Int J Adv Manuf Technol

DOI 10.1007/s00170-015-7215-y

ORIGINAL ARTICLE

Scheduling unrelated parallel machines with sequence-dependent

setup times
Javad Rezaeian Zeidi 1 & Samir MohammadHosseini 1

Received: 2 July 2012 / Accepted: 26 April 2015

# Springer-Verlag London 2015

Abstract In this study, the problem of scheduling jobs on distinctive relationship between the processing times of jobs
unrelated parallel machines with sequence-dependent setup on different machines. This case is more realistic than the
times under due-date constraints is considered to minimize other two cases of parallel machine environments named uni-
the total cost of tardiness and earliness. A new mathematical form and identical parallel machines [1].
model is presented for considered problem and due to the The majority of scheduling problems deal with environ-
complexity of the problem; an integrated meta-heuristic algo- ments where the setup time considered negligible or assumed
rithm is designed to solve the problem. The proposed algo- to be a part of job processing times [2]. Such problems are
rithm consists of genetic algorithm as the basic algorithm and modeled with these assumptions that setup times are ignored
simulated annealing method as local search procedure that while they are not very large in comparison with job process-
follows the genetic algorithm to improve the quality of solu- ing times or they are included in processing times when they
tions. The performance of the proposed algorithm is evaluated are independent of job processing sequences. However, a
by solving a set of test problems. The results show that the sequence-dependent setup time is usually incurred in some
proposed integrated algorithm is effective. industrial environments when switching between two differ-
ent jobs [3]. In such cases, setup time as a separate part of
Keywords Unrelated parallel machines . processing time depends on the precedence of jobs. Consider-
Sequence-dependent setup times . Integrated meta-heuristic . ing setup time as a separate part of the processing time allows
Due-date constraints operations to be performed simultaneously, and hence, the
performance of production systems will be improved signifi-
cantly [4].
1 Introduction On the other hand, by increasing the time delivery require-
ments for customers, measures based on tardiness have be-
Unrelated parallel machine scheduling is a generalization of come one of the most active research criteria for different
classical parallel machine environments. The classical parallel scheduling problems [5]. The total tardiness is one of the most
machine scheduling problem consists of a set of independent important factors related to job due-dates. When a job is not
jobs that each of them has to be processed on one of any completed by its due-date, certain costs and penalties will be
available identical parallel machines. Generally, machines incurred [6]. Minimization of the total tardiness of orders from
are considered unrelated; the processing time of each job on primary customers is the main requirement of primary cus-
any machine is related to both job and machine types. It means tomers that have strict (or hard) due date constraints [7].
that machine’s characteristics are not identical, and there is no In this study, the problem of scheduling jobs on unrelated
parallel machines with sequence-dependent setup times under
due-date constraints is considered to minimize the total cost of
* Javad Rezaeian Zeidi tardiness and earliness of all jobs. Following the standard
[email protected]
notation of Graham et al. [8], the problem addressed in this
paper can be represented by Rm|sijk , dj|∑ej *Ej +tj *Tj notation.
1
Mazandaran University of Science and Technology, Babol, Iran This problem is NP-hard in the strong sense since the classic
 Int J Adv Manuf Technol

parallel machine problem with sequence-dependent setup weighted earliness and tardiness and developed a mixed inte-
time belongs to the class of NP-hard problems [9]. Due to ger programming formulation. Suresh and Chaudhuri [21] and
the complexity of this problem, the exact methods may not Adamopoulos and Pappis [22] suggested several new heuris-
be able to find optimal solutions for large-sized problems of tics for the problem with various due-date and processing time
this type within reasonable computational times. Here, a math- combinations to minimize maximum lateness. Kim et al. [23]
ematical model is presented for considered problem and an developed a simulated annealing (SA) method to minimize the
integrated meta-heuristic (IMH) algorithm is designed to solve total tardiness. Lopes and De-Carvalho [24] presented a
the problem. The performance of the algorithm is evaluated branch-and-price algorithm to minimize total weighted tardi-
using a set of large-sized problems. ness. Chen and Chen [25] proposed several hybrid meta-
The rest of this paper is structured as follows: a review on heuristics based on variable neighborhood descent approach
the related literature is given in section 2. Section 3 provides a and TS method to minimize the number of tardy jobs. Their
description of the problem. In section 4, the proposed IMH computational results indicated that the proposed hybrid meta-
algorithm is presented. Computational results are reported in heuristics outperform several basic TS heuristics. Vallada and
section 5, and finally, conclusions are followed in section 6. Ruiz [1] developed a GA for the make-span objective function
which includes fast local search and a local search enhanced
crossover operator.
2 Literature review Recently, Ying and Lin [3] tackled the Rm|sijk , dj|∑Tj prob-
lem with an artificial bee colony (ABC) algorithm and showed
The parallel machine scheduling problem has received a that their proposed algorithm outperforms existing algorithms
great deal of attention since McNaughton’s study [10]. for most problem combinations. As mentioned by Ying and
In the literature, there are several reviews dealing with Lin [3], the problem remains under presented in the research
parallel machine problems like those of Cheng and Sin literature, and their paper is the last study that works on this
[11], Lam and Xing [12], and Mokotoff [13]. The ma- problem preceded by Chen [7] and Lin et al. [26] works. Here,
jority of these studies considered identical parallel ma- an integrated meta-heuristic of GA and SA approaches is pro-
chine cases while unrelated parallel machine problems posed to deal with this problem and the performance of pro-
remained less studied. Moreover, the consideration of posed algorithm is evaluated to show its convergence to near
sequence-dependent setup times between jobs has not optimal solutions.
been considered until recently [1]. For example, Lee
and Pinedo [14] considered the identical parallel ma-
chine problem with sequence-dependent setup time and 3 The problem formulation
proposed a three-phase heuristic method to minimize the
sum of weighted tardiness of the jobs. Park et al. [15] 3.1 Problem description
applied neural network to obtain some look-ahead pa-
rameters, which were used to determine the priority in- In this study, a set of N jobs is considered to be proc-
dex of each job. An integer programming formulation essed on a set of M unrelated parallel machines. Each
was presented by Kurz and Askin [16] for a problem job can be processed only on one out of M machines
with non-zero release dates and minimization of make- and preemption of jobs is not allowed. Each machine
span. Fowler et al. [17] proposed a hybrid genetic algo- can process only one job at the same time. Since ma-
rithm (GA) for a problem with non-zero release date chines are unrelated, the processing time pik of ith job
jobs and several objective functions including make- on kth machine depends on both job and machine type.
span, total weighted completion time, and total weighted A sequence-dependent setup time sijk incurred when ith
tardiness. Kim and Shin [18] proposed a restricted tabu job switches to jth job on kth machine. A due date dj is
search (TS) method to minimize the maximum lateness assigned to job j and it is desirable to finish jobs no
of jobs. Nessah et al. [19] presented a necessary and later than their due-dates. Moreover, the jobs from pri-
sufficient condition for a local optimal solution and pro- mary customers have strict due date constraints and
posed a heuristic method based on this condition to must not be delayed. For a given processing order of
minimize the total completion time. the jobs, the completion time Cj and the tardiness Tj can
The above studies have considered the identical parallel be computed for each job where tardiness is obtained
machines problems with sequence-dependent setup times, by Tj =max (Cj −dj,0). Also, earlier completion time of
but the unrelated parallel machines problem with sequence- a job than its due date lead to early costs such as
dependent setup time has been less studied and only a few holding cost, etc. The objective of the problem is se-
studies can be found in the literature [1]. Zhu and Heady quencing the jobs on all machines that the total cost of
[20] considered this problem with minimizing of total earliness and tardiness of all jobs is minimized.
Int J Adv Manuf Technol

Sequence of jobs on machines

3.2 Notations
Machine 1: 2
The following notations are used throughout the paper: Machine 2: 4-3-6
Indices Machine 3: 5-1

i,j Indices for jobs (i,j=1,2,.., N) Chromosome:

k Index for machines (k=1,2,..,M) 4 5 3 2 1 6
2 3 2 1 3 2
Parameters
Fig. 1 Chromosome encoding
pik Processing time of ith job on kth machine
sijk Setup time of jth job when processed after ith job on C i þ E i −T i ¼ d i i ¼ 1; …; N ð7Þ
kth machine
di Due date of ith job C j −C i −L*Y i jk ≥ p jk þ si jk −L ð8Þ
ei Earliness cost of ith job
ti Tardiness cost of ith job
zik = If ith job can be processed on kth machine; 0, j ≠ i ; i ¼ 0; …; N ;
1 otherwise j ¼ 1; …; M ; k ¼ 1; …; M
X i j ∈ f0; 1g i ¼ 1; …; M ; j ¼ 1; …; N
L A large number
Y i jk ∈f0; 1g i ¼ 0; …; N ; j ¼ 1; …; N ; k ¼ 1; …; M
Decision variables C i ; Ei ; T i ≥0 i ¼ 1; …; N

Xik = If ith job is processed on kth machine; 0, otherwise

1 The objective function (1) represents the total cost of earliness
yijk = If ith job is processed before jth job on kth machine; and tardiness. Eq. (2) ensures that each job will be assigned to
1 0, otherwise one machine only. Eqs. (3) and (4) show that there is a job
Ci Completion time of ith job exactly before and after each job, respectively. Eq. (5) satisfies
Ei Earliness of ith job the initial setup time on every machine. Eq. (6) represents the
Ti Tardiness of ith job machine eligibility restriction. Eq. (7) calculates the earliness and
tardiness of each job. Eq. (8) executes the setup time between the
completion time of predecessor jobs and successor jobs.
3.3 The problem formulation

The objective function and constraints can be formulated as 4 Proposed algorithm

follows.
X In this study, an integrated meta-heuristic (IMH) is presented
min n
j¼1 e j E j þ t jT j ð1Þ which consists of GA as main optimization algorithm and SA

1 3 2 5 4 2 5 1 4 3
Subject to
1 2 1 3 3 3 2 3 1 1
X Parent 1 Parent 2
N
i¼1 X ik ¼1 k ¼ 1; …; M ð2Þ (Stage 1)

X 1 3 2 2 5 1
j¼1 Y i jk ≤ X ik j≠i ; i ¼ 1; …; N ;
N
k 1 2 1 3 2 3

¼ 1; …; M ð3Þ
1 3 2 4 3 2 5 1 5 4
1 2 1 1 2 3 2 3 3 3
X
N
i¼0 Y i jk ¼ X jk j≠i ; j ¼ 1; …; N ; k (Stage 2)

¼ 1; …; M ð4Þ 1 3 2 4 5 2 5 1 3 4

X 1 2 1 1 2 3 2 3 2 3
N
j¼1 Y 0 jk ¼1 k ¼ 1; …; M ð5Þ Off-spring 1 Off-spring 2
(Stage 3)
X ik ≤ zik i ¼ 1; …; N ; j ¼ 1; …; M ð6Þ Fig. 2 Crossover operator
 Int J Adv Manuf Technol

method as local search procedure that follows the GA to im- 4 1 3 5 2

prove the quality of solutions. The IMH algorithm consists of
following steps: 3 1 2 3 1

4.1 Proposed GA

GA is a powerful and extensively applicable stochastic search

4 5 3 1 2
and optimization technique based on evolutionary computa- 3 1 2 3 1
tion concepts. GA begins with a population of chromosomes
which represents a sample of the problem solution and moves Fig. 4 Job swap
from this population to a new one by using a kind of natural
selection together with genetic-inspired operations [27]. A accepted with a penalty cost (Pcost) per unit time of
detailed description of main factors for the proposed GA is tardiness for each delayed job with respect to the pri-
reported as follows: mary customer [3].

Initialization The initial population consists of Pop size Evaluation The fitness value of each chromosome is evalu-
chromosomes of solutions that each chromosome is related ated by relation (9).
to a candidate solution of the problem. A suitable design of
chromosome structure can reduce the computing effort re- 1
f ð hÞ ¼ Xn h ¼ 1; …; Popsize ð9Þ
quired by operations of the algorithm. Generally, a typical 1þ j¼1
T j þ P ð hÞ
chromosome applied in most of the parallel machine schedul-
ing problems is a vector of integer numbers considering some where f(h) is the fitness value of h th chromosome in a popu-
heuristic rules represent the sequence of the jobs based on n
their assigned machines. Here, a two by N array is used to lation, ∑ T j is the total tardiness of all jobs, and P(h) is the
j¼1
represent each chromosome where N is the number of jobs. total penalty cost of tardy jobs from primary customers.
The first row of the array contains a permutation of job num-
bers from one to N, and each element of the second row con- Parent selection Each chromosomes is selected based on it’s
tains the machine number related to the job of first row. For probability proportionated to their fitness value through the
each two jobs that have the same machine number in the roulette wheel method. It is measured by relation (10), where
second row, the one which appears first in the first row of PV(h) shows the selection probability of hth chromosome. In
the array is processed earlier then next job and so on. In this this method, the chromosome with higher fitness value has
manner, each job assigned to a machine and the sequence of more chance to be in the pool of parents for creation of off-
jobs is determined. A typical chromosome is shown in Fig. 1. springs and a chromosome can be selected as a parent one
Considering this fact that sometimes accepting an infeasible more time.
solution may help the algorithm to jump out of a local optimal,
in this study, an infeasible solution can be temporarily f ð hÞ
PV ðhÞ ¼ X Pop h ¼ 1; …; Popsize ð10Þ
size
f ð hÞ
1 2 5 3 4 h¼1

Crossover The main goal of the crossover operation is to

3 1 2 1 1 explicitly combine the chromosomes in order to explore the

Table 1 Efficient ranges

1 2 5 3 4 of control factors Factor Efficient range

3 2 2 1 1 Popsize
Pc
80-60
0.8-0.6
Pm 0.12-0.08
Gmax 90-50
1 5 2 3 4 Ps 0.25-0.15
T0 350-150
3 3 2 1 1 α 0.99-0.95
Iter_max 140-100
Fig. 3 Job reallocation
Int J Adv Manuf Technol

Table 2 Data of experimental design

Run No. Control Factors Responses

Popsize Pc Pm Gmax Ps T0 α Iter_max Quality of solutions CPU (s)

1 60 0.6 0.08 50 0.15 150 0.95 20 1255.52 13.39

2 60 0.6 0.08 50 0.2 220 0.97 25 1236.96 22.61
3 60 0.6 0.08 50 0.25 350 0.99 30 1198.08 27.64
4 60 0.7 0.1 70 0.15 150 0.95 25 1210.04 18.30
5 60 0.7 0.1 70 0.2 220 0.97 30 1273.04 21.56
6 60 0.7 0.1 70 0.25 350 0.99 20 1282.26 21.15
7 60 0.8 0.12 90 0.15 150 0.95 30 1312.26 22.51
8 60 0.8 0.12 90 0.2 220 0.97 20 1273.7 23.08
9 60 0.8 0.12 90 0.25 350 0.99 25 1198.64 28.32
10 70 0.6 0.1 90 0.15 220 0.99 20 1321.53 22.94
11 70 0.6 0.1 90 0.2 350 0.95 25 1201.62 30.7
12 70 0.6 0.1 90 0.25 150 0.97 30 1202.46 35.96
13 70 0.7 0.12 50 0.15 220 0.99 25 1266.89 13.96
14 70 0.7 0.12 50 0.2 350 0.95 30 1225.4 25.36
15 70 0.7 0.12 50 0.25 150 0.97 20 1219.58 25.66
16 70 0.8 0.08 70 0.15 220 0.99 30 1194.78 22.82
17 70 0.8 0.08 70 0.2 350 0.95 20 1326.31 23.26
18 70 0.8 0.08 70 0.25 150 0.97 25 1191.5 29.60
19 80 0.6 0.12 70 0.15 350 0.97 20 1202.43 25.74
20 80 0.6 0.12 70 0.2 150 0.99 25 1256.97 31.49
21 80 0.6 0.12 70 0.25 220 0.95 30 1168.3 43.61
22 80 0.7 0.08 90 0.15 350 0.97 25 1243.85 27.91
23 80 0.7 0.08 90 0.2 150 0.99 30 1192.83 34.25
24 80 0.7 0.08 90 0.25 220 0.95 20 1246.57 34.07
25 80 0.8 0.1 50 0.15 350 0.97 30 1246.03 27.97
26 80 0.8 0.1 50 0.2 150 0.99 20 1235.49 25.34
27 80 0.8 0.1 50 0.25 220 0.95 25 1221.38 31.40

search space and to find good solutions. Once a pair of parents chromosome, and therefore, some other jobs are absent.
is selected, these are recombined together through a single The number of repeated jobs, if any, is exactly equal to
point crossover operator with probability of Pc as cross rate the number of absent jobs. Hence, the chromosomes are
(Fig. 2) and two off-springs are produced. The crossover op- modified to obtain feasible off-springs. For this purpose,
erator generates two off-springs out of every pair of parents. each repeated job in the second part is replaced by one of
At first, a position in each chromosome is randomly cho- the absent jobs randomly, and the required machines relat-
sen, and the parts before this position remain constant. ed to each repeated jobs are also selected randomly. A
The parts of chromosomes after this position are replaced typical instance is depicted in Fig. 2 which consists of
together. After this stage, some jobs may occur twice in a three machines and five jobs.

Table 3 Analysis of variance for means Table 4 Analysis of variance for S/N ratios

Level Popsize Pc Pm Gmax Ps T0 α Iter_max Level Popsize Pc Pm Gmax Ps T0 α Iter_max

1 25.5 28.57 27.52 26.57 25.69 27.56 28.04 26.6 1 27.86 28.65 28.48 28.25 27.93 28.42 28.46 28.24
2 27.64 27.17 27.63 27.66 27.81 28.05 27.79 27.73 2 28.67 28.47 28.62 28.54 28.75 28.65 28.76 28.68
3 30.19 27.58 28.17 29.09 29.82 27.71 27.49 28.99 3 29.32 28.73 28.75 29.07 29.17 28.78 28.64 28.93
Delta 4.69 1.4 0.65 2.52 4.14 0.48 0.54 2.38 Delta 1.46 0.25 0.28 0.82 1.24 0.36 0.3 0.69
Rank 1 5 6 3 2 8 7 4 Rank 1 8 7 3 2 5 6 4
 Int J Adv Manuf Technol

Fig. 5 Main effect plot for means

Main Effects Plot (data means) for Means
Pop_size Pc Pm
30.0

27.5

25.0
60 70 80 0.60 0.70 0.80 0.08 0.10 0.12

Mean of Means
G max Ps T0
30.0

27.5

25.0
50 70 90 0.15 0.20 0.25 150 220 350
A lfa Iter_max
30.0

27.5

25.0
0.95 0.97 0.99 20 25 30

Mutation Mutation operation is a simple mechanism to 4.2 Selection

avoid convergence to local optimal solutions and per-
forms a small random perturbation on chromosomes to In this step, a number of solutions from the last generated
diversify the search process. Here, this operator selects population of GA are selected randomly. Therefore, a selec-
a job from the first row of the chromosome and re- tion ratio denoted by Ps is considered here for the number of
places its corresponding machine with another machine solutions. This ratio has a critical impact on optimal solutions
from its processing set randomly based on mutation rate of the algorithm and must be effectively determined (see
(Pm). section 5.2).

Fig. 6 Main effect plot for S/N

ratios Main Effects Plot (data means) for SN ratios
P op_size Pc Pm

29.0

28.5

28.0

60 70 80 0.60 0.70 0.80 0.08 0.10 0.12

Mean of SN ratios

G max Ps T0

29.0

28.5

28.0

50 70 90 0.15 0.20 0.25 150 220 350

A lfa Iter_max

29.0

28.5

28.0

0.95 0.97 0.99 20 25 30

Signal-to-noise: Larger is better

Int J Adv Manuf Technol

Table 5 Analysis of 3.2.4 If ΔE≥0 and Random[0,1)<Exp(−ΔE/Tr) then

variance for S/N ratios Factor Level
set: Xi =Xnew, i=i+1.
Popsize 80 3.3 Set: r=r+1, Tr =α Tr−1.
Pc 0.6 4. Return Xbest.
Pm 0.12
Gmax 90 The elements which characterized this method are de-
Ps 0.25 scribed as follows:
T0 350
α 0.97 Initial solution The SA procedure gets the initial solution
Iter_max 30 from the solutions of last generated population by GA (XGA).

Energy function The energy function is calculated as objec-

4.3 SA integration tive function of problem, i.e., total tardiness of all jobs plus
total penalty cost of tardy jobs from primary customers.
SA is a local search method that is presented by Kirkpatric
et al. [28] as a random search algorithm for combinatorial Neighborhood generation Two methods from Chen [7]
optimization problems. The idea behind the algorithm is based are implemented to generate neighborhood solutions.
on annealing process of metals and glass in statistical mechan- At first, a job from a machine with higher tardiness
ics science [29]. In this step, each initial solution chosen from value is reallocated to an intermediate machine. Then,
the last generated population of the proposed GA is fed to SA a job from this intermediate machine is reallocated to a
method. These solutions lead to different trajectory in the machine with lower tardiness value. Both selections are
search space and different preliminary good solutions. The performed randomly. For example, suppose that job 2
SA method is presented here as following pseudo code: from machine 1 reallocated to the first position in ma-
chine 2. Then, job 5 from machine 2 reallocated to the
1. Get XGA, T0 second position in machine 3 (Fig. 3). In second meth-
2. Set: X0 =XGA, Xbest =XGA, T=T0. od, a job from a machine with higher tardiness value is
3. Repeat the following steps until stop criterion is met: swapped with a job from a machine with lower tardi-
ness value (Fig. 4). These two methods consecutively
3.1 Set: i=0, j=0, r=0. are executed to generate neighbor solutions at each iter-
3.2 Repeat following steps until i>I or j>J: ation of the algorithm.

3.2.1 j=j+1. Cooling ratio The temperature is gradually reduced during

3.2.2 Xnew = generate (Xi) and set: ΔE = E(Xnew) − search process through the relation (11):
E(Xbest).
3.2.3 If ΔE<0 then set: Xbest =Xnew, Xi =Xnew, i=i+1. T r ¼ α T r−1 0 < α < 1 ; r ¼ 1; 2; …: ð11Þ

Table 6 Parameter settings for

Chen’s benchmark problems [7] Parameter Level

Number of jobs (N) 30, 50, 70, 90

Number of machines (M) 4, 6, 8
Number of job families (F) N/7 +1, N/8 +1
Due-date priority factor (τ) 0.4, 0.8
Due-date range factor (R) 0.4, 1
Proportion of jobs from primary customers (P) 0.2, 0.3
Speed factor for jobs of family q on machine m, fmq 1/U [5, 15]
Processing time for job j of family q on machine m, pmj 1/fmq × U [10, 40]
Setup time for a job of family k following a job of family U [10, 100]
q (k≠q) on machine m, smqk
 
Due-date U ½C max 1−τ− R2 ; C max ð1−τ þ R2 

 
N M N
C max ¼ ∑ ∑ pm j þ ∑ smi j =N =M Þ= M
j¼1m¼1 i¼1
 Int J Adv Manuf Technol

Table 7 Comparison between the results of the ABC and IMH algorithms

ABC IMH

Min. sol. Max. sol. Ave. sol. Min. sol. Max. sol. Ave. sol. No. better No. equal No. worse

N 30 1389.2 1402.8 1393.5 1340.8 1560.9 1435.1 77 114 49

50 2903.7 3001.0 2939.7 2666.9 3174.6 2882.6 81 120 39
70 5891.2 6129.2 5992.3 5256.1 6224.5 5619.7 94 122 24
90 9649.3 10054.1 9838.3 8403.9 9618.2 8915.1 95 123 22
M 4 7590.2 7819.7 7693.7 6382.8 7460.3 6808.5 162 156 2
6 4634.8 4826.5 4715.6 4271.8 4981.7 4558.4 117 155 48
8 2650.1 2794.0 2713.6 2596.2 2991.7 2774.5 68 168 84
F N/7 +1 4956.8 5149.3 5042.1 4369.1 5073.8 4661.8 180 240 60
N/8 +1 4959.9 5144.2 5039.8 4464.7 5215.4 4764.4 167 239 74
τ 0.4 1.3 1.6 1.4 0.66 6.9 2.9 6 471 3
0.8 9915.4 10291.9 10080.5 8833.2 10282.2 9423.3 341 8 131
R 0.4 8733.4 8948.9 8828.3 7891.1 8864.0 8268.8 195 240 45
1 1183.3 1344.6 1253.6 942.7 1425.1 1157.4 152 239 89
P 0.2 4799.2 4960.3 4869.9 4279.5 4894.5 4547.5 178 239 63
0.3 5117.5 5333.2 5212.0 4554.3 5394.7 4878.8 169 240 71
Total 4958.4 5146.7 5041.0 4416.9 5144.6 4713.1 347 479 134

where Tr is the current temperature after r th iteration and α is number of machines (M) due to a certain coefficient (β)
a real number between one and zero as the cooling ratio. and determined by the relation (12):

The length of Markov chain The length of Markov chain as I ¼βN M ð12Þ
the number of movement from the current solution to However, the neighbor solutions are accepted with decreas-
the neighbor solution is determined, and so the number ing probability during search process, and this may lead to a
of accepted neighbor solutions at each temperature is a too long length of the Markov chain for small values of tem-
certain parameter [30]. This parameter is denoted by I perature at the end of the search. Preventing this event, the
which is proportional to the number of jobs (N) and the number of neighbor solutions generated at each temperature is

Table 8 Comparison between the results of the SA_HEU and IMH algorithms

Subset SA_HEU IMH No. better No. equal No. worse

Min. sol. Max. sol. Avg. sol. Min. sol. Max. sol. Avg. sol.

N 30 1404.25 1419.19 1412.54 1340.80 1560.99 1435.10 87 113 40

50 2961.08 3004.90 2986.67 2666.86 3174.6 2882.56 92 120 28
70 6040.07 6140.02 6106.60 5256.13 6224.53 5619.73 101 123 16
90 10020.51 10254.30 10106.76 8403.93 9618.23 8915.10 100 123 17
M 4 7861.48 7995.00 7941.49 6382.78 7460.31 6808.47 162 156 2
F 6 4737.57 4835.69 4795.58 4271.81 4981.70 4558.41 123 156 41
8 2720.38 2783.12 2759.10 2596.19 2991.73 2774.49 95 167 58
[N/7]+1 5104.22 5206.68 5166.52 4369.15 5073.19 4661.80 197 240 43
[N/8]+1 5108.74 5202.53 5164.27 4464.7 5215.38 4764.45 183 239 58
τ 0.4 1.30 1.52 1.41 0.66 1.7 1.3 7 470 3
0.8 10211.66 10407.69 10329.38 8833.19 10282.26 9423.33 373 9 98
R 0.4 8972.78 6100.69 9047.71 7891.12 8864.02 8268.84 213 239 28
1 1240.18 1308.52 1283.08 942.73 1425.14 1157.41 167 240 73
Total 5106.48 5204.60 5165.39 4416.13 5144.58 4713.12 380 479 101
Int J Adv Manuf Technol

limited to a parameter denoted by J which is determined by the 5.3 Computational experiments

relation (13):
The proposed IMH algorithm is coded in MATLAB 7.1 soft-
J ¼γN M ð13Þ
ware, and the computational experiments are carried out on a
Stopping criterion The search process is terminated when the Pentium IV 3.4 MHz computer with 512 MB of RAM mem-
best solution is not improved after Iter_max iterations. ory. All of the 960 test problems are solved by IMH algorithm
ten times. A detailed survey is performed on the computation-
al results regarding the six parameters included: Min. sol, Avg.
sol., Max. sol., No. better, No. equal and No. worse, in which
5 Computational results the first three parameters are the minimum, average, and max-
imum values of ten replications of each test problem in a
5.1 Parameters setting, Taguchi methods (TM) specific subset of the test problem set, respectively, and the
second three parameters are the number of problems that have
The proposed IMH algorithm contains several factors which better, equal, and worse minimum values of ten replications
control the effectiveness and efficiency of the algorithm. The- for each problem than the best solution in a specific subset of
se factors have to be adjusted to the problem in order to get the test problem set. Tables 7 and 8 show the results related to
best solutions of problems. The values of the control factors these parameters.
such as Popsize, Pc , Pm , Gmax , Ps , T0 , α,and Iter_max are According to Tables 7 and 8, the total average of Min. Sol.,
determined by a set of experiments which are designed by Max. Sol., Ave. Sol. of the proposed IMH algorithm is 4416.9,
Taguchi method. For this reason, at first, the efficient range 5144.6, and 4713.1, respectively.
of each parameter is determined using the experience by initial
test problems relatively (Table 1), and then the exact values
are tuned by analyzing the interaction between factors.
6 Conclusions
The designed experiments of control factors are shown in
Table 2 where the quality of solutions and computational time
This paper presented scheduling problem of unrelated parallel
(CPU) are considered as responses. The analysis of variance is
machines to minimize the total cost of tardiness and earliness
done for the signal to noise (S/N) ratio and mean of responses
considering sequence-dependent setup times and due-date
of experiments. Finally, the averages of responses and S/N for
constraints. This problem belongs to the class of NP-hard
factors are summarized in Table 3.
problems which can be applied for industries with parallel
Based on Tables 3 and 4, popsize, ps factors have the first
machines or lines such as automotive industry, packaging in-
and second ranks in both tables. But other factors has different
dustry especially for stations by workers efforts. A new math-
ranks in these tables; hence, the interactions of factors are
ematical model and an integrated meta-heuristic (IMH) ap-
depicted in Figs. 5 and 6. Finally, the optimum levels of fac-
proach is proposed to solve this problem. The performance
tors are listed in Table 5.
of the proposed algorithm is evaluated by comparing its com-
putational results with ABC and SA_HEU algorithms under
5.2 Benchmark of instances the relevant benchmark available in the literature. This model
can be developed for future studies by considering learning
The performance of the proposed IMH algorithm is evaluated effect and deterioration constraints when labor effort and ma-
using a set of test problem. Chen [7] presented a comprehen- chine capabilities affect on the objectives.
sive benchmark of instances for the Rm|sijk , dj|∑Tj problem.
The test problems are generated randomly considering six
parameters involved: number of jobs (N), number of machines
(M), number of job families (F), due-date priority factor (τ),
References
due-date range factor (R), and proportion of jobs from primary
customers (P). Considering this parameters, the input factors
1. Vallada E, Ruiz R (2011) A genetic algorithm for the unrelated
including processing times, setup times, and due-dates are parallel machine scheduling problem with sequence-dependent set-
generated randomly. Table 6 shows the levels of the factors up times. Eur J Oper Res 211:612–622
and the methods used to generate the input parameters. There 2. Allahverdi A, Ng CT, Cheng TCE, Kovalyov MY (2008) A survey
are four and three levels for N and M, respectively, and two of scheduling problems with setup times or costs. Eur J Oper Res
187:985–1032
levels for other four factors. For each combination of the fac- 3. Ying K-C, Li S-W (2012) Unrelated parallel machine scheduling
tors, five random problems are generated which bring out with sequence- and machine-dependent setup times and due date
totally 960 test problems. constraints. Int J Innov Comput Inf Control 8:1–19
 Int J Adv Manuf Technol

4. Allahverdi A, Gupta JND, Aldowaisan T (1999) A review of sched- 18. Kim SS, Shin HJ (2003) Scheduling jobs on parallel machines: a
uling research involving setup considerations. OMEGA Int J restricted tabu search approach. Int J Adv Manuf Technol 22:278–
Manag Sci 27:219–239 287
5. Lin S-W, Chou S-Y, Ying K-C (2007) A sequential exchange ap- 19. Nessah F, Yalaoui F, Chu C (2005) New heuristics for identical
proach for minimizing earliness-tardiness penalties of single- parallel machine scheduling with sequence dependent setup times
machine scheduling with a common due date. Eur J Oper Res and dates. In: Proceedings of the International Conference on
177:1294–1301 Industrial Engineering and Systems Management: 32–41
6. Arroyo JEC, Ottoni RS, Oliveira AP (2011) Multi-objective vari- 20. Zhu Z, Heady RB (2000) Minimizing the sum of earliness-tardiness
able neighborhood search algorithms for a single machine schedul- in multi-machine scheduling: a mixed integer programming ap-
ing problem with distinct due windows. Electron Notes Theor proach. Comput Ind Eng 38:297–305
Comput Sci 281:5–19 21. Suresh V, Chaudhuri D (1994) Minimizing maximum tardiness for
7. Chen JF (2009) Scheduling on unrelated parallel machines with unrelated parallel machines. Int J Prod Econ 34:223–229
sequence- and machine-dependent setup times and due-date con- 22. Adamopoulos GI, Pappis CP (1998) Scheduling under a common
straints. Int J Adv Manuf Technol 44:1204–1212 due-date on parallel unrelated machines. Eur J Oper Res 105:494–
8. Graham R, Lawler E, Lenstra J, Kan AR (1979) Optimization and 501
approximation in deterministic sequencing and scheduling: a sur- 23. Kim D, Kim K, Jang W, Chen F (2002) Unrelated parallel machine
vey. Ann Discret Math 5:287–326 scheduling with setup times using simulated annealing. Robot
9. Ullman JD (1975) NP-complete scheduling problem. J Comput Comput Integr Manuf 18:223–231
Syst Sci 10:384–393 24. Pereira Lopes MJ, De-Carvalho JMV (2007) A branch-and-price
algorithm for scheduling parallel machines with sequence depen-
10. McNaughton R (1959) Scheduling with deadlines and loss func-
dent setup times. Eur J Oper Res 176:1508–1527
tions. Manag Sci 6:1–12
25. Chen C-L, Chen C-L (2009) Hybrid meta-heuristics for unrelated
11. Cheng TCE, Sin CCS (1990) A state-of-the-art review of parallel
parallel machine scheduling with sequence-dependent setup times.
machine scheduling research. Eur J Oper Res 47:271–292
Int J Adv Manuf Technol 43:161–169
12. Lam K, Xing W (1997) New trends in parallel machine scheduling. 26. Lin S-W, Lu C-C, Ying K-C (2011) Minimization of total tardiness
Int J Oper Prod Manag 17:326–338 on unrelated parallel machines with sequence- and machine-
13. Mokotoff E (2001) Parallel machine scheduling problems: a survey. dependent setup times under due date constraints. Int J Adv
Asia Pac J Oper Res 18:193–242 Manuf 53:353–361
14. Lee Y, Pinedo M (1997) Scheduling jobs on parallel machines with 27. Mitchell M (1999) An introduction to genetic algorithms. MIT
sequence-dependent setup times. Eur J Oper Res 100:464–474 Press, Massachusetts
15. Park Y, Kim S, Lee Y (2000) Scheduling jobs on parallel machines 28. Kirkpatric S, Gelatt JCD, Vecci MP (1983) Optimization by simu-
applying neural network and heuristic rules. Comput Ind Eng 38: lated annealing. Science 220:671–680
189–202 29. Blum C, Roli A (2003) Meta-heuristics in combinatorial optimiza-
16. Kurz M, Askin R (2001) Heuristic scheduling of parallel machines tion: overview and conceptual comparison. ACM Comput Surv 35:
with sequence-dependent setup times. Int J Prod Res 39:3747–3769 268–308
17. Fowler JW, Horng SM, Cochran JK (2003) A hybridized genetic 30. Figielska E (2009) A genetic algorithm and a simulated annealing
algorithm to solve parallel machine scheduling problems with se- algorithm combined with column generation technique for solving
quence dependent setups. Int J Ind Eng Theory Appl Pract 10:232– the problem of scheduling in the hybrid flow-shop with additional
243 resources. Comput Ind Eng 56:142–151