Download as pdf or txt
Download as pdf or txt
You are on page 1of 10

Computers & Industrial Engineering 145 (2020) 106505

Contents lists available at ScienceDirect

Computers & Industrial Engineering


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

Meta-heuristics for unrelated parallel machines scheduling with random T


rework to minimize expected total weighted tardiness
Xiaoming Wang, Zhantao Li , Qingxin Chen, Ning Mao

State Key Laboratory of Precision Electronic Manufacturing Technology and Equipment, Guangdong University of Technology, Guangzhou 510006, PR China
Provincial Key Laboratory of Computer Integrated Manufacturing, Guangdong University of Technology, Guangzhou 510006, PR China

ARTICLE INFO ABSTRACT

Keywords: Unrelated parallel machines scheduling problems with random rework have many industrial applications. Since
Unrelated parallel machines scheduling the problem has been proven to be NP hard in a strong sense, we concentrate on the implementation of efficient
Random rework approximate methods based on deterministic optimization techniques. Two mixed integer programming models
Mixed integer programming are formulated based on aggregate and separate task estimation, respectively. In order to obtain an approximate
Total weighted tardiness
solution of a large-scale instance, we further propose modified genetic algorithm and simulated annealing al-
Meta-heuristics
gorithm. The initial solutions of the algorithms are obtained by effective priority rules. Computational experi-
ments based on randomly generated instances demonstrate that the proposed aggregate task estimation is more
MSC 2019:
00-01 efficient and more stable than the existing separate task estimation. The proposed meta-heuristics are superior to
99-00 classical priority rules and close to the exact method. Specifically, simulated annealing algorithm is preferred
due to higher computational efficiency than that of genetic algorithm.

1. Introduction Flapper, 2006), etc. Since frequent rework can easily cause the original
scheduling to fail, the delivery time of the task is often difficult to
Parallel machines scheduling problem (PMSP) is one of the most control. Therefore, the scheduling problem with random rework is very
popular studies in the field of scheduling due to it being common in important for production management. However, at present, relevant
engineering practices, such as manufacturing, computers, etc. In the research is relatively few. The complexity and randomness of the cor-
PMSP, a set of independent tasks has to be processed on a set of ma- responding scheduling problem prevent us from applying existing de-
chines. Each task should be processed only on one machine and each terministic solution methods.
machine can process only one task at the same time. Most of the ex- In this study, we focus on unrelated PMSP with random rework to
isting literature on unrelated PMSP assume that the problem environ- minimize expected total weighted tardiness (TWT) because it is closer
ment is deterministic (Cheng & Sin, 1990). to the reality of manufacturing engineering. The rest of this paper is
However, in reality unrelated PMSP may face a lot of uncertainties organized as follows. The next section provides a relevant literature
(Durasevic & Jakobovic, 2018). For instance, random rework is review. In Section 3, we will give a description and a formal statement
common in engineering practices due to high quality requirements of of the problem under investigation. In Section 4, we present a metho-
products or tasks. Specifically, a great number of tasks in many high- dology framework and two task estimation methods. In Section 5, two
tech industries need to be inspected before completion. In the custo- MIP models are formulated based on different task estimation methods.
mized manufacturing industry (such as mold and die), some key design In Section 6, approximate methods that based on priority rules (PRs)
tasks in new product development should be confirmed by customer or and meta-heuristics are proposed. In Section 7, we present a compu-
reviewed by managers (Browning & Yassine, 2016; Luh, Liu, & Moser, tational experiment to verify the reliability of the proposed model and
1999; Wang, Leus, Creemers, Chen, & Mao, 2018; Yassine, Mostafa, & procedures. Finally, Section 8 draws overall conclusions, and puts for-
Browning, 2017). Furthermore, task failures are also common in the ward suggestions for future research.
electronics manufacturing industries (Raghavan, Yoon, & Srihari, 2015,
2018; Rambod & Rezaeian, 2014; Sha, Hsu, Che, & Chen, 2006; Shin & 2. Literature review
Kang, 2010), the transfer-line production systems (Liberopoulos,
Kozanidis, & Tsarouhas, 2007), the pharmaceutical industry (Teunter & The single machine scheduling problem to minimize TWT has been


Corresponding author at: No. 100 Waihuanxi Road, University Town, Guangdong University of Technology, Guangzhou 510006, PR China.
E-mail address: lzt0@163.com (Z. Li).

https://doi.org/10.1016/j.cie.2020.106505
Received 23 July 2019; Received in revised form 21 April 2020; Accepted 23 April 2020
Available online 04 May 2020
0360-8352/ © 2020 Elsevier Ltd. All rights reserved.
X. Wang, et al. Computers & Industrial Engineering 145 (2020) 106505

proven to be NP-hard in the strong sense (Lawler, 1977). Obviously, also based on classical PRs (Browning & Yassine, 2016) or GA (Yassine
unrelated PMSP with random rework to minimize expected TWT is et al., 2017).
strongly NP-hard as well. Due to the complexity of the problem, ap- Although an optimal solution of unrelated PMSP with random re-
proximate methods are more favored than exact algorithms in en- work can be solved theoretically using stochastic dynamic program-
gineering practice. ming, it can easily become intractable due to the curse of dimension-
A variety of approximate methods for unrelated PMSP has been ality. Since the research on this topic is still very rare, there is a lack of
proposed during the past decades, such as heuristics, meta-heuristics, exploration and comparison of effective methods. In this study, two MIP
matheuristics, decomposition, etc. Heuristics are problem-dependent formulations for the unrelated PMSP with random rework are proposed,
techniques, which are often too greedy and get trapped in a local op- two task estimation approaches and several approximate scheduling
timum. Heuristics have been widely applied to unrelated PMSP to methods are compared.
minimize makespan Villa, Vallada, and Fanjul-Peyro (2018) and mini-
mize total tardiness Perez-Gonzalez, Fernandez-Viagas, García, and 3. Problem description
Framinan (2019). Meta-heuristics are problem-independent techniques,
which allowed to explore the solution space more thoroughly and thus The considered nonpreemptive unrelated PMSP with random re-
to hopefully get a better solution. One of the most popular meta- work consists of a set of original tasks = {1, 2, …, N } and an unrelated
heuristics for unrelated PMSP is genetic algorithm (GA) and its variants, machine set R = {1, 2, …, M } . Task j has a known due date Dj ,
which are usually referred to as biologically-inspired meta-heuristics tardiness weight wj and maximum number of rework Lj . A finished task
(Joo & Kim, 2015; Rambod & Rezaeian, 2014). Another popular meta- will be considered qualified and leave the system directly if the max-
heuristics in this area is simulated annealing (SA) algorithm, in which imum number of rework has been reached, otherwise inspection is re-
non-improving moves are probabilistically accepted in an attempt to quired. The inspection result is either qualified or need to be reworked.
avoid local optimum (Bektur & Saraç, 2019). Generally, SA is more The task will return to the waiting queue if it needs to be reworked. The
efficient but GA can get better results over a long run time especially in above working process is illustrated in Fig. 1.
a large-scale instance. The other meta-heuristics for unrelated PMSP Generally, the duration and rework probability of a rework task is
include tabu search (Bektur & Saraç, 2019; Sels, Coelho, Dias, & different from that of the original task. We extend the set to set ,
Vanhoucke, 2015), bees algorithm (Rambod & Rezaeian, 2014), iter- which is composed of both the original and rework tasks, as illustrated
ated greedy algorithm (Fanjul-Peyro & Ruiz, 2010; Rodriguez, Lozano, in Eq. (1).
Blum, & García-Martínez, 2013), etc. Matheuristics is made by the in-
teroperation of meta-heuristics and mathematical programming tech-
niques, which may outperform the MIP models in large-scale instances
Reworks of task 1
(Fanjul-Peyro, Perea, & Ruiz, 2017). Besides, decomposition method is
= 1, …, N , N + 1, , N + L1 , … ,
another effective way to deal with large-scale unrelated PMSP. For in-
stance, the original unrelated PMSP is decomposed into multiple in-
dependent single-machine scheduling problems, so that the existing Reworks of task N

efficient algorithms for single-machine scheduling problems can be N+


N 1
Lj + 1, ,N +
N
Lj
applied (Bülbül & Şen, 2017; Sen & Bülbül, 2015). j=1 j =1

Although a number of models and algorithms for the deterministic (1)


unrelated PMSP have been proposed, little attention has been paid to
The duration and rework probability that task j is processed on
unrelated PMSP with random rework. To the best of our knowledge,
machine k R is denoted by djk and pjk , respectively. The inspection
there is only one research on this topic (Rambod & Rezaeian, 2014).
duration of task j is denoted by hj . We suppose that the above
The authors transform the original stochastic problem to a deterministic
parameters can be obtained from the statistics of historical data. Thus,
problem using separate task estimation and proposed a GA and two
rework is the only uncertain factor in the above problem. The weighted
bees algorithms for minimizing the makespan, results shows that the
tardiness of task j is wj (cj Dj )+ , where cj is its completion time
bees algorithm is outperform GA (Rambod & Rezaeian, 2014). In ad- and (cj Dj ) = max {cj Dj , 0} . The problem is to get an assignment of
+
dition, there is several research on single-machine scheduling problem tasks to the machines to minimize expected TWT.
and identical PMSP with random rework. For instance, a shortest total
estimated processing time (STEPT) algorithm (without machine idle)
and a modified STEPT (MSTEPT) algorithm (with machine idle) for 4. Methodology and task estimation
single-machine TWT minimization problem were introduced by
Raghavan et al. (2015). Subsequently, the authors proposed a modified 4.1. Approximate decision-making procedure
genetic algorithm (GA) which shows to be outperform the MSTEPT
algorithm (Raghavan, Yoon, & Srihari, 2018). The shortest processing In any given state x, we know the waiting task set Wx , processing
time insert heuristic for the single-machine rework rescheduling pro- task set Px and inspecting task set Ix . Theoretically, an optimal solution
blem was proposed by Guo, Huang, Wang, and Leon (2016). An in- can be obtained by considering all possible scenarios of the remaining
tegrated rework-dispatching rule for the scheduling of defective wafers tasks (Luh et al., 1999; Wang, Chen, Mao, Chen, & Li, 2015; Wang et al.,
was developed by Sha et al. (2006). As for the identical PMSP with 2018). However, it can easily become intractable due to the curse of
random rework, two heuristics were proposed by (Kang & Shin, 2010), dimensionality. In order to avoid a complete enumeration of scenarios,
one of which called the minimum rework probabilities with due-dates we first transform the original stochastic problem to multiple de-
was also employed by Shin and Kang (2010). terministic sub-problems through approximate task estimation. On this
Some research on project scheduling is also considering random basis, the transformed deterministic unrelated PMSP in each state can
rework. For the resource-constrained project scheduling problem with
random rework, an optimal solution to minimize expected makespan
can be obtained using stochastic dynamic programming (Wang et al.,
2018). For the resource-constrained multi-project scheduling problem
with finite random rework, an approximate method based on La-
grangian relaxation and stochastic dynamic programming was pro-
posed by Luh et al. (1999). The approximate methods in this area are Fig. 1. The working process of design tasks.

2
X. Wang, et al. Computers & Industrial Engineering 145 (2020) 106505

Fig. 2. The remaining process of each current task.

be solved relatively easily using the existing deterministic optimization denote the last rework and the probability of occurrence of task j Ux ,
approaches. respectively. It is obvious that p j = 1 for j Wx and p j = pir for
Our main contribution lies in the task estimation methods and sub- i Px Ix if i is processed on machine r. Let Sx Ux denote the set of
problem solving methods, which will be described in detail in sub- start tasks that consists of task in Wx or its last rework in Px Ix . Then,
sequent Sections. Let and denotes the employed task estimation we have p j = p i p̄i for task j Ux Sx .
method and sub-problem solving method, respectively. The macro ap- The duration of each virtual task is consists of the processing time of
proximate decision procedure is as follows: current waiting task and all its potential inspections and rework time in
the future. Suppose that task j Ux is the l j -th rework of its original
Algorithm 1. Macro approximate decision-making procedure task. Let d j denotes the expected duration of the virtual task that
correspond to task j Ux . As illustrated in Fig. 2, d j is the sum of
1: Initialize x x 0 , Wx , Px and Ix ;
expected duration of task j Ux and its subsequent tasks, which is
2: while Wx and Px and Ix do
3: Approximate task estimation with method represented by
4: Get an action a x by solving the corresponding unrelated PMSP with method
p j d¯j , if l j = Lj
based on the task estimations; dj=
5: Apply solution a x ; p j (d¯j + hj ) + d j , otherwise (4)
6: Get the next decision point and the corresponding completed task set;
7: Randomly generate the results of completed inspection tasks; If task j Ux will be processed on machine r R , then its remaining
8: Get the next state x by adding new rework tasks to Wx , removing completed duration d jr can be obtained from formula (4) by replacing d̄j with djr .
tasks from Px and adding completed tasks that need to be inspected to Ix , x x.
In addition, we can estimate the latest finish time of task j by
9: end while
LFj = Dj d j + p j d̄j (5)
The work flow of any task is illustrated in Fig. 2.
4.2. Task estimation
5. Mathematical model
We consider two methods for task estimation, namely the aggregate
estimation and separate estimation. The aggregate estimation is to As mentioned previously, the PMSP has been extensively studied
merge the remaining process of each current waiting task into a virtual and different MIP formulations were proposed (Unlu & Mason, 2010).
task and assumes that all current processing and inspecting tasks will no The formulation based on time-indexed (TI) variables is widely used
longer need to be reworked. In addition, each virtual task is assumed to because its LP-relaxation provides a strong bound though a large
be continuously processed on the same machine. This method is similar number of variables and constraints required (Bigras, Gamache, &
to the estimation of expected total processing time by Raghavan et al. Savard, 2008; Sousa & Wolsey, 1992). Another popular formulation is
(2015, 2018). based on linear ordering (LO) variables which arise in the linear or-
There are three drawbacks of the aggregate estimation method. First dering problem (Dyer & Wolsey, 1990; Unlu & Mason, 2010).
of all, the inspections are merged into the virtual task which over- Generally speaking, the TI model is sensitive to task duration while
estimates the workload of relevant machines. Secondly, all current LO model is sensitive to the number of tasks (Unlu & Mason, 2010).
processing and inspecting tasks are assumed to no longer need to be Therefore, we construct a TI model for the instances with small dura-
reworked, which might be an over-optimistic estimate. Thirdly, each tion, and construct a LO model for the instances with large duration.
current waiting task and all its potential rework tasks are assumed to be Considering that there might be several optimal solutions of the un-
continuously processed on the same machine. However, assigning them related PMSP that results in no tardiness, we construct a bi-objective
to different machines might yield better results. These issues may be optimization model to make the decisions in a greedy manager. The
overcome with the separate estimation method, which considers each first objective is to minimize TWT and the second objective is to
potential task individually. This method is also used by Rambod and minimize total weighted completion time. These two objectives are
Rezaeian (2014). combined through a large enough coefficient C +
.
Let Ux denotes the set of all remaining tasks that consists of Wx and
its potential reworks, as well as the potential reworks of task in Px and 5.1. TI formulation based on aggregate task estimation
Ix . Since we are not sure which machine will process a remaining task
before the decision is made, we just estimate its expected duration Under the aggregate task estimation, we can construct both the TI
under the assumption that each machine has the same probability of model and LO model. Since the unrelated PMSP based on aggregate task
processing it. Let d̄j and p̄j denote the average immediate processing estimation is a special case of it based on separate task estimation, we
time and rework probability of task j Ux , respectively. Then, d̄j and p̄j only describe the LO model based on separate task estimation in the
are given by Eq. (2) and (3), respectively. next Section. The TI model based on aggregate task estimation is as
1 follows.
djr , if j Wx
Rx Let ETr t x denotes the earliest idle time of machine r R in cur-
r Rx
d¯j = 1 rent state x , G denotes a large enough time that all waiting tasks can be
djr , otherwise
R started, binary decision variable jtr = 1 if task j Wx is processed on
r R (2)
machine r R at time t and jtr = 0 otherwise. The proposed TI model
1 is represented by
Rx
pjr , if j Wx
p¯j =
r Rx min wj (C (cj Dj )+ + cj )
1
pjr , otherwise j Wx (6)
R
r R (3)
G
In order to estimate the expected duration of a remaining task, we s. t. jtr = 1, j Wx
have to further consider its probability of occurrence. Let i and p j r R t = ETr (7)

3
X. Wang, et al. Computers & Industrial Engineering 145 (2020) 106505

t c rj cir + p j djr (x ijr + z ir + z rj 2) C (1 xijr ), i, j Ux , r R


jt r 1, r R, t ETr , G (20)
j Wx t = max{ETr , t d jr + 1} (8)
cj c rj , j Ux , r R (21)
G
cj = jtr t+d jr , j Wx p j = 1, j Wx (22)
r R t = ETr (9)
p j = pjr , j Px Ix (23)
jtr {0, 1}, j Wx , r R, t [ETr , G]. (10)
p j =pi z ir pir , i, j Ux , =1
Constraints (7) indicate that each task must be processed by one of the ij
(24)
r R
machines when it become idle. Constraints (8) indicate that each ma-
chine can only handle no more than one task at any time point. x ijr = 0, i Ux , j Prei , r R (25)
There is no doubt that a tighter G is essential to reduce the number
of decision variables. Let dmax = max{d jr j Wx , r R} denote the c rj +, cj +, j Ux , r R (26)
maximum estimated duration of waiting tasks, tI = max{ETr r R} de-
notes the earliest time that all machines become idle. Suppose that both x ijr 0, 1 , yijr 0, 1 , z ir 0, 1 , i, j Ux , r R.
the machines and the waiting tasks are identical, and the duration is (27)
dmax . Then, all waiting tasks will be finished after u = Wx / R times of
Constraints (12) indicate that there must be a sequence if two different
dmax , where · is a ceiling function. Based on this approximate esti-
tasks are processed on the same machine. Constraints (13) means there
mation, we set G = tI + udmax .
is no circle between any three tasks on any machine. Constraints (14)
indicate that each task should be processed on one of the machines.
5.2. LO formulation based on separate task estimation
Constraints (15) and (16) are used to ensure that both z ir and z rj are
equal to 0 if yijr = 1 and both of them are equal to 1 otherwise. Con-
Since the estimated durations are sequence-dependent under the
straints (17) indicate that each task can only be started when the ma-
separate task estimation, the TI model is difficult to be construct in this
chine is available. Constraints (18) indicate that each task in Sx can only
case. Under the separate task estimation, Rambod and Rezaeian (2014)
be started when it is eligible. Constraints (19) indicate that each task
proposed a MIP model with binary decision variable x ir and yij , where
can only be started after the expected completion of its immediate
x ir indicates whether task i is processed on machine r or not and yij
predecessor inspection. Constraints (20) indicate that each task can
indicate whether task j is processed exactly after task i. In this section,
only be started after the completion of tasks that precede it on the same
we will introduce a MIP formulation based on LO variables which is
machine. Constraints (22)–(24) are used to calculate the probability of
relatively simpler.
occurrence of each task. Constraints (25) indicate that the processing
In order to build the LO model, we further define the following
order should respect the precedence constraint.
variables:
We can get the LO model based on aggregate task estimation from
c rj +
denotes the completion time of task j on machine r;
the above model by mainly removing constraint (19) and (22)–(24).
z rj {0, 1} taking value 1 if task j is processed on machine r and 0
otherwise;
6. Approaches
yijr 0, 1 taking value 0 if both task i and j are processed on
Due to the NP hardness of the problem, approximate methods (such
machine r and 1 otherwise;
as PRs, meta-heuristics, etc.) are more preferred to the medium and
x ijr {0, 1} taking value 1 if task i is processed before j on machine r
and 0 otherwise; large-scale problems. Although PRs for the dynamic unrelated PMSP
ej denotes the earliest time that task j Sx can be started, i.e., ej = t x (Durasevic & Jakobovic, 2018) and meta-heuristics for the scheduling
for j Wx and ej is equal to the completion time of last inspection for problem with random rework (Raghavan et al., 2018; Rambod &
j Sx Wx ; Rezaeian, 2014) has been studied, we will propose and compare some
Prej is the set of predecessors of task j; approximate methods for unrelated PMSP with random rework to
Fx Ux denotes the set of remaining final rework tasks in state x; minimize expected TWT. Specifically, the proposed approximate
{0, 1} taking value 1 if task j is the next rework of i and 0 methods are based on PRs, GA and SA, respectively.
ij
otherwise.
Finally, the proposed LO model is represented by 6.1. Priority rule

min wj (C (cj Dj )+ + cj ) The decision-making process in each state can be divided into two
j Fx (11) phases. The first phase is to select some waiting tasks to start. The
s. t. x ijr + x jir + yijr = 1, i, j Ux , i < j, r R second phase is to allocate it to idle machines. In the proposed PRs-
(12)
based approximate approach, we first select the waiting tasks to start by
x ijr + r
x jk + xkir 2, i, j, k Ux , i < j < k , r R (13) using PRs, and then let the tasks be processed on the machine they will
be completed the soonest. This approach is also described in Durasevic
z jr = 1, j Ux and Jakobovic (2018).
r R (14) Although a variety of PRs were proposed during the past decades,
none of them can dominate all other PRs in all instances. In the present
z ir + z rj + yijr 2, i, j Ux , i < j, r R (15) work, we only focus on the PRs that are suitable for minimizing TWT. It
z ir + z rj + 2yijr 2, i, j Ux , i < j, r R has been shown that the earliest due date (EDD), minimal latest finish
(16)
time (MinLFT), weighted shortest processing time (WSPT), modified
c rj ETr + p j djr C (1 z rj ), j Ux , r R (17) due date (MDD) (Baker & Bertrand, 1982), weighted MDD (WMDD)
(Kanet & Li, 2004), cost over time (COVERT) (Carroll, 1965) and ap-
c rj ej + p j djr C (1 z rj ), j Sx , r R (18) parent tardiness cost (ATC) (Vepsalainen & Morton, 1987) performs
well for solving a flowshop or a jobshop scheduling problem to mini-
c rj ci + p i hi + p j djr C (1 z rj ), i, j Ux , ij = 1, r R (19) mize TWT. The selected PRs are shown in Table 1. where:

4
X. Wang, et al. Computers & Industrial Engineering 145 (2020) 106505

Table 1 Yassine et al., 2017). Generally, the initial solutions generated by ef-
The selected seven classical PRs. fective PRs are better than random solutions, which can also speedup
PRs Rank Formulation the convergence of GA.
The procedure of the proposed modified GA is as follows:
EDD Min Dj
MinLFT Min LFj Step 1. Initial population. Generate part of the initial individuals by
MDD Min (Dj , d j + t )+ PRs as listed in Table 1 and generate remaining individuals ran-
WMDD Min
domly.
1
(Dj t , d j )+
wj
ATC Max (Dj d j t )+
Step 2. Fitness evaluation. Decoding and evaluate the fitness function
wi
d j
exp
1 d¯ value fi of each individual i pop with objective function (6) or
COVERT Max (Dj d j t )+
+ (11). Sort the individuals in ascending order of fitness function
wi
d j
1
2 Wj value.
WSPT Max wi Step 3. Loop until the maximum number of generations is reached.
d j
Step 3.1. Elitism mechanism. Copy the elites to the next generation.
Step 3.2. Selection strategy. Randomly select an elite individual
elite _ size
with the selection probability pi = fi / k = 1 fk , randomly select
LFj is the latest finish time of task j as described in Eq. (5);
another individual with the selection probability
t is current decision time; pop _ size
pi = fi / k = 1 fk .
d j is the estimated duration of task j as described in Eq. (4); Step 3.3. Crossover operator. Generate two offspring with single
d̄ is the average estimated duration of all waiting tasks; point crossover operator. Check and modify the illegal offspring
1 is a look-ahead parameter that scales the slack according to the due to the lost or redundant genes (interchange the repetitive
expected number of competing jobs (Though Vepsalainen & Morton tasks in the first offspring with the repetitive tasks in the second
(1987) suggested that 1 = 2 for static flow shops and 1 = 3 for offspring).
dynamic job shops, we take different values of 1 and find that the Step 3.4. Mutation operator. Randomly select a chromosome and its
best results can be obtained with 1 = 1); two genes. Swap the two task sequence genes. The machine as-
Wj is the average waiting time of task j (we find that its best value is signment genes of all chromosomes are mutated uniformly.
equal to d j );
Step 3.5. Fitness evaluation. Decoding and evaluate the fitness
2 is another look-ahead parameter that adjusting the expected
function value of each new chromosome with objective function
waiting time to the worst case, say the 99% limit of the cumulative
(6) or (11). Sort the chromosomes in ascending order of fitness
probability distribution (In accordance with Vepsalainen & Morton
function value.
(1987), we let 2 = 2 ).

6.2. Genetic algorithm 6.3. Simulated annealing

Although the GA algorithms have been applied to other scheduling The structure of representation of the SA solution is illustrated in
problems with random rework (Raghavan et al., 2018; Rambod & Fig. 4, which is similar to the one used by Bektur and Saraç (2019).
Rezaeian, 2014; Yassine et al., 2017), our GA algorithm differs from Here, rework tasks are also labeled with the same number as its original
these algorithms in the following three aspects. task.
First of all, the encoding method (as illustrated in Fig. 3) adopted The procedure of the proposed SA is as follows:
here is similar to it in Rambod and Rezaeian (2014). The difference is
that we label all rework tasks with the same number as the original Step 1. Initialize temperature T, final temperature Tf and cooling rate .
task. The actual task number in each position of the task sequence is Step 2. Select the best solution by PRs as the initial solution S.
determined in the decoding stage. For example, all rework tasks of Step 3. Loop until T < Tf .
original task 2 are labeled with number 2 in Fig. 3. No matter which Step 3.1. Generate a neighborhood solution S of S by randomly selecting one of the
following three typical local search methods.
swap operations are executed, the decoded sequence of task 2 will al-
Move. Randomly move a task from one machine to another machine.
ways be j2 , j6 and j7 , where j6 and j7 are the first and second rework task Swap 1. Randomly swap two tasks of a machine.
of j2 , respectively. In other words, the precedence constraints between Swap 2. Randomly swap two tasks of two different machines.
the original task and rework tasks will never be violated by any swap Step 3.2. Set = f (S ) f (S ) . Let S be S if < 0 , otherwise let S be S with pr-
operation. Therefore, the proposed GA algorithm can be used in com- obability exp( /T ).
bination with both aggregate task estimation and separate task esti- Step 3.3. Set T = T .

mation.
The second difference is that an elitism mechanism is adopted to
prevent losing non-dominated individuals generated during the evolu-
7. Computational study
tionary process and speedup the convergence of GA, while it is not
adopted by Yassine et al. (2017), Rambod and Rezaeian (2014) and
To assess the performance of the constructed models and methods,
Raghavan et al. (2018).
we design a number of computational experiments based on randomly
The third difference is that part of initial individuals are generated
generated instances. The program is coded in C# and runs on a work-
by the effective PRs as listed in Table 1 and the remaining are generated
station that has an Intel Core CPU 2.1 GHz and 128 GB of RAM. All of
randomly, while all of the initial individuals are generated randomly in
the MIP models are solved by Gurobi® (version 8.1) with an academic
relevant research (Raghavan et al., 2018; Rambod & Rezaeian, 2014;
license and default parameter settings.

Fig. 3. The schematic structure of representation of the GA solution. Fig. 4. The schematic structure of representation of the SA solution.

5
X. Wang, et al. Computers & Industrial Engineering 145 (2020) 106505

p 4 = p 2 p̄2 = 0.2.
According to the aggregate task estimation, we merge the remaining
process of each current waiting task and get two virtual tasks. Then, we
calculate the expected duration of each virtual task based on formula
(4). Specifically, we have d 3 = p 3 d¯3 = 0.3, d 4 = p 4 d¯4 = 0.4, d 1 =
p 1 (d¯1 + h1) + d 3 = 7.3 and d 2 = p 2 (d¯2 + h2 ) + d 4 = 4.4 . In other
words, the expected duration of the two virtual tasks are 7.3 and 4.4,
Fig. 5. An illustrative example with N = 2, M = 1 and L1 = L 2 = 1. respectively.
According to the separate task estimation, we keep all current
waiting tasks and all its potential rework tasks. In this example, there
Table 2
are four tasks and the expected duration of task j is equal to p j d̄j . Note
Parameter settings of durations and rework probabilities.
that the time interval between a task with its next rework needs to be
Parameters Distribution greater than the expected time for inspection.
Under the above task estimation, the proposed methods can be
Small Duration Large Duration
applied to solve the transformed unrelated PMSP. Since there are only
Duration of original tasks U {5, 10} U {50, 150} four scenarios of random rework in this example, an optimal solution
Rework probability of original tasks U (0.1, 0.3) U (0.1, 0.3) can be easily obtained using SDP. The minimum expected TWT is 1.03
Inspection duration of original tasks U {2, 5} U {20, 60} and the optimal action at the initial time is to start task 1. Due to the
Duration of 1st time rework tasks U {3, 5} U {30, 90}
limitation of space, we only present the results obtained by EDD and
Rework probability of 1st time rework tasks U (0.05, 0.15) U (0.05, 0.15)
WMDD rule with one scenario (task 1 rework and task 2 not rework).
Inspection duration of 1st time rework tasks U {1, 3} U {10, 30}
Duration of 2nd time rework tasks U {1, 3} U {10, 50}
According to the formulation of EDD and WMDD as listed in Table 1,
task 2 will be started first when using EDD because D1 > D2 and task 1
will be started first when using WMDD because
Table 3 (D1, d 1)+/ w1 < (D2 , d 2 )+/ w2 . Finally, the TWT obtained by EDD and
ANOVA table for the experiment on tuning the parameters of SA. WMDD in this scenario are 1.5 and 1.3, respectively.
Source Sum of squares df Mean square F-Ratio p-Value
7.2. Experimental data
EM 5261.664 1 5261.664 49.231 0.000
N 167.951 1 167.951 1.415 0.235
DT 1591.603 1 1591.603 13.791 0.000 We set up a variety of different problem environments to test the
DDM 1056.876 1 1056.876 9.060 0.003 proposed methods more comprehensively. Specifically, we consider
727.976 2 363.988 3.093 0.046 three instance sizes (small, medium and large), two types of duration
T0 4.062 2 2.031 0.017 0.983 (small and large), three due date tightness (tight, medium and loose)
10030.271 2 5015.136 52.235 0.000
and two due date setting approaches. Hence, there are in total
Total 51218.700 431
3 × 2 × 3 × 2 = 36 combinations of environment parameters. For each
combination of environment parameters, we randomly generate five
7.1. An illustrative example instances. Finally, we get 36 × 5 = 180 instances.
The number of original tasks and machines corresponding to dif-
In order to illustrate the problem and how the task estimation and ferent instance sizes is N = {6, 20, 50} and M = {2, 5, 10} , respectively.
schedule procedures proposed in the paper work, we introduce a small The tardiness weight of each original task is randomly generated ac-
example as illustrated in Fig. 5. cording to the continuous uniform distribution U (0.1, 0.5) . The max-
In this example, the original task set = {1, 2} and its extended set imum number of reworks of each original task is set to 2. Durations and
= {1, 2, 3, 4} . At initial time t = 0 , we get the waiting task set rework probabilities are randomly generated according to the dis-
W0 = {1, 2} , processing task set P0 = , inspecting task set I0 = and tribution as showed in Table 2, where U {a, b} and U (a, b) denotes the
remaining task set U0 = {1, 2, 3, 4} . On this basis, we get S0 = W0 . Since discrete and continuous uniform distribution, respectively.
there is only one machine, we have d¯j = dj for We set the due date coefficient = {0.6, 0.8, 1.0} . The due dates are
j {1, 2, 3, 4}, p¯1 = p1 = 0.3, p¯2 = p2 = 0.2 and p¯3 = p¯4 = 0 . In addi- determined by the following two approaches. The first due date setting
tion, we have pj=1 for j W0, p 3 = p 1 p¯1 = 0.3 and approach is to randomly generate the due date of task j according
to the discrete uniform distribution U {1.2 d j , 2.4 d j } . The second due

Fig. 6. Confidence interval for Dev (TWT ) by parameter T0 and , respectively.

6
X. Wang, et al. Computers & Industrial Engineering 145 (2020) 106505

Table 4
A comparison of methods on average RDI.
Method DDM DT N Total

RND WIQ Small Large 6 20 50 0.6 0.8 1.0

ATC 50.26 36.65 42.60 44.31 64.46 37.10 28.81 53.03 32.15 45.19 43.45
COVERT 46.38 40.18 42.68 43.88 60.13 39.69 30.01 55.07 36.62 38.14 43.28
WSPT 57.83 99.59 82.96 74.47 97.73 72.49 65.92 71.09 78.69 86.37 78.71
MinLFT 84.09 31.23 52.99 62.33 52.39 62.87 57.72 72.93 53.10 46.94 57.66
EDD 83.71 29.83 52.75 60.79 55.13 57.18 58.01 70.19 53.29 46.83 56.77
MDD 56.48 23.66 44.81 35.33 59.16 33.29 27.76 46.73 39.64 33.84 40.07
WMDD 43.15 31.02 38.05 36.12 42.53 39.70 29.02 46.71 30.26 34.27 37.08
GA-A 9.04 7.64 10.35 6.33 15.67 3.35 6.00 7.72 8.58 8.72 8.34
GA-S 19.57 7.44 16.53 10.48 8.64 16.04 15.83 17.65 11.05 11.82 13.51
SA-A 8.70 7.57 10.84 5.43 16.41 3.49 4.51 7.18 8.03 9.19 8.14
SA-S 18.93 8.97 16.46 11.45 9.76 16.07 16.03 18.28 11.96 11.62 13.95
MIP-A 3.84 4.25 3.42 5.93 7.44 0.86 0.46 3.31 3.53 5.31 4.05
MIP-S 4.04 7.73 7.97 3.79 5.88 –a – 5.22 4.19 8.24 5.88
SDP 0.00 1.34 1.34 0.00 0.67 – – 0.00 0.00 2.01 0.67

a
Symbol ’–’ indicates that not all instances can be solved by the method due to memory limitations.

date setting approach using task characteristic information together (gensize), population size (popsize), crossover operator, probability of
with shop status information, which is similar to the work in queue mutation ( pm ) and percentage of elite ( pe ). By comparing one-point
(WIQ) method (Ragatz & Mabert, 1984). In WIQ model, the flowtime of crossover, two-point crossover and uniform crossover, we find that the
task j is estimated based on the total processing time of all tasks in single-point crossover operator is the best choice for this experiment,
the queue on j’s routing. In this paper, we set the due date of task j which was also employed in Rambod and Rezaeian (2014), Raghavan
as et al. (2018). According to (Raghavan et al., 2018), we further set the
parameters gensize = 20, pe = 5%, pm = 1% and popsize = 50z , where z
Dj = (d j + Qj ) (28)
is the number of estimated remaining tasks.
where Qj is the expected cumulative work of the tasks arrived before As described in Section 6.3, the main parameters of the proposed SA
task j. Let j denotes the tasks arrived before task j, we approximate Qj algorithm include the initial temperature T0 , final temperature Tf and
as cooling rate . We fix Tf to 0.001, and carried out a full factorial design
using the parameter T0 and as factors at the following levels (values):
d j T0 = {300, 400, 500}, = {0.9, 0.95, 0.99} . Furthermore, we consider the
Qj =
j j
R
(29) uncontrolled problem environment factors, which include the two task
estimation methods (EM), two instance sizes (N = 20 and N = 50 ), two
It is noted that the due dates assigned by the RND-based method are duration types (DT), two due date settings (DDS) and three due date
concentrated in a certain range, which will lead to a relatively tighter coefficients. We solve all instances by SA under each combination of the
due date environment than the WIQ-based method. above factors. Once all computations are done, we use the following
In this experiment, the methods to be compared include the PRs as relative deviation as the response variable in ANOVA.
listed in Table 1, GA, SA and MIP based on aggregate estimation (called
Givensol Bestsol
GA-A, SA-A and MIP-A, respectively), GA, SA and MIP based on sepa- Dev (TWT ) = × 100%
Bestsol (30)
rate estimation (called GA-S, SA-S and MIP-S, respectively) and sto-
chastic dynamic programming (SDP). These methods will be applied to where Givensol is the average TWT for a given method, Bestsol is the best
different instances. Specifically, PRs, GA-A, GA-S, SA-A, SA-S can be average TWT among all methods.
applied to all of the instances, but MIP-S and SDP cannot be applied to There are three assumptions for applying ANOVA, namely nor-
the medium and large-scale instances, MIP-A cannot be applied to the mality, homogeneity of variance and independence of samples. The
medium and large-scale instances with large durations. samples resulting from this experiment were analyzed and all three
Small-scale instances will be solved by enumerating all possible hypotheses could be accepted. The ANOVA results are listed in Table 3.
scenarios, and the medium and large-scale instances will be solved by The analysis indicates that the factors EM, DT, DDM, and are
Monte Carlo simulation (each method will solve 1000 scenarios of each significant. The magnitude of the F-ratio is a clear indicator of sig-
instance). During the computation, the indicators of average TWT and nificance. Surprisingly, the factors N and T0 do not yield statistically
computation time will be collected for comparisons. significant. Although a better combination of parameters might be de-
termined by considering the uncontrollable factors, it will also com-
7.3. Parameter selection of GA and SA plicate the SA algorithm. Therefore, we only focus on the analysis of the
best level of parameter . The interval plot of the parameters T0 and
The performance of GA and SA is sensitive to the predefined para- are shown in Fig. 6.
meters. Unfortunately, there are no choices of parameters that will be As a result of the experimental analysis, we set the parameters
good for all problems, and there is no general way to find the best T0 = 300 and = 0.99. It is noted that we chose parameters T0 = 300
choices for a given problem. To achieve better results of GA and SA, the because it is statistically similar to T0 = 400 and T0 = 500 but more ef-
algorithm’s parameter tuning is required. Since some GA algorithms ficient.
have been applied to the scheduling problem with random rework
(Raghavan et al., 2018; Rambod & Rezaeian, 2014; Yassine et al., 7.4. Comparison and discussion
2017), we will refer to these studies to set the parameters of our GA. In
this section, we focus on the parameter setting of SA by means of a According to the above experimental design, we obtain the average
multi-factor Analysis of Variance (ANOVA) technique. TWT of each method under each combination of environment para-
In case of GA, independent parameters including generation size meters. Since the optimization effects of PRs are significantly worse

7
X. Wang, et al. Computers & Industrial Engineering 145 (2020) 106505

Fig. 7. Confidence interval for average RDI.

than the other methods, the objective value for each method is com- to analyse the influence of the levels of the factor on the average RDI.
pared using the following relative deviation index (Perez-Gonzalez From Table 4 and Fig. 7, we draw the following conclusions.
et al., 2019) rather than Dev (TWT ) . First of all, there is a significant gap between the PRs and other
Givensol Bestsol methods, and the performance of PRs varied with the problem en-
RDI = × 100% vironment. Generally speaking, ATC, COVERT, MDD and WMDD per-
Worstsol Bestsol (31)
form well and are stable in most cases, while MinLFT and EDD only
where Worstsol is the worst average TWT among all methods. perform well in a loose due date environment especially when the WIQ-
Table 4 shows the average RDI values for each levels of the factors based due date setting approach is applied. WSPT has the worst
and Fig. 7 contains five graphics with Tukey 95% confidence intervals

8
X. Wang, et al. Computers & Industrial Engineering 145 (2020) 106505

Table 5 heuristics for unrelated PMSP with random rework to minimize ex-
Comparison of methods on average computation time in seconds. pected TWT. To the best of our knowledge, this is the first paper to
Method Small Duration Large Duration compare approximate methods for unrelated PMSP with random re-
work with the exact method SDP. Although this comparison can only be
N=6 N = 20 N = 50 N=6 N = 20 N = 50 carried out in a small-scale instance due to the curse of dimensionality,
it proves that the approximate method is effective. Specifically, we
PRs 0.00 0.00 0.00 0.00 0.00 0.00
GA-A 0.24 0.18 1.37 0.13 0.26 3.05 proposed a methodology framework that first transforms the original
GA-S 2.65 6.60 97.81 1.76 9.90 235.32 stochastic optimization problem to multiple deterministic unrelated
SA-A 0.13 0.01 0.04 0.09 0.02 0.11 PMSP through task estimation. Two MIP formulations that based on TI
SA-S 0.35 0.16 0.96 0.27 0.26 2.24 and LO variables and two meta-heuristics (GA and SA) are proposed for
MIP-A 0.85 1.77 18.99 0.58 – –
solving the transformed unrelated PMSP in each state. These models
MIP-S 155.17 – – 81.96 – –
SDP 54.00 – – 289.42 – – and algorithms are different from those in the existing literature.
Computational results show that the proposed aggregate task estima-
tion approach is more efficient and more stable than the existing se-
performance. In addition, the trend of results under large and small parate task estimation. In addition, the meta-heuristics are superior to
durations are almost the same. These conclusions are consistent with classical PRs and close to the exact method SDP. There is no significant
those described in existing research (Durasevic & Jakobovic, 2018; difference in the optimization effect between SA and GA, but the
Wang et al., 2015). This suggested that the weighted PRs (WMDD, ATC, computational efficiency of SA is much higher than that of GA.
and COVERT) is a good choice for minimizing TWT, the due date-based The limitations of this work are that the proposed TI model only
PRs (MinLFT, EDD and MDD) is also a reasonable strategy for decision- performs well with small task duration (Bigras et al., 2008; Dyer &
making in a loose due date environment (e.g. when WIQ-based due date Wolsey, 1990) and only the gap between the approximate solution and
setting approach is applied). the optimal solution of small-scale instances can be evaluated. The
Secondly, the task estimation approach has a great impact on the work could be extended in several research directions. First of all, more
overall optimization effect. Similar to the performance of PRs, the meta-heuristics (such as tabu search, iterated greedy algorithm, etc.)
performance of task estimation approach also varied with the problem can be implemented to the problem considered. Secondly, machine
environment. Generally, the proposed aggregate task estimation per- learning methods can be employed to automatically choose an appro-
forms well in most cases, especially in a tight due date environment priate task estimation approach in any given state.
(e.g. the RND-based due date setting approach is applied or = 0.6 ). By
contrast, separate task estimation only performs well in a very loose due CRediT authorship contribution statement
date environment (e.g. the WIQ-based due date setting is applied with
= 1.0 ). To the best of our knowledge, this is the first paper to make a Xiaoming Wang: Conceptualization, Methodology, Software,
detailed comparison between the aggregate and separate task estimate Investigation, Formal analysis, Writing - original draft. Zhantao Li:
approach though similar methods were applied in Raghavan et al. Validation, Formal analysis, Data curation, Visualization, Writing - re-
(2015, 2018) and Rambod and Rezaeian (2014). view & editing. Qingxin Chen: Resources, Writing - review & editing,
Thirdly, there is only a small difference between the optimization Supervision. Ning Mao: Resources, Writing - review & editing,
effects of GA and SA under the same task estimation approach, and the Supervision.
best results obtained by the meta-heuristics (GA or SA) are close to the
optimal solutions. Specifically, the gaps between the best results ob- Acknowledgement
tained by GA or SA with the optimal results obtained by SDP or MIP are
within 20% in most cases. It is also noted that the results obtained by This research is jointly supported by the National Natural Science
MIP or SDP might be worse than the results obtained by meta-heuristics Foundation of China under Contract No. 71972053, 61973089,
(e.g. the case with N = 6, = 1.0 and WIQ-based due date setting ap- 51505090, 51675107 and 51775120.
proach). There are two reasons to explain this. The first reason is that
the SDP we used only considers the non-delay schedules, while the References
meta-heuristics will consider feasible delay schedules. The second
reason is due to the unavoidable task estimation errors. In addition, the Baker, K. R., & Bertrand, J. (1982). A dynamic priority rule for scheduling against due-
above trend of results under large and small duration is almost the dates. Journal of Operations Management, 3, 37–42.
Bektur, G., & Saraç, T. (2019). A mathematical model and heuristic algorithms for an
same. unrelated parallel machine scheduling problem with sequence-dependent setup
Table 5 shows the average computation time in seconds for each times, machine eligibility restrictions and a common server. Computers & Operations
method. The computational time corresponding to N = 6 refers to the Research, 103, 46–63.
Bigras, L.-P., Gamache, M., & Savard, G. (2008). Time-indexed formulations and the total
average time required to solve an instance, while that corresponding to weighted tardiness problem. INFORMS Journal on Computing, 20, 133–142.
N = 20 and N = 50 refers to the average time required to complete a Browning, T. R., & Yassine, A. A. (2016). Managing a portfolio of product development
single simulation. It shows that the small-scale instances can be solved projects under resource constraints. Decision Sciences, 47, 333–372.
Bülbül, K., & Şen, H. (2017). An exact extended formulation for the unrelated parallel
by all proposed methods. However, as the problem instance size grows,
machine total weighted completion time problem. Journal of Scheduling, 20, 373–389.
the SDP method becomes intractable due to the curse of dimensionality Carroll, D. C. (1965). Heuristic Sequencing of Jobs with Single and Multiple Components.
and the MIP model can not be solved because the problem is NP-hard. Thesis.
Cheng, T. C. E., & Sin, C. C. S. (1990). A state-of-the-art review of parallel-machine
Fortunately, the proposed meta-heuristics are applicable to large-scale
scheduling research. European Journal of Operational Research, 47, 271–292.
instances. For the same meta-heuristic, it runs much faster under ag- Durasevic, M., & Jakobovic, D. (2018). A survey of dispatching rules for the dynamic
gregate task estimation than under separate task estimation. In addi- unrelated machines environment. Expert Systems with Applications, 113, 555–569.
tion, the SA algorithm is preferred because the optimization effect of SA Dyer, M. E., & Wolsey, L. A. (1990). Formulating the single machine sequencing problem
with release dates as a mixed integer program. Discrete Applied Mathematics, 26,
is not significantly different from that of GA, but SA is more efficient. 255–270.
Fanjul-Peyro, L., Perea, F., & Ruiz, R. (2017). Models and matheuristics for the unrelated
8. Summary and conclusions parallel machine scheduling problem with additional resources. European Journal of
Operational Research, 260, 482–493.
Fanjul-Peyro, L., & Ruiz, R. (2010). Iterated greedy local search methods for unrelated
The most significant contribution of this paper lies in the develop- parallel machine scheduling. European Journal of Operational Research, 207, 55–69.
ment of aggregate task estimation approach and improved meta- Guo, Y., Huang, M., Wang, Q., & Leon, V. J. (2016). Single-machine rework rescheduling

9
X. Wang, et al. Computers & Industrial Engineering 145 (2020) 106505

to minimize maximum waiting-times with fixed sequence of jobs and ready times. Sels, V., Coelho, J., Dias, A. M., & Vanhoucke, M. (2015). Hybrid tabu search and a
Computers & Industrial Engineering, 91, 262–273. truncated branch-and-bound for the unrelated parallel machine scheduling problem.
Joo, C. M., & Kim, B. S. (2015). Hybrid genetic algorithms with dispatching rules for Computers & Operations Research, 53, 107–117.
unrelated parallel machine scheduling with setup time and production availability. Sen, H., & Bülbül, K. (2015). A strong preemptive relaxation for weighted tardiness and
Computers & Industrial Engineering, 85, 102–109. earliness/tardiness problems on unrelated parallel machines. INFORMS Journal on
Kanet, J. J., & Li, X. (2004). A weighted modified due date rule for sequencing to mini- Computing, 27, 135–150.
mize weighted tardiness. Journal of Scheduling, 7, 261–276. Sha, D., Hsu, S., Che, Z., & Chen, C. (2006). A dispatching rule for photolithography
Kang, Y. H., & Shin, H. J. (2010). An adaptive scheduling algorithm for a parallel machine scheduling with an on-line rework strategy. Computers & Industrial Engineering, 50,
problem with rework processes. International Journal of Production Research, 48, 233–247.
95–115. Shin, H. J., & Kang, Y. H. (2010). A rework-based dispatching algorithm for module
Lawler, E. L. (1977). A “pseudopolynomial” algorithm for sequencing jobs to minimize process in tft-lcd manufacture. International Journal of Production Research, 48,
total tardiness. In P. Hammer, E. Johnson, B. Korte, & G. Nemhauser (Eds.), Studies in 915–931.
Integer Programming. Annals of Discrete Mathematics (Vol. 1, pp. 331–342). Elsevier. Sousa, J. P., & Wolsey, L. A. (1992). A time indexed formulation of non-preemptive single
Liberopoulos, G., Kozanidis, G., & Tsarouhas, P. (2007). Performance evaluation of an machine scheduling problems. Mathematical Programming, 54, 353–367.
automatic transfer line with wip scrapping during long failures. Manufacturing & Teunter, R. H., & Flapper, S. D. P. (2006). A comparison of bottling alternatives in the
Service Operations Management, 9, 62–83. pharmaceutical industry. Journal of Operations Management, 24, 215–234.
Luh, P. B., Liu, F., & Moser, B. (1999). Scheduling of design projects with uncertain Unlu, Y., & Mason, S. J. (2010). Evaluation of mixed integer programming formulations
number of iterations. European Journal of Operational Research, 113, 575–592. for non-preemptive parallel machine scheduling problems. Computers & Industrial
Perez-Gonzalez, P., Fernandez-Viagas, V., García, M. Z., & Framinan, J. M. (2019). Engineering, 58, 785–800.
Constructive heuristics for the unrelated parallel machines scheduling problem with Vepsalainen, A. P. J., & Morton, T. E. (1987). Priority rules for job shops with weighted
machine eligibility and setup times. Computers & Industrial Engineering, 131, 131–145. tardiness costs. Management Science, 33, 1035–1047.
Ragatz, G. L., & Mabert, V. A. (1984). A simulation analysis of due date assignment rules. Villa, F., Vallada, E., & Fanjul-Peyro, L. (2018). Heuristic algorithms for the unrelated
Journal of Operations Management, 5, 27–39. parallel machine scheduling problem with one scarce additional resource. Expert
Raghavan, V. A., Yoon, S. W., & Srihari, K. (2015). Heuristic algorithms to minimize total Systems with Applications, 93, 28–38.
weighted tardiness with stochastic rework and reprocessing times. Journal of Wang, X., Chen, Q., Mao, N., Chen, X., & Li, Z. (2015). Proactive approach for stochastic
Manufacturing Systems, 37, 233–242. rcmpsp based on multi-priority rule combinations. International Journal of Production
Raghavan, V. A., Yoon, S. W., & Srihari, K. (2018). A modified genetic algorithm approach Research, 53, 1098–1110.
to minimize total weighted tardiness with stochastic rework and reprocessing times. Wang, X., Leus, R., Creemers, S., Chen, Q., & Mao, N. (2018). A ctmdp-based exact
Computers & Industrial Engineering, 123, 42–53. method for rcpsp with uncertain activity durations and rework. In N. Kliewer, J. F.
Rambod, M., & Rezaeian, J. (2014). Robust meta-heuristics implementation for unrelated Ehmke, & R. Borndörfer (Eds.). Operations Research Proceedings 2017 (pp. 559–565).
parallel machines scheduling problem with rework processes and machine eligibility Cham: Springer International Publishing.
restrictions. Computers & Industrial Engineering, 77, 15–28. Yassine, A. A., Mostafa, O., & Browning, T. R. (2017). Scheduling multiple, resource-
Rodriguez, F. J., Lozano, M., Blum, C., & García-Martínez, C. (2013). An iterated greedy constrained, iterative, product development projects with genetic algorithms.
algorithm for the large-scale unrelated parallel machines scheduling problem. Computers & Industrial Engineering, 107, 39–56.
Computers & Operations Research, 40, 1829–1841.

10

You might also like