A deep reinforcement learning model based on Node2vec-GraphSAGE for multi-agent path finding

Abstract Multi-agent path finding (MAPF) problem seeks collision-free and efficient paths for a set of agents from their start locations to designated goal locations, where effective inter-agent coordination is key. Advanced reinforcement learning (RL)-based MAPF planner often employ convolutional neural networks (CNN) to extract features from observed states; however, this approach typically fails to capture an efficient representation of the global topological structure. To address this limitation, we propose a novel graph modeling approach for MAPF. Within this framework, we explicitly model the observed state by constructing a NodeSAGE model. This model integrates node2vec-based fused-node random walks with a Graph sample and AggregatE network to capture graph structural features. Subsequently, decentralized policies are trained via RL augmented with expert knowledge. To enhance agent-environment coordination, the policy network incorporates dedicated attention-based modules: a static-obstacle aggregation module and a dynamic-obstacle aggregation module. Simulation experiments demonstrate that the learned policies achieve good performance.