Sim2O: Efficient Offline-to-Online MARL via Joint Action Composition

Sim2O is a novel framework for offline-to-online Multi-Agent Reinforcement Learning (MARL) that addresses the challenge of online exploration costs by leveraging offline datasets. It employs a compositional approach to adapt joint actions by blending offline and online proposals, using a centralized value function to evaluate these combinations without additional training overhead. Empirical results show that Sim2O outperforms existing baselines, highlighting its efficiency and effectiveness for practitioners working on coordinated decision-making in MARL environments.