Abstract
Multiagent reinforcement learning (MARL) has achieved impressive success across various domains, yet the neural network policies produced with MARL fail to explain their behaviors in a human-understandable way. To address this challenge, we propose global and coalition cognition graph modeling (GC2GM), a novel framework that equips agents with explicit cognitive representations integrating both coalition-level relational reasoning and individualized global state cognition. Specifically, we first introduce an information exchange graph network (IEGN) to build coalition cognition that enables agents to dynamically reason about the intentions and behaviors of teammates. We further enhance individualized global cognition of agents via imposing a mutual-information regularization that aligns the global state with its local perspective, which produces globally informed yet locally grounded representations of the environment. During execution, agents rely solely on their local information and consciousness to make decisions, adhering to the centralized training with decentralized execution (CTDE) paradigm. Additionally, GC2GM is modular and can be seamlessly integrated into existing CTDE MARL methods. Experiments on a wide range of challenging MARL benchmarks demonstrate that GC2GM achieves desirable results, striking a favorable balance between performance and interpretability.