Multiple Oracle Algorithm to Solve Continuous Games

TL;DR

This paper introduces a generalized double oracle algorithm for computing approximate Nash equilibria in multiplayer continuous games, extending applicability beyond two-player zero-sum cases with proven convergence in Wasserstein distance.

Contribution

It extends the double oracle algorithm to multiplayer general-sum continuous games and proves convergence in Wasserstein distance, broadening the scope of equilibrium computation methods.

Findings

Algorithm converges to an approximate equilibrium in finitely many steps.

Method performs well on various classes of games, including random examples.

Wasserstein distance is effective for measuring strategy convergence.

Abstract

Continuous games are multiplayer games in which strategy sets are compact and utility functions are continuous. These games typically have a highly complicated structure of Nash equilibria, and numerical methods for the equilibrium computation are known only for particular classes of continuous games, such as two-player polynomial games or games in which pure equilibria are guaranteed to exist. This contribution focuses on the computation and approximation of a mixed strategy equilibrium for the whole class of multiplayer general-sum continuous games. We vastly extend the scope of applicability of the double oracle algorithm, initially designed and proved to converge only for two-player zero-sum games. Specifically, we propose an iterative strategy generation technique, which splits the original problem into the master problem with only a finite subset of strategies being considered, and the subproblem in which an oracle finds the best response of each player. This simple method is guaranteed to recover an approximate equilibrium in finitely many iterations. Further, we argue that the Wasserstein distance (the earth mover's distance) is the right metric for the space of mixed strategies for our purposes. Our main result is the convergence of this algorithm in the Wasserstein distance to an equilibrium of the original continuous game. The numerical experiments show the performance of our method on several classes of games including randomly generated examples.