(Data-Driven) Development of dynamic scheduling in semiconductor manufacturing using a Q-learning approach

Abstract

ABSTRACT In accordance with the operational characteristics of semiconductor wafer fabrication (FAB), multiple-dynamic-scheduling-rule (MDSR) selection mechanisms must be developed for dynamic scheduling in FAB. The main disadvantage of the classic MDSR selection approach is that it does not immediately reflect changes in production requirements in a dynamic FAB environment. A dynamic scheduling knowledge base (KB) is generally not static; therefore, if a major change occurs in the production system, a mechanism for modifying the KB must be established. Accordingly, we develop a dynamic scheduling scheme that uses a Q-learning-based MDSR selection mechanism to support shop floor control in FAB. This selection mechanism involves two phases: initial generation of an MDSR KB and revision of the MDSR KB. On the basis of various performance criteria, the proposed scheme exhibited superior system performance to a classic MDSR selection approach and an approach that applies fixed scheduling decisions.