First- and Zeroth-Order Learning in Asynchronous Games

TL;DR

This paper studies asynchronous game dynamics where agents update at different times, providing convergence conditions, new algorithms, and analysis for both first- and zeroth-order learning methods, validated through economic simulations.

Contribution

It introduces new convergence conditions and algorithms for asynchronous games with partial asynchronism, including first- and zeroth-order learning methods.

Findings

Derived tight convergence conditions for quadratic games.

Provided a quasidominance condition for general convex games.

Validated algorithms through economic market simulations.

Abstract

This paper investigates the discrete-time asynchronous games in which noncooperative agents seek to minimize their individual cost functions. Building on the assumption of partial asynchronism, i.e., each agent updates at least once within a fixed-length time interval, we explore the conditions to ensure convergence of such asynchronous games. The analysis begins with a simple quadratic game from which we derive tight convergence conditions through the lens of linear control theory. Then, we provide a quasidominance condition for general convex games. Our results demonstrate that this condition is stringent since when this condition is not satisfied, the asynchronous games may fail to converge. We propose both first- and zeroth-order learning algorithms for asynchronous games, depending on the type of available feedback, and analyze their last-iterate convergence rates. Numerical experiments are presented on economic market problems to verify our results.