Graph Neural Network-Based Relation Modeling for Multi-Agent Reinforcement Learning
Tingting Wei, Xueqiang Gu, Zhangling Wang, Zhiheng Zhang, Yi Cheng, Lina Lu · 2025
Cooperative multi-agent reinforcement learning (MARL) seeks to learn optimal team strategies in complex environments. However, current MARL approaches often fail to adequately model interactions among multiple agents, particularly in intricate settings where algorithmic performance is jeopardized. In addition, the traditional MARL algorithm simply takes other agent status information as input for action selection, and does not further process these information, resulting in low information utilization efficiency. Given the inherent advantages of graph neural networks in relation modeling and information aggregation, this paper proposes a novel MARL algorithm based on graph neural network relation modeling (GRM). We use the self-attention mechanism to realize dynamic graph learning, and the designed agent network can dynamically update the multi-agent topology according to the environment change. The proposed method is verified on the SMAC environment. The experimental results show that the proposed method performs better than the baseline algorithm, indicating that the proposed method can deal with more complex and variable environments.