Genetics based machine learning approach to lot release scheduling in a semiconductor manufacturing system.

Fujii, N. ; Takasu, R. ; Kobayashi, M. 等

Abstract: This paper presents a new approach to lot release scheduling using genetics based machine learning (GBML). A scheduling agent with a classifier system (CS), which is one GBML method, is employed to generate product-dispatching rules that release materials onto the production floor according to the state of the production floor. Feasibility of the proposed CS based scheduling is assessed using results of the computer simulation in terms of turn around time (TAT), throughput, and the work in process (WIP) number.

Key words: Biological manufacturing systems, Genetics based machine learning, Classifier system, Lot release scheduling, Semiconductor manufacturing system

1. INTRODUCTION

A semiconductor manufacturing system is a representative example of complex manufacturing systems because it usually comprises hundreds of machines. A product consists of more than hundreds of recipes and its process flow is also complex. The huge problem size and the complex process relationship imply a huge number of combinations; also, the objective function is dynamic, not static. It is important to plan a proper production schedule to achieve high productivity and low cost in the production system. However, the specific properties of semiconductor manufacturing make it difficult to determine an optimal solution using traditional static optimization.

This study mainly addresses lot release scheduling problems in the semiconductor manufacturing. Some lot release scheduling methods have been proposed: constant work in process number (CONWIP; Spearman et al., 1990), workload regulation (WR; Wein, 1988) and starvation avoidance (SA; Glassey & Resende, 1988). To stabilize the work in process (WIP) throughout the production floor or part of it, CONWIP manages product-dispatching timing. In turn, WR regulates product release timing depending on the WIP number in the whole production floor to keep the workload of the bottleneck facility at a certain level. Products are released by SA so that the WIP number of a bottlenecked facility can maintain a proper level.

Current proposed lot release scheduling methods are reportedly effective and can determine a proper schedule to realize short turn around time (TAT) of products on a static production floor. However, it is difficult to find the premeditated static threshold of the scheduling criteria in a large-scale and complex production floor. Moreover, even if a suitable threshold value is found, the threshold value is no more effective in cases of a production situation change in a turbulent semiconductor-manufacturing environment. A new method must be developed that can plan a proper schedule without using static predetermined threshold of criteria.

This study adopts a biological manufacturing systems approach (BMS; Ueda, 1992), and proposes to employ genetics based machine learning (GBML) to control lot release timing and thereby realize adaptive product release scheduling using no premeditated thresholds in the scheduling criteria.

2. GBML-BASED LOT RELEASE SCHEDULING

2.1 GBML Based Scheduling

It is proposed that a product release scheduling agent have a learning classifier system (Holland et al., 1985), which is one GBML method to decide product dispatching timing that is introduced to the production floor. It is expected that the proposed scheduling agent can plan a proper lot release schedule without premeditated thresholds because the scheduling agent can find the threshold by itself through trial-and-error learning. Because of the learning characteristics, it is also expected that the proposed scheduling method is applicable to large-scale scheduling problems and can adapt properly to dynamic fluctuation on the production floor.

Figure 1 depicts the proposed system using the scheduling agent with a classifier system. The input to the agent is the state of the production floor, i.e. the WIP number of each machine group that consists of alternative machines. On the other hand, the output from the agent is the action: lot release. The agent action is evaluated using the evaluation criteria. Consequently, the agent obtains a suitable state-action chain.

[FIGURE 1 OMITTED]

2.2 Modeling Scheduling Agent

The scheduling agent is modeled using the classifier system. The classifier system consists of <if / then> fashioned n-bit production rules. In this paper, the condition part of the rule consists of a 2m-bit binary string that represents the WIP number of m machine groups. Condition {10} means that WIP is larger than twice the number of machines at the machine group. Condition {11} represents that WIP is less than twice, but more than the machine number, condition {01} means that WIP is less than the machine number, and condition {00} means that no product exists at the machine group. The action part consists of an n-bit string in the case that n kinds of product are produced. Action {1} shows that it dispatches one lot of each product; action {0} means it dispatches no lots.

The objective of the scheduling agent is to obtain a rule to fulfill high productivity and a preset throughput rate among product kinds. The action of the scheduling agent is evaluated as following: the scheduling agent is rewarded if a finished product keeps its due-date and is penalized if not. Credit values are apportioned among classifiers by the bucket brigade algorithm. The action is also evaluated for the throughput rate among product kinds: the scheduling agent receives a reward if the throughput rate meets or exceeds the target throughput rate, or it is penalized. The credit values are assigned among classifiers by a profit sharing plan (Grefennstette, 1988). A genetic algorithm (Goldberg, 1989) is used to create new classifiers to maintain a variety of classifiers and to avoid entrapment in a local minimal solution.

2.2 Dynamic Scheduling Using Self-organization

The production-floor scheduling problem must be solved properly. A dynamic scheduling method using self-organization (Vaario & Ueda, 1998) solves the scheduling problem. In self-organization based scheduling, production proceeds through local interaction among machines and AGVs transporting a product without global control. A potential field, which is superposition of attraction and repulsion fields, is employed to realize the interaction on the production floor.

3. SIMULATION RESULTS AND DISCUSSION

3.1 Simulation Settings

Computer simulations are conducted to verify the feasibility of the proposed classifier system based scheduling. Figure 2 illustrates the modelled production floor. One lot release scheduling agent is introduced onto the production floor. Two kinds of product that are produced have 1/4 the process steps of the actual process flow; TYPE1 has 90 production steps and its total production time is 1.95 (day); TYPE2 has 60 steps and 2.49 (day) production time. The respective due dates of products are set to about twice their production times. The production floor has 40 machines, which are classifiable into 22 kinds as alternative machine groups. Consequently, the scheduling agent has 44-bit input condition string to detect the situations of machine groups and 2-bit output action string to dispatch each product. The target throughput rate is set as 1:1. The number of AGVs is set to four; six stockers are also introduced to stock the products waiting for the next process. This system has one product dispatching point and one collecting point. The simulation runs 15 trials in which each trial has 1000 days simulation. The trial is stopped and moved to the next situation if irrelevant dispatching fills the floor.

[FIGURE 2 OMITTED]

3.2 Results and Discussion

Figure 3 shows the transition of WIP number during the simulation trials. The horizontal axis shows simulation steps; p_q represents p-trial and q-days. The WIP number accomplishes a better and steady value at the 15th trial, although it achieves wrong and fluctuating values in initial trials. Table 1 shows average throughput and TAT during the last 50 days at the 15th trial. It is readily apparent that the throughput rate of two product types accomplishes almost 1:1. The availability of the machine with highest value also obtains 96.9% at the last trial. The result clarifies that obtained scheduling rules can achieve not only a target throughput rate, but also high productivity on the production floor.

Observation and analysis of the executed rules by the scheduling agent at the final trial revealed that the scheduler switches product dispatching on/off timing according to the conditions of specific machines: machines with highest availability, machines with high availability and no alternative machines, and machines with much accessed time by product.

[FIGURE 3 OMITTED]

4. CONCLUDING REMARKS

This paper presented a new approach to lot release scheduling in semiconductor manufacturing: a scheduling agent found proper scheduling rules independently using a classifier system without premeditated thresholds. Current scheduling methods require them before being introduced. Computer simulation results verified the new method's feasibility.

This study was supported in part by Grants-in-Aid for Young Scientists (B), 17760326, and for Scientific Research on Priority Areas (2), 16016228 from the Ministry of Education, Culture, Sports, Science and Technology.

6. REFERENCES

Spearman, M.L.; Woodruff, D.L. & Hopp, W.J. (1990). CONWIP: a pull alternative to kanban, International Journal of Production Research, Vol.28, No.5, pp. 879-894

Wein, L.M. (1988). Scheduling Semiconductor Wafer Fabrication, IEEE Transactions on Semiconductor Manufacturing, Vol.1, No.3, pp. 115-130

Glassey, C.R. & Resende, M.G.C. (1988). Closed-Loop Job Release Control for VLSI Circuit Manufacturing, IEEE Transactions on Semiconductor Manufacturing, Vol.1, No.1, pp. 36-46

Ueda, K. (1992). A concept for bionic manufacturing systems based on DNA-type information, Proceedings of 8th International PROLAMAT Conference, IFIP, pp. 853-863

Holland, J.H.; Holyoak, K.J.; Nisbett, R.E. & Thagard, P.R. (1985). Induction-process of inference, learning and discovery, The MIT Press

Goldberg, D.E. (1989). Genetic Algorithm in Search, Optimization & Machine Learning, Addison Wesley

Grefennstette, J.J. (1988). Credit Assignment in Rule Discovery System Based on Genetic Algorithms, Machine Learning, Vol.3, pp. 225-245

Vaario, J. & Ueda, K. (1998). An emergent modeling method for dynamic scheduling, Journal of Intelligent Manufacturing, Vol. 9, pp. 129-140

Table 1. Average of throughput and TAT at the 15th trial.

 TYPE1 TYPE2

Throughput 5.46 5.50
TAT (day) 2.82 3.11