Independent Multi-Agent Reinforcement Learning with Counterfactual Semantic-Social World Models
CASTLE is a framework for fully decentralized multi-agent reinforcement learning that uses two complementary world models to guide agents. It separates offline pretraining from online control. The Semantic-Social World Model predicts short-term consequences of actions based on plausible teammate/opponent behavior, while the Local Dynamics World Model provides local trajectory context. This allows agents to prospectively compare candidate actions without relying on centralized critics or communication.