Generalized Intention Modeling in Multi-Agent Reinforcement Learning

Abstract

An opponent modeling framework for competitive and general-sum multi-agent reinforcement learning that learns a task-dependent mixture of intent representations rather than fixing one in advance. Mixer of Intention eXperts (MIX) uses a learned gate to blend embeddings predictive of opponent actions, opponent observations, and future states with a new embedding that maximizes mutual information with the agent's own future returns. Evaluated on Kuhn Poker, Predator-Prey, Level-Based Foraging, and Google Research Football, where it broadly matches or exceeds state-of-the-art baselines.

Paper