Abstract: Currently,multi-agent reinforcement learning is widely applied in various cooperative tasks.In real environments,agents always have access to only partial observations,leading to inefficient exploration of cooperative strategies.Moreover,sharing reward values among agents makes it challenging to accurately assess individual contributions.To address these issues,a novel graph attention-based grouped multi-agent reinforcement learning framework is proposed,which improves cooperation efficiency and enhances the evaluation of individual contributions.Firstly,a multi-agent system with graph structure is constructed,which learning relationships among the individual agents and their neighbors for sharing information.This approach expands individual agents’ perceptual fields to mitigate constraints from partial observability and assess individual contributions.Secondly,an action reference module is designed to provide joint action reference information for individual action selection,enabling agents to explore more efficiently and diversely.Experimental results in two different scales of multi-agent control scenarios demonstrate significant advantages over baseline methods.Detailed ablation studies further verify the effectiveness of the graph attention grouping approach and communication settings.

Key words: Multi-agent reinforcement learning, Graph attention network, Centralized training decentralized execution, Multi-agent cooperation, Multi-agent communication