Unrelated Parallel Machine Scheduling With Setup Times Using Simulated Annealing
Unrelated Parallel Machine Scheduling With Setup Times Using Simulated Annealing
Unrelated Parallel Machine Scheduling With Setup Times Using Simulated Annealing
Abstract
This paper presents a scheduling problem for unrelated parallel machines with sequence-dependent setup times, using simulated
annealing (SA). The problem accounts for allotting work parts of L jobs into M parallel unrelated machines, where a job refers to a
lot composed of N items. Some jobs may have different items while every item within each job has an identical processing time with
a common due date. Each machine has its own processing times according to the characteristics of the machine as well as job types.
Setup times are machine independent but job sequence dependent. SA, a meta-heuristic, is employed in this study to determine a
scheduling policy so as to minimize total tardiness. The suggested SA method utilizes six job or item rearranging techniques to
generate neighborhood solutions. The experimental analysis shows that the proposed SA method significantly outperforms a
neighborhood search method in terms of total tardiness. r 2002 Elsevier Science Ltd. All rights reserved.
0736-5845/02/$ - see front matter r 2002 Elsevier Science Ltd. All rights reserved.
PII: S 0 7 3 6 - 5 8 4 5 ( 0 2 ) 0 0 0 1 3 - 3
224 D.-W. Kim et al. / Robotics and Computer Integrated Manufacturing 18 (2002) 223–231
Many researchers have studied the identical parallel the SA are also discussed. The suggested SA approach is
machine scheduling. Pourbabai [5] introduced the compared with a neighborhood search (NS) method to
optimal batch size to minimize total tardiness. Shutten show the performance of our proposed approach.
and Leussink [6] presented a branch-and-bound algo-
rithm to minimize maximum lateness considering
release dates, due dates, and family setup times. 2. Problem definition and characteristics
Akkiraju [7] investigated identical parallel machine
scheduling to minimize job tardiness. Balakrisnan et al. Suppose there are M unrelated parallel machines and
[8] studied uniform parallel machine scheduling with L jobs, where a job refers to a lot composed of N items.
ready times and sequence-dependent setup times using Each job may have different processing times
a compact mathematical model for small sized depending on the assigned machine, but items in the
problems. same lot are assumed to have the same processing times
Several studies discuss unrelated parallel machine with the same machine. We also assume that each
scheduling problems for the purpose of minimizing machine can process one item at a time, and the
tardiness. Koulamas [9] reviewed existing heuristic processing is non-preemptive. Items in a job lot may be
approaches in this area. Suresh and Chaudhuri [10] processed in multiple machines. The completion time
and Adamopoulos and Pappis [11] suggested new of the last item in a lot becomes the completion time of
heuristics for various due date and processing time the lot.
combinations. Slowinski [12] and Lee and Guignard [13] The setup time might have different values depending
considered different setup times in unrelated machines. on job sequences. This means setup times depend on
Both studies used Lagrange functions and two-phase both the job just completed and the job about to be
methods in order to find good solutions. Weng et al. [14] processed. Because a job lot has only identical items,
addressed the toll plaza repair scheduling problem with setup times between these items are equal to zero. We
setup times to minimize a weighted mean completion assume setup times obey the widely accepted triangle
time. inequality [8,16]. We also assume each job lot has its
Family or batch scheduling models are increasingly own due date. The following notations shall be used to
being studied. Bruno and Sethi [15] introduced a task define our problem:
sequencing problem in a batch environment with setup
i lot index
times. Ghosh and Gupta [16] addressed a single machine
j item index in a lot
batch scheduling problem to minimize maximum
k machine index
lateness. Liaee and Emmons [17] reviewed scheduling
L number of lots
theories concerning the processing of several families of
M number of machines
jobs on single or parallel facilities. Potts and Kovalyov
N number of items in each lot
[18] also reviewed lots of literature on scheduling with
cij completion time of item j in lot i
batch decisions, according to scheduling models as well
Ci completion time of the last item in the ith lot,
as basic algorithms.
(Ci ¼ maxcij;j¼1 ; y; N)
Recently, meta-heuristics, which were developed to
di due date of the ith lot
solve complicated combinatorial problems, have been
Ti tardiness of lot i; (Ti ¼ maxfCi 2di ; 0g)
applied to scheduling problems. Tamimi and Rajan [19]
sk job sequence on machine k
used a Genetic algorithm to find a scheduling policy for
lk ðiÞ the index of the lot in position i on machine k
identical parallel machines with setup times. Koulamas
nk the number of lots assigned to machine k:
[20] used simulated annealing (SA) in an identical
pijk the operation time of the item j of lot i in
parallel machine scheduling problem. He used SA to
machine k
exchange jobs assigned to machines using decomposi-
sijk sequence-dependent setup time in machine k when
tion. Applying Tabu search into parallel machine
a lot i is changed to a lot j; where the setup times
scheduling was reported by Suresh and Chaudhuri [21]
satisfy the triangle inequality: sijk þ sjlk Xsilk
and Armentano and Yamashita [22]. Park and Kim [23]
compared SA and Tabu search under the objective of The scheduling problem considered in this study is to
minimizing maintenance costs with ready time and due minimize the total maximum tardiness:
date constraints. Jozefowska et al. [24] also suggested XL
and compared SA, Tabu, and Genetic algorithm Objective ¼ Minimize maxðCi di ; 0Þ:
approaches. i¼1
We present the unrelated parallel machine scheduling A solution is represented as a set of (s1 ; y; sk ; y; sM ),
problem using SA. Based on the characteristics of our where sk is a sequence of jobs on machine k: The
problem, six different methods of neighborhood solu- sequence sk on machine k can be represented by a set of
tion generation and appropriate parameter values for ordered lots, flk ð1Þ; lk ð2Þ; y; lk ðnk Þg:
D.-W. Kim et al. / Robotics and Computer Integrated Manufacturing 18 (2002) 223–231 225
M1 ... P321 P351 S351 P511 P521 S511 P111 ... 3.2.4. Lot split
Items of one lot scheduled in one machine are split
M2 ... P252P242 P272 S262 P632 S652 P542 ... into two machines. This decreases tardiness by dividing
items that cause large completion times. The selection
rule of a lot is the same as before. The selected set of
M1 ... P321 P351 S321 P251 P241 P271 S211 P111 ... items is divided arbitrarily and scheduled randomly to a
machine that currently does not have items from the
M2 ... P632 S652 P512 P522 P542 ...
same lot. However, the new schedule should not cause a
Fig. 2. Lot interchange.
setup time increase by dividing items from the same lot.
Fig. 5 shows selection and split of lot 1 in machine 1,
and resulting new schedules of machine 1 and 2. This
M1 ... P321 P351 S351 P511 P521 S511 P111 ... rule can decrease tardiness by dividing items in a lot that
may cause large completion times when the items are
M2 ... P252 P242 P272 S242 P412 ...
scheduled to the same machine.
3.2.3. Lot merge M2 ... P752 P732 S 742 P412 S482 P862 ...
This selects two sets of items in the same lot scheduled
to process in two different machines, and merges them
M1 ... P321 P351 S351 P511 ...
so that they are processed in one machine. This
procedure decreases additional setup times that occur ... P752 P732 P722
M2 S742 P412 S482 P862 ...
when items in the same lot are distributed to more than
one machine. The lot selection rule is the same as the Fig. 4. Lot merge.
one described in the lot interchange method. Once a lot
is selected, two sets of items, one with the largest
completion time and the other randomly selected, in this M1 ... P321 P351 S351 P511 S511 P111 P121 P151 ...
lot are merged. For example, lot 7 is first selected in
M2 ... P252 S242 P412 S482 P862 ...
Fig. 4. If an item in lot 7 has the largest completion time
in machine 1 then items of lot 7 in machine 1 should be
merged with randomly selected items of lot 7, which are,
M1 ... P321 P351 S351 P511 S511 P111 ...
in this case, items in machine 2. We need this rule
because merging lots may decrease unnecessary setup M2 ... P252 S212 P122 P152 S142 P412 S482 P862 ...
times that occur when the same types of items are
distributed into more than one machine. Fig. 5. Lot split.
D.-W. Kim et al. / Robotics and Computer Integrated Manufacturing 18 (2002) 223–231 227
Perturbation Procedure
Begin; Fig. 6. Initial temperature and solution quality.
Y 1=Lot-Interchange(X );
Y 2=Lot-Insert(X ); preliminary experiments before actual SA [26]. We used
Y 3=Lot-Merge(X ); a 60% acceptance ratio in this paper after 10,000
Y 4=Lot Split(X ); experiments.
Y 5=Item Interchange(X );
Y 6=Item Insert(X ); 3.3.2. Cooling ratio
Find X corresponding to MinðY 1; Y 2; Y 3; Y 4; There are two widely accepted common ratios such as
Y 5; Y 6Þ; the log ratio, which use the temperature of kth exterior
Y ¼ X loop Tk :
End;
T0 logðk0 Þ
Tk ¼ ; k0 > 1
logðk þ k0 Þ
3.3. Determination of parameters
or geometric ratio Tk :
SA is a meta-heuristic that has proven to be very Tk ¼ aTk1 ; k ¼ 1; 2; y; 0oao1:
effective for solving complicated combinatorial pro-
Log ratio guarantees the convergence but is slow to
blems. The theory of SA indicates that SA converges to
converge. Geometric ratio, used more commonly in
a globally optimal solution with probability 1 if all
practice, use a between 0.5 and 0.99.
theoretical conditions are satisfied [20,25]. However, the
conditions for the asymptotic convergence, such as
3.3.3. Epoch length
cooling schedule and stop criterion, cannot be met in
Epoch length is often proportional to the number of
practice [20]. Thus, it is often critical to adjust values of
possible neighborhood solutions. It along with cooling
parameters such as initial temperature, cooling schedule,
ratio significantly affects the performance of algorithm.
repetition number, and ending condition based on
We use epoch length, which is proportional to the
problem characteristics.
number of lots, because the number of possible
We first use standard parameters, which are applic-
neighborhood solutions is too large:
able to our problem, and then determine parameters to
use considering the performance and running time of the epoch length ¼ number of lots ðLÞ lot size ðNÞ b:
algorithm. We used the standard setting of a problem
with 10 machines, 50 lots, and 20 items per lot to 3.3.4. End count
determine the parameters. End count is the condition that finishes the algorithm,
and it finishes theoretically at the point that converges to
3.3.1. Temperature temperature 0. Because the algorithm has to spend a
Theoretically speaking, initial temperature should be long time under low temperatures in this case, it is
high enough so that all movements are acceptable. typical to stop the algorithm if the external loop does
However, it is necessary to control initial temperature not improve after some iteration. This value is propor-
more efficiently because very high initial temperature tional to the number of lots in our study. The parameter
could consume too much time in preparation stage of g is set to be 14, and the algorithm also finishes when the
our problems (Fig. 6). temperature reaches 0.1, which is usually low enough:
Therefore, initial temperature is often determined by
end count ¼ number of lots ðLÞ lot size ðNÞ g:
the minimum temperature that is larger than or equal to
the acceptance ratio, the number of accepted movements The next step adjusts initial parameters so that they
divided by the number of total movements, tested by are more appropriate to our problems. We consider not
228 D.-W. Kim et al. / Robotics and Computer Integrated Manufacturing 18 (2002) 223–231
only the quality of solutions but also computation time, 4. Performance evaluation
which is the drawback of SA compared to other
approaches. We tested the algorithm using randomly The suggested SA algorithm (SA S) has to be
generated variables 10 times while all the other compared to other algorithms to prove its efficiency.
parameters are fixed except the one we wanted to adjust. One algorithm is a conventional SA (SA C), which
The box-whisker plot, which provides center point, exchanges and inserts an item as a unit without
variability, and tail length of a distribution, is used for considering problem characteristics such as lots and
data analysis. setup times. We selected the SA C since we have to show
Fig. 7 represents the relationship between the solution the relative advantage of the SA S considering the six-
acceptance ratio, which is set to be 54%, 55%, 57%, neighborhood solution generation to find the next move.
60%, 70%, and 80%, and the quality of solutions under The other is a neighborhood search (NS) method
the preliminary tests. Fig. 8 represents the relationship which is sometimes called a descent technique because
between acceptance ratio and computation time. While each new seed represents a lower value of the objective
solutions have high quality when acceptance ratio is as introduced in Baker [27]. One of the problems of NS
relatively low such as 54%, 55%, and 57%, computa- procedures is their tendency to become trapped at local
tional times are similar except when it is 80%. optima [27]. The method of obtaining the initial seed
We have found from Fig. 8 that computational times and the neighborhood generating mechanism for our NS
do not influence significantly on the quality of solutions method are the same as the SA S, except selecting a
if we take the acceptance ratio lower than 60%. Hence, particular sequence to be a new seed. The NS finds the
we use initial temperature 261 that induces 57% best solution (the new seed) inside a searchable region,
acceptance ratio, which shows the best results in both while it does not allow a movement to a worse solution.
average and variance. Similarly, we decided on a cooling We apply the same conditions including the avoidance
ratio using a ¼ 0:95; epoch length using b ¼ 1:1; and of unnecessary searches to all algorithms for the validity
end count using g ¼ 14: of the comparison. We use the same order as shown in
the perturbation procedure in Section 3.2 to apply for
the generation of neighborhood solutions based on our
past experience.
Our study is based on the production data for 1 week
obtained from a compound semiconductor manufac-
turing company located in Iksan, Chonbuk, Korea.
The processing time for each item (wafer) depends
on the machine used and it is generated from Uniform
[30, 60]. Given the number of job lots to be processed,
for each job pair (i; j), the sequence-dependent setup
times sijk on a machine k were randomly chosen
from Uniform [10, 90]. The resulting setup time matrix
was then repeatedly corrected to satisfy the triangular
property. Setup times between the items in a job
lot are equal to zero because a job lot has only identical
Fig. 7. Acceptance ratio and quality of solutions. items.
Due dates of jobs are integer values generated from
Uniform ½Pð1 t r=2Þ; Pð1 t þ r=2Þ as suggested
by Potts and Van Wassenhove [28]. P; t; and r control
make-span, priority factor, and due date range factor,
respectively. Because P cannot be calculated accurately
it is estimated to be the smallest of processing times and
setup times of a job. r is fixed to be 0.8, and t is 0.4
(tight: T), 0.45 (moderate: M), and 0.5 (loose: L).
In addition to the standard setting of 10 machines, 50
lots, and 20 items per lot, we considered various
combinations of the number of machines (M ¼ 5; 10,
15), the number of lots (L ¼ 25; 50, 75), the size of lots
(N ¼ 10; 20, 30), and due dates t=(T; M; L). We
compared the suggested SA S, the conventional SA C,
Fig. 8. Acceptance ratio and times. and the NS through the combinations of these factors.
Because SA C uses only two out of six different
D.-W. Kim et al. / Robotics and Computer Integrated Manufacturing 18 (2002) 223–231 229
Table 1
Computational results: 25 lots
Final sol. (min) Com. time (s) Final sol. (min) Com. time (s) Final sol. (min) Com. time (s)
Table 2
Computational results: 50 lots
Final sol. (min) Com. time (s) Final sol. (min) Com. time (s) Final sol. (min) Com. time (s)
Table 3
Computational results: 75 lots
Final sol. (min) Com. time (s) Final sol. (min) Com. time (s) Final sol. (min) Com. time (s)
neighborhood generation methods, the iteration of its quence-dependent setup times and lot processing. We
internal loop is increased 3 times. suggested new parameter values appropriate to our
Algorithms are written in C++, and a Pentium problem through preliminary experiments based on
500 MHz computer is used to run computations. Tables standard parameter setting. Neighborhood solution
1–3 represent the results of computational experiments generation to perturb initial or existing solutions was
when the number of lots is 25, 50, and 75, respectively. suggested. The generation includes six methods such as
The parameter most significantly affecting computation lot interchange, lot insert, lot merge, lot split, item
time is the number of lots, and the size of lot, the interchange, and item insert so that the characteristics of
number of machines, and the due date follow next. NS is our problem are well reflected.
faster for small size problems, but the computation A suggested SA approach (SA S) was compared with
speed of SA S and SA C becomes relatively faster as the a conventional SA method (SA C), which used neigh-
size gets larger. It is also shown that SA C needs longer borhood solutions generated just by item interchange or
computational time due to the increased number of by item insert without considering setup times and lot
neighborhood solution comparison. processing. The purpose was to show the superiority of
However, SA C prematurely finishes early in some SA S considering the six neighborhood solutions based
cases because a solution is not improved for a while at on the characteristics of our problem over the SA C
the high temperature. The solutions of SA S are always method. The SA S was also compared with a neighbor-
better than the other two, while SA C is better than NS, hood search (NS) method that appears to be a
37 times out of 81 tests under different settings. promising heuristic procedure for solving sequencing
problems. The NS always finds a best solution, or a new
seed, inside a searchable region without allowing a move
5. Conclusion to a worse solution.
The performance of SA S is always better than the
This paper presented a simulated annealing (SA) other two when they are compared based on the number
approach for an unrelated parallel machine scheduling of machines, the lot size, the number of lots, and the
problem to minimize total tardiness considering se- urgency of due date. Although NS shows faster
D.-W. Kim et al. / Robotics and Computer Integrated Manufacturing 18 (2002) 223–231 231
computation for smaller problems used in this study, it [8] Balakrishnan N, Kanet JJ, Sridharan SV. Early/tardy scheduling
actually takes longer as the size of a problem increases. with sequence dependent setups on uniform parallel machines.
This might be caused by the main property of NS, which Comput Oper Res 1999;26:127–41.
[9] Koulamas C. Total tardiness problem: review and extensions.
searches all possibilities and selects the best. Oper Res 1993;42:1025–41.
It might be desirable to apply other meta-heuristics [10] Suresh V, Chaudhuri D. Minimizing maximum tardiness for
such as Tabu search and Genetic algorithm in addition unrelated parallel machines. Int J Prod Econ 1994;34:223–9.
to SA to our problem and compare outcomes. When we [11] Adamopoulos GI, Pappis CP. Scheduling under a common
determine the values of parameters in this study, a box- due-date on parallel unrelated machines. Eur J Oper Res
1998;105(3):494–501.
whisker plot under the assumption of independence of [12] Slowinski R. Production scheduling on parallel machines subject
parameters was used. However, it will also be interesting to staircase demands. Eng Costs Prod Econ 1988;14:11–17.
to simultaneously consider those parameters without [13] Lee H, Guignard M. Hybrid bounding procedure for the
assumptions. Finally, other practical situations such as workload allocation problem on parallel unrelated machines with
setups. J Oper Res Soc 1996;47:1247–61.
different ready times of lots and priority lots could be
[14] Weng MX, Lu J, Ren H. Unrelated parallel machine scheduling
incorporated into our future research. with setup consideration and a total weighted completion time
objective. Int J Prod Econ 2001;70:215–26.
[15] Bruno J, Sethi R. Task sequencing in a batch environment with
Acknowledgements setup times. Found Control Eng 1978;3:105–17.
[16] Ghoshi JB, Gupta JND. Batch scheduling to minimize maximum
lateness. Oper Res Lett 1997;21:77–80.
This paper was partially supported by the research [17] Liaee MM, Emmons H. Scheduling families of jobs with setup
funds of Chonbuk National University. The authors are times. Int J Prod Econ 1997;51:165–76.
indebted to Dr. M.W. Park for his helpful suggestions [18] Potts CN, Kovalyov MY. Scheduling with batching: a review. Eur
on this research. They are also grateful for the J Oper Res 2000;120:228–49.
anonymous referees whose suggestions helped improve [19] Tamimi SA, Rajan VN. Reduction of total weighted tardiness on
uniform machines with sequence dependent setups. Industrial
the presentation of this paper. Engineering ResearchFConference Proceedings, 1997. p. 181–5.
[20] Koulamas C. Decomposition and hybrid simulated annealing
heuristics for the parallel-machine total tardiness problem. Nav
References Res Logistics 1997;44:105–25.
[21] Suresh V, Chaudhuri D. Bicriteria scheduling problem for
unrelated parallel machines. Comput Ind Eng 1996;30:77–82.
[1] Na DG, Jang W, Kim DW. Resource planning for compound
semiconductor manufacturing. Proceedings of the Ninth Interna- [22] Armentano VA, Yamashita DS. Tabu search for scheduling on
tional Manufacturing Conference in China, The Hong Kong identical parallel machines to minimize mean tardiness. J Intell
Manuf 2000;11:453–60.
Polytechnic University, Hong Kong, August 16–17, 2000.
[2] Cheng TCE, Sin CCS. A State-of-the-art review of parallel- [23] Park MW, Kim YD. Search heuristics for a parallel machine
machine scheduling research. Eur J Oper Res 1990;77:271–92. scheduling problem with ready times and due dates. Comput Ind
[3] Karp RM. Reducibility among combinatorial problems: complex- Eng 1997;33:793–6.
[24] Jozefowska J, Mika M, Rozycki R, Waligora G, Weglarz J. Local
ity of computer computations. New York: Plenum Press, 1972.
p. 85–103. search meta-heuristics for discrete-continuous scheduling pro-
[4] Ho JC, Chang YL. Minimizing the number of tardy jobs for blems. Eur J Oper Res 1998;107:354–70.
m-parallel machines. Eur J Oper Res 1995;84:343–55. [25] Kirkpatric S, Gelatt Jr. CD, Vecci MP. Optimization by
simulated annealing. Science 1983;220:671–80.
[5] Pourbabai B. One stage scheduling of preemptive jobs on parallel
machines with setup times and due dates. Proc Am Inst Ind Eng, [26] Johnson DS, Aragon CR, Mageoch LA, Schevon C. Optimization
Annu Conf Conv 1985;258:525–8. by simulated annealing: an experimental evaluation; part 1, graph
partitioning. Oper Res 1989;37:865–92.
[6] Schutten JMJ, Leussink RAM. Parallel machine scheduling with
release dates, due dates, and family setup times. Int J Produc Econ [27] Baker KR. Elements of sequencing and scheduling. Amos Tuck
1996;46–47:119–25. School of Business Administration, Dartmouth College, Hanover,
[7] Akkiraju R, Murthy S, Keskinocak P, Wu F. Multi machine NH, 1995.
[28] Potts CN, Van Wassenhove L. Decomposition algorithm for the
scheduling: an agent-based approach. Innovative Applications
of Artificial IntelligenceFConference Proceedings, July 26–30, single machine total tardiness problem. Oper Res Lett 1982;
1998. p.1013–18. 1(5):177–81.