New AI algorithm solves 'fundamental open problem' in multi-agent systems

Researchers have developed a novel two-timescale Actor-Critic algorithm that achieves provable global convergence for learning stationary policies in general-sum Markov games—a long-standing open problem in Multi-Agent Reinforcement Learning (MARL). The method leverages Risk-averse Quantal response Equilibria (RQE), incorporating risk aversion and bounded rationality. Empirical tests show it outperforms risk-neutral baselines, offering the first finite-sample guarantees for this class of problems and making complex multi-agent coordination practically learnable.