Multi-Agent Incentive Communication via Decentralized Teammate Modeling

Effective communication can improve coordination in cooperative multi-agent reinforcement learning (MARL). One popular scheme is exchanging agents' local observations or latent embeddings and using them to augment individual policy input. Such a paradigm reduce uncertainty for decision-making induce implicit coordination. However, it enlarges spaces increases complexity, leading poor complex settings. To handle this limitation, paper proposes novel framework named Multi-Agent Incentive Communication (MAIC) that allows each agent learn generate incentive messages bias other value functions directly, resulting effective explicit Our method firstly learns targeted teammate models, with which anticipate the teammate's action selection tailored specific agents. We further introduce regularization leverage interaction sparsity efficiency. MAIC agnostic MARL algorithms be flexibly integrated different function factorization methods. Empirical results demonstrate our significantly outperforms baselines achieves excellent performance on multiple tasks.