Equilibrium in Multi-Agent Reinforcement Learning
Standard solution concepts for stochastic games, such as Markov perfect equilibrium and Markov coarse correlated equilibrium, are computationally difficult, and thus, standard decentralized reinforcement-learning algorithms should not generally be expected to converge to them. In this paper, we study the equilibrium generated by such algorithms. In particular, we introduce a new solution concept for stochastic games, Markov Bayes coarse correlated equilibrium (MBCCE), defined as a distribution over states and stationary policy profiles such that, after observing the state but before observing her recommended action, no player can gain by choosing a different current action, with the sampled policy profile governing play thereafter. We discuss the parallels between MBCCE and coarse correlated equilibrium (CCE) in finite normal-form games and show that MBCCE retains several of its key properties. We then introduce a corresponding regret notion, adaptive Markov coarse regret (AMCR), and show that vanishing AMCR implies that every accumulation point of the empirical distribution of realized states and policy profiles is an MBCCE. Crucially, we show that achieving AMCR reduces to two standard learning tasks: minimizing external regret at each state and accurately evaluating the current joint policy. We then prove that under mild conditions these properties hold for two natural RL algorithmic designs: a decentralized asynchronous actor--critic algorithm through a new two-timescale stochastic-approximation analysis, and a standard episodic multi-agent projected policy-gradient method. Hence, both algorithms generate approximate MBCCEs, and we establish explicit finite-time convergence rates for both.