Computing correlated equilibria in multi-player games

Abstract

We develop polynomial-time algorithms for finding correlated equilibria—a well-studied notion of rationality that generalizes the Nash equilibrium—in a broad class of succinctly representable multiplayer games, encompassing graphical games, anonymous games, polymatrix games, congestion games, scheduling games, local effect games, as well as several generalizations. Our algorithm is based on a variant of the existence proof due to Hart and Schmeidler, and employs linear programming duality, the ellipsoid algorithm, Markov chain steady state computations, as well as application-specific methods for computing multivariate expectations over product distributions. For anonymous games and graphical games of bounded tree-width, we provide a different polynomial-time algorithm for optimizing an arbitrary linear function over the set of correlated equilibria of the game. In contrast to our sweeping positive results for computing an arbitrary correlated equilibrium, we prove that optimizing over correlated equilibria is NP-hard in all of the other classes of games that we consider.