This article was downloaded by: [Aston University]

On: 27 August 2014, At: 08:31

Publisher: Taylor & Francis
Informa Ltd Registered in England and Wales Registered Number: 1072954
Registered office: Mortimer House, 37-41 Mortimer Street, London W1T 3JH, UK

International Journal of Production

Research
Publication details, including instructions for authors and
subscription information:
http://www.tandfonline.com/loi/tprs20

Tabu-search simulation optimization

approach for flow-shop scheduling
with multiple processors — a case
study
T. Yang , Y. Kuo & I. Chang
a
Institute of Manufacturing Engineering , National Cheng
Kung University , Tainan 701, Taiwan, Republic of China
b
Institute of Manufacturing Engineering , National Cheng
Kung University , Tainan 701, Taiwan, Republic of China E-
mail:
Published online: 21 Feb 2007.

To cite this article: T. Yang , Y. Kuo & I. Chang (2004) Tabu-search simulation optimization
approach for flow-shop scheduling with multiple processors — a case study, International
Journal of Production Research, 42:19, 4015-4030, DOI: 10.1080/00207540410001699381

To link to this article: http://dx.doi.org/10.1080/00207540410001699381

PLEASE SCROLL DOWN FOR ARTICLE

Taylor & Francis makes every effort to ensure the accuracy of all the information (the
“Content”) contained in the publications on our platform. However, Taylor & Francis,
our agents, and our licensors make no representations or warranties whatsoever
as to the accuracy, completeness, or suitability for any purpose of the Content. Any
opinions and views expressed in this publication are the opinions and views of the
authors, and are not the views of or endorsed by Taylor & Francis. The accuracy
of the Content should not be relied upon and should be independently verified
with primary sources of information. Taylor and Francis shall not be liable for any
losses, actions, claims, proceedings, demands, costs, expenses, damages, and other
liabilities whatsoever or howsoever caused arising directly or indirectly in connection
with, in relation to or arising out of the use of the Content.

This article may be used for research, teaching, and private study purposes. Any
substantial or systematic reproduction, redistribution, reselling, loan, sub-licensing,
systematic supply, or distribution in any form to anyone is expressly forbidden.
 Terms & Conditions of access and use can be found at http://www.tandfonline.com/
page/terms-and-conditions
Downloaded by [Aston University] at 08:31 27 August 2014
 int. j. prod. res., 1 october 2004,
vol. 42, no. 19, 4015–4030

Tabu-search simulation optimization approach for flow-shop scheduling

with multiple processors — a case study

T. YANG*, Y. KUO and I. CHANG

The flow shop with multiple processors (FSMP) environment is relatively

common and has a variety of applications. The majority of academic authors
solve the scheduling problem of FSMP using deterministic data that ignore the
stochastic nature of a real-world problem. Discrete-event simulation can model a
non-linear and stochastic problem and allows examination of the likely behaviour
of a proposed manufacturing system under selected conditions. However, it does
not provide a method for optimization. The present paper proposes to solve the
Downloaded by [Aston University] at 08:31 27 August 2014

FSMP scheduling problem by using a tabu-search simulation optimization

approach. It features both the stochastically modelling capability of the
discrete-event simulation and the efficient local-search algorithm of tabu search.
A case study from a multilayer ceramic capacitor manufacturing illustrates the
proposed solution methodology. Empirical results show promise for the practical
application of the proposed methodologies. Future research opportunities are
then addressed.

1. Introduction
The flow shop with multiple processors (FSMPs) is also known as ‘flexible flow
line’ and ‘hybrid flow shop’. The FSMP scheduling problem involves the sequencing
of jobs in a flow shop in which, at any stage, more than one identical machine
might exist. A machine can process exactly one job at a time, and the jobs are subject
to precedence constraints (Salvador 1973). This type of scheduling environment
is relatively common and has a variety of applications — including semiconductor
and electronics manufacturing, and petrochemical production (Santos et al. 1996,
Botta-Genoulaz 2000). First-in-first-out (FIFO) is used as the sequencing procedure
in the subsequent stages because it is the strategy usually used in FSMP environments
(Pinedo 1995).
The FSMP scheduling problem is NP-hard and its optimal solution is computa-
tionally prohibitive for a practical-size problem (Gupta 1988, Brah 1992). In the past
few decades there has been a significant number of reports in the literature that have
discussed the single-processor flow-shop scheduling problem (Campbell et al. 1970,
Townsend 1977, Nawas et al. 1983, Proust et al. 1991). Studies on FSMP scheduling
problems are relatively recent. Of these, most have dealt with the makespan criterion
and have usually been limited to two stages (Gupta 1988, Gupta and Tunc 1991,
Deal et al. 1994, Uetake et al. 1995). The two-stage problem is a special case of the
FSMP scheduling problem and is not practical.

Revision received March 2004.

Institute of Manufacturing Engineering, National Cheng Kung University, Tainan 701,
Taiwan, Republic of China.
*To whom correspondence should be addressed. e-mail: tyang@mail.ncku.edu.tw

International Journal of Production Research ISSN 0020–7543 print/ISSN 1366–588X online # 2004 Taylor & Francis Ltd
http://www.tandf.co.uk/journals
DOI: 10.1080/00207540410001699381
 4016 T. Yang et al.

The majority of reports in the literature solve the FSMP scheduling problem
using deterministic data that ignore the stochastic nature of a real-world problem
— such as process, set-up, load/unload, machine failure times, etc. Discrete-event
simulation can model a non-linear and stochastic problem and allows examination
of the likely behaviour of a proposed manufacturing system under selected condi-
tions. However, it does not provide a method for optimization (Yang and Tseng
2002).
Optimization-based simulation in solving an FSMP scheduling problem is rare
in the literature. Grangeon et al. (1999) used a generic FSMP simulation model to
evaluate the performances of an FSMP system using different dispatching rules.
Hsieh et al. (2001) solved the semiconductor wafer fabrication-scheduling problem
using rule-based simulation. Jeong (2000) proposed a conceptual framework for
the development of optimized simulation-based scheduling systems. This framework
integrated a discrete-event simulator and a rule-based system to handle condition-
Downloaded by [Aston University] at 08:31 27 August 2014

based events for a given environment. Azzaro-Pantel et al. (1998) solved a batch–
plant-scheduling problem using an iterative discrete-event simulation and genetic
algorithm methodology.
Tabu search (TS) is a local search-based optimization method that has been
successfully applied to solve many difficult combinatorial optimization problems.
Its applications to sequencing-type problems are impressive and have exhibited
considerable robustness (Ben-Daya and Al-Fawzan 1998). For many scheduling
optimization problems, the empirical results from Taillard (1989), Widmer
and Hertz (1989), Barnes and Laguna (1993) and Dorn et al. (1996) showed that
TS can outperform simulated annealing, branch-and-bound method, genetic algo-
rithms and random search methods. Although the comparison is only possible
between implementations of metaheuristics for a given problem, the potential of
TS for the present study is clear by examining the empirical results from existing
literature.
The present paper proposes to solve the FSMP scheduling problem by a TS
simulation optimization approach. It explores the hybrid TS and discrete-event
simulation methodologies in solving a complex manufacturing system problem,
which is not well addressed in literature. It features both the stochastic modelling
capability of the discrete-event simulation and the efficiency of the TS algorithm.
In addition, a practical case study from a multilayer ceramic capacitor (MLCC)
manufacturing is used to illustrate the proposed solution methodology. For this
type of problem, the case study presents the MLCC manufacturing application for
its first time in literature.

2. Case study
Passive components are used in all electronics to regulate the flow of electricity
and to store electric charge for peak power needs, for frequency control and for
filtering. Among the passive components, capacitors (mostly tantalum capacitors
and MLCCs) account for approximately 70% of the USA’s revenue. The use of
tantalum capacitors and MLCCs is growing quickly because they are surface moun-
table and are used primarily in communication and computing (Harris and Roesch
2001).
The MLCC manufacturing process begins from ceramic powder preparation
and ends with reel taping. Figure 1 illustrates a generic MLCC manufacturing
process flow. Figure 2 illustrates the appearance of an MLCC. The dimensions
 Tabu-search simulation optimization approach for flow-shop scheduling 4017
Downloaded by [Aston University] at 08:31 27 August 2014

Figure 1. Generic MLCC process flow.

Figure 2. A MLCC example.

for a fine-pitch MLCC are about 1.0, 0.5 and 0.5 mm for length, width and height,
respectively.
The MLCC manufacturing has a sole process flow with precedence constraints.
In addition, there are multiple machines in each stage. Thus, it is an FSMP problem.
However, this is a more complex modelling problem than is a conventional FSMP
problem for the following reasons. First, it has more than 10 stages in its process
flow. Second, the binder burnout and sintering stages are batch-processing proces-
sors. Finally, its product mix is determined by both material type and product size,
the combination of which decides the set-up requirements.
The case study presented here is of a leading MLCC manufacturing company
located in Tainan, Taiwan. Its annual sales are expected to exceed US$229 million
in 2003. An MLCC plant usually has different inventory-carry policies in the before-
stacking and after-testing stages. The raw materials in the before-stacking stage are
primarily powder and foil, and are received in bulk. Therefore, batch production is
 4018 T. Yang et al.

a common practice and this results in highly variable work-in-process inventory. The
taping stage usually uses a late customization policy and is not part of the process
flow. Thus, this research models the case study problem from the stacking stage to
the testing stage. In total, the case study problem entails nine stages. The detailed
modelling data — including stage names, number of machines in each stage, set-up
time, machine failure information and processing time — are shown in appendices
A1–5. The present study has modified some data to protect proprietary information
from the case study company.
There are 24 product types consisting of the various combinations of six material
types and four product sizes. The smallest processing unit is termed a ‘lot’. The lot
size is product specific and the various sizes are collected in appendix A2. Due to
the batch-processing requirements of the binder burnout and sintering stages, the
study assigns a batch size of six lots for experimental purposes. Hereafter, each
customer order represents one batch. This study adopts the FIFO method for the
Downloaded by [Aston University] at 08:31 27 August 2014

interstage job sequencing. It assumes that all production lots are available before
the production starts, and then tries to decide the job sequence for entering the
production line. The proposed methodology solves the order-sequencing problem
for MLCC manufacturing and thus minimizes the total tardiness that is one of the
key performance indices of the case study.

3. Proposed methodology
The proposed TS simulation optimization approach is an iterative search algo-
rithm. It follows the efficient TS procedure and uses computer simulation to evaluate
system performance for each search iteration.
The key requirement of any global optimization method is that it should be able
to avoid entrapment in seeking a locally optimal solution while continuing the search
to provide a near-optimal final solution. TS has been proposed by Glover (1989) as
an iterative process that explores the solution space by moving from one solution
to another neighbouring solution. By using the short-term memory of recent solu-
tions, the optimization routines can effectively escape from local optimal solutions
(Glover and Laguna 1993).
TS begins with an initial solution, and then moves from one solution S to another
solution S’, which is located in its neighbourhood N(S ). A solution S’, which is
worse than the current solution S, might be accepted to escape from the entrapment
of a local optimum. The most recent moves are classified as tabu for a particular num-
ber of iterations — i.e. tabu tenure or tabu list size — to avoid the cyclic searching
path. Aspiration criterion is used when a move is in tabu list, but it has a better
solution. In this instance, the move is released from the tabu list. The search will
terminate when it satisfies the stopping criterion. When TS uses the above short-term
flexible memory structures, it is usually referred to as a ‘simple TS’ (Glover 1989),
which is hereafter denoted as ‘TS1’.
A more advanced type of TS includes frequency-based memory intensification
and frequency-based memory diversification. The former seeks to reinforce moves
that incorporate attributes of good solutions found in the past, whereas the latter
seeks to drive the search into unexplored regions (Wen and Huang 1996). The TS
with intensification and diversification strategies will hereafter be denoted by ‘TS2’.
Details of the proposed methodology are discussed below.
 Tabu-search simulation optimization approach for flow-shop scheduling 4019

3.1. Initial solution

An initial solution to solving a scheduling-type problem often involves, a priori,
the use of a heuristic approach, and this often influences the quality of the final
solution. In the case study under consideration here, the earliest-due-date (EDD)
rule is used to generate the initial solution. EDD is an effective rule for minimizing
production tardiness in most instances (Hopp and Spearman 2001).

3.2. Move
For TS operation, a neighbourhood solution can be reached from the current
solution. The pairwise-exchange (or swap) method is often used as a move to con-
struct a neighbourhood solution in a permutation-type problem (Glover et al. 1995).
This study randomly exchanges two orders to reach a neighbourhood solution. The
underlying rationale for a random exchange is to avoid entrapment in local optimum
and to have the chance to explore a different solution search area.
Downloaded by [Aston University] at 08:31 27 August 2014

3.3. Neighbourhood size

The neighbourhood size is the number of candidate solutions to be evaluated in
each iteration along the search process. The existing literature discussed three dif-
ferent neighbour selection strategies (Ben-Daya and Al-Fawzan 1998):
. Choose the first neighbourhood solution that improves the current solution
(referred to as ‘first-fit strategy’ hereafter).
. Consider a subset of neighbours and search the best solution.
. Search the whole neighbourhood and choose the best solution.
In general, it is computationally too expensive to explore the whole neighbourhood
of the current solution. A subset of neighbours is often used to alleviate the compu-
tational efforts (Alvarez-Valdes et al. 2002). In addition, the size of the subset is too
large to be useful (Higgins 2001, Negenman 2001). In other words, the improvement
to solution quality will become negligible when the size of the subset exceeds a
certain level. The larger the subset size, the longer the computational time.
When the computational efficiency is a concern, the first-fit strategy can be used
(Ben-Daya and Al-Fawzan 1998). In this research, we adopted the first-fit strategy
in consideration of the computational burden associated with a simulation
approach.

3.4. Performance evaluation

A discrete-event simulator capable of modelling a complicated stochastic prob-
lem is used to evaluate the performance of each TS move. A commercial software
program, eM-Plant (Tecnomatix Technologies 2000), was used as the simulation
tool. The TS algorithms were coded using the eM-plant embedded programming
language (called Method) for customized applications. Its syntax is similar to the
programming language Cþþ.
Due to the embedded random variations in a discrete-event simulator, a certain
number of replications might be required to obtain an adequate confidence interval
for estimating the population mean. A smaller confidence interval is a result of an
increased number of replications. However, there is a trade-off between an improved
confidence interval and computing time. Thus, we perform a preliminary study to
investigate the adequate number of replications, as described below.
 4020 T. Yang et al.

Due date
Scenario Number of Batch size Number of Number of distribution
number batches (Lot) material types product sizes (days)

1 9 6 3 3 UNIF (8, 14)

2 12 6 4 3 UNIF (8, 19)
3 16 6 4 4 UNIF (8, 26)
4 20 6 5 4 UNIF (8, 31)
5 24 6 6 4 UNIF (8, 35)

Table 1. Production scenarios.

Production scenario:

1 2 3 4 5
Downloaded by [Aston University] at 08:31 27 August 2014

0.16
0.14
0.12
0.10
CV

0.08
0.06
0.04
0.02
0.00
1 2 3 4 5 6 7 8 9
Number of replication

Figure 3. CV analysis.

Five production scenarios are generated (table 1). Those scenarios represent a
wide range of product mixes. In addition, the due date for each batch is randomly
generated using a uniform distribution, as shown in the far-right column of table 1.
Each scenario uses the same sequence from the EDD rule, but collates the outputs
(tardiness values) from different numbers of replications. The result of each replica-
tion represents one sample. Then, the coefficient of variation (CV), which is defined as
the ratio of the sample standard deviation to the sample mean, is used as an indicator
of the magnitude of variance. Figure 3 shows the relationship between CV and
the number of replications for each of those five production scenarios defined
in table 1.
Figure 3 shows minor CVs (< 0.1) in most instances. A system is considered to
be stable when its CV<0.75 (Hopp and Spearman 2001). In view of these results and
taking into account computational efficiency, four replications are taken as being
adequate for the case study.

3.5. Aspiration criterion

The proposed TS algorithm searches only for non-tabu moves except it exhausts
all the non-tabu moves and cannot find an improved solution that is better than
the current solution. If the best search result from tabu list outperforms the best
solution found so far, then the aspiration criterion overrides the tabu rule. The pro-
posed search scheme has been adopted by Yang et al. (1999) to generate promising
results.
 Tabu-search simulation optimization approach for flow-shop scheduling 4021

3.6. Intensification and diversification

TS2 uses a long-term memory structure — intensification and diversification —
to generate a new initial search sequence, and then restarts the TS procedure. The
methodology proposed by Ben-Daya and Al-Fawzan (1998) is adapted for the long-
term memory as follows.
The long-term memory starts with an n
m frequency matrix in which the n
columns and the m rows represent the n positions in a sequence and the m produc-
tion batches, respectively. Note that n ¼ m for this study. Each entry in the matrix is
denoted as fij and is incremented by 1 whenever batch i visits position j. Let k be an
index for the iteration number. Then, the proposed intensification and diversification
method has five steps as follows:
Step 0. Let k ¼ 1.
Step 1. Find the entry fij of the frequency matrix having the largest value.
Step 2. Assign order i to position j.
Downloaded by [Aston University] at 08:31 27 August 2014

Step 3. Delete row i and column j.

Step 4. Set k ¼ k þ 1.
Step 5. If k > n, stop, otherwise go to Step 1.
The idea of the above procedure is to restart the algorithm not only in a different
region of the solution space, but also in a region that hopefully contains good
solutions, thus combining intensification and diversification — which can potentially
guide the TS to a better solution region.

3.7. Tabu tenure size

Tabu tenure is a basis for preventing the search from repeating move performed
in the recent past, potentially reversing the effects of previous moves by interchanges
that might return to previous positions. For example, if we classify as tabu for the
three most recent moves, the tabu tenure size is 3. Tabu tenure size is an important
parameter for TS. Yang et al. (1999) proposed a variable tenure size to be an integer
between n/3 and 3n/2, where n is the problem size. For the proposed FSMP schedul-
ing problem, n is the total number of production batches. Four alternative tenure
sizes were set for this research: 0.6n, 0.9n, 1.2n and 1.5n. The second scenario from
table 1 is used to test the performance of the four tenure sizes. The results are shown
in figures 4 and 5 for TS1 and TS2, respectively. The empirical results suggest the

0.6n 0.9n 1.2n 1.5n

850
Average tardiness (min)

800
750
700
650
600
550
1 51 101 151
Number of iterations

Figure 4. TS1 Performance versus four Tabu tenure sizes.

 4022 T. Yang et al.

0.6n 0.9n 1.2n 1.5n

850

Average tardiness (min)

800
750
700
650
600
550
1 51 101 151
Number of iterations

Figure 5. TS2 Performance versus four Tabu tenure sizes.

Downloaded by [Aston University] at 08:31 27 August 2014

tabu tenure-sizes for TS1 and TS2 are 1.2n and 0.9n, respectively. This set-up is then
used hereafter.

3.8. Stopping criteria

The stopping criterion is the maximum number of search iterations. For a given
sequence, the total number of possible swaps is equal to n(n  1)/2. The present
authors feel that the number of iterations should be a multiple of the total
number of possible swaps. As a result of a consideration of figures 4 and 5, the
stopping criteria for TS1 and TS2 were both set as n(n  1). Given the following
notation, the overall TS procedure is illustrated in figure 6:
T tabu tenure size,
Mcount stopping criterion (the maximal number of search iterations),
count index of search iteration,
Z* current best solution,
X* best solution,
Xnow current solution,
Xnext neighbourhood solution.
In figure 6, the ‘restart’ module uses the EDD rule and intensification–
diversification long-term memory to determine the scheduling sequence for its
next iteration for TS1 and TS2, respectively. In other words, TS1 uses the same
restart sequence (as determined by EDD), whereas TS2 has a changing one (as
determined by the long-term memory). Due to the inherent random moves of
TS, TS1 can still have improved solutions given the same restart sequence as the
search progresses. This restart strategy may impose a constraint to the search
algorithm as contrast to TS2. In general, the long-term memory strategy is
more robust to obtain a quality solution. It may have a trade-off between
the increased memory size and improved solution quality depending on its
application areas.

4. Empirical illustrations
The five production scenarios in table 1 are used for the empirical illustrations.
A steepest descent pairwise interchange (SDPI) heuristic is used to solve the
 Tabu-search simulation optimization approach for flow-shop scheduling 4023

Define Tabu tenure,

Mcount, Count=0

Select an initial
solution X now

Let X *=X now,

Z *=Z(X now)
Yes

Randomly make a No X now better

Restart Count=Count+1 non-tabu move then than X *?
get X next
No
Downloaded by [Aston University] at 08:31 27 August 2014

All moves not No Z(X next)

in tabu list better than
have been Z(X now)?
searched?
Yes
Yes
Search all moves in
Tabu list

No Aspiration
criterion
satisfied?

Yes
Xnow=Xnext,
Select the best X next Count=Count+1
in Tabu list

Update tabu
attributes

No
Count>Mcount?

Yes
Stop

Figure 6. The proposed TS procedure.

problem for benchmarking the quality of the solution against the proposed meth-
odologies. For a detailed discussion of the SDPI procedure, see Tompkins et al.
(1996). Since the different due-date settings will influence the resulting performance
measure, there is a need to collect more testing samples to illustrate the effective-
ness and efficiency of the proposed methodologies. We decided to collect five
samples through some preliminary studies. Thus, for each production scenario,
 4024 T. Yang et al.

Average tardiness (min)

Scenario number EDD SDPI TS1 TS2

1 3744.93 2907.70 1751.44 1835.77

2 2929.88 2504.44 504.02 797.67
3 2642.37 2137.33 991.07 730.70
4 491.10 445.30 221.96 206.57
5 1294.12 905.29 508.25 564.12

Table 2. Summary of tardiness performance.

Scenario SDPI versus TS1 versus TS2 versus TS2 versus

number EDD SDPI SDPI TS1
Downloaded by [Aston University] at 08:31 27 August 2014

1 22.36 39.77 36.87 4.82

2 14.52 79.88 68.15 58.26
3 19.11 53.63 65.81 26.27
4 9.33 50.16 53.61 6.93
5 30.05 43.86 37.69 10.99

Data are percentages.

Table 3. Summary of tardiness improvement.

Average computational time (s) Ratio

Scenario number (1) SDPI (2) TS1 (3) TS2 (2)/(1) (3)/(1)

1 107 109 109 1.02 1.01

2 335 254 266 0.76 0.79
3 883 632 718 0.72 0.81
4 1106 1031 974 0.93 0.88
5 2051 1549 1335 0.76 0.65

Table 4. Summary of computational time for the tardiness measure.

five different due-date settings are generated to illustrate the robustness of

the proposed solution methods. The distributions in the last column of
table 1 are used for the due-date generation. The solution by the simple EDD
rule is also collated. Tables 2 and 3 show the performance of the proposed
methodologies.
Table 2 shows the resulting performances from different sequencing methods.
Table 3 shows the performance improvement between different methods. Both
tables 2 and 3 show that the TS algorithms consistently outperform the other two
algorithms, EDD and SDPI. There is not an obvious difference in performance
between TS1 and TS2. The solution quality by TS is significantly better than that
by SDPI. The results also show that the simple EDD rule cannot provide a quality
solution, despite its ease of implementation.
Table 4 and figure 7 show the computational efficiency of the proposed TS
methodologies that require less computing time than the time for SDPI for most
instances. The longest TS computing time is 1549 s (i.e. 25.82 min) for the largest
 Tabu-search simulation optimization approach for flow-shop scheduling 4025
2500

2000

Time(sec) SDPI
1500
TS1
1000 TS2
500

0
1 2 3 4 5
Scenario
Figure 7. Trend of computational time for tardiness measure.
Downloaded by [Aston University] at 08:31 27 August 2014

Flow time (min)

Scenario number EDD SDPI TS1 TS2

1 22 404.75 21 798.56 19 362.52 19 126.91

2 31 690.67 30 301.57 26 168.12 26 465.11
3 35 842.92 34 271.44 30 898.63 30 063.39
4 44 105.79 42 392.48 36 341.82 36 470.89
5 51 219.19 48 566.98 42 569.32 42 841.55

Table 5. Summary of flow time performance.

Scenario SDPI versus TS1 versus TS2 versus TS2 versus

number EDD SDPI SDPI TS1

1 2.71 11.18 12.26 1.22

2 4.38 13.64 12.26 1.13
3 4.38 9.84 12.28 2.70
4 3.88 14.27 13.97 0.36
5 5.18 12.35 11.79 0.64
Data are percentages.
Table 6. Summary of flow time improvement.

problem size (production scenario 5). In general, the proposed TS methodologies

are quite efficient in solving a practical size problem. However, their efficiency can
be expected to deteriorate as problem size increases.
To gain further insights into the proposed case study, we used flow time (or
makespan) as another performance measure to test the proposed methodologies.
This experiment kept the same tabu parameters set-up and the same production
scenarios used in the earlier experiments. The summary of flow time performance
measure is shown as tables 5 and 6. Table 7 and figure 8 show the summary of
computational time.
The experiment for the flow time measure also illustrated the effectiveness
and efficiency of the proposed methodologies. Accordingly, it shows promise
 4026 T. Yang et al.

Average computational time (s) Ratio

Scenario number (1) SDPI (2) TS1 (3) TS2 (2)/(1) (3)/(1)

1 121 74 116 0.61 1.96

2 303 258 256 0.85 0.84
3 696 609 614 0.88 0.88
4 1021 995 997 0.97 0.98
5 1828 1636 1554 0.89 0.85

Table 7. Summary of computational time for flow time measure.

2000

1500
Time(sec)

SDPI
Downloaded by [Aston University] at 08:31 27 August 2014

1000 TS1
TS2
500

0
1 2 3 4 5
Scenario

Figure 8. Trend of computational time for flow time measure.

for the proposed methodologies in solving a complex FSMP sequencing

problem.

5. Conclusions and future research opportunities

The present paper has presented a TS simulation optimization approach to
solving an FSMP scheduling sequence problem. A case study from an MLCC
manufacturing system was used to illustrate the proposed methodologies.
Empirical results showed promise for the capacity of the proposed methodologies
to solve a practical application. Accordingly, it would seem that the proposed meth-
odologies could potentially solve many other planning and control problems in
FSMP-type manufacturing systems. Although the scheduling issue was solved in
this research, the dispatching decision was not addressed here, and this might repre-
sent an opportunity to improve the quality of the solution. A future research oppor-
tunity would be to investigate the impact on the FSMP problem from the
dispatching decision onwards. An integrated scheduling and dispatching decision
is another issue that is worthy of investigation. Finally, the link between a shop
floor control system and the proposed decision module would be an important step
towards the implementation of a real-time planning and control system. This repre-
sents another future research direction.

Appendices
The appendices contain the data for the case study. Appendix A1 collects infor-
mation on the machine inventory, set-up time and downtime. Appendix A2
shows lot-size information. Appendices A3–5 collect the detailed processing-time
information.
 Tabu-search simulation optimization approach for flow-shop scheduling 4027

A1. Basic data

Number of
No. Stage name machines Set-up time (min) MTBF1 (min) MTTR2 (min)

1 Stacking 17 uniform (27, 33)* Expo (30) Expo (5)

2 Pressing 1 0 0 0
3 Cutting 4 uniform (25, 33)y Expo (2) Expo (0.17)
4 Binder burn out 16 0 0 0
5 Sintering 2 constant (1440)* 0 0
6 Tumbling 8 0 Expo (54) Expo (15)
7a Dipping 4 uniform (29, 35)y Expo (25) Expo (3)
7b Curing 4 constant (1440)* Expo (35) Expo (2)
8 Plating 1 uniform (1, 2)*y Expo (72) Expo (7)
9 Testing 12 uniform (20,29)y Expo (2) Expo (0.17)
*Set-up needed for the material change.
Downloaded by [Aston University] at 08:31 27 August 2014

ySet-up needed for the size change.

1
Mean time between failures.
2
Mean time to repair.

A2. Lot-size data

The notation for product size and material types is a convention of the case study
company. To protect proprietary information, the company’s definitions are not
provided.

Material types
4
Lot size (
10 ) K12000 K3400 K3000 K90 K60 K30

Product sizes
0402 149.5 89.7 77.5 114.4 131.3 95.2
0603 79.3 41.6 39.6 67.6 53.3 45.3
0805 31.0 16.0 13.0 26.0 21.0 18.1
1005 21.7 9.6 7.4 21.9 15.1 12.0

A3. Material-type-dependent process-time data

The process time for the following stages is dependent on material type, regard-
less of product size. The unit of measure is minute-per-lot.

Material types

Stage name K30 K60 K90 K3000 K3400 K12000

Stage 1: stacking 393.00 555.85 694.80 296.73 446.58 823.60

Stage 2: pressing 39.03 44.05 44.05 39.05 39.05 22.05
Stage 3: cutting 132.00 176.16 220.20 147.00 147.00 344.88
Stage 4: binder burn out 2714.17 2114.17 2534.17 3434.17 2714.17 3434.17
 4028 T. Yang et al.

A4. Product-type-dependent process-time data

The process time for the following stages is dependent on product type. The unit
of measure is minute-per-lot.

Material types

Product sizes K12000 K3400 K3000 K90 K60 K30

Stage 6: tumbling
0402 399.17 361.90 374.67 555.83 525.00 690.00
0603 489.17 459.17 450.00 689.17 689.17 749.17
0805 541.82 543.67 450.00 753.83 772.50 772.50
1005 579.17 603.67 483.67 818.83 872.50 832.50
Stage 8: plating
0402 336.00 336.00 336.00 296.00 336.00 336.00
0603 276.00 306.00 306.00 316.00 306.00 306.00
Downloaded by [Aston University] at 08:31 27 August 2014

0805 296.00 256.00 256.00 286.00 296.00 296.00

1005 316.00 276.00 276.00 366.00 316.00 316.00
Stage 9: testing
0402 836.58 560.58 560.58 674.58 752.58 752.58
0603 512.58 338.58 338.58 458.58 392.58 392.58
0805 332.58 242.58 242.58 302.58 272.58 272.58
1005 242.58 216.58 216.58 242.58 212.58 212.58

A5. Single process-time data

The process time for the following stages is independent of product-type. The
unit of measure is minute-per-lot.

Stage 5: sintering Stage 7a: dipping Stage 7b: curing

1593.00 105.02 62.28

Acknowledgements
The authors thank the anonymous company for providing the case study.
Work was supported in part by the National Science Council of Taiwan, Republic
of China, under Grant NSC91-2622-E006-035-CC3.

References
ALVAREZ-VALDES, R., CRESPO, E. and TAMARIT, J. M., 2002, Design and implementation of
a course scheduling system using tabu search. European Journal of Operational
Research, 137, 512–523.
AZZARO-PANTEL, C., BERNAL-HARO, L., BAUDET, P., DOMENECH, S. and PIBOULEAU, L., 1998,
A two-stage methodology for short-term batch plant scheduling: discrete-event simula-
tion and genetic algorithm. Computers and Chemical Engineering, 22, 1461–1481.
BARNES, J. W. and LAGUNA, M., 1993, Solving the multiple-machine weighted flow time
problem using tabu search. IIE Transactions, 25, 121–128.
BEN-DAYA, M. and AL-FAWZAN, M., 1998, A tabu search approach for the flowshop schedul-
ing problem. European Journal of Operational Research, 109, 88–95.
BOTTA-GENOULAZ, V., 2000, Hybrid flowshop scheduling with precedence constrains and time
lags to minimize maximum lateness. International Journal of Production Economics, 64,
101–111.
 Tabu-search simulation optimization approach for flow-shop scheduling 4029

BRAH, S. A., 1992, Complexity of the flow shop with multiple processors scheduling
problem, and some dominance conditions. In K. H. Phua, C. M. Wang, W. Y.
Yeong, T. Y. Leong, H. T. Loh, K. C. Tan and F. S. Chou (eds), Optimization:
Techniques and Applications, Vol. 1 (Singapore: World Scientific), pp. 538–545.
CAMPBELL, H. G., DUDEK, R. A. and SMITH, M. L., 1970, A heuristic algorithm for the n-job,
m-machine sequencing problem. Management Science, 16, B630–B637.
DEAL, D. E., YANG, T. and HALLQUIST, S. J., 1994, Job scheduling in petrochemical produc-
tion: two-stage processing with finite intermediate storage. Computers and Chemical
Engineering, 18, 333–344.
DORN, J., GIRSH, M., SKELE, G. and SLANY, W., 1996, Comparison of iterative improvement
techniques for schedule optimization. European Journal of Operational Research, 94,
349–361.
GLOVER, F., 1989, Tabu search — Part I. ORSA Journal on Computing, 1, 190–206.
GLOVER, F., KELLY, J. P. and LAGUNA, M., 1995, Genetic algorithms and tabu search: hybrids
for optimization. Computers and Operations Research, 22, 111–134.
GLOVER, F. and LAGUNA, M., 1993, Tabu search. In C. Reeves (ed.), Modern Heuristic
Techniques for Combinatorial Problems (Oxford: Blackwell), pp. 70–150.
Downloaded by [Aston University] at 08:31 27 August 2014

GRANGEON, N., TANGUY, A. and TCHERNEV, N., 1999, Generic simulation model for hybrid
flow-shop. Computer and Industrial Engineering, 37, 207–210.
GUPTA, J. N. D., 1988, Two-stage, hybrid flowshop scheduling problem. Journal of the
Operational Research Society, 39, 359–364.
GUPTA, J. N. D. and TUNC, E. A., 1991, Schedules for a two-stage hybrid flowshop with
parallel machines at the second stage. International Journal of Production Research,
29, 1489–1502.
HARRIS, J. M. and ROESCH, E. B., 2001, US Passive Component Suppliers (New York: UBS
Warburg LLC).
HIGGINS, A. J., 2001, A dynamic tabu search for large-scale generalized assignment problems.
Computers and Operations Research, 28, 1039–1048.
HOPP, W. J. and SPEARMAN, M. L., 2001, Factory Physics, 2nd edn (New York: McGraw-Hill).
HSIEH, B. W., CHEN, C. H. and CHANG, S. C., 2001, Scheduling semiconductor wafer fabrica-
tion by using ordinal optimization-based simulation. IEEE Transactions on Robotics and
Automation, 17, 599–608.
JEONG, K. Y., 2000, Conceptual frame for development of optimized simulation-based sched-
uling systems, Expert Systems with Applications, 18, 299–306.
NAWAS, M., ENSCORE, E. and HAM, I., 1983, A heuristic algorithm for the m machine, n job
flow shop sequence problem. Omega, 11, 91–95.
NEGENMAN, E. G., 2001, Local search algorithms for multiprocessor flow shop scheduling
problem. European Journal of Operational Research, 128, 147–158.
PINEDO, M., 1995, Scheduling: Theory, Algorithms, and Systems (Englewood Cliffs: Prentice
Hall).
PROUST, C., GUPTA, J. N. D. and DESCHAMPS, V., 1991, Flowshop scheduling with set-up,
processing and removal times separated. International Journal of Production Research,
29, 479–493.
SALVADOR, M. S., 1973, A solution to a special case of flow-shop scheduling problems. In S. E.
Elmaghraby (ed.), Symposium of the Theory of Scheduling and Applications (New York:
Springer), pp. 83–91.
SANTOS, D. L., HUNSUCKER, J. L. and DEAL, D. E., 1996, An evaluation of sequencing
heuristics in flow shops with multiple processors. Computers and Industrial
Engineering, 30, 681–692.
TAILLARD, E. 1989, Parallel taboo search technique for the job shop scheduling
problem. Research Report ORWP 89/11, Ecole Polytechnique Federale de Lausanne,
Department de mathematiques.
TECNOMATIX TECHNOLOGIES, 2000, eM-Plant User’s Manual, Version 4.6.1 (Stuttgart:
Tecnomatix Technologies).
TOMPKINS, J. A., WHITE, J. A., BOZER, Y. A., FRAZELLE, E. H., TANCHOCO, J. M. A. and
TREVINO, J., 1996, Facilities Planning, 2nd edn (New York: Wiley).
TOWNSEND, W., 1977, Sequencing n-jobs on m-machines to minimize maximum tardiness:
a branch-and-bound solution. Management Science, 23, 1016–1019.
 4030 T. Yang et al.

UETAKE, T., TSUBONE, H. and OHBA, M., 1995, A production scheduling system in a hybrid
flow shop. International Journal of Production Economics, 41, 395–398.
WEN, U. P. and HUANG, A. D., 1996, A simple tabu search method to solve the mixed-integer
linear bilevel programming problem. European Journal of Operational Research, 88,
563–571.
WIDMER, M. and HERTZ, A., 1989, A new method for the flow sequencing problem. European
Journal of Operational Research, 41, 186–193.
YANG, T. and TSENG, L., 2002, Solving a multi-objective simulation model using a hybrid
response surface method and lexicographical goal programming approach — a case
study on integrated circuit ink-marking machines. Journal of the Operational Research
Society, 53, 211–221.
YANG, T., RAJASEKHARAN, M. and PETERS, B. A., 1999, Semiconductor fabrication facility
design using a hybrid search methodology. Computer and Industrial Engineering, 36,
565–583.
Downloaded by [Aston University] at 08:31 27 August 2014