Speaker: Yuze Han, Peking University

Time: 16:00-17:00 pm, December 19, 2022, GMT+8

Venue: Tecent Meeting ID: 723 1564 5542

Abstract:

The optimistic gradient (OG) method has received growing attention due to its favorable performance in saddle-point optimization problems, which include two-player zero-sum games (a.k.a. matrix games) as a special case. Compared to another classical method, Extragradient, OG is more suitable for repeated games since it is a no-regret algorithm and only requires one gradient call in each iteration. Although the sublinear average-iterate convergence of OG has been provided for years, the more appealing last-iterate convergence rate has only been established recently.

In this talk, we will introduce OG as well as its several variants, and discuss the convergence rate of the last iterate in saddle-point optimization problems. Moreover, we will pay particular attention to matrix games.

Source: School of Mathematical Sciences