Dynamic Scheduling of Maintenance by a Reinforcement Learning Approach - A Semiconductor Simulation Study

Scheduling in a semiconductor back-end factory is an extremely sophisticated and complex task. In semiconductor industry, more often than not, the scheduling of maintenance is underexposed to production scheduling. This is a missed opportunity as maintenance and production activities are deeply intertwined. This study considers the dynamic scheduling of maintenance activities on an assembly line. A policy is constructed to schedule a cleaning activity on the last machine of an assembly line such that the average production rate is maximized. The policy takes into account the given flexibility and the buffer content of the buffers in-between the machines in the assembly line. A Markov Decision Process is formulated for the problem and solved using Value Iteration and Reinforcement Learning Algorithms. In addition, for a real world case study, a simulation analysis is performed to evaluate the potential practical benefits.