Reinforcement learning-based adaptive production control of pull manufacturing systems

Abstract

ABSTRACTThis research proposes a methodology based on Reinforcement Learning for deriving optimal adaptive control policies for CONWIP-type manufacturing systems. The Reinforcement Learning agent interacts with the production system and learns the best actions for each state. The goal is to minimize the total, average cost that consists of holding and backorder costs. The Reinforcement Learning-based approach is compared to the standard CONWIP, the Dynamic CONWIP and a special case of the Generalized Kanban system in a series of simulation experiments. The proposed approach is found to outperform all alternative control policies in all cases. The Generalized Kanban policy is reported to be a good approximation of the control policies that are derived by the Reinforcement Learning approach in situations where the demand rate is low. A numerical investigation of the total cost function properties for the CONWIP, Dynamic CONWIP and Generalized Kanban is also provided.