RIDRL: A Deep Reinforcement Learning Based on Multiple Dispatching Rules and IGA Algorithm for JSP

Abstract

Job shop scheduling problem (JSP) is a very general and complex scheduling problem in the manufacturing industry. The traditional priority dispatching rule (PDR) can get the approximate solution for some specific problems. Nevertheless, for the complex and changing realistic factory floor, the quality of the existing solution fluctuates significantly. To solve the problem, this paper fuse multiple dispatching rules and the Insertion Greedy Algorithm (IGA) to deep reinforcement learning (DRL), namely RIDRL, to solve the job shop scheduling problem. In this method, we manually choose five generalizable state features as the states of the workshop environment. Employing 18 scheduling rules as the action space in the agent while designing a quick converge reward function. Additionally, we use a Proximal Policy Optimization Algorithm (PPO) to train the DRL agent with minimizing makespan as the optimization objective. Several simulation experiments on many standard instances indicate that the proposed method obtains competitive solutions for problems of different sizes.