An improved disjunctive graph and deep reinforcement learning for FJSP

Abstract In the current industrial manufacturing environment, the flexible job shop scheduling problem (FJSP), which is an extension of the job shop scheduling problem, holds significant research significance. While intelligent algorithms are commonly used to solve the FJSP problem, this paper proposes the use of reinforcement learning algorithms. The FJSP is first transformed into a Markov decision process, and an improved FJSP disjunction graph structure is introduced. Subsequently, 8 state features are designed, and a combination of gated attention network and multi-layer perceptron is employed to extract these features. For the decision-making processes of FJSP, a 3-tuple is designed as the action space. Finally, a reward function is formulated to optimize the objective of minimizing the maximum makespan. The training algorithm utilizes the Advantage Actor-Critic algorithm to establish an end-to-end deep reinforcement learning framework. Experimental results demonstrate that this algorithm exhibits significant advantages in terms of FJSP solution quality and efficiency. In conclusion, the reinforcement learning method proposed in this article presents a novel approach to address the FJSP problem and holds promising application prospects.