ARTICLE IN PRESS

Int. J. Production Economics 107 (2007) 104114 www.elsevier.com/locate/ijpe

An evolutionary-simulation approach for the optimization of multi-constant work-in-process strategyA case study
Taho Yanga,, Hsin-Pin Fub, Kuang-Yi Yanga
b

Institute of Manufacturing Engineering, National Cheng Kung University, Tainan 701, Taiwan Department of Marketing and Distribution Management, National Kaohsiung First University of Science and Technology, Kaohsiung, Taiwan Received 18 May 2005; accepted 19 February 2006 Available online 28 November 2006

Abstract This research addressed an evolutionary-simulation optimization approach in solving a multi-constant work-in-process (multi-CONWIP) pull strategy problem. A case study is adopted to illustrate the performance of the applied methodology. Both the single-loop CONWIP and the just-in-time (JIT) control strategies are a special case of the proposed multiCONWIP strategy. Empirical results showed that the proposed multi-CONWIP strategy outperformed the single-loop CONWIP and JIT for the case study problem. In general, the proposed methodology was effective and robust for the proposed problem. However, computational complexity existed for some instances; and these computational efciency problems are addressed as potential future research opportunities. r 2006 Elsevier B.V. All rights reserved.
Keywords: Evolutionary algorithms; Discrete-event simulation; Constant work-in-process; Just-in-time; Integrated circuit packaging

1. Introduction An effective production control system is a critical factor in achieving manufacturing excellence. A pull control strategy authorizes a job release, whilst a push control schedules it. In general, the pull control strategy is more robust than a push strategy in terms of cycle time and work-in-process (WIP) performances. Cycle time is the average time for a lot to traverse the production lines. The advantages of the pull over the push may

Corresponding author. Tel.: +886 6 2090780; fax: +886 6 2085334. E-mail address: tyang@mail.ncku.edu.tw (T. Yang).

trade-off with throughput performance if it is not well planned in advance. Kanban and Constant WIP (CONWIP) are the two most commonly seen pull strategies. Kanban is a method for the implementation of the well-known just-in-time (JIT) production. Golhar and Stamm (1991) identify over 860 JIT articles in professional journals since 1970. JIT has evolved into a management philosophy rather than a specic method. The CONWIP strategy aims at combining the low WIP levels of Kanban with the high throughput of push systems (Spearman et al., 1990). There is a signicant body of literature dedicated to the discussions of JIT and CONWIP strategies (for examples, Spearman and Zazanis, 1992; Paris and Pierreval, 2001; Cao and Chen, 2005; Takahashi

0925-5273/$ - see front matter r 2006 Elsevier B.V. All rights reserved. doi:10.1016/j.ijpe.2006.02.014

ARTICLE IN PRESS
T. Yang et al. / Int. J. Production Economics 107 (2007) 104114 105

WIP upper limit = 10

WIP upper limit = 20

Raw materials

1 Stage

3 Inventory

5 Finished goods Inventory

Fig. 1. Multi-CONWIP illustration.

and Hirotani, 2005). Hopp and Spearman (2000) concluded that the WIP cap set by a pull strategy is the essence that makes the difference between pull and push strategies. In fact, most of the existing literature has been dedicated to the WIP setting decisions in a pull environment. Bonvik et al. (1997) proposed the concept of a hybrid Kanban/CONWIP strategy whose objective is to combine the advantages of CONWIP with those of Kanban by adding JIT cells to CONWIP (they call it hybrid KanbanCONWIP). They compared the difference policies in a four-stage tandem production line using simulation. Each of the policies was compared using constant demand and demand that has a stepped increase/decrease. They found that the CONWIP/pull policy decreased inventories by 1020% over the Kanban policy while maintaining the same service levels. Gaury et al. (2000) and Gaury et al. (2001) proposed a generic model for the CONWIP/pull control strategy. It incorporates the characteristics of all pull strategies that can model a Kanban, a CONWIP, or a hybrid Kanban/CONWIP system, also allows customized pull control policies to be developed. Simulation and evolutionary algorithm (EA) were used to study the generic model. Geraghty and Heavey (2004) performed a comprehensive comparison of hybrid push/pull and CONWIP/pull production control policies. In a hybrid push/pull system, some production stages are controlled by push-type control and other stages by a pull-type control (Cochran and Kim, 1998). Through a simple analysis, Geraghty and Heavey (2004) showed the optimal hybrid push/pull system is effectively CONWIP/pull. The majority of the existing literature for the CONWIP/pull control strategies used a conceptual model or a simplied problem to illustrate its effectiveness. Whilst the potential benets of a CONWIP/pull system is well-addressed in literature, its applicability in solving a practical size problem is not well-documented in literature, particularly, for a complex manufacturing system.

The present research aims at solving a multiCONWIP design which is illustrated by Fig. 1. In Fig. 1, the ve-stage production line is divided into two segmentsstages 12 and stages 35, and then forms two CONWIP loops. The rst and the second loops have WIP upper limits of 10 and 20, respectively. The WIP upper limit regulates the WIP level for each CONWIP loop and triggers the pull mechanism. As opposed to the single CONWIP design in literature, the present study proposes a multiCONWIP conguration. When the number of CONWIP loops is known, the present study solves both the segmentation and WIP upper limit decisions to optimize the system performances. When the number of CONWIP loops is equivalent to the number of process stages, it is equivalent to the Kanban system. An applied case study is adopted for empirical illustrations. This required many realistic constraints to be considered that included: machine failures, setup requirements, and multiple machines in a workstation. By the use of a practical application, the conceptual pull control strategy can be realized in a practical sense. This research proposes to solve the pull strategy optimization problem by using an evolutionarysimulation-optimization approach. Simulation can model non-linear and stochastic problems and allow examination of the likely behavior of a proposed manufacturing system under selected conditions. It can take into account many details and constraints in evaluating the performance of a system. However, it does not provide a method for optimization. EA is known as one of the metaheuristics and is proven to be an effective tool in solving an optimization problem (Reeves, 2003). EA has been used to solve a variety of industrial applications, and is one of the active research areas in computational intelligence (Ovaska et al., 2002). The combination of simulation and EA embodies both the modeling capability of simulation and the efcient local search algorithms of EA.

ARTICLE IN PRESS
106 T. Yang et al. / Int. J. Production Economics 107 (2007) 104114

The remainder of this paper is organized as follows. In Section 2, the details of the proposed case study are provided. The proposed EA is discussed in Section 3. In Section 4, numerical results are provided to illustrate the proposed methodology. Section 5 contains conclusions and discussions of future research opportunities.

2. The case study problem An anonymous, market leader, integrated circuit (IC)-packaging company, located in Kaohsiung, Taiwan that provides an IC-packaging service to customers worldwide, was used to illustrate the proposed methodology. The manufacture of an IC has three major stages: (i) wafer manufacturing; (ii) wafer processing; and (iii) packaging and nal test. In the rst stage, wafer is manufactured from high-purity molten silicon. The wafer-processing stage then achieves full integration of the original circuit design onto a die on a wafer, as illustrated in Fig. 2 (Yang and Tseng, 2002). Each die in Fig. 1 represents an individual circuit chip. A nished wafer invariably contains 100200 dies, depending on the size of the wafer and the size of the die. A circuit chip must be sealed for mechanical and environmental protection. IC-packaging processing separates a die from the wafer, which is then put on a lead frame for a connecting circuit and an application device, as illustrated in Fig. 3(a). It is then encapsulated with plastic, as illustrated in

Fig. 3(b). Next, the chip is separated from the lead frame, as illustrated in Fig. 3(c). Finally, it is tested for its required functions. The company case had three product types: TF, TG, and TK. There are 18 processing stages for the encapsulation of an IC chip. The process stages associated with process time information are depicted in Appendix A. All products have the same process ow that starts from stage 1 of wafer mount to stage 18 of packaging. However, different products might require different process times at the same stage. The number of machines is varying among different stages due to the variation in single-machine throughput. Note that all of the primary data were physically collected directly from the companys shop oor. The raw data were analyzed using commercial statistical software, Expert Fit (1998), to nd their associated statistical distributions (for succinctness these are not shown in this paper). Some proprietary information collected was modied; and additionally the following assumptions were made. Moreover, the present data are considered to be sufciently informative for the development of the proposed methodology.

  

All the machines in a workstation (stage) are identical. There is no handling time between consecutive workstations. There is no yield loss and rework.

Fig. 2. A nished wafer.

The present study aims to apply the proposed multi-CONWIP strategy in solving the case study problem. The order due dates are not explicitly considered, which is similar to the practice of a commonly seen pull-based control strategy. The raw materials for the planned orders are available at the beginning of the planning period. Then, the orders are processed sequentially. The different orders will represent different products for the present study

Fig. 3. IC packaging.

ARTICLE IN PRESS
T. Yang et al. / Int. J. Production Economics 107 (2007) 104114 107

and there is a setup for product changeover at a machine. The required setup time for different machines can also be found in Appendix A. Currently, the company adopts its Manufacturing Resources Planning system to generate its production schedule, and hence is using a push-based production control strategy. The company would like to know whether the pull-based strategy is a better alternative method. The proposed multiCONWIP strategy represents a generic pull-based control method since both Kanban and single-loop CONWIP are a special case of the multi-CONWIP strategy. For the case study, the number of CONWIP loops is ranging from 1 to 18. It is the instance of the single CONWIP loop when the loop comprises all of the 18 process stages. On the other hand, it is the case of Kanban when there are 18 CONWIP loops. The multi-CONWIP strategy allows the system conguration to be between the above two extremes, and thus has better modeling exibility. The proposed method requires that the number of loops is known a priori, the loop segmentation and WIP upper limit for each segmented loop are then solved accordingly. The evolutionary-simulation method, therefore, needs to solve each of the 18 segmentation scenarios to nd the optimal control strategy. 3. Proposed methodology

and are often real numbers in the case of EA (Michalewicz, 1992). EA makes the initial population evolve towards a population that is expected to contain the optimum solution. They use the following reproduction evaluation cycle for each iterationreferred to as generation. Individuals from the current population are selected with a given probability and copies of these individuals are created. The selection of individuals is based on their tness relative to the current populationthat is, the stronger individuals will have a higher probability of being copied. The tness is a function of the simulation models response (Pierreval and Tautou, 1997). Selected individuals are subjected to mutation and to recombination (cross-over in genetic algorithms). The processwhich contains selection, mutation, and recombinationis called reproduction. The recombination mechanism allows for the mixing of parental information in passing it on to their descendants (offspring). Mutation introduces innovation into the population. From one generation to the next, the population tends, globally, to have a better t (Paris and Pierreval, 2001). The following EA schema is used in this study to solve the proposed simulation-optimization problem (Muhlenbein, 1993).

 
The section describes the general principle of EA followed by the details of the modeling methodology. 3.1. General EA principles EA includes various methodssuch as genetic algorithms, evolution strategies, evolutionary programming, and genetic programming. Back and Schewefel (1993) discussed the classication of the different methods. However, there are many evolutionary systems that are more difcult to classify. One of the principles of EA is that the strongest individuals of a population have more chance of survival. EA begins their search for the optimal solution from a set of potential solutions (p solutions), which is called the population with size of p. The kth solution from the population is coded through a chromosome, which is represented by a vector Xk, whose components (genes) are binary elements in the particular case of genetic algorithms,

  

Step 0: Dene appropriate coding of the solutions. Step 1: Create an initial population Pt X 1 ; X 2 ; . . . ; X p ; set t 0. Step 2: Evaluation: compute the tness f(Xk) for each individual Xk, k 1; 2; . . . ; p. Step 3: Selection: select individuals from P(t); this gives the mating pool S(t). Any given individual from P(t) can appear several times in S(t). Step 4: Recombination and mutation: pair individuals of S(t); and, for each pair of individuals: recombine the pair with probability pcross and copy the resulting offspring into the set S(t+1) (the pair of chromosomes being eliminated and replaced by its offspring); and copy the pair of individuals in the set S(t+1) with probability 1pcross. For each individual of S(t+1): apply mutation to the individual with probability pmut, and copy the transformed individuals in P(t+1); and copy the individual into P(t+1) with probability 1pmut.

ARTICLE IN PRESS
108 T. Yang et al. / Int. J. Production Economics 107 (2007) 104114

Step 5: Set t t+1, and return to step 2 until the stopping criterion is reached.

In step 2, the tness of a chromosome is evaluated by simulation. The interactions between EA and simulation can be illustrated in Fig. 4. 3.2. EA coding The present study solves the multi-CONWIP decision in a sequential manner. Let S be the number of CONWIP loops in a production system and be known a priori. The proposed EA then optimizes the multi-CONWIP design problem. The notation used in this paper is summarized as follows. Input parameters N S Wj the total number of process stages the number of CONWIP loops the WIP cap for the jth CONWIP loop

Eq. (1) represents the kth solution from the population size of p. For Fig. 1 example, S is equal to 2. Then, the kth solution, X k fck1 ; wk1 ; wk2 g f2; 10; 20g. The three data elds represent the three genes in a chromosome. The rst CONWIP loop contains stages 1 and 2 (since ck1 2) and has WIP upper limit of 10. The second CONWIP loop includes stages 35, and has WIP cap of 20. Note that the WIP upper limit must be less or equal to its associated WIP cap. If the WIP caps for loops 1 and 2 are 15 and 30, respectively. Then, wk1 can be ranging from 1 to 15 and wk2 can be ranging from 1 to 30. The performance of a chromosome is evaluated by computer simulation. The optimum solution that determines both the segmentation and WIP upper limits are solved by EAsimulation methodology. 3.3. WIP cap Since wkj belongs to the set of f1; 2; . . . ; W j g, the value of Wj is one of the determinants for the EA solution space. The Wj value setup is a trade-off between solution space and computational efciency. Thus, the larger the value of Wj, the larger the solution space. However, it may trade-off with increased computational time. Thus, it is important to set an adequate WIP cap for each CONWIP loop. This research uses Littles Law as Eq. (2) to estimate each Wj value. Note that the details of Eq. (2) can be found in Hopp and Spearman (2000, pp. 223224) WIP cycle time throughtput rate: (2)

Decision variables cki the ith separation stage for the kth solution, cki 2 f1; 2; . . . ; Ng, k 1; 2; . . . ; p, i 1; 2; ...;S 1 the WIP upper limit for the jth CONWIP for the kth solution, wkj 2 f1; 2; . . . ; W j g, k 1; 2; . . . ; p, j 1; 2; . . . ; S

wkj

The index k indicates the kth solution (chromosome) out of the population size of p. Then the chromosome that represents a solution is shown as below X k fck1 ; ck2 ; . . . ; ckS1 ; wk1 ; . . . ; wkS g.
Termination

(1)

The historical cycle time performance at the case study company was approximately 80 h. The bottleneck, wire-bonding workstation (stage 6), was used to represent the line throughput rate. The overall
Initial population generation

EA encoding

Mutation

EA search procedure

Evaluation

Simulation

Recombination

Selection

Fig. 4. EAsimulation methodology.

ARTICLE IN PRESS
T. Yang et al. / Int. J. Production Economics 107 (2007) 104114 Table 1 Normalized process times TK No. of dies per lot Process time (1000 dies per hour) Process time per lot (h) Normalized process time 640 0.063 10.16 1.000 TG 720 0.114 6.32 0.628 TF 720 0.135 5.33 0.525 109

3.5. Recombination This research adopted a two-point strategy, as shown in Fig. 5 (adopted from eM-Plant Users Manual, 2001). Whilst a single-point recombination strategy is often used, its performance is usually relatively poor compared with one that has more cut points (Man et al., 1996). The recombination operation is applied to two chromosomes. Initially, two intersecting points are selected randomly, and the ranges between these two points are then exchanged. The recombination operators can guide the direction to search for an improved solution because there is a high probability that short and good ranges are preserved, and will therefore be reproduced in a growing number of solutions. 3.6. Mutation The mutation principle randomly selects a gene in a solution and replaces it with a new gene. Several selection strategies can be found in the literature (Michalewicz, 1992; Reeves, 2003). The mutation operation in this research exchanged two randomly chosen genes, as illustrated in Fig. 6 (adopted from eM-Plant Users Manual, 2001). Using the mutation operation, any solution of the search space can be randomly generated in the initial population. Additionally, any solution of the search space can be reached from any other solution using a nite sequence of mutations. 4. Empirical results The study horizon is 288 h (or 12 days), which is considered to be adequate to address the performance of the case study problem, since the current cycle time is approximately 80 h. The present study, rst, builds the proposed simulation-optimization model. It is then followed by numerical analyses. 4.1. Simulation-optimization modeling The proposed problem was modeled by eM-Plant Users Manual (2001)a form of discrete-event simulation software that has a hierarchical modeling structure and graphical interface. The EA was coded using the eM-plant embedded programming language (called SimTalk) for customized applications. Its syntax is similar to the popular programming language C++.

WIP cap for the line can then be estimated by Eq. (2). In order to take into account the difference products, we adopted the notion of weighted product as follows. The unit load is dened as a lot. The normalized lot process time for the three different products are summarized in Table 1. Subsequently, the throughput of the line in terms of normalized lot can be found by
WIPlot=line 49machines 80hr 63die=hr machine . 640dies=lot 385lot 3

Based on Eq. (3), the value of Wj is determined by W j lot 385lot S 90 for j 1; 2; . . . ; S. (4)

There are several underlining reasons for Eq. (4). Firstly, the summation of Wj, for j 1; 2; . . . ; S, must be less than or equal to the overall line WIP cap of 385 lots. Thus, it is unlikely to have a wkj that approaches the line WIP cap. An equally distributed WIP among S CONWIP loops seems to be a good start. The addition of 90 lots to each Wj is a good trade-off between solution quality and computational efciency. This claim has been veried through extensive preliminary numerical analyses. 3.4. Fitness and selection The production service level criterion was adopted as the responses. The service level is the percentage of demand completed by due date. The objective of this research is to maximize the service level by a simulation-optimization system. The most commonly used selection technique appears to be the wheel principle (Goldberg, 1989). The selection can be elitist if a given number of the best chromosomes are systematically retained in the next generation.

ARTICLE IN PRESS
110 T. Yang et al. / Int. J. Production Economics 107 (2007) 104114

Fig. 5. Recombination operation.

Table 2 Experiment for stopping criterion No. of generations 50 100 150 200 Service level (%) 80 83 91 91

Fig. 6. Mutation operation.

Although there is no doubt about the modeling capability of simulation, the modeling process often incurs signicant efforts. Both the simulation model and EA are conceptually straightforward; however, their implementations are application-specic. An experienced programmer is often a must to complete the coding process, particularly, for a complex manufacturing system, which is the instance of the present study. When stochastic simulation models are used, the tness should be computed from several replications to achieve an adequate condence interval. Preliminary study showed that 10 replications could achieve an acceptable condence interval of less than 5% of the sample mean. This was considered to be a conservative number for the tness evaluation. To commence the EA, it was rst necessary to determine several EA parameters. These parameters were: (i) population size; (ii) recombination rate; (iii) mutation rate; and (iv) stopping criterion. The present research adopted a local optimum approach in determining the EA parameters. Following an extensive literature review and preliminary exploratory study, the authors rst set the population size at 20. This is considered to be conservative in seeking a quality solution, and is adequate for a typical EA application.

The next step was to determine the proper stopping criterionthat is, the number of search generations. Both the recombination and mutation rates were 0.5, and the population size was 20. An initial production scenario is generated for the experimental purpose as follows. An assumption was made that the product mix for the planning horizon consisted of 500 lots for TF, 500 lots for TG, and 500 lots for TK. (Note that this production plan represents approximately 78% of line capacity.) Given the same initial production scenario, the present study increased the number of generations from 50 in increments of 50, until a total of 200 were reached. The results are collated in Table 2. Table 2 shows that an increase in the number of generations after 150 made no contribution to solution quality. The stopping criterion was therefore set at 150 generations. Given the above population size and stopping criterion, the study determined the recombination rate and the mutation rate using a three-level experimental design, as shown in Table 3. This led to probabilities of 0.8 and 0.5 for recombination and mutation rates, respectively. Finally, the EA was tested against a different population size of 30 using the previous parameters as shown in Table 4: that lead to the use of a population size of 30. Since the size of 30 is considered to be conservative and the computing time is becoming to be a burden, we did not explore a larger population size.

ARTICLE IN PRESS
T. Yang et al. / Int. J. Production Economics 107 (2007) 104114 Table 3 EA parameters setup Recombination rate Mutation rate (%) 0.3 0.3 0.5 0.8
a

111

Table 5 Production scenarios Scenario no. 0.8 92 87 71 1 2 3 Product type (lots/12-day) TF 500 100 1000 TG 500 700 0 TK 500 700 500

0.5 89 91 93a

91 91 92 The best solution.

Table 4 Experiment for population size Population size 20 30 Service level (%) 91 93

Table 6 Summary of experimental results S value Service level for different scenarios (%) 1 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18
a

2 93 97 98a 96 97 97 97 95 97 97 87 89 87 68 55 55 17 9

3 97 97 97 96 97 98a 98a 97 97 98a 97 91 95 90 90 90 35 29

4.2. Experimental results Three product-mix scenarios for a 12-day planning horizon were designed to illustrate the robustness of the proposed methodology in solving the pull control problem as shown in Table 5. For the initial product mix, the rst case had a production plan of 500 lots for TF, 500 lots for TG, and 500 lots for TK. It represented an evenly distributed production start for each product type. The second case had a production plan of 100 lots for TF, 700 lots for TG, and 700 lots for TK. The third case had only two product types: 1000 lots for TF and 500 lots for TK. The value of S ranged from 1 to 18; the results are summarized in Table 6. For scenario 1, the optimal solution has the design of either a single-loop CONWIP with WIP level of 287 lots or three CONWIP loops with WIP levels of 168, 87, and 82 lots. The optimal solution has there CONWIP loops for scenario 2. For scenario 3, the optimal design has 6, 7, or 10 CONWIP loops. The details of each design are shown as Table 7. The different production scenarios required alternative CONWIP strategies. The two extreme values of S are 1 and 18, that represent single-loop CONWIP and JIT pull strategies, respectively. Their comparisons with the proposed multi-CONWIP strategy are summarized in Table 8. In addition to the results of service level, we explored a different performance measurecycle time, for the case study problem. The EA process is

98a 96 98a 97 96 97 97 93 96 97 96 87 87 86 69 42 53 26 The best service level.

the same and is not repeated here. Table 9 summarizes the comparisons for cycle time. The JIT strategy has a very low service level, although its cycle time performance is good. The reason arises from the signicant setup time for some workstations: such as molding (Stage 8), electro deash (Stage 12), and solder plating (Stage 13). It appears that it is not suited to solve the proposed problem for the single-loop CONWIP strategy. From Tables 8 and 9, the proposed methodology was deemed to be effective and robust in solving the proposed problem. However, the computational efciency is a potential concern. The average computing time for each scenario of the above experiment ranged from approximately 13 min to 24 h; which is considered to be onerous. If the problem size were to increase, the proposed

ARTICLE IN PRESS
112 Table 7 The optimal solutions Scenarios 1 S 1 3 3 6 7 10 X* fC n ; wn g 0; 287 1 1 fC n ; C n ; wn ; wn ; wn g 10; 17; 168; 87; 82 1 2 1 2 3 fC n ; C n ; wn ; wn ; wn g 6; 15; 143; 53; 130 1 2 1 2 3 n n n n n n n n n n n C 1 ; C 2 ; C 3 ; C 4 ; C 5 ; w1 ; w2 ; w3 ; w4 ; w5 ; w6 7; 8; 10; 16; 17; 62; 91; 73; 41; 28; 12 fC n ; C n ; C n ; C n ; C n ; C n ; wn ; wn ; wn ; wn ; wn ; wn ; wn g 7; 8; 12; 14; 15; 17; 55; 101; 45; 23; 7; 17; 96 1 2 3 4 5 6 1 2 3 4 5 6 7 fC n ; C n ; C n ; C n ; C n ; C n ; C n ; C n ; C n ; wn ; wn ; wn ; wn ; wn ; wn ; wn ; wn ; wn ; wn g 1 2 3 4 5 6 7 8 9 5 7 1 2 3 4 6 8 9 10 8; 9; 10; 11; 12; 13; 14; 15; 17; 69; 12; 40; 28; 42; 16; 10; 26; 25; 47 T. Yang et al. / Int. J. Production Economics 107 (2007) 104114

2 3

Table 8 Comparisons for service level Scenario Service level (%) Single-loop CONWIP 1 2 3 Average 98 93 97 96 JIT 26 9 29 21 Multi-CONWIP 98 98 98 98

Table 9 Comparisons for cycle time Scenario Cycle time (min) Single-loop CONWIP 1 2 3 Average 3979 3494 4683 4052 JIT 2514 1854 1901 2090 Multi-CONWIP 3064 2851 2043 2653

methodology might become computationally prohibitive, and this might well require further development if used to solve the identied problem in such circumstances. 5. Conclusions and future research opportunities This research addressed an evolutionary-simulation optimization approach in solving a multiCONWIP pull strategy problem. A company-based case study was adopted to illustrate the performance of the proposed methodology. Both the single-loop CONWIP and the JIT control strategies are a special case of the proposed multi-CONWIP strategy. For the service level performance, empirical

results showed that the proposed multi-CONWIP strategy outperformed the single-loop CONWIP and JIT for the case study under research. The prominent cycle time performance of the JIT strategy traded off with its lowest service level performance; that is far below an acceptable level. In general, the proposed methodology was effective and robust for the proposed problem. However, its computational efciency may prohibit a larger size problem. The unwieldy computational execution existed for some instances of the study. The efciency of the built-in eM-Plant programming language to code the EA could be one of the problems since the runtime of its interpreter mode is very slow. It may be a better alternative to use an external programming language and to embed the code into eM-Plant. This requires a further investigation in the future. The adoption of a novel solution search accelerator to facilitate the solution search procedure is also worth investigation. For example, neural-network models have been adopted as an accelerator for simulation-optimization problems (Laguna and Mart , 2002). Another opportunity may simplify the proposed problem, i.e., eliminating some nonbottleneck workstations from the simulation model may reduce the size of the case study problem without scarifying solution quality. Finally, the merits of a simulation model were off set by its modeling complexity. It required a signicant amount of modeling time and experiences to build a simulation model to address the case study problem. This, in turn, becomes a barrier when adopting a simulation-optimization approach in solving a practical problem. Future research might seek alternative modeling approaches, and so lower this particular modeling barrier.

ARTICLE IN PRESS
T. Yang et al. / Int. J. Production Economics 107 (2007) 104114 Table A1
No. Stage MTBF (h) MTTR (h) No. machines Setup (min) Throughput (1000 pcs/h) TF 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 Wafer mount Die saw Inspection 1 Die bond Curing Wire bond Inspection 2 Molding First side mark Dejunk/trim Post mold cure Electro deash Solder plating Second side mark Inspection 3 Forming/ singulation Leg trimming Packaging 8 6 0 5+ERLA (5.88,2) 0.001+WEIB(12.3,1.02) Empirical N/A 0.01+WEIB(103,1.14) 5+69*BETA(0.81,0.996) 1+LOGN(5.57,3.84) Empirical 2+LOGN(5.68,4.5) 2+WEIB(6.21,1.95) Empirical N/A 2+Gamm(2.21,2.21) 0.001+WEIB(17.6,1.111) N/A 1.5 1 0 0.02+Gamm (0.1,5.77) Empirical Empirical N/A LOGN(0.323,0.258) LOGN(0.265,0.146) 1+LOGN(0.279,0.105) 0.001+0.691*BETA(0.305,0.927) 0.12+LOGN(0.235,0.126) 0.11+LOGN(0.265,0.13) NORM(0.299,0.147) N/A 0.06+Gamm(0.0427,5.92) Empirical N/A 2 4 3 9 4 49 10 8 3 8 8 3 3 6 N/A 12 7 N/A N/A 30 N/A 1520 N/A 30 N/A 480 5 3045 N/A 240 240 15 N/A 4560 N/A N/A 69.55 7.01 1.745 1.91 2.1 0.135 0.584 1.43 7.3 3.22 1.7 5.414 7.662 2.88 6.304 1.03 1.07 24.686 TG 69.55 7.01 1.309 1.91 2.1 0.114 0.534 1.57 7.3 3.22 1.7 14.4 14.4 2.88 6.893 1.85 1.07 27.491 TK 30.92 3.96 8.727 1.34 1.86 0.063 0.441 0.91 5.31 2.1 1.34 3.61 3.8 2.13 2.092 0.885 0.96 7.68

113

Acknowledgments The authors wish to thank the anonymous company for providing the data to allow development of this case study research. The National Science Council of Taiwan, ROC, supported this work, in part, under Grants NSC90-2212-E006-076 and NSC93-2213-E006-065. The authors also thank Ihui Chang for coding the eM-Plant program. Appendix A. IC-packaging stages This Appendix contains the detailed data from the case study company, which are used to build the simulation model (see Table A1). References
Back, T., Schewefel, H.P., 1993. An overview of evolutionary algorithms for parameter optimization. Evolutionary Computation 1, 123. Bonvik, A.M., Couch, C.E., Gershwin, S.B., 1997. A comparison of production line control mechanism. International Journal of Production Research 25, 789804. Cao, D., Chen, M., 2005. A mixed integer programming model for a two line CONWIP-based production and assembly system. International Journal of Production Economics 95, 317326.

Cochran, J.K., Kim, S.S., 1998. Optimum junction point location and inventory levels in serial hybrid push/pull production systems. International Journal of Production Research 36, 11411155. eM-Plant Users Manual, version 5.5, 2001. Tecnomatix Technologies, Stuttgart, Germany. ExpertFits users guide, 1998. Averill M. Law & Associates, Tucson, AZ. Gaury, E.G.A., Pierreval, H., Kleijnen, J.P.C., 2000. An evolutionary approach to select a pull system among Kanban, Conwip, and Hybrid. Journal of Intelligent Manufacturing 11, 157167. Gaury, E.G.A., Kleijnen, J.P.C., Pierreval, H., 2001. A methodology to customize pull control systems. Journal of the Operational Research Society 52, 789799. Geraghty, J., Heavey, C., 2004. A comparison of hybrid push/ pull and CONWIP/pull production inventory control policies. International Journal of Production Economics 91, 7590. Goldberg, D., 1989. Genetic Algorithms in Search Optimization and Machine Learning. Addison-Wesley Publishing, Harlow, England. Golhar, D.Y., Stamm, C.L., 1991. The just-in-time philosophy: A literature review. International Journal of Production Research 29, 657676. Hopp, J.P., Spearman, M.L., 2000. Factory Physics, second ed. Irwin, Chicago, IL. Laguna, M., Mart , R., 2002. Neural network prediction in a system for optimizing simulations. IIE Transactions 34, 273282. Man, K.F., Tang, K.S., Kwong, S., 1996. Genetic algorithms: Concepts and applications. IEEE Transactions on Industrial Electronics 43, 519534.

ARTICLE IN PRESS
114 T. Yang et al. / Int. J. Production Economics 107 (2007) 104114 Reeves, C., 2003. Genetic algorithms. In: Glover, F., Kochenberger, G.A. (Eds.), Handbook of Metaheuristics. Kluwer Academic, Boston, pp. 5582. Spearman, M.L., Zazanis, M.A., 1992. Push and pull production systems: Issues and comparisons. Operations Research 40, 521532. Spearman, M.L., Woodruff, D.L., Hopp, W.J., 1990. CONWIP: A pull alternative to Kanban. International Journal of Production Research 28, 879894. Takahashi, K., Hirotani, M.D., 2005. Comparing CONWIP, synchronized CONWIP, and Kanban in complex supply chains. International Journal of Production Economics 93-94, 2540. Yang, T., Tseng, L.-P., 2002. Solving a multiple objective simulation model using a hybrid response surface method and lexicographical goal programming approacha case study on IC ink marking machines. Journal of the Operational Research Society 53, 211221. Michalewicz, Z., 1992. Genetic Algorithms+Data StructureEvolution Programs. Verlag, Berlin. Muhlenbein, H., 1993. Evolutionary algorithms: Theory and applications. In: Narb, E.H.L., Lenstra, J.K. (Eds.), Local Search in Combinatorial Optimization. Wiley, New York. Ovaska, S.J., VanLandingham, H.F., Kamiya, A., 2002. Fusion of soft computing and hard computing in industrial applications: An overview. IEEE Transactions on Systems, Man, and CyberneticsPart C: Applications and Reviews 32, 7279. Paris, J.-L., Pierreval, H., 2001. A distributed evolutionary simulation optimization approach for the conguration of multiproduct Kanban systems. International Journal of Computer Integrated Manufacturing 14, 421430. Pierreval, H., Tautou, L., 1997. Using evolutionary algorithms and simulation for the optimization of manufacturing systems. IIE Transactions 29, 181189.