Existence of $ε$-Nash Equilibria in Nonzero-Sum Borel Stochastic Games and Equilibria of Quantized Models

Abstract: Establishing the existence of exact or near Markov or stationary perfect Nash equilibria in nonzero-sum Markov games over Borel spaces remains a challenging problem, with few positive results to date. In this paper, we establish the existence of approximate Markov and stationary Nash equilibria for nonzero-sum stochastic games over Borel spaces, assuming only mild regularity conditions on the model. Our approach involves analyzing a quantized version of the game, for which we provide an explicit construction under both finite-horizon and discounted cost criteria. This work has significant implications for emerging applications such as multi-agent learning. Our results apply to both compact and non-compact state spaces. For the compact state space case, we first approximate the standard Borel model with a finite state-action model. Using the existence of Markov and stationary perfect Nash equilibria for these finite models under finite-horizon and discounted cost criteria, we demonstrate that these joint policies constitute approximate Markov and stationary perfect equilibria under mild continuity conditions on the one-stage costs and transition probabilities. For the non-compact state space case, we achieve similar results by first approximating the model with a compact-state model. Compared with previous results in the literature, which we comprehensively review, we provide more general and complementary conditions, along with explicit approximation models whose equilibria are $\epsilon$-equilibria for the original model.