Empirical Distribution of Equilibrium Play and Its Testing Application

TL;DR

This paper demonstrates that in large normal-form games, approximate equilibria can be efficiently obtained and tested through sampling, with bounds that improve previous support size estimates.

Contribution

It provides new bounds on the number of samples needed to find and test approximate equilibria for Nash, correlated, and coarse correlated equilibria.

Findings

Sample-based methods can efficiently approximate equilibria.

Supports of approximate equilibria are polylogarithmic in game size.

New bounds improve previous polynomial support size estimates.

Abstract

We show that in any $n$-player $m$-action normal-form game, we can obtain an approximate equilibrium by sampling any mixed-action equilibrium a small number of times. We study three types of equilibria: Nash, correlated and coarse correlated. For each one of them we obtain upper and lower bounds on the number of samples required for the empirical distribution over the sampled action profiles to form an approximate equilibrium with probability close to one. These bounds imply that using a small number of samples we can test whether or not players are playing according to an approximate equilibrium, even in games where $n$ and $m$ are large. In addition, our results substantially improve previously known upper bounds on the support size of approximate equilibria in games with many players. In particular, for all the three types of equilibria we show the existence of approximate equilibrium with support size polylogarithmic in $n$ and $m$, whereas the previously best-known upper bounds were polynomial in $n$.