GRDC: A Unified Graph-Driven Framework for Role Discovery and Communication in Multi-Agent Reinforcement Learning

Effective coordination in Multi-Agent Reinforcement Learning (MARL) is particularly challenging under partial observability, where agents must identify and coordinate with task-relevant collaborators based solely on local information. Existing methods can be categorised into communicationbased approaches, which allow message exchange but either rigidly predefine or misidentify collaborators, and rolebased approaches, which promote functional specialization based on observed behavioural similarity. However, both paradigms overlook the dynamic and context-specific cooperative dependencies induced by the task, which determine which agents should collaborate, thereby leading to miscommunication or role misassignment under partial observability. We introduce GRDC (Graph-driven Role Discovery and Communication), a unified framework that approximates these dependencies by dynamically constructing local interaction graphs from trajectory embeddings, then uses these graphs to infer roles via prototype matching and to restrict communication to intra-role agents with attention-based aggregation. In addition to role inference and communication, GRDC promotes a structured and compact role space by maximising role entropy, decorrelating prototypes, and dynamically pruning redundant prototypes. Experimental results on Predator Prey, Cooperative Navigation, and SMACv2 demonstrate that GRDC consistently outperforms state-of-the-art communication-and role-based baselines, improving coordination efficiency and training stability across tasks.