Learning to Schedule Job-Shop Problems via Hierarchical Reinforcement Learning

Zijun Liao, Qiwen Li, Yuanzhi Dai, Zizhen Zhang

Abstract: The job-shop scheduling problem (JSSP) is a classic combinatorial optimization problem in the areas of computer science and operations research. It is closely associated with many industrial scenarios. In today’s society, the demand for efficient and stable scheduling algorithms has significantly increased. More and more researchers have recently tried new methods to solve JSSP. In this paper, we effectively formulate the scheduling process of JSSP as a Semi-Markov Decision Process. We then propose a method of using hierarchical reinforcement learning with graph neural networks to solve JSSP. We also demonstrate that larger-sized instances require the support of a bigger number of sub-policies and different scheduling phases require using different sub-policies.