Behavior-Guided Identity Learning for Multiagent Cooperation

Effective cooperation in multiagent systems often requires agents to assume different responsibilities, leading to distinct behavior patterns for complementary roles. To characterize such behavioral differences in cooperative multiagent reinforcement learning, we propose Behavior-Guided Identity Learning (BGIL), which learns identity representations from behavior. First, we introduce an identity contrastive objective that maximizes the mutual information between behavior and identity representations, improving inter-agent discriminability. Second, we design a temporal identity consistency mechanism to ensure that different behaviors of the same agent over time can be consistently associated with the same identity. Extensive experiments on the SMAC benchmark demonstrate that BGIL consistently outperforms competitive baselines, especially on heterogeneous tasks.