Multi-agent reinforcement learning based on graph convolutional network for flexible job shop scheduling

With the development of Internet of manufacturing things, decentralized scheduling in flexible job shop is arousing great attention. To deal with the challenges confronted by personalized manufacturing, such as high level of flexibility, agility and robustness for dynamic response, we design a centralized-learning decentralized-execution (CLDE) multi-agent reinforcement learning scheduling structure based on Graph Convolutional Network (GCN), namely graph-based multi-agent system (GMAS), to solve the flexible job shop scheduling problem (FJSP). Firstly, according to the product processing network and job shop environment, the probabilistic model of directed acyclic graph for FJSP is constructed. It models the FJSP as the process of topology graph structure predicting, and the scheduling strategy is adjusted by predicting the connection probability among edges. Then, the multi-agent reinforcement learning system consisting of environment module, job agent module, and machine agent module is constructed. The job agents execute scheduling actions by interacting with environment and machine agents in a decentralized way. Meanwhile, the interaction between job agents is extracted as an abstract global action based on GCN. The experimental results demonstrate that GMAS outperforms its rivals on FJSP, especially in complicated situations. Our results thus shed light on a novel direction for FJSP in dynamic and complex scenarios.