Real-Time Decision Support with Reinforcement Learning for Dynamic Flowshop Scheduling

Abstract:
The dynamic flowshop scheduling problem has attracted a lot of attention from both academia and industry because of its nature of NP-hardness in computation on the one hand, and its great value for optimization of manufacturing systems on the other. Traditionally, effective heuristic methods have been widely adopted in industry to solve the problem. In recent years, machine learning has also exhibited great potential in the field. In this paper, we propose a reinforcement learning approach to this problem. We discuss settings for orders, performance measurements, and learning methods in detail to construct a controlled environment for this research. We then establish a manufacturing simulation system to compare the performance of the reinforcement learning approach and three heuristic approaches. While the experimental results revealed the superiority of the reinforcement learning method, investigations into dispatching decisions exposed its limitations in actual industry applications. Hence, two strategies were designed and employed to improve the reinforcement learning approach. After validating the reinforcement learning method with the improved strategies, we summarized and presented the dispatching rules of the reinforcement learning method as one type of complementary decision for supporting real-time flowshop scheduling problems.