Real-Time Scheduling for Flexible Job Shop With AGVs Using Multiagent Reinforcement Learning and Efficient Action Decoding
针对含自动导引车（AGV）的动态柔性作业车间调度问题，提出一种基于多智能体强化学习的实时调度方法，通过高效动作解码算法提升学习效率，实验表明该方法优于多种基准方法，性能提升超10%。
The application of automated guided vehicle (AGV) greatly improves the production efficiency of workshop. However, machine flexibility and limited logistics equipment increase the complexity of collaborative scheduling, and frequent dynamic events bring uncertainty. Therefore, this article proposes a real-time scheduling method for dynamic flexible job shop scheduling problem with AGVs using multiagent reinforcement learning (MARL). Specifically, a real-time scheduling framework is proposed in which a multiagent scheduling architecture is designed for achieving task selection, machine allocation and AGV allocation. Then, an action space and an efficient action decoding algorithm are proposed, which enable agents to explore in the high-quality solution space and improve the learning efficiency. In addition, a state space with generalization, a reward function considering machine idle time and a strategy for handling four disturbance events are designed to minimize the total tardiness cost. Comparison experiments show that the proposed method outperforms the priority dispatching rules, genetic programming and four popular reinforcement learning (RL)-based methods, with performance improvements mostly exceeding 10%. Furthermore, experiments considering four disturbance events demonstrate that the proposed method has strong robustness, and it can provide appropriate scheme for uncertain manufacturing system.