Multi-Agent Deep Reinforcement Learning for Solving Flexible Job Shop Scheduling Problem

Based on the research on solving the flexible job shop scheduling problem with the objective of minimizing the maximum completion time, a multi-agent deep reinforcement learning algorithm (MADRL) adopting the centralized training and decentralized execution (CTDE) framework was proposed. Job agent and Machine Agent selected operation tasks and processing machines, and heterogeneous graph neural networks were utilized to extract the state features of the scheduling environment. The negative value of the maximum completion time was used as a common reward for both agents, while a central controller assisted in training the two agents to further improve the solving performance. The algorithm’s trained agents are verified to have good generalization and strong robustness through several standard instances.