Graph-Based Multi-Agent Reinforcement Learning for Dynamic Flexible Flow Shop Scheduling

The Dynamic Flexible Flow Shop Scheduling Problem (DFFSP) is a combinatorial optimization problem involving strong NP-hard and dynamics. To meet the requirements of realtime performance and robustness in high-frequency dynamic order insertion scenarios, this paper proposes an intelligent scheduling method based on graph state modeling and multiagent reinforcement learning, the state features are extracted through graph neural networks and multi-head attention mechanisms, and the multi-agent strategy optimization module is utilized to achieve the collaborative control of job selection and machine allocation, thereby realizing rapid and efficient scheduling decisions in a production environment with frequent disturbances. Extensive experiments have shown that our method demonstrates superior performance in scheduling problems of various scales, especially in high-frequency dynamic order insertion scenarios, and has good universality.