From Average-Iterate to Last-Iterate Convergence in Games: A Reduction and Its Applications

TL;DR

This paper introduces a simple reduction transforming average-iterate convergence into last-iterate convergence in certain games, enabling improved convergence rates for learning algorithms in multi-player zero-sum polymatrix games.

Contribution

The authors present a black-box reduction that converts average-iterate to last-iterate convergence for a broad class of games, enhancing convergence analysis and rates.

Findings

Achieved exponential improvement in convergence rate dependence on dimension under gradient feedback.

Improved last-iterate convergence rates under bandit feedback for polymatrix games.

Applied reduction to optimize the performance of the Optimistic Multiplicative Weights Update algorithm.

Abstract

The convergence of online learning algorithms in games under self-play is a fundamental question in game theory and machine learning. Among various notions of convergence, last-iterate convergence is particularly desirable, as it reflects the actual decisions made by the learners and captures the day-to-day behavior of the learning dynamics. While many algorithms are known to converge in the average-iterate, achieving last-iterate convergence typically requires considerably more effort in both the design and the analysis of the algorithm. Somewhat surprisingly, we show in this paper that for a large family of games, there exists a simple black-box reduction that transforms the average iterates of an uncoupled learning dynamics into the last iterates of a new uncoupled learning dynamics, thus also providing a reduction from last-iterate convergence to average-iterate convergence. Our reduction applies to games where each player's utility is linear in both their own strategy and the joint strategy of all opponents. This family includes two-player bimatrix games and generalizations such as multi-player polymatrix games. By applying our reduction to the Optimistic Multiplicative Weights Update algorithm, we obtain new state-of-the-art last-iterate convergence rates for uncoupled learning dynamics in multi-player zero-sum polymatrix games: (1) an $O(\frac{\log d}{T})$ last-iterate convergence rate under gradient feedback, representing an exponential improvement in the dependence on the dimension $d$ (i.e., the maximum number of actions available to either player); and (2) an $\widetilde{O}(d^{\frac{1}{5}} T^{-\frac{1}{5}})$ last-iterate convergence rate under bandit feedback, improving upon the previous best rates of $\widetilde{O}(\sqrt{d} T^{-\frac{1}{8}})$ and $\widetilde{O}(\sqrt{d} T^{-\frac{1}{6}})$.