Policy-based reinforcement learning algorithm and its application to semiconductor fab dispatching rule

Abstract

This paper deals with the scheduling optimization based on dispatching rules for the efficient operation of the semiconductor factory. The modern semiconductor plant has a large scale and complex structure. It requires a very high computational cost in the design of sophisticated scheduling. These problems make it difficult to apply popular scheduling methods such as mathematical formulation and meta-heuristic, which require a high computational cost, to semiconductor factory. Therefore, almost manufacturers have been designing the schedule using very simple dispatching rules. However, in order to be a more efficient operation, more sophisticated dispatching rules should be designed for the factory. Thus, in this paper, we proposed the reinforcement learning-based algorithms for the design of more effective and sophisticated dispatching rules. As a first study, we proposed the per-machine linear dispatching rule learning approach for different multi-machines using population-based search. As a second study, to achieve higher data efficiency, we proposed per-machine dispatching rule learning approach using policy gradient, in which actors are decentralized, and critic is centralized. As a third study, we proposed a hybrid algorithm that takes both the advantage of a policy gradient method and a population-based search. Experiments showed that the proposed methods have better performance or data efficiency than existing methodologies.