A Multi-Agent Reinforcement Learning Strategy Based on Graph Neural Networks and Credit Similarity

This paper presents GACS, a novel graph reinforcement learning-based cooperative multi-agent strategy, addressing complex inter-agent interactions and behavioral homogenization in multi-agent systems. GACS leverages graph neural networks (GNNs) to model agent relational structures and integrates a credit similarity perception module to enhance credit allocation differentiation, thereby boosting individual decision-making. Experiments across two environments and ablation studies confirm that the proposed components significantly improve overall system performance. This work highlights the potential of GNN-reinforcement learning integration in multi-agent scenarios, providing a valuable direction for advancing cooperative strategies.