A multi-action deep reinforcement learning based on BiLSTM for flexible job shop scheduling problem with tight time

The Flexible Job Shop Scheduling Problem (FJSP) with tight time is a significant challenge in both academic and industrial fields of production scheduling. This paper addresses the FJSP with tight time using a Multi-action Deep Reinforcement Learning (MDRL) method. First, a multi-action Markov Decision Process (MDP) is formulated, integrating operation and machine sets into a unified multi-action space. Then, a scheduling policy is developed using a Bi-Directional Long Short-Term Memory Network (BiLSTM) to extract intrinsic scheduling information. Finally, Proximal Policy Optimization (PPO) enhanced with reward shaping is employed to train the model, enabling intelligent decision-making in action selections. Extensive experiments are conducted on four problem instances of varying scales. Comparisons among 20 priority dispatch rules and two closely rated DRL methods demonstrate the superior performance of the proposed MDRL approach.