International Journal of Production Research, 2021

Vol. 59, No. 14, 4403–4415, https://doi.org/10.1080/00207543.2020.1764125

Minimising makespan in job-shops with deterministic machine availability constraints

Shih-Wei Lina,b,c and Kuo-Ching Ying d∗

a Department
of Information Management, Chang Gung University, Taoyuan, Taiwan; b Department of Industrial Engineering and
Management, Ming Chi University of Technology, Taipei, Taiwan; c Department of Neurology, Linkou Chang Gung Memorial Hospital,
Taoyuan, Taiwan; d Department of Industrial Engineering and Management, National Taipei University of Technology, Taipei, Taiwan
(Received 8 July 2019; accepted 28 April 2020)

This paper proposes an effective and efficient multi-temperature simulated annealing (MTSA) algorithm to minimise the
makespan of a job-shop under the constraint that machines are not continuously available for processing during the whole
scheduling horizon. The proposed MTSA algorithm uses an embedded multi-temperature mechanism to vary the thermal
transition probabilities of the simulated annealing algorithm. This mechanism can help prevent the algorithm from becom-
ing trapped in a local minimum and ensures its movement towards a broad region of the search space containing optimal
solutions. An effective and robust lower bound is developed for the problem to evaluate the quality of solutions. Extensive
computational results show that the proposed MTSA algorithm significantly outperforms the state-of-the-art meta-heuristic
algorithms reported in the literature. The proposed algorithm and lower bound can assist further research in the schedul-
ing research field as it is both effective and efficient in handling job-shop scheduling problems with machine availability
constraints.
Keywords: scheduling; job-shop; machine availability constraints; meta-heuristic

1. Introduction
The job-shop scheduling problem (JSSP) is one of the most significant branches of scheduling problems, and has been
the subject of a great number of previous studies. The vast majority of JSSP studies have adopted the assumption that
machines are available for processing during the whole scheduling period. However, in numerous practical manufacturing
environments, machines may become unavailable for different reasons, such as machine breakdowns, unexpected quality
control problems, scheduled maintenance, and previous scheduling decisions in a rolling horizon framework (Tamssaouet,
Dauzère-Pérès, and Yugma, 2018). Therefore, planners usually carry out preventive maintenance tasks on the machines to
avoid the risk of unplanned downtime caused by machine failure. A significant real-life application is the semiconductor
fabrication plant scheduling problem, where machines are very expensive, and machine availability limitations must be
considered. In such cases, since the integration of availability constraints in the JSSP yields better modelling of industrial
reality and produces feasible schedules, it has thus received a lot of research attention over the last decade.
Although the JSSP with machine availability constraints plays an important role in many industrial settings, there is
still relatively little available literature on the problem. This work aims to solve the JSSP with deterministic machine
availability constraints, i.e. the unavailability periods (sometimes named holes) of machines are known in advance and
fixed, and preventive maintenance activities due to prior scheduling decisions in a rolling horizon framework. The problem
under consideration can be formally defined as a set of n pre-determined jobs J = {Ji |i = 1, 2, . . . , n} that must be pro-
cessed on a set of m dedicated machines M = {Mj |j = 1, 2, . . . , m}. Each job Ji (i = 1, 2, . . . , n) consists of a linear sequence
(routeing) of oi operations {Oi1 , . . . , Oioi }. Each machine may process at most one operation at a time, and each operation
Oij (i = 1, 2, . . . , n; j ∈ M) is performed exactly once by machine Mj during uninterrupted processing time units pij . For each
machine Mj (j = 1, 2, . . . , m), there are kj unavailability periods (holes) Hj = {hj1 , . . . , hjki }. The scheduling objective is to
find a schedule under the machine unavailability constraints to minimise the makespan Cmax , i.e. the total time required to
process all jobs. Also, the following assumptions are made in this study:
• The processing time for the operation of job i on the machine j, pij (i = 1, 2, . . . , n; j ∈ M) is known in advance.
• The beginnings and durations of unavailability periods are fixed and known in advance.
• The ready times of all jobs start at the beginning of the scheduling horizon.

*Corresponding author. Email: [email protected]

© 2020 Informa UK Limited, trading as Taylor & Francis Group
4404 S.-W. Lin and K.-C. Ying

• All machines are idle at the beginning of the scheduling horizon.

• Each machine operates independently of other machines, and is always available except during unavailability
periods.
• Operations may not be preempted by other operations or unavailability periods.
• A job may require different processing times on different machines.

Using the standard three-field α|β|γ notation, this problem can be denoted as Jm |hjk |Cmax , where Jm describes the shop
type as an m-machine job-shop; hjk specifies the JSSP with deterministic machine unavailability constraints; and Cmax
indicates that the objective is to minimise the makespan. The Jm |hjk |Cmax problem is known to be strongly NP-hard, as it
generalises the classical JSSP, which is strongly NP-hard even if there are only two machines and each job has at most three
operations (Lawler et al., 1993).
Considering the inherent complexities involved in the Jm |hjk |Cmax problem, efficient and robust solution algorithms
are necessary to facilitate industry applications of this scheduling problem. Among modern meta-heuristics, the simulated
annealing (SA) algorithm is very popular and has been successfully applied to solve different combinatorial optimisation
problems (Lin and Ying, 2015). Despite its merits, the basic SA algorithm may become trapped in local optimality when
solving large-scale complex problems (Lin et al., 2011). To improve the algorithm’s performance, this study proposed an
effective and efficient multi-temperature simulated annealing (MTSA) algorithm. The proposed MTSA algorithm combines
the advantages of the basic SA algorithm for effectively achieving search convergence with that of the multi-temperature
method for escaping local optimality. The remainder of this paper is arranged as follows. Section 2 reviews relevant studies
regarding the jm | hik | C max problem. Section 3 presents a lower bound for the jm | hik | C max problem. Section 4 describes the
procedure of the proposed MTSA algorithm. Section 5 verifies the performance of MTSA by comparing its experimental
results with those of state-of-the-art meta-heuristic algorithms on two benchmark problem sets. Finally, Section 6 presents
conclusions and recommendations for future research.

2. Literature review
Detailed literature reviews of different scheduling problems with machine availability constraints are given in Schmidt
(2000), Saidy and Taghavi-Fard (2008), and Ma et al. (2010). Unlike classical scheduling problems, the literature dedi-
cated to JSSP with machine availability constraints is relatively sparse. The earliest study on the JSSP involving machine
availability constraints was presented by Aggoune (2002), who studied the JSSP with machine unavailability periods for
either fixed or flexible preventive maintenance activities and proposed a heuristic algorithm and a branch-and-bound (B&B)
method to solve the problem. The experimental results obtained using randomly generated instances demonstrated the effec-
tiveness of the proposed approaches. Mauguière et al. (2003) also proposed a B&B algorithm for the JSSP with availability
constraints.
In the scheduling literature, JSSPs involving machine availability constraints can be divided into two traditional cases:
resumable and non-resumable. In the first case, the processing of operations can be interrupted by unavailability periods, and
resumed afterward; in the second, the production of operations cannot be interrupted by unavailability periods. Mauguiere
et al. (2005) proposed a B&B algorithm to deal with a general case of machine availability constraints in JSSPs with both
resumable and non-resumable operations. Computational results obtained using three modified benchmark sets showed that
the novel B&B algorithm outperformed the B&B algorithm presented in Mauguière et al. (2003).
The unavailability constraints resulting from preventive maintenance can be divided into two main cases: deterministic
and flexible. In deterministic cases, the maintenance periods are fixed in advance; in flexible cases, the maintenance periods
are decision variables scheduled in given time windows. Azem et al. (2007) used disjunctive and time-indexed formulations
to present two mixed-integer linear programming (MILP) models for the Jm |hjk |Cmax problem. Their experimental results
showed that the disjunctive formulation outperformed the time-indexed one. Subsequently, Zribi et al. (2008) proposed a
two-phase algorithm to minimise the makespan for the JSSP with limited machine availability. In the first phase, a priority
rule-based heuristic was used to solve the assignment problem. Then, in the second phase, a genetic algorithm (GA) was
applied to solve the sequencing problem; computational results showed that the two-phase algorithm gave interesting results
compared to existing algorithms, as well as the theoretical lower bound. Aggoune et al. (2009) extended the geometric
algorithm to present a polynomial solvable approach for solving the two-job JSSP with an arbitrary number of unavailability
periods on each machine; test results showed that this algorithm outperformed existing ones in terms of complexity as well
as its representation of real instances.
Ploydanai and Mungwattana (2010) proposed an algorithm based on a non-delay scheduling heuristic to solve the JSSP
with machine availability constraints. In the same year, Mati (2010) proposed taboo thresholding (TT) meta-heuristic that
used a new block-based neighbourhood function and some sufficient conditions to eliminate non-improving moves to solve
 International Journal of Production Research 4405

the Jm |hjk |Cmax problem. Simulation results carried out on the benchmark problem set of Azem et al. (2007), and newly
generated problem instances demonstrated the efficiency of the TT meta-heuristic method. After this, Azem et al. (2012)
proposed construction heuristics for the JSSP with limited resource availability constraints, preemption of an operation by an
unavailability period, and flexible unavailability periods. These construction heuristics could quickly determine a schedule
based on decision strategies, and be used in improving approximation methods (such as meta-heuristics). Benttaleb et al.
(2016) treated a non-preemptive case of a two-machine JSSP with an availability constraint on the first machine; they
proposed two mixed integer programming (MIP) models and a B&B algorithm to minimise the makespan for this problem.
Their computational results showed that the B&B algorithm was much more effective than the two MIP models.
Recently, Benttaleb et al. (2018a, 2018b) developed some MIP models and B&B algorithms based on different schemes
to minimise the makespan for the two-machine JSSP with availability constraints on one machine and each machine, respec-
tively. Their computational results, obtained using benchmark instances from the literature, validated the proposed schemes,
and proved the efficiency of the B&B algorithms. Tamssaouet et al. (2018) proposed a move evaluation that permitted
ignoring a great proportion of non-improving moves, and then presented a simulated annealing (SA) meta-heuristic and a
Tabu search (TS) meta-heuristic for solving the Jm |hjk |Cmax problem. They embedded diversification structures and specific
neighbourhood functions in the SA and TS meta-heuristics to improve their performance. Their computational experiments,
conducted using benchmark problem sets, showed that TS outperforms SA, Mati’s (2010), and Azem et al.’s MIP model
(2007). To the best of the authors’ knowledge, Tamssaouet et al.’s TS meta-heuristic (2018) is the current state-of-the-art
approach in the literature. Therefore, to demonstrate the performance of the MTSA algorithm proposed in this study, its
computational results are compared with those obtained using Tamssaouet et al.’s TS meta-heuristic.

3. Lower bound for the Jm |hjk |Cmax problem

In this section, we present a lower bound on the makespan, denoted by LBLY , for the Jm |hjk |Cmax problem to evaluate the
performance of all compared algorithms in this study. The LBLY is calculated using the following formula.

LBLY = max{LB1 , LB2 , LB3 },

where LB1 denotes the machine-based lower bound, LB2 represents the job-based lower bound, and LB3 is the lower bound
obtained by solving the disjunctive MILP formulation of Azem et al. (2007) via Gurobi version 8.1, while the maximal
computational time for each test instance is set to an elapsed CPU time of 1800 s.
The following two subsections describe the detailed steps to calculate LB1 and LB2 .

3.1. Machine-based lower bound

The machine-based lower bound, LB1 , is calculated based on the completion times of the machines. To calculate LB1 , the
following steps must be performed:
Step 1: Use the following equations to calculate the earliest starting time ESTi,[j] of each operation Oi,[j]
(i = 1, 2, . . . , n; j = 1, 2, . . . , oi ) in the order of operations for every job, and then form the data into a matrix ESTO
of the possible earliest starting times for every operation:

ESTi,[1] = 0, ∀i = 1, . . . , n, (1)
ESTi,[j+1] = ESTi,[j] + pi,[j] , ∀i = 1, . . . , n; j = 1, . . . , oi − 1. (2)

Step 2: Rearrange each row of ESTO according to the order of the machines (i.e. from M1 to Mm ) and form a matrix
ESTM of the possible earliest starting times for every machine.
Step 3: Select the minimum value of each column in ESTM , which is set as the earliest starting time of each machine,
ESTMj (j = 1, . . . , m).  
Step 4: LB1 = maxj=1,...,m {ESTMj + i pij + Hj }, where i pij (j = 1, . . . , m) is the total processing time of the jobs
on each machine. Hj is the total machine unavailable time of machine j before its completion time, which can be
calculated as follows:

Let kj be the hole number of machine j, and set the initial value of t = ESTMj + i pij .
FOR hole 1 TO hole kj ;
IF the starting time of the hole is less than t, then t := t+ the period of the hole;
OTHERWISE, STOP; 
OUTPUT Ht = t − (ESTMj + i pij ).
4406 S.-W. Lin and K.-C. Ying

Table 1. Lower bound example data.

Jobs Machine sequence Processing times Machines Unavailability periods
1 1, 3, 4, 5, 2 130, 140, 100, 120, 120 1 [206, 305], [662, 761]
2 5, 4, 2, 1, 3 60, 50, 50, 130, 150 2 [482, 591], [782, 891]
3 1, 4, 5, 2, 3 90, 60, 110, 80, 140 3 [311, 428], [635, 752]
4 2, 1, 3, 4, 5 150, 80, 60, 140, 110 4 [159, 256], [575, 672]
5 2, 5, 3, 4, 1 150, 120, 100, 140, 70 5 [313, 416], [882, 985]

3.2. Job-based lower bound

The job-based lower bound, LB2 , is calculated based on the completion times of the jobs. To calculate LB2 , the following
steps must be performed:
Step 1: Let ECTi,[1] = pi,[1] (∀i = 1, . . . , n) and ki,[j] (i = 1, 2, . . . , n; j = 1, 2, . . . , oi ) be the hole number of the
machine corresponding to the jth operation of job i. For each job i(i = 1, 2, . . . , n), use the following procedure to
calculate the earliest completion time ECTi,[j] of each operation Oi,[j] (i = 1, 2, . . . , n; j = 1, 2, . . . , oi ) in the order
of operations of that job.

FOR j =1 TO oi − 1;
FOR hole 1 TO hole ki,[j] ;
IF the starting time of the hole is less than ECTi,[j] and the end time of
the hole is larger than ECTi,[j] , then ECTi,[j] = the end time of the hole;
END FOR;
Let ECTi,[j+1] = ECTi,[j] + pi,[j+1] ;
END FOR;
OUTPUT ECTi,[oi ] .
Step 2: LB2 = maxi=1,...,n {ECTi,[oi ] }.

3.3. A numerical example

The proposed LBLY is illustrated by applying it to the data in Table 1. To calculate LB1 , a matrix ESTO as follows is obtained
using Equations (1) and (2).
⎡ ⎤
0 130 270 370 490
⎢0 60 110 160 290⎥
⎢ ⎥
ESTO = ⎢ ⎢0 90 150 260 340⎥
⎥
⎣0 150 230 290 430⎦
0 150 270 370 510
Rearrangement each row of ESTo according to the processed machines from M1 to M5 , a matrix ESTM is formed as
follows.
⎡ ⎤
0 490 130 270 370
⎢160 110 290 60 0 ⎥
⎢ ⎥
ESTM = ⎢ ⎢ 0 260 340 90 150 ⎥
⎥
⎣150 0 230 290 430⎦
510 0 270 370 150
Select the minimum value of each column in ESTM , we can get ESTM1 = 0, ESTM2 = 0, ESTM3 = 130, ESTM4 = 60, and
ESTM5 = 0. Then, according to Step 4, H1 = (305 − 206) = 99, H2 = (591 − 482) = 109, H3 = (428 − 311) + (752 −
635) = 234, H4 = (256 − 159) + (672 − 575) = 194, and H5 = (416 − 313) = 103 can be calculated. Finally, LB1 is
computed as follows.

LB1 = maxj=1,...,5 {ESTMj + pij + Hj }

i

= maxj=1,...,5 {0 + 500 + 99, 0 + 550 + 109, 130 + 590 + 234, 60 + 490 + 194, 0 + 520 + 103}
= 954
 International Journal of Production Research 4407

According to the equation, ECTi,[1] = pi,[1] (∀i = 1, . . . , n), we can set ECT1,[1] = 130,, ECT2,[1] = 60,, ECT3,[1] = 90,,
ECT4,[1] = 150, and ECT5,[1] = 150. To determine LB2 , we firstly calculate ECTi,[j] (i = 1, 2, . . . , n; j = 1, 2, . . . , oi ) as
follows.
ECT1,[2] = ECT1,[1] + p1,[2] = 130 + 140 = 270,

ECT1,[3] = ECT1,[2] + p1,[3] = 270 + 100 = 370,

ECT1,[4] = ECT1,[3] + p1,[4] = 416 + 120 = 536 (∵ 313 < 370 < 416),

ECT1,[5] = ECT1,[4] + p1,[5] = 591 + 120 = 711 (∵ 482 < 536 < 591).
In the same way, we can get ECT2,[5] = 585, ECT3,[5] = 568, ECT4,[5] = 1095, and ECT5,[5] = 882. Finally, LB2 is
determined as follows.

LB2 = maxi=1,...,5 {ECTi,[oi ] } = max{711, 585, 568, 1095, 882} = 1095.

Since LB3 = 1095 is obtained by solving the disjunctive MILP formulation via Gurobi, LBLY = max{LB1 , LB2 , LB3 } =
max{954, 1095, 1095} = 1095, which is the same as the optimal solution of this example.

4. Proposed MTSA algorithm

This study presents an SA-based meta-heuristic, called multi-temperature simulated annealing (MTSA), to solve the
Jm |hjk |Cmax problem. SA is a very simple but effective meta-heuristic first introduced by Kirkpatrick et al. in 1983. Different
SA-based algorithms have successfully derived approximate global optima in a large search space for many combinatorial
optimisation problems (e.g. Ceschia et al., 2017; Lin et al., 2018; Lin and Ying, 2013, 2015; Lin and Yu, 2017; Shaabani
and Kamalabadi, 2016; Wang et al., 2015; Ying et al., 2011). The traditional SA was simulated from annealing in metal-
lurgy, a technique involving heating and controlled cooling of materials to diminish their defects and increase the size of
their crystals. During the SA searching processes, the temperature is progressively decreased from an initial temperature.
This notion of slow cooling implemented in the SA is interpreted as a slow reduction in the probability of accepting worse
solutions while the solution space is explored. The initial temperature and cooling rate are two significant SA parameters
because they play a crucial role in controlling an extensive search for the global optimum. However, determining how to
obtain suitable initial temperature and cooling rate values remain significant challenges.
The major problem of setting initial temperature and cooling rates is due to the non-robustness of their corresponding
solutions. Unfortunately, a parameter combination of initial temperature and cooling rate that outperforms other parameter
combinations for all of the test instances does not exist. Therefore, this study proposes an MTSA algorithm, which has
an embedded multi-temperature mechanism to improve the solution quality. The following subsections further discuss the
solution representation, neighbourhood, and the detailed MTSA algorithm procedure for solving the Jm |hjk |Cmax problem.

4.1. Solution representation


In this study, a solution is coded to a list of ordered operations that contains ni=1 oi elements, where oi is the operation
number of job Ji (i = 1, 2, . . . , n). Each job appears in the list exactly oi times. Based on the list, the solution is decoded by a
one-pass heuristic (Cheng et al., 1996) as follows. If a number i appears on the list for the first time, the first operation of the
job Ji is carried out on its corresponding machine. The next time number i appears on the list; the second operation of the
job Ji is carried out on its corresponding machine. When scheduling an operation of a certain job, the earliest possible start
time of the operation is the larger of the complete time of its previous operation of the same job and the complete time of
its previous operation on the processed machine. If operation of a certain job is carried out at the earliest possible start time
and its processing time will overlap an unavailability period of the corresponding machine, the earliest possible start time
of the operation will be set to the end time of that unavailability period. It will then be checked again until the operation of
a certain job will not overlap any unavailability periods of the corresponding machine. Once the last operation is scheduled,
it is easy to determine the makespan Cmax () of a given solution .
The solution representation and the makespan calculation are illustrated by applying it to the data in Table 1 as follows.
Suppose a solution is coded to a list of ordered operations 2-5-5-2-1-3-2-3-4-4-1-3-4-5-3-1-1-3-2-5-4-1-4-5-2. Because the
first number in the list is 2, the first operation of the job J2 is processed on machine M5 and 60 units of time are required.
The second number in the list is 5; therefore, the first operation of the job J5 is processed on machine M2 and 150 units
4408 S.-W. Lin and K.-C. Ying

Figure 1. Gantt chart of an example solution.

of time are required. Because the third number in the list is 5, the second operation of the job J5 is processed on machine
M5 and 120 units of time are required. Other numbers in the list are decoded in the same way. Figure 1 shows the Gantt
chart of this solution. It can be seen that when scheduling the fourth operation of job J3 , its possible earliest start time is
the maximum between 575 (the complete time of its previous operation of J3 ) and 350 (the complete time of its previous
operation on machine M2 ). However, if the forth operation of job J3 is begun at 575, it will overlap the first unavailability
period on machine M2 . Therefore, the starting time of job J3 on machine M2 is set to 591, which is the end time of the first
unavailability period on machine M1 .

4.2. Neighbourhood
Let P = {1 , . . . , Psize } denote the population of current solutions, where q (q = 1, . . . , Psize ) is the current solution and
Psize represents the number of current solutions in P. Let N(q ) (q = 1, . . . , Psize ) denote the sets of solutions that neighbour
the current solution q (q = 1, . . . , Psize ). In this study, N(q ) (q = 1, . . . , Psize ) are generated using the Swap, Insertion,
2-Swap, 2-Insertion or Inversion operators, where the Swap operator randomly selects two operations in q and switches
their positions. The Insertion operator randomly selects two operations in q and inserts the first selected operation in the
position immediately before the second selected operation. The 2-Swap/2-Insertion operator performs the aforementioned
Swap/Insertion operator twice, and the Inversion operator randomly selects a substring of operations in q , and reverses its
order.
For example, suppose solution q is coded to a list of ordered operations: 1-3-3-2-1-5-4-5-3-1-5-4-4-3-2-2-1-5-2-3-5-4-
1. If the swap operator is performed by selecting the seventh and 14th operations in q , N(q ) will become 1-3-3-2-1-5-3-
5-3-1-5-4-4-4-2-2-1-5-2-3-5-4-1. If the insertion operator is performed by selecting the fourth and eighth operations in q ,
N(q ) will become 1-3-3-2-1-5-4-2-5-3-1-5-4-4-3-2-2-1-5-2-3-5-4-1. If the inversion operator is performed by selecting
the fourth and ninth operations in q , N(q ) will become 1-3-3-3-5-4-5-1-2-1-5-4-4-3-2-2-1-5-2-3-5-4-1. Since there are
numerous combinations of these operators, the probability of selecting any one of the operations to generate N(q ) is fixed
at 0.2 in order to simplify the parameter settings.

4.3. MTSA algorithm procedure

There are two phases to the proposed MTSA algorithm procedure: the initialisation phase and the iteration phase.
Figure 2 shows the pseudo-code of the proposed MTSA algorithm. In the initialisation phase, the initial population
P = {1 , . . . , Psize } is generated randomly in Step 1, and the best solution among P is recorded as the incumbent best
solution best . Then, in Steps 2 and 3, the number of unimproved solutions for each solution, NUSq (q = 1, . . . , Psize ), and
the current temperature for each solution, Tq (q = 1, . . . , Psize ), are set to NUSq = 0 and Tq = q × T0 /Psize (q = 1, . . . , Psize ),
respectively, where T0 is the initial temperature parameter.
In the iteration phase, a solution q is randomly selected from the population P using a roulette wheel selec-
tion method in Step 4. Then, in Step 5, a neighbourhood solution q ∈ N(q ) is generated by applying the Swap,
Insertion, 2-Swap, 2-Insertion, or Inversion operators with a selection probability of 20% for each operator. Let q =
(Cmax (q ) − Cmax (q ))/Cmax (q ), where Cmax (q ) and Cmax (q ) denote the makespan of q and q , respectively. In
Step 6, if q ≤ 0, the solution q is replaced by q ; otherwise, the number of unimproved solutions NUSq is increased
by 1. The probability of replacing q with q , given that q > 0, is exp(−q /Tq ). This replacement is performed by
generating a random number r ∈ U(0, 1), and replacing solution q with q if r < exp(−q /Tq ). In Step 7, if q is better
than the incumbent best solution best found so far, it will replace best and NUSq will be reset to zero, and all solutions in
population P will simultaneously be set to best . Finally, in Step 8, Tq is reduced after Iiter successive iterations in which q
is worse than q from the last reduction, using the formula Tq =: Tq × α, where 0 < α < 1. If the running time exceeds the
 International Journal of Production Research 4409

Figure 2. Pseudo-code of the proposed MTSA algorithm.

maximal time allowed, tmax , the algorithm is terminated. Upon the termination of the MTSA procedure, the (near-) global
optimal schedule will be obtained.

5. Results and discussion

This section evaluates the performance of the proposed MTSA algorithm on two benchmark problem sets. The computa-
tional results obtained by the proposed MTSA algorithm are compared with those obtained by the Mati’s TT meta-heuristic
(Mati, 2010), Tamssaouet et al.’s SA and TS meta-heuristics (Tamssaouet et al., 2018), and Azem et al.’s disjunctive MILP
(Azem et al., 2007). The MILP model was resolved via Gurobi version 8.1, and the proposed MTSA algorithm was coded
using C + + . Both Gurobi and MTSA were running on a personal computer (PC) with an Intel® Core™ i7-920 2.67-
GHz CPU, and 20 GB RAM. The computational results obtained by TT, SA and the TS meta-heuristics were cited from
Tamssaouet et al. (2018), who implemented these meta-heuristics in C + + running on a PC with a 1.90-GHz CPU, and 8
GB RAM. The maximal computational times of TT, SA and the TS meta-heuristics were set to 60 s for solving each prob-
lem instance. In this study, the maximal computational time of the proposed MTSA meta-heuristic was set to 30 s, while
the maximal computational time of the MILP model was set to 1800 s. Since the computing times varied with hardware,
software and programming skill, a completely fair comparison of the compared algorithms’ efficiencies is not viable. The
following subsections discuss the test problems, parameter calibration, and computational results.

5.1. Test problems

To evaluate the proposed MAST algorithm, the benchmark problem sets presented by Azem et al. (2007) and Mati (2010)
were used. Azem et al.’s benchmark problem set consists of four classes of problem instances. The number of jobs and
number of machines, (n, m), of each class is (5, 5), (5, 10), (10, 10) and (10, 15), in which each class has five problem
instances. Therefore, there are a total of 5 × 4 = 20 problem instances in the Azem et al.’s benchmark problem set.
4410 S.-W. Lin and K.-C. Ying

Table 2. Orthogonal array and the obtained ARPDs.

Experiment no. Psize T0 Iiter α tmax ARPD
1 3 0.050 250nma 0.93 20 1.5509
2 3 0.075 500nm 0.95 30 1.4160
3 3 0.100 750nm 0.97 35 1.3867
4 3 0.125 1000nm 0.99 40 1.1328
5 5 0.050 500nm 0.97 40 1.2955
6 5 0.075 250nm 0.99 35 1.2545
7 5 0.100 1000nm 0.93 30 1.3508
8 5 0.125 750nm 0.95 20 1.6270
9 7 0.050 750nm 0.99 30 1.2830
10 7 0.075 1000nm 0.97 20 1.5634
11 7 0.100 250nm 0.95 40 1.3113
12 7 0.125 500nm 0.93 35 1.0972
13 9 0.050 1000nm 0.95 35 1.5197
14 9 0.075 750nm 0.93 40 1.5205
15 9 0.100 500nm 0.99 20 1.8132
16 9 0.125 250nm 0.97 30 1.4768
an and m denote the number of jobs and the number of machines,
respectively.

Mati’s benchmark problem set (2010) is generalised based on Lawrence’s 20 medium-sized problem instances
(Lawrence, 1984), which also consist of four classes of problem instances. The number of jobs and number of machines,
(n, m), of each class is (15, 10), (20, 10), (30, 10) and (15, 15), in which each class has five problem instances. Mati
extended these instances by adding different numbers of unavailability periods, which yielded three subsets of problem
instances: ldata, mdata and hdata. The average number of unavailability periods for each machine is less than two, equal to
two, and between two and three in the ldata, mdata and hdata problem instances, respectively. Therefore, there are a total
of 20 × 3 = 60 problem instances in Mati’s benchmark problem set.

5.2. Parameter calibration

The five key parameters of the proposed MTSA algorithm are Psize , T0 , Iiter , α, and tmax , where Psize denotes the popula-
tion size; T0 represents the initial temperature; Iiter denotes the maximum number of consecutive iterations in which q
is worse than q from the last temperature reduction; α represents the cooling rate; and tmax is the allowed maximum
computational time. Since parameter values may influence the quality of the computational results, it is important to set
them carefully. As shown in Table 2, each of the five parameters had four levels to be calibrated; hence, there were 45
possible combinations of parameter values. The four levels for each of the five parameters were initially set by referring
to several related studies (Lin, 2013; Lin and Ying, 2013, 2015; Lin and Yu, 2017), and were then fine-tuned via a pilot
experiment conducted on six randomly generated test instances. To more efficiently determine the best combination of the
five parameter values, Type B of the Taguchi L16 orthogonal experiment design, which involves 16 parameter combinations
(see https://www.york.ac.uk/depts/maths/tables/l16b.htm), was applied to 6 additional randomly generated test instances. In
reference to related studies (Lin, 2013; Lin and Ying, 2013, 2015; Lin and Yu, 2017), for each parameter combination, the
proposed MTSA algorithm was independently executed 20 times for each of the 6 generated test instances. The average
relative percentage deviation (ARPD) from the best solution of the six generated instances is given in Tables 2 and 3 as the
response variable, in which the relative percentage deviation (RPD) was computed as follows:

RPDi = (Cmax
i
− Cmax
best
)/Cmax
best
× 100%
i
where Cmax denotes the average makespan of a specific problem instance i (i = 1, . . . , 6) obtained in 20 runs using the
best
proposed MTSA algorithm with a specific parameter combination and Cmax is the minimum makespan of a specific problem
instance obtained in 20 × 16 = 320 runs using the proposed MTSA algorithm with all parameter combinations.
Table 3 presents the ARPDs obtained by different levels of each parameter. As shown in Table 3, the most significant
parameter among the five parameters was tmax , which had the largest ARPD range. That is, when the allowed maximum
computation time, tmax , was increased, better solutions could be obtained at the expense of requiring higher computational
time. The second significant parameter among the five parameters was the population size, Psize , which indicated that the
multi-temperature mechanism would vary the thermal transition probabilities to help prevent the algorithm from becoming
 International Journal of Production Research 4411

Table 3. ARPDs obtained by different levels of each parameter.

Level Psize T0 Iiter α tmax
1 1.3716 1.4123 1.3984 1.3799 1.6386
2 1.3819 1.4386 1.4055 1.4685 1.3816
3 1.3137 1.4655 1.4543 1.4306 1.3145
4 1.5825 1.3334 1.3916 1.3709 1.3150
Range 0.2688 0.1321 0.0626 0.0976 0.3241
Rank 2 3 5 4 1

trapped in a local minimum and would ensure its movement towards better solutions. It was difficult to assess the effect of
Psize , T0 , Iiter , and α separately. In general, if the initial temperature T0 was too high, more computational time would be
needed to obtain better solutions. On the other hand, if T0 was too low, the proposed approaches would tend to converge
prematurely because of the low probability of acceptance. According to the experimental results, a better solution could
be obtained by increasing the computing time using the proposed MTSA algorithm. To establish an equilibrium between
solution quality and computational time in this study, the parameter values of Psize , T0 , Iiter , α and tmax were set to 5, 0.125,
1000nm, 0.99 and 30 s, respectively.

5.3. Computational results and discussion

As the TT and TS meta-heuristics were compared by Tamssaouet et al. (2018), the proposed MTSA algorithm was executed
five times on each problem instance with a different random seed. The best makespan value of each problem instance among
five runs was then recorded for comparison. To verify the performance of the proposed MTSA algorithm, the best makespan
values of each problem instance obtained by MTSA, TT, TS, and MILP among five runs were compared using the relative
percentage deviation (RPD) as follows:

RPDh = (Cmax
h
− Cmax
BKS
)/Cmax
BKS
× 100%
h BKS
where Cmax and Cmax denote the best makespan value obtained by a specific compared algorithm h among five runs and the
makespan value of the best-known solution (BKS) obtained by all compared algorithms, respectively.
h
Table 4 presents the best makespan value Cmax and the RPDs of each problem instance of Azem et al.’s benchmark prob-
lem set (2007) that obtained by the five compared approaches. The computational times of MILP are listed in the second-last
column, and the LB values are listed in the last column. From Table 4, MTSA obtained 19 out of 20 (19/20 = 95%) BKSs,
while TT, SA, TS and MILP obtained 13 (13/20 = 65%), 13 (13/20 = 65%), 14 (14/20 = 70%) and 15 (15/20 = 75%)
out of 20 BKSs, respectively. Furthermore, the total average RPD of the 20 problem instances obtained using the proposed
MTSA algorithm was 0.109%, while the corresponding values obtained using TT, SA, TS, and MILP were 0.752%, 0.752%,
0.348%, and 0.454%, respectively. These analytical results revealed that the proposed MTSA algorithm outperformed the
other four state-of-the-art algorithms when solving Azem et al.’s benchmark problem set. Note that there were 15 out of 20
LB values that were the same as the makespans of the BKSs/optimums, and the average deviation between LB values using
the makespan values of the BKSs were very small. These experimental results showed that the proposed LB approach was
generally satisfactory and promising for solving Azem et al.’s benchmark problem set.
h
Table 5 presents the best makespan value Cmax and the RPDs of each problem instance of the Mati’s benchmark problem
set (2010) that obtained by the five compared approaches. The LB values are listed in the last column. For the 20 problem
instances of the ldata subset, MTSA, TT, TS, and MILP obtained 13 (13/20 = 65%), 1 (1/20 = 5%), 1 (1/20 = 5%), and
8 (8/20 = 40%) out of 20 BKSs, respectively. For the 20 problem instances of the mdata subset, MTSA, TT, TS and
MILP obtained 12 (12/20 = 60%), 3 (3/20 = 15%), 0 (0/20 = 0%), and 5 (5/20 = 25%) out of 20 BKSs, respectively. For
the 20 problem instances of the hdata subset, MTSA, TT, TS, and MILP obtained 12 (12/20 = 60%), 4 (4/20 = 20%),
2 (2/20 = 10%), and 5 (5/20 = 25%) out of 20 BKSs, respectively. The proposed MTSA algorithm obtained more BKSs
than the other three compared algorithms. Furthermore, the total average RPD of the 60 test instances in Mati’s benchmark
problem set obtained using MTSA was 0.458%, whereas the corresponding values obtained using TT, TS and MILP were
14.914%, 1.967%, and 2.803%, respectively. It was thus clear that the proposed MTSA algorithm could yield better solutions
than the other best algorithms when solving Mati’s benchmark problem set. Note that for some test instances, the MILP
formulation may not find the optimal solution within 1800 s; however, the lower bound value can be validated to be optimal
if it is the same as the best solution obtained by other approaches. As revealed in Table 5, the best solutions obtained by the
MSTA algorithm were very close to the lower bound values. These results confirmed the effectiveness and robustness of the
proposed lower bound approach and MTSA algorithm for solving the Jm |hjk |Cmax problem.
4412 S.-W. Lin and K.-C. Ying

Table 4. Computational results obtained for the Azem et al.’s benchmark problem set.
MTSA TT SA TS MILP LB
Instance h
Cmax RPD h
Cmax RPD h
Cmax RPD h
Cmax RPD h
Cmax RPD Time

5m_5n_1 894a 0.000 894 0.000 894 0.000 894 0.000 894 0.000 0.10 894
5m_5n_2 1095 0.000 1095 0.000 1095 0.000 1095 0.000 1095 0.000 0.08 1095
5m_5n_3 1069 0.000 1069 0.000 1069 0.000 1069 0.000 1069 0.000 0.11 1069
5m_5n_4 1146 0.000 1146 0.000 1146 0.000 1146 0.000 1146 0.000 0.23 1146
5m_5n_5 1201 0.000 1201 0.000 1201 0.000 1201 0.000 1201 0.000 0.12 1201
5m_10n_1 1360 0.000 1360 0.000 1360 0.000 1360 0.000 1360 0.000 1.41 1360
5m_10n_2 1454 0.000 1454 0.000 1454 0.000 1454 0.000 1454 0.000 12.65 1454
5m_10n_3 1606 0.000 1606 0.000 1606 0.000 1606 0.000 1606 0.000 37.63 1606
5m_10n_4 1506 0.000 1536 1.992 1506 0.000 1506 0.000 1506 0.000 5.68 1506
5m_10n_5 1414 0.000 1414 0.000 1414 0.000 1414 0.000 1414 0.000 3.03 1414
10m_10n_1 1947 0.000 1947 0.000 1947 0.000 1947 0.000 1947 0.000 15.39 1947
10m_10n_2 1916 0.000 1916 0.000 1916 0.000 1916 0.000 1916 0.000 4.35 1916
10m_10n_3 1874 0.000 1974 5.336 1874 0.000 1874 0.000 1874 0.000 7.76 1874
10m_10n_4 1771 0.000 1771 0.000 1786 0.847 1786 0.847 1771 0.000 7.07 1771
10m_10n_5 1974 2.174 1932 0.000 1960 1.449 1932 0.000 1932 0.000 14.02 1932
10m_15n_1 2062 0.000 2118 2.716 2120 2.813 2106 2.134 2088 1.261 1800 1989
10m_15n_2 2232 0.000 2259 1.210 2284 2.330 2252 0.896 2254 0.986 1800 2132
10m_15n_3 2180 0.000 2182 0.092 2212 1.468 2182 0.092 2182 0.092 1800 2014
10m_15n_4 2082 0.000 2119 1.777 2131 2.354 2117 1.681 2131 2.354 1800 1903
10m_15n_5 2143 0.000 2184 1.913 2224 3.780 2171 1.307 2237 4.386 1800 1992
Ave RPD 0.109 0.752 0.752 0.348 0.454
a Bold font denotes the solution is optimum.

Table 5. Computational results obtained for the Mati’s benchmark problem set.
ldata mdata hdata
Inst n m MTSA TT TS MILP LB MTSA TT TS MILP LB MTSA TT TS MILP LB
La21 15 10 1202 1296 1205 1200 1127 1182 1182 1185 1176 1078 1311 1296 1320 1296 1296
La22 15 10 1069a 1206 1084 1069 1069 1066 1148 1082 1082 1017 1141 1127 1162 1127 1127
La23 15 10 1177 1184 1182 1177 1168 1202 1131 1202 1218 1168 1253 1257 1250 1256 1236
La24 15 10 1064 1206 1072 1060 1027 1119 1270 1118 1091 1091 1172 1319 1172 1172 1172
La25 15 10 1091 1402 1108 1082 1082 1114 1277 1115 1114 1114 1214 1201 1214 1201 1201
La26 20 10 1318 1289b 1334 1346 1301 1346 1457 1348 1386 1338 1384 1334 1427 1417 1360
La27 20 10 1367 1391 1379 1484 1306 1374 1615 1391 1452 1306 1409 1462 1446 1480 1365
La28 20 10 1365 1483 1391 1400 1276 1376 1534 1389 1416 1336 1393 1733 1416 1448 1336
La29 20 10 1281 1696 1330 1327 1160 1303 1844 1324 1338 1215 1391 1577 1430 1413 1202
La30 20 10 1423 1630 1466 1464 1422 1515 1828 1519 1542 1489 1522 1851 1543 1561 1489
La31 30 10 1843 2159 1843 1905 1843 1902 2235 1946 1954 1902 1902 2141 1927 2075 1902
La32 30 10 1972 2300 2059 2125 1972 1973 2218 2030 2012 1972 2035 2704 2068 2112 2033
La33 30 10 1776 2642 1816 1952 1776 1776 2447 1897 1920 1776 1835 2041 1924 2064 1833
La34 30 10 1835 2085 1853 1985 1835 1835 2308 1917 2035 1835 1896 2291 1973 1990 1892
La35 30 10 2012 2352 2033 2180 2012 2013 2652 2018 2112 2012 2076 2322 2110 2090 2074
La36 15 15 1396 1657 1431 1388 1287 1485 1533 1490 1457 1442 1466 1615 1465 1453 1393
La37 15 15 1605 1672 1586 1558 1401 1552 1481 1557 1552 1436 1697 1960 1718 1701 1518
La38 15 15 1338 1804 1357 1359 1231 1358 1518 1374 1333 1333 1446 1716 1475 1418 1418
La39 15 15 1382 1492 1415 1381 1312 1418 1360 1446 1361 1361 1472 1479 1491 1472 1383
La40 15 15 1225 1620 1246 1266 1055 1225 1229 1233 1269 1055 1225 1229 1233 1261 1055
Ave RPD 0.252 17.735 1.790 3.109 0.792 15.741 2.157 2.793 0.329 11.267 1.955 2.508
a Bold font denotes the solution is optimum.
b Bold font with underline denotes the solution is incorrect, because its makespan value is less than LB.
 International Journal of Production Research 4413

Table 6. Paired t-tests in terms of the ARPD using Azem et al.’s

benchmark problem set.
MTSA vs. TT SA TS MILP
Difference 0.6431 0.6433 0.2391 0.3452
Degree of freedom 19 19 19 19
t-value − 1.8779 − 2.3579 − 1.2371 − 1.2350
One-tail significant 0.0379 0.0146 0.1156 0.1159

Table 7. Paired t-tests in terms of the ARPD on the

benchmark problem set of Mati (2010).
MTSA vs. TT TS MILP
Difference 13.2430 1.5123 2.3493
Degree of freedom 59 59 59
t-value − 8.2331 − 8.0331 − 4.9887
One-tail significant 0.0000 0.0000 0.0000

To confirm whether the MTSA algorithm significantly outperformed the TT, SA, TS, and MILP, a group of one-sided
paired t-tests in terms of the ARPD was conducted. As shown in Tables 6 and 7, at confidence level α = 0.05, the statistical
results demonstrated significant differences between MTSA and the TT and SA for solving Azem’s benchmark problem set,
and between MTSA and the TT, TS, and MILP for solving Mati’s benchmark problem set. The proposed MTSA algorithm
thus significantly outperformed the TT, SA, TS and MILP algorithms for solving the Jm |hjk |Cmax problem because the
proposed MTSA meta-heuristic uses an embedded multi-temperature mechanism, which can increase the perturbation of
solutions by varying the thermal transition probabilities. This innovative mechanism enables the MTSA algorithm to escape
local optima and to converge quickly toward the global optimum.

6. Conclusion and recommendations for future Research

This paper investigated the Jm |hjk |Cmax problem, which is found in many real-world situations. Because this problem is
strongly NP-hard, an MTSA meta-heuristic algorithm with the advantages of simplicity, rapidity, understandability, and
insensitivity to problem size was proposed to solve it. The MTSA algorithm included an embedded multi-temperature
mechanism to fluctuate the thermal transition probabilities, which enables the algorithm to escape local minima and bet-
ter move towards optimal solutions. To evaluate the solution quality of all compared algorithms, a new lower bound was
presented for the Jm |hjk |Cmax problem. The experimental and statistical results obtained in this study showed that MTSA
outperformed the state-of-the-art meta-heuristic algorithms in the literature, and confirmed the effectiveness and robustness
of the proposed lower bound. Due to the universality and extreme complexity of the Jm |hjk |Cmax problem, this work elim-
inates the gap between practical applications and scheduling theory by providing an effective and efficient algorithm for
making realistic scheduling decisions.
Since the JSSP with machine availability constraints is the least studied topic compared to other scheduling problems,
some interesting avenues are worthy of future research. First, more effective and efficient meta-heuristics for solving the
Jm |hjk |Cmax problem provides a venue for future research. Second, extensions of the JSSP with machine availability con-
straints involving different performance criteria and multi-criteria are promising topics of future study. Third, an online
version of the corresponding problem would be of significant value in real-world industry applications. Fourth, an extension
of the proposed algorithm to the general flexible JSSP with machine availability constraints would also be an interest-
ing avenue for future study. Finally, the JSSP with preemptive, crossable or non-crossable unavailability periods is more
complex, but certainly a worthy topic of further study.

Acknowledgments
The work of the first author was supported in part by the Ministry of Science and Technology of the Republic of China, Taiwan, under
Grant MOST107-2410-H-182-005-MY2, and in part by the Linkou Chang Gung Memorial Hospital under Grant CMRPD3G0011. The
work of the corresponding author was supported in part by the Ministry of Science and Technology of the Republic of China, Taiwan,
under Grant MOST 108-2221-E-027-025.
4414 S.-W. Lin and K.-C. Ying

Disclosure statement
No potential conflict of interest was reported by the author(s).

Funding
The work of the first author was supported in part by the Ministry of Science and Technology of Taiwan [grant number MOST107-
2410-H-182-005-MY2], and in part by the Chang Gung Memorial Hospital, Linkou [grant number CMRPD3G0011]. The work of the
corresponding author was supported in part by the Ministry of Science and Technology of Taiwan [grant number MOST 108-2221-E-
027-025].

ORCID
Kuo-Ching Ying http://orcid.org/0000-0002-9549-5290

References
Aggoune, R. (2002). Ordonnancement D’Ateliers Sous Contraintes de Disponibilite’ des Machines. Ph.D. Thesis, Universite’ de Metz,
France.
Aggoune, R., Y. Mati, and S. Dauzère-Pérès. 2009. “A Reduced-Complexity Algorithm for Two-Job Shop Scheduling Problems with
Availability Constraints.” IFAC Proceedings Volumes 42 (4): 1190–1195.
Azem, S., R. Aggoune, and S. Dauzère-Pérès. 2007. “Disjunctive and Time-Indexed Formulations for Non-preemptive Job Shop
Scheduling with Resource Availability Constraints. IEEE International Conference on Industrial Engineering and Engineering
Management.” Industrial Engineering and Engineering Management 2007: 787–791.
Azem, S., R. Aggoune, and S. Dauzère-Pérès. 2012. “Heuristics for Job Shop Scheduling with Limited Machine Availability.” IFAC
Proceedings Volumes 45 (6): 1395–1400.
Benttaleb, M., F. Hnaien, and F. Yalaoui. 2016. “Two-machine Job Shop Problem for Makespan Minimization Under Availability
Constraint.” IFAC-PapersOnLine 49 (28): 132–137.
Benttaleb, M., F. Hnaien, and F. Yalaoui. 2018a. “Two-machine Job Shop Problem Under Availability Constraints on One Machine:
Makespan Minimization.” Computers & Industrial Engineering 117 (February): 138–151.
Benttaleb, M., F. Hnaien, and F. Yalaoui. 2018b, in press. “Minimising the Makespan in the Two-Machine Job Shop Problem Under
Availability Constraints.” International Journal of Production Research.
Ceschia, S., L. Di Gaspero, and A. Schaerf. 2017. “Solving Discrete Lot-sizing and Scheduling by Simulated Annealing and Mixed
Integer Programming.” Computers & Industrial Engineering 114 (December): 235–243.
Cheng, R., M. Gen, and Y. Tsujimura. 1996. “A Tutorial Survey of Job-shop Scheduling Problems Using Genetic Algorithms – I.
Representation.” Computers & Industrial Engineering 30 (4): 983–997.
Kirkpatrick, S., C. D. Gelatt, and M. P. Vecchi. 1983. “Optimization by Simulated Annealing.” Science 220 (4598): 671–680.
Lawler, E., J. Lenstra, A. Rinnooy Kan, and D. Shmoys. 1993. “Sequencing and Scheduling: Algorithms and Complexity.” Handbook in
Operations Research and Management Science 4: 445–522.
"
Lawrence, S. (1984). Supplement to Resource Constrained Project Scheduling: An Experimental Investigation of Heuristic Scheduling
"
Techniques.GSIA, Carnegie Mellon University, Pittsburgh, PA, 4(7), 4411–4417.
Lin, S. W. 2013. “Solving the Team Orienteering Problem Using Effective Multi-Start Simulated Annealing.” Applied Soft Computing 13
(2): 1064–1073.
Lin, S. W., C. Y. Huang, K. C. Ying, and D. L. Chen. 2018. “Decreasing the System Testing Makespan in a Computer Manufacturing
Company.” IEEE Access 6 (March): 16464–16473.
Lin, S. W., and K. C. Ying. 2013. “Minimizing Makespan and Total Flowtime in Permutation Flowshops by a Bi-objective Multi-Start
Simulated-annealing Algorithm.” Computers & Operations Research 40 (6): 1625–1647.
Lin, S. W., and K. C. Ying. 2015. “A Multi-point Simulated Annealing Heuristic for Solving Multiple-objective Unrelated Parallel
Machine Scheduling Problem.” International Journal of Production Research 53 (4): 1065–1076.
Lin, S. W., K. C. Ying, C. C. Lu, and J. N. D. Gupta. 2011. “Applying Multi-start Simulated Annealing to Schedule a Flowline
Manufacturing Cell with Sequence Dependent Family Setup Times.” International Journal of Production Economics 130 (2):
246–254.
Lin, S. W., and V. F. Yu. 2017. “Solving the Team Orienteering Problem with Time Windows and Mandatory Visits by Multi-start
Simulated Annealing.” Computers & Industrial Engineering 114 (December): 195–205.
Ma, Y., C. Chu, and C. Zuo. 2010. “A Survey of Scheduling with Deterministic Machine Availability Constraints.” Computers &
Industrial Engineering 58 (2): 199–211.
Mati, Y. 2010. “Minimizing the Makespan in the Non-preemptive Job-shop Scheduling with Limited Machine Availability.” Computers
& Industrial Engineering 59 (4): 537–543.
Mauguière, P., J. C. Billaut, and J. L. Bouquard. 2005. “New Single Machine and Job-shop Scheduling Problems with Availability
Constraints.” Journal of Scheduling 8 (3): 211–231.
 International Journal of Production Research 4415

Mauguière, P., J. L. Bouquard, and J. C. Billaut. 2003. “A Branch and Bound Algorithm for a Job Shop Scheduling Problem with
Availability Constraints. In Proceedings of the Sixth Workshop on Models and Algorithms for Planning and Scheduling Problems.”
Models and Algorithms for Planning and Scheduling 2003: 147–148.
Ploydanai, K., and A. Mungwattana. 2010. “Algorithm for Solving Job Shop Scheduling Problem Based on Machine Availability
Constraint.” International Journal on Computer Science and Engineering 2 (5): 1919–1925.
Saidy, H. R. D., and M. T. Taghavi-Fard. 2008. “Study of Scheduling Problems with Machine Availability Constraint.” Journal of
Industrial and Systems Engineering 1 (4): 360–383.
Schmidt, G. 2000. “Scheduling with Limited Machine Availability.” European Journal of Operational Research 121 (1): 1–15.
Shaabani, H., and I. N. Kamalabadi. 2016. “An Efficient Population-based Simulated Annealing Algorithm for the
Multi-product Multi-retailer Perishable Inventory Routing Problem.” Computers & Industrial Engineering 99 (September):
189–201.
Tamssaouet, K., S. Dauzère-Pérès, and C. Yugma. 2018. “Metaheuristics for the Job-shop Scheduling Problem with Machine Availability
Constraints.” Computers & Industrial Engineering 125 (November): 1–8.
Wang, C., D. Mu, F. Zhao, and J. W. Sutherland. 2015. “A Parallel Simulated Annealing Method for the Vehicle Routing Problem with
Simultaneous Pickup–Delivery and Time Windows.” Computers & Industrial Engineering 83 (May): 111–122.
Ying, K. C., S. W. Li, and C. C. Lu. 2011. “Cell Formation Using a Simulated Annealing Algorithm with Variable Neighbourhood.”
European Journal of Industrial Engineering 5 (1): 22–42.
Zribi, N., A. El Kamel, and P. Borne. 2008. “Minimizing the Makespan for the mpm Jobshop with Availability Constraints.” International
Journal of Production Economics 112 (1): 151–160.
Copyright of International Journal of Production Research is the property of Taylor & Francis
Ltd and its content may not be copied or emailed to multiple sites or posted to a listserv
without the copyright holder's express written permission. However, users may print,
download, or email articles for individual use.