
[14] Riane, F., Artiba, A., & Elmaghraby, S. E. (1998). A hybrid three-stage flowshop
problem: efficient heuristics to minimize makespan. European Journal of Operational
Research, 109(2), 321-329.
https://doi.org/10.1016/S0377-2217(98)00060-5
[15] Hunsucker, J. L., & Shah, J. R. (1994). Comparative performance analysis of priority
rules in a constrained flow shop with multiple processors environment. European
Journal of Operational Research, 72(1), 102-114.
https://doi.org/10.1016/0377-2217(94)90333-6
24
ACCEPTED MANUSCRIPT
[16] Kyparisis, G. J., & Koulamas, C. (2006). Flexible flow shop scheduling with uniform
parallel machines. European journal of operational research, 168(3), 985-997.
https://doi.org/10.1016/j.ejor.2004.05.017
[17] Moccellin, J. V., Nagano, M. S., Neto, A. R. P., & de Athayde Prata, B. (2018).
Heuristic algorithms for scheduling hybrid flow shops with machine blocking and setup
times. Journal of the Brazilian Society of Mechanical Sciences and Engineering, 40(2),
40.
https://doi.org/10.1007/s40430-018-0980-4
[18] Tan, Y., Mönch, L., & Fowler, J. W. (2018). A hybrid scheduling approach for a two-
PT
stage flexible flow shop with batch processing machines. Journal of Scheduling, 21(2),
209-226.
https://doi.org/10.1007/s10951-017-0530-4
RI
[19] Shao, W., Pi, D., & Shao, Z. (2018). A hybrid discrete teaching-learning based meta-
heuristic for solving no-idle flow shop scheduling problem with total tardiness
criterion. Computers & Operations Research, 94, 89-105.
SC
https://doi.org/10.1016/j.cor.2018.02.003
[20] Jin, Z. H., Ohno, K., Ito, T., & Elmaghraby, S. E. (2002). Scheduling hybrid flowshops
in printed circuit board assembly lines. Production and Operations Management, 11(2),
U
216-230.
https://doi.org/10.1111/j.1937-5956.2002.tb00492.x
AN
[21] Besbes, W., Teghem, J., & Loukil, T. (2010). Scheduling hybrid flow shop problem with
non-fixed availability constraints. European Journal of Industrial Engineering, 4(4),
413-433.
M
https://doi.org/10.1504/EJIE.2010.035652
[22] “Costa, A., Cappadonna, F. A., & Fichera, S. (2014). A novel genetic algorithm for the
hybrid flow shop scheduling with parallel batching and eligibility constraints. The
D
[23] Choi, S. H., & Wang, K. (2012). Flexible flow shop scheduling with stochastic
processing times: A decomposition-based approach. Computers & Industrial
Engineering, 63(2), 362-373.
EP
https://doi.org/10.1016/j.cie.2012.04.001
[24] Wardono, B., & Fathi, Y. (2004). A tabu search algorithm for the multi-stage parallel
machine problem with limited buffer capacities. European Journal of Operational
Research, 155(2), 380-401.
C
https://doi.org/10.1016/S0377-2217(02)00873-1
AC
[25] Chen, L., Bostel, N., Dejax, P., Cai, J., & Xi, L. (2007). A tabu search algorithm for the
integrated scheduling problem of container handling systems in a maritime
terminal. European Journal of Operational Research, 181(1), 40-58.
https://doi.org/10.1016/j.ejor.2006.06.033
[26] Zhang, B., Pan, Q. K., Gao, L., Zhang, X. L., Sang, H. Y., & Li, J. Q. (2017). An
effective modified migrating birds optimization for hybrid flowshop scheduling problem
with lot streaming. Applied Soft Computing, 52, 14-27.
https://doi.org/10.1016/j.asoc.2016.12.021
[27] Meng, T., Pan, Q. K., Li, J. Q., & Sang, H. Y. (2018). An improved migrating birds
optimization for an integrated lot-streaming flow shop scheduling problem. Swarm and
Evolutionary Computation, 38, 64-78.
25
ACCEPTED MANUSCRIPT
https://doi.org/10.1016/j.swevo.2017.06.003
[28] Peng, K., Pan, Q. K., Gao, L., Zhang, B., & Pang, X. (2018). An Improved Artificial Bee
Colony Algorithm for Real-World Hybrid Flowshop Rescheduling in Steelmaking-
Refining Continuous Casting Process. Computers & Industrial Engineering, 122, 235-
250.
https://doi.org/10.1016/j.cie.2018.05.056
[29] Riahi, V., & Kazemi, M. (2018). A new hybrid ant colony algorithm for scheduling of
no-wait flowshop. Operational Research, 18(1), 55-74.
https://doi.org/10.1007/s12351-016-0253-x
PT
[30] Tavakkoli-Moghaddam, R., Safaei, N., & Sassani, F. (2009). A memetic algorithm for
the flexible flow line scheduling problem with processor blocking. Computers &
Operations Research, 36(2), 402-414.
RI
https://doi.org/10.1016/j.cor.2007.10.011
[31] Tang, L., Liu, W., & Liu, J. (2005). A neural network model and algorithm for the
SC
hybrid flow shop scheduling problem in a dynamic environment. Journal of Intelligent
Manufacturing, 16(3), 361-370.
https://doi.org/10.1007/s10845-005-7029-0
U
[32] Zandieh, M., Ghomi, S. F., & Husseini, S. M. (2006). An immune algorithm approach to
hybrid flow shops scheduling with sequence-dependent setup times. Applied
AN
Mathematics and Computation, 180(1), 111-127.
https://doi.org/10.1016/j.amc.2005.11.136
M
[33] Li, J., Duan, P., Sang, H., Wang, S., Liu, Z., & Duan, P. (2018). An efficient
optimization algorithm for resource-constrained steelmaking scheduling problems. IEEE
Access.
D
https://doi.org/10.1109/ACCESS.2018.2840512
[34] Li, J. Q., Pan, Q. K., & Mao, K. (2016). A hybrid fruit fly optimization algorithm for the
TE
[36] Marichelvam, M. K., & Geetha, M. (2018). Cuckoo Search Algorithm for Solving Real
AC
26
ACCEPTED MANUSCRIPT
Singapore.
https://doi.org/10.1007/978-981-10-7212-3_17
[39] Gong, D., Han, Y., & Sun, J. (2018). A novel hybrid multi-objective artificial bee colony
algorithm for blocking lot-streaming flow shop scheduling problems. Knowledge-Based
Systems, 148, 115-130.
https://doi.org/10.1016/j.knosys.2018.02.029
[40] Li, J. Q., Sang, H. Y., Han, Y. Y., Wang, C. G., & Gao, K. Z. (2018). Efficient multi-
objective optimization algorithm for hybrid flow shop scheduling problems with setup
PT
energy consumptions. Journal of Cleaner Production, 181, 584-598.
https://doi.org/10.1016/j.jclepro.2018.02.004
[41] Nowicki, E., & Smutnicki, C. (1998). The flow shop with parallel machines: A tabu
RI
search approach. European Journal of Operational Research, 106(2-3), 226-253.
https://doi.org/10.1016/S0377-2217(97)00260-9
[42] Marichelvam, M. K., Prabaharan, T., Yang, X. S., & Geetha, M. (2013). Solving hybrid
SC
flow shop scheduling problems using bat algorithm. International Journal of Logistics
Economics and Globalisation, 5(1), 15-29.
https://doi.org/10.1504/IJLEG.2013.054428
U
[43] Alaykýran, K., Engin, O., & Döyen, A. (2007). Using ant colony optimization to solve
hybrid flow shop scheduling problems. The international journal of advanced
AN
manufacturing technology, 35(5-6), 541-550.
https://doi.org/10.1007/s00170-007-1048-2
[44] Engin, O., & Döyen, A. (2004). A new approach to solve hybrid flow shop scheduling
M
[45] Pan, Q. K., Wang, L., Li, J. Q., & Duan, J. H. (2014). A novel discrete artificial bee
colony algorithm for the hybrid flowshop scheduling problem with makespan
TE
solving the hybrid flow shop scheduling problem. Applied Soft Computing, 24, 63-77.
https://doi.org/10.1016/j.asoc.2014.07.005
[47] Liao, C. J., Tjandradjaja, E., & Chung, T. P. (2012). An approach using particle swarm
C
https://doi.org/10.1016/j.asoc.2012.01.011
[48] Duan, J. H., Meng, T., & Pan, Q. K. (2017, July). An effective fruit fly optimization
algorithm for the hybrid flowshop scheduling problem. In 2017 13th International
Conference on Natural Computation, Fuzzy Systems and Knowledge Discovery (ICNC-
FSKD) (pp. 930-934). IEEE.
https://doi.org/10.1109/FSKD.2017.8393402
[49] Marichelvam, M. K., Prabaharan, T., & Yang, X. S. (2014). Improved cuckoo search
algorithm for hybrid flow shop scheduling problems to minimize makespan. Applied Soft
Computing, 19, 93-101.
https://doi.org/10.1016/j.asoc.2014.02.005
27
ACCEPTED MANUSCRIPT
[50] Hidri, L., & Gharbi, A. (2017). New Efficient Lower Bound for the Hybrid Flow Shop
Scheduling Problem With Multiprocessor Tasks. IEEE Access, 5, 6121-6133.
https://doi.org/10.1109/ACCESS.2017.2696118
[51] Oğuz, C., Ercan, M. F., Cheng, T. E., & Fung, Y. F. (2003). Heuristic algorithms for
multiprocessor task scheduling in a two-stage hybrid flow-shop. European Journal of
Operational Research, 149(2), 390-403.
https://doi.org/10.1016/S0377-2217(02)00766-X
[52] Ying, K. C. (2009). An iterated greedy heuristic for multistage hybrid flowshop
scheduling problems with multiprocessor tasks. Journal of the Operational Research
PT
Society, 60(6), 810-817.
https://doi.org/10.1057/palgrave.jors.2602625
[53] Kahraman, C., Engin, O., Kaya, İ., & Öztürk, R. E. (2010). Multiprocessor task
RI
scheduling in multistage hybrid flow-shops: A parallel greedy algorithm approach.
Applied Soft Computing, 10(4), 1293-1300.
https://doi.org/10.1016/j.asoc.2010.03.008
SC
[54] Lahimer, A., Lopez, P., & Haouari, M. (2013). Improved bounds for hybrid flow shop
scheduling with multiprocessor tasks. Computers & Industrial Engineering, 66(4), 1106-
1114.
U
https://doi.org/10.1016/j.cie.2013.08.004
[55] Ying, K. C., & Lin, S. W. (2009). Scheduling multistage hybrid flowshops with
AN
multiprocessor tasks by an effective heuristic. International Journal of Production
Research, 47(13), 3525-3538.
https://doi.org/10.1080/00207540701871085
M
[57] Wang, H. M., Chou, F. D., & Wu, F. C. (2011). A simulated annealing for hybrid flow
shop scheduling with multiprocessor tasks to minimize makespan. The International
TE
[58] Chou, F. D., Wang, H. M., & Wu, F. C. (2010). A restricted simulated annealing for
hybrid flow-shop scheduling with multiprocessor tasks. In Proceedings of the 2010
International Conference on Mechanical, Industrial, and Manufacturing Technologies
(MIMT 2010). ASME Press.
C
https://doi.org/10.1115/1.859544.paper17
AC
[59] Oğuz, C., Zinder, Y., Janiak, A., & Lichtenstein, M. (2004). Hybrid flow-shop
scheduling problems with multiprocessor task systems. European Journal of
Operational Research, 152(1), 115-131.
https://doi.org/10.1016/S0377-2217(02)00644-6
[60] Sivrikaya-Şerifoglu, F., &Ulusoy, G. (2004). Multiprocessor task scheduling in
multistage hybrid flow-shops: a genetic algorithm approach. Journal of the Operational
Research Society, 55(5), 504-512.
https://doi.org/10.1057/palgrave.jors.2601716
[61] Oğuz, C., & Ercan, M. F. (2005). A genetic algorithm for hybrid flow-shop scheduling
with multiprocessor tasks. Journal of Scheduling, 8(4), 323-351.
https://doi.org/10.1007/s10951-005-1640-y
28
ACCEPTED MANUSCRIPT
[62] Engin, O., Ceran, G., & Yilmaz, M. K. (2011). An efficient genetic algorithm for hybrid
flow shop scheduling with multiprocessor task problems. Applied Soft Computing, 11(3),
3056-3065.
https://doi.org/10.1016/j.asoc.2010.12.006
[63] Oguz, C., Fung, Y. F., Ercan, M. F., & Qi, X. T. (2003, June). Parallel genetic algorithm
for a flow-shop problem with multiprocessor tasks. In International conference on
computational science (pp. 548-559). Springer, Berlin, Heidelberg.
https://doi.org/10.1007/3-540-44863-2_54
[64] Tseng, C. T., & Liao, C. J. (2008). A particle swarm optimization algorithm for hybrid
PT
flow-shop scheduling with multiprocessor tasks. International Journal of Production
Research, 46(17), 4655-4670.
https://doi.org/10.1080/00207540701294627
RI
[65] Chou, F. D. (2013). Particle swarm optimization with cocktail decoding method for
hybrid flow shop scheduling problems with multiprocessor tasks. International Journal
of Production Economics, 141(1), 137-145.
SC
https://doi.org/10.1016/j.ijpe.2012.05.015
[66] Singh, M. R., & Mahapatra, S. S. (2012). A swarm optimization approach for flexible
flow shop scheduling with multiprocessor tasks. The International Journal of Advanced
U
Manufacturing Technology, 62(1-4), 267-277.
https://doi.org/10.1007/s00170-011-3807-3
AN
[67] Singh, M. R., Mahapatra, S. S., & Mishra, K. (2013). A novel swarm optimiser for
flexible flow shop scheduling. International Journal of Swarm Intelligence, 1(1), 51-69.
https://doi.org/10.1504/IJSI.2013.055802
M
[68] Akkoyunlu, M. C., Engın, O., & Büyuközkan, K. (2015, May). A harmony search
algorithm for hybrid flow shop scheduling with multiprocessor task problems. In
Modeling, Simulation, and Applied Optimization (ICMSAO), 2015 6th International
D
[69] Zini, H., & ElBernoussi, S. (2017, June). Minimizing makespan in hybrid flow shop
scheduling with multiprocessor task problems using a discrete harmony search. In
EP
hybrid flow shop problems using discrete firefly algorithm. International Journal of
Advanced Intelligence Paradigms, 8(4), 377-391.
AC
https://doi.org/10.1504/IJAIP.2016.080191
[71] Xu, Y., Wang, L., Liu, M., & Wang, S. Y. (2013). An effective shuffled frog-leaping
algorithm for hybrid flow-shop scheduling with multiprocessor tasks. The International
Journal of Advanced Manufacturing Technology, 68(5-8), 1529-1537.
https://doi.org/10.1007/s00170-013-4940-y
[72] Jouglet, A., Oğuz, C., & Sevaux, M. (2009). Hybrid flow-shop: a memetic algorithm
using constraint-based scheduling for efficient search. Journal of Mathematical
Modelling and Algorithms, 8(3), 271-292.
https://doi.org/10.1007/s10852-008-9101-1
[73] Choong, F., Phon-Amnuaisuk, S., & Alias, M. Y. (2011). Metaheuristic methods in
29
ACCEPTED MANUSCRIPT
hybrid flow shop scheduling problem. Expert Systems with Applications, 38(9), 10787-
10793.
https://doi.org/10.1016/j.eswa.2011.01.173
[74] Ying, K. C. (2012). Minimising makespan for multistage hybrid flowshop scheduling
problems with multiprocessor tasks by a hybrid immune algorithm. European Journal of
Industrial Engineering, 6(2), 199-215.
https://doi.org/10.1504/EJIE.2012.045605
[75] Xu, Y., Wang, L., Wang, S., & Liu, M. (2013). An effective immune algorithm based on
novel dispatching rules for the flexible flow-shop scheduling problem with
PT
multiprocessor tasks. The International Journal of Advanced Manufacturing
Technology, 67(1-4), 121-135.
https://doi.org/10.1007/s00170-013-4759-6
RI
[76] Lin, S. W., Ying, K. C., & Huang, C. Y. (2013). Multiprocessor task scheduling in
multistage hybrid flowshops: A hybrid artificial bee colony algorithm with bi-directional
planning. Computers & Operations Research, 40(5), 1186-1195.
SC
https://doi.org/10.1016/j.cor.2012.12.014
[77] Dorigo, M., & Stützle, T. (2003). The ant colony optimization metaheuristic:
Algorithms, applications, and advances. In Handbook of metaheuristics (pp. 250-285).
U
Springer, Boston, MA.
https://doi.org/10.1007/0-306-48056-5_9
AN
[78] M. Dorigo and T. Stützle. (2004). Ant Colony Optimization. MIT Press, Cambridge,
MA.
[79] Dorigo, M., & Blum, C. (2005). Ant colony optimization theory: A survey. Theoretical
M
https://doi.org/10.1016/S0167-5060(08)70356-X
[81] Dorigo, M., & Gambardella, L. M. (1997). Ant colony system: a cooperative learning
approach to the traveling salesman problem. IEEE Transactions on evolutionary
EP
30
ACCEPTED MANUSCRIPT
Table 1
An example of a 5× 2 HFSPMT.
Stage 1 Stage 2
m1 = 3 m2 = 2
Job pij sizeij Job pij sizeij
1 10 2 1 15 1
2 13 3 2 12 2
3 8 2 3 15 1
PT
4 10 3 4 20 2
5 15 1 5 10 2
RI
U SC
AN
M
D
Table 2
TE
ARD-BS ARD-AvS ACT ARD-BS ARD-AvS ACT ARD-BS ARD-AvS ACT ARD-BS ARD-AvS ACT ARD-BS ARD-AvS ACT
Type-1
5x2 15,01 15,01 6,54 14,44 14,44 7,24 14,44 14,68 7,28 14,44 14,60 7,34 14,44 14,44 7,26
5x5 21,95 21,95 9,76 20,73 21,00 19,42 20,73 21,00 19,90 20,73 21,00 19,68 20,73 21,00 19,06
5x8 22,05 22,05 15,34 18,16 19,59 29,92 18,16 19,59 29,68 18,16 19,59 29,64 18,16 19,59 29,20
10 x 2 1,57 1,58 8,54 1,57 1,57 9,66 1,57 1,57 9,00 1,57 1,57 9,14 1,57 1,57 9,42
10 x 5 10,23 10,83 20,76 9,54 9,92 23,86 9,76 10,32 21,26 9,63 10,12 24,84 9,30 9,34 24,80
C
10 x 8 17,23 17,82 33,58 16,58 16,96 80,96 17,29 17,58 82,22 16,44 16,94 80,04 15,91 15,93 81,02
20 x 2 2,06 2,35 11,80 1,71 1,80 27,06 1,83 1,95 27,24 1,74 1,80 27,46 1,67 1,74 27,70
20 x 5 5,47 6,44 30,34 3,36 3,79 60,50 4,04 4,42 56,46 3,52 3,88 61,60 2,63 3,09 61,58
20 x 8 12,12 13,18 67,50 6,58 7,86 255,24 6,87 8,18 333,90 6,86 7,98 349,92 6,30 7,38 273,64
AC
50 x 2 3,84 4,18 57,28 1,93 2,44 110,88 2,11 2,59 114,06 2,08 2,47 111,20 1,93 2,34 112,92
50 x 5 3,13 3,54 157,04 1,44 1,66 265,88 1,55 1,75 271,13 1,49 1,69 260,30 1,31 1,58 259,14
50 x 8 6,50 7,43 205,02 1,87 2,72 721,98 2,01 2,94 715,92 1,88 2,74 696,98 1,71 2,55 694,04
100 x 2 1,51 1,68 122,52 0,45 0,53 375,54 0,47 0,57 382,64 0,45 0,56 380,48 0,45 0,54 377,32
100 x 5 2,55 2,94 299,78 0,65 0,95 993,74 0,68 0,99 1058,96 0,65 0,94 996,44 0,65 0,95 996,78
100 x 8 3,60 4,15 799,76 1,02 1,34 2769,54 1,03 1,35 2729,98 0,98 1,31 2748,60 0,95 1,25 2658,14
Type-2
5x2 34,54 34,54 5,64 33,71 33,71 10,98 33,71 33,71 10,78 33,71 33,71 11,12 33,71 33,71 10,82
5x5 43,81 43,81 9,60 41,92 42,59 30,10 41,92 42,62 28,94 41,92 42,59 29,72 41,92 42,59 29,06
5x8 29,63 29,63 15,86 27,78 28,15 44,02 27,78 28,15 47,88 27,78 28,15 44,80 27,78 28,15 44,54
10 x 2 12,89 13,09 10,46 13,58 13,58 12,42 14,15 14,22 11,50 13,46 13,46 12,14 12,26 12,26 11,94
10 x 5 26,84 27,50 26,68 27,92 27,99 72,68 28,59 29,04 73,40 28,15 28,21 72,32 25,87 26,01 71,94
10 x 8 26,70 27,78 39,06 24,25 24,83 94,36 26,25 26,62 93,16 25,02 25,16 91,10 23,69 23,71 75,54
20 x 2 11,19 11,58 15,42 10,61 10,78 31,70 10,89 11,20 31,34 10,81 10,97 33,36 10,46 10,54 33,88
20 x 5 19,48 20,94 32,04 13,94 15,90 150,04 15,90 17,47 166,26 15,06 16,34 161,34 14,25 15,81 158,50
20 x 8 31,31 32,77 74,04 23,94 25,77 233,46 24,47 26,59 246,16 23,84 25,70 231,38 23,58 25,72 243,84
50 x 2 9,32 9,80 70,40 7,40 7,95 111,82 7,53 8,16 111,91 7,35 7,88 109,64 7,11 7,70 110,44
50 x 5 17,41 18,59 171,38 10,91 12,50 264,34 11,40 12,98 268,30 10,96 12,56 272,36 10,91 12,51 263,66
50 x 8 29,17 30,42 256,92 18,99 21,44 453,98 19,26 21,82 458,12 18,95 21,28 451,48 18,96 21,34 450,62
100 x 2 6,94 7,48 140,80 3,25 3,89 394,92 3,29 3,97 406,62 3,33 3,93 399,60 3,21 3,87 398,48
100 x 5 18,27 19,00 327,08 11,16 12,16 987,96 11,56 12,43 988,92 11,43 12,32 923,48 11,33 12,21 921,00
100 x 8 23,53 24,59 845,74 15,14 16,34 3219,24 15,28 16,48 3213,36 15,04 16,27 3177,98 14,80 16,09 3230,44
1
ACCEPTED MANUSCRIPT
Table 3
Results of ACSNDPs with different Non-DaemonActions procedures on the second
benchmark problem set.
n × k ACSNDP-SN ACSNDP-SJ ACSNDP-SP ACSNDP-SJP ACSNDP
ARD-BS ARD-AvS ACT ARD-BS ARD-AvS ACT ARD-BS ARD-AvS ACT ARD-BS ARD-AvS ACT ARD-BS ARD-AvS ACT
Type-A
5x2 9,26 9,26 3,67 9,26 9,26 3,76 9,26 9,26 3,72 9,26 9,26 3,78 9,26 9,26 3,68
5x5 25,44 25,44 6,33 23,63 23,63 8,06 23,63 23,63 8,11 23,59 23,59 8,12 23,59 23,94 8,10
5x8 26,30 26,30 9,55 24,55 25,33 10,50 24,55 25,33 10,50 24,55 25,33 11,00 24,55 25,33 10,00
5 x 10 26,13 26,13 12,21 23,80 24,45 13,00 23,80 24,45 11,00 23,80 24,45 12,50 23,80 24,45 12,50
10 x 2 2,88 2,89 6,88 2,90 2,96 8,40 2,90 3,00 8,47 2,88 2,88 8,38 2,88 2,88 12,14
10 x 5 11,37 12,22 17,90 10,74 11,10 20,24 11,74 12,10 19,80 10,90 11,16 20,51 9,90 10,41 20,28
10 x 8 19,73 21,54 27,96 19,96 20,47 33,58 20,38 21,50 33,00 19,66 20,17 33,56 18,53 18,73 32,06
10 x 10 15,18 15,59 33,34 13,34 13,74 40,86 14,49 15,31 40,32 13,24 14,23 40,76 13,33 13,75 40,18
PT
20 x 2 6,31 6,63 16,30 5,74 6,10 34,26 6,17 6,95 33,76 5,39 5,91 33,04 5,36 5,43 33,74
20 x 5 9,51 10,60 32,34 5,79 6,14 75,32 7,59 8,48 75,34 5,81 6,62 74,20 5,81 6,38 76,80
20 x 8 10,36 11,42 76,96 4,98 5,71 128,50 7,37 8,29 130,80 5,63 6,27 130,34 5,15 6,10 129,66
20 x 10 14,24 15,89 85,06 8,03 9,21 154,20 10,37 11,29 156,96 9,13 9,97 156,41 8,13 10,26 155,59
50 x 2 4,08 4,53 47,96 2,50 3,09 94,34 2,40 3,31 95,74 2,16 2,56 93,48 2,16 2,61 95,70
50 x 5 4,15 4,66 121,62 1,71 1,92 220,98 1,94 2,24 218,00 1,45 1,77 220,36 1,45 1,97 222,44
50 x 8 8,70 9,40 221,68 3,62 3,89 370,76 3,78 5,17 376,84 3,36 3,96 379,12 3,00 3,83 376,52
50 x 10 9,78 10,62 340,86 2,95 3,44 473,92 4,56 4,78 474,32 3,17 3,90 458,64 2,91 3,73 485,26
1,60 1,60
RI
100 x 2 2,33 2,56 170,88 2,15 2,58 325,88 2,56 2,78 334,00 1,17 324,56 1,53 327,86
100 x 5 4,83 5,41 383,06 3,42 3,60 784,82 3,31 3,61 781,56 2,31 2,59 760,06 2,21 2,60 758,10
100 x 8 4,66 5,31 594,40 2,42 2,60 1256,36 2,31 2,61 1317,66 1,31 1,59 1311,44 1,21 1,47 1299,02
100 x 10 5,21 5,72 977,32 0,99 1,75 1568,38 1,44 1,79 1638,54 1,12 1,68 1633,86 1,01 1,64 1578,60
Type-B
5x2 24,62 24,62 3,00 24,35 24,35 3,20 24,35 24,35 3,20 24,35 24,35 3,16 24,35 24,35 3,26
5x5 40,26 40,26 5,76 36,96 37,32 7,00 37,96 38,32 7,20 36,96 37,32 7,92 36,96 37,32 7,38
SC
5x8 32,76 32,76 10,43 28,07 28,10 12,00 28,20 28,99 11,00 28,07 28,10 12,67 28,07 28,10 11,00
5 x 10 36,86 36,86 13,33 32,41 32,99 17,00 32,41 32,99 17,00 32,41 32,99 15,50 32,41 32,99 14,50
10 x 2 12,73 13,25 10,78 12,27 12,38 39,26 12,75 12,98 10,30 12,25 12,78 10,52 12,25 12,28 10,88
10 x 5 21,29 22,33 21,66 20,15 21,61 25,28 23,47 23,58 24,64 20,09 20,44 24,82 18,92 19,25 25,34
10 x 8 36,45 37,25 35,36 32,98 34,04 42,62 35,68 35,79 40,60 33,90 35,57 40,26 33,37 33,69 41,12
10 x 10 32,75 34,29 47,72 30,05 30,88 49,92 31,05 32,24 48,96 30,21 30,94 49,38 29,67 29,95 50,08
20 x 2 7,30 8,30 19,20 6,32 7,24 22,54 7,49 8,59 22,12 6,30 7,18 22,04 5,77 6,14 22,10
20 x 5 17,77 20,31 52,84 11,18 13,70 49,98 12,75 14,46 51,06 12,38 13,05 50,20 11,59 13,57 50,50
20 x 8 31,42 32,88 83,68 23,60 24,24 112,20 27,55 29,99 113,18 23,81 25,48 111,74 23,75 25,81 111,06
U
20 x 10 37,20 39,26 108,78 29,31 30,52 181,30 31,37 34,74 181,42 28,98 31,62 183,00 29,41 32,73 182,32
50 x 2 8,74 9,46 56,16 6,48 7,33 84,12 6,36 7,63 80,72 5,27 5,89 83,08 5,45 5,95 82,66
50 x 5 19,25 20,47 145,00 13,14 14,18 209,48 15,07 16,04 214,72 12,05 13,16 210,20 11,59 13,33 210,92
50 x 8 28,25 29,82 263,52 22,16 23,01 354,42 22,67 23,53 368,18 20,18 21,24 358,24 19,37 21,58 354,20
50 x 10 33,81 35,32 414,78 24,28 25,31 610,12 25,29 26,08 611,20 23,66 24,77 615,20 23,61 24,76 615,04
AN
100 x 2 8,12 8,57 204,32 6,57 7,28 341,02 8,13 9,15 351,26 5,29 6,09 340,70 5,55 6,00 343,28
100 x 5 16,33 16,90 408,60 10,20 11,78 792,70 11,67 12,43 823,34 9,63 11,95 791,36 9,26 10,41 799,62
100 x 8 22,91 23,58 713,78 14,92 16,71 1246,84 17,36 17,43 1255,00 14,81 16,94 1257,12 14,04 15,43 1257,72
M
D
TE
Table 4
Relative improvements by ACSNDP to ACSNDPs with alternatives Non-DaemonActions
procedures on both of the benchmark problem sets.
EP
2
ACCEPTED MANUSCRIPT
Table 5
Computational results of AV-BS obtained by the compared works for the first benchmark
problem set.
n × k GAOE GAJOS ACSYL MAJOS SAWCW HIAYing HABCLYH ACSNDP
Type-1
5x2 267,6 267,6 267,0 267,0 267,3 267,0 267,0 267,0
5x5 472,1 472,1 469,8 466,6 469,2 469,2 468,7 468,1
5x8 641,6 641,6 628,9 616,7 634,7 628,7 636,8 621,4
PT
10 x 2 451,1 451,1 451,1 451,1 451,1 451,1 451,1 451,1
10 x 5 639,1 648,4 645,1 637,8 643,0 641,4 641,3 644,1
10 x 8 836,9 850,5 852,4 840,3 847,8 843,0 851,3 848,7
20 x 2 876,7 876,5 877,1 877,7 875,3 875,9 875,9 875,9
RI
20 x 5 1072,1 1077,7 1074,7 1070,3 1064,9 1066,5 1066,6 1070,3
20 x 8 1319,3 1319,9 1315,6 1315,2 1303,2 1305,3 1302,4 1309,8
50 x 2 2049,4 2048,5 2052,1 2046,1 2038,5 2042,3 2038,6 2045,0
SC
50 x 5 2604,0 2574,8 2577,2 2571,7 2566,6 2564,6 2564,2 2573,0
50 x 8 2669,6 2634,3 2644,8 2623,1 2609,3 2609,7 2599,1 2623,8
100 x 2 4355,0 4351,5 4363,8 4348,6 4349,2 4349,0 4346,8 4348,2
100 x 5 4755,0 4755,9 4697,1 4746,9 4670,7 4696,7 4664,0 4682,6
U
100 x 8 5327,7 5260,2 5231,3 5267,2 5206,4 5229,1 5191,4 5213,3
Type-2
AN
5x2 256,4 256,4 254,1 253,5 253,5 253,5 253,5 253,5
5x5 423,8 423,8 421,4 418,4 421,1 419,1 422,6 418,4
5x8 614,2 614,2 605,8 599,9 607,4 603,0 607,9 602,1
10 x 2 426,3 423,7 422,1 422,8 422,2 422,0 422,9
M
409,5
10 x 5 616,3 606,8 600,3 594,9 598,8 595,1 599,9 599,5
10 x 8 813,1 840,4 842,2 824,0 829,9 828,3 827,1 832,7
20 x 2 757,4 809,5 808,3 807,6 806,0 806,3 806,1 806,7
D
1671,1
50 x 5 2074,6 1971,7 1978,1 1960,7 1912,5 1884,9 1850,5 1940,6
50 x 8 2321,8 2292,2 2268,5 2315,7 2194,4 2175,5 2128,0 2210,6
100 x 2 3205,3 3242,5 3268,3 3215,0 3191,6 3215,0 3166,6 3194,5
100 x 5 4145,3 3822,3 3635,5 3802,0 3530,0 3616,1 3415,7 3599,8
EP
3
ACCEPTED MANUSCRIPT
Table 6
Computational results of AV-BS obtained by the compared works for the second benchmark
problem set.
n × k GASU GAJOS ACSYL MAJOS SAWCW HIAYing HABCLYH ACSNDP
Type-A
5x2 275,3 275,3 275,3 275,3 275,3 275,3 275,3 275,3
5x5 452,9 452,9 449,8 466,0 449,9 447,5 450,8 446,0
5x8 632,4 631,0 629,5 621,4 629,7 629,1 631,1 624,2
PT
5 x 10 809,9 810,0 803,7 794,3 807,9 803,2 812,5 800,3
10 x 2 506,9 506,9 506,9 506,9 506,9 506,9 506,9 506,9
10 x 5 668,2 669,2 667,6 661,6 665,8 664,5 666,0 665,5
10 x 8 854,2 860,1 850,7 840,5 850,5 845,5 849,0 850,9
RI
10 x 10 999,9 1010,8 998,6 988,1 994,0 991,4 995,1 996,8
20 x 2 820,2 821,5 821,6 820,6 819,7 819,7 819,7 819,7
20 x 5 1109,6 1110,4 1109,9 1109,2 1106,6 1107,7 1105,9 1108,0
SC
20 x 8 1337,3 1345,0 1337,1 1336,8 1319,6 1321,9 1321,9 1326,9
20 x 10 1459,3 1455,9 1451,5 1462,7 1438,1 1440,8 1441,7 1450,3
50 x 2 2024,3 2007,1 2009,7 2004,2 1999,8 2004,0 2000,0 2005,4
50 x 5 2486,0 2472,6 2481,5 2468,3 2463,2 2456,8 2452,2 2463,1
U
50 x 8 2777,5 2741,2 2751,4 2742,6 2717,8 2709,1 2696,1 2718,3
50 x 10 2922,0 2899,4 2893,6 2893,3 2858,9 2855,8 2846,7 2874,2
AN
100 x 2 4183,4 4127,4 4138,8 4135,4 4121,6 4126,8 4119,6 4123,9
100 x 5 4665,2 4636,7 4590,2 4622,2 4552,0 4574,6 4528,8 4567,9
100 x 8 5142,3 5155,7 5098,4 5157,7 5063,3 5092,0 5041,5 5067,3
100 x 10 5358,8 5329,1 5266,6 5323,7 5238,8 5265,8 5229,8 5251,6
M
Type-B
5x2 252,2 252,2 251,7 251,7 251,7 251,7 251,7 251,7
5x5 438,6 438,6 432,9 429,2 432,9 431,7 432,8 429,2
D
4
ACCEPTED MANUSCRIPT
Table 7
Relative improvements by ACSNDP in AV-BS to the other works for the available instances.
PT
ACSYL 700 1750,44 1736,05 0,82
RI
SAWCW 700 1728,17 1736,05 -0,46
U SC
AN
M
D
Table 8
TE
Computational results of ARD-BS obtained by the compared works for the first benchmark
problem set.
n × k Type-1 Type-2
EP
TSOZHJL ACSYL PSOCFA IGYing HIAYing ACSNDP TSOZHJL ACSYL PSOCFA IGYing HIAYing ACSNDP
5x2 11,42 14,64 - 10,88 14,79 14,44 20,22 34,25 - 18,35 34,91 33,71
5x5 38,98 21,29 - 18,62 21,68 20,73 52,45 43,27 - 38,49 42,51 41,92
5x8 38,49 19,58 - 17,53 21,74 18,16 94,56 28,54 - 26,87 29,34 27,78
C
10 x 2 3,00 1,60 1,70 1,49 1,60 1,57 10,82 12,62 10,11 8,85 12,29 12,26
10 x 5 29,42 9,51 9,78 8,73 9,25 9,30 45,14 26,09 11,32 23,49 25,42 25,87
10 x 8 46,53 16,50 16,14 13,91 15,39 15,91 77,21 25,14 25,83 22,17 23,46 23,69
AC
20 x 2 2,88 1,92 1,12 1,87 1,81 1,67 7,25 10,73 9,59 6,93 10,58 10,46
20 x 5 24,40 3,07 3,19 2,97 2,59 2,63 35,13 15,11 10,77 12,64 13,54 14,25
20 x 8 42,47 6,77 6,69 5,83 6,39 6,30 62,99 25,18 24,34 22,79 22,51 23,58
50 x 2 2,23 2,37 2,40 2,21 1,87 1,93 5,80 8,17 7,02 5,66 7,41 7,11
50 x 5 10,51 1,51 2,06 1,49 1,17 1,31 28,64 13,11 13,24 11,29 9,51 10,91
50 x 8 21,04 2,59 3,00 2,47 2,09 1,71 54,25 22,23 23,07 20,71 19,11 18,96
100 x 2 0.94 0,91 0,82 0,89 0,48 0,45 5,19 5,66 3,21 5,04 3,90 3,21
100 x 5 7.07 1,05 1,05 1,03 1,03 0,65 26,49 12,45 12,45 10,53 11,87 11,33
100 x 8 14.58 1,33 2,11 1,23 1,42 0,95 36,05 13,90 14,43 12,85 16,01 14,80
5
ACCEPTED MANUSCRIPT
Table 9
Computational results of ARD-BS obtained by the compared works for the second
benchmark problem set.
n × k Type-A Type-B
GASU ACSYL PSOTL HIAYing IGYing ACSNDP GASU ACSYL PSOTL HIAYing IGYing ACSNDP
5x2 9,26 9,26 9,26 9,35 7,97 9,26 24,89 24,58 24,89 24,58 19,35 24,35
5x5 25,71 24,62 25,71 24,16 22,45 23,59 40,20 38,42 40,20 38,78 35,83 36,96
5x8 26,30 25,72 26,30 26,10 23,71 24,55 32,42 29,62 32,42 30,13 27,65 28,07
5 x 10 24,14 24,29 24,14 24,93 21,63 23,80 34,56 33,41 34,56 35,03 31,78 32,41
10 x 2 2,88 2,88 2,88 2,91 2,79 2,88 12,75 13,19 12,75 12,50 11,49 12,25
PT
10 x 5 10,45 10,26 10,45 10,12 9,38 9,90 19,58 18,93 19,58 17,96 17,69 18,92
10 x 8 19,01 18,48 19,01 18,27 16,39 18,53 33,92 33,57 33,92 32,53 31,37 33,37
20 x 10 13,68 13,46 13,68 13,34 11,47 13,33 30,75 30,34 30,75 30,42 28,45 29,67
20 x 2 5,46 5,59 5,40 5,40 5,44 5,36 6,17 5,99 6,05 5,66 5,77 5,77
RI
20 x 5 6,17 6,17 6,04 6,09 6,08 5,81 12,62 12,87 12,33 10,79 12,09 11,59
20 x 8 5,89 5,92 5,76 5,41 5,43 5,15 25,20 25,17 24,98 23,73 23,19 23,75
20 x 10 8,79 8,20 8,23 8,71 8,06 8,13 30,20 30,97 30,11 28,72 28,74 29,41
50 x 2 3,40 2,45 2,30 2,19 2,15 2,16 7,36 5,89 5,69 5,18 5,39 5,45
SC
50 x 5 2,56 2,32 1,44 1,55 2,30 1,45 15,84 14,70 12,47 10,96 13,27 11,59
50 x 8 5,44 4,42 2,91 3,65 4,16 3,00 23,08 19,85 19,41 17,96 18,44 19,37
50 x 10 4,43 3,61 2,53 3,33 3,27 2,91 26,58 25,70 23,31 22,17 22,16 23,61
100 x 2 3,31 1,98 1,62 1,62 1,59 1,53 9,34 6,98 6,56 6,03 6,74 5,55
100 x 5 4,81 2,83 2,80 2,45 2,77 2,21 16,24 11,56 11,49 10,75 10,93 9,26
U
100 x 8 2,95 1,96 1,53 1,79 1,84 1,21 20,89 16,02 15,93 15,26 14,96 14,04
100 x 10 3,07 1,31 0,98 1,27 1,23 1,01 28,01 21,25 20,93 21,08 20,07 19,98
AN
M
D
TE
Table 10
Relative improvements by ACSNDP in ARD-BS to the other works on both of benchmark
problem sets.
EP
6
ACCEPTED MANUSCRIPT
PT
RI
SC
m1 m2 mk
U
AN
Fig. 1. A hybrid flow shop scheduling environment.
M
D
Processor
TE
Stage
C EP
AC
1
ACCEPTED MANUSCRIPT
β
τ iu (t )ηiu (t )
PT
RI
U SC
Fig. 3. The effects of the proposed Non-DaemonActions procedure.
AN
M
D
TE
5000
4900 5 10
EP
4800 20 40
4700
Makespan
4600
4500
AC
4400
4300
4200
4100
0 200 400 600 800 1000
Number of iterations
2
ACCEPTED MANUSCRIPT
5000
0.7 0.5
4900
4700
Makespan
4600
4500
PT
4400
4300
RI
4200
4100
SC
0 200 400 600 800 1000
Number of iterations
U
AN
5000
M
0.7 0.5
4900
4800
0.3 0.1
D
4700
Makespan
TE
4600
4500
4400
EP
4300
4200
C
4100
0 200 400 600 800 1000
AC
Number of iterations
3
ACCEPTED MANUSCRIPT
5000
1 3
4900
4800 5 7
4700
Makespan
4600
4500
PT
4400
4300
RI
4200
4100
SC
0 200 400 600 800 1000
Number of iteration
U
Fig. 7. ACSNDP with various values of ߚ.
AN
5000
M
0.3 0.5
4900
4700
Makespan
TE
4600
4500
4400
EP
4300
4200
C
4100
0 200 400 600 800 1000
AC
Number of iterations
4
ACCEPTED MANUSCRIPT
5000
0 3
4900
6 9
4800
12
4700
Makespan
4600
4500
PT
4400
4300
RI
4200
4100
SC
0 200 400 600 800 1000
Number of iterations
U
Fig. 9. ACSNDP with different neighbourhood sizes.
AN
5000 ACSNDP-SN ACSNDP-SP
M
4700
Makespan
TE
4600
4500
EP
4400
4300
4200
C
4100
AC
Fig. 10. Convergence curves for the five different versions of the ACSNDP algorithm.
5
ACCEPTED MANUSCRIPT
40
35
30
95% CI for the mean
25
20
PT
15
10
RI
5
0 1 OZHJL
TS 2 YL
ACS 3 Ying
IG 4 Ying
HIA 5
ACSNDP 6
SC
p-value=0.000019
Fig. 11. Means and 95% confidence intervals on the first benchmark problem set.
U
AN
M
40
D
35
30
95% CI for the mean
TE
25
20
EP
15
10
C
5
AC
0 GA1SU 2 YL
ACS IG3Ying 4 Ying
HIA 5 TL
PSO 6
ACSNDP 7
p-value=0.951754
Fig. 12. Means and 95% confidence intervals on the second benchmark problem set.
6
ACCEPTED MANUSCRIPT
PT
• Computational results validate the effectiveness of the proposed algorithm.
RI
U SC
AN
M
D
TE
C EP
AC