A Novel Reinforcement Learning Approach for Enhancing Flexible Job-Shop Scheduling with Dual Gated-Attention Network

The flexible job-shop scheduling problem (FJSP) is a classical problem in combinatorial optimization. For the FJSP, each job operation can be processed on multiple machines, which requires determining the sequence of job operations and selecting a suitable machine. Deep reinforcement learning (DRL) has recently emerged as a trend for improving the quality of FJSP solutions. However, the quality of solutions still has room for improvement compared to those obtained by exact methods like OR-Tools. To address this issue, this paper proposes a novel end-to-end DRL scheduling model for enhancing FJSP. The model utilizes a disjunctive graph to represent the scheduling environment and innovatively employs a dual gated-attention network (DGAN) encoder to capture and embed features of the disjunctive graph, revealing the complex relationships between operations and machines. A concise multi-layer perceptron (MLP) is designed to generate optimal scheduling sequences as the decision-making network. During the training phase, the proximal policy optimization (PPO) algorithm is applied to adjust network parameters, minimizing the job completion time. Experiments demonstrate that our model outperforms heuristic rules and the latest DRL baseline models on synthetic and public benchmark datasets.