![Learning and games_hi-res logo](/sites/default/files/styles/workshop_banner_sm_1x/public/2023-03/Learning%20and%20games_hi-res_RGB.jpg?h=5f9e7f71&itok=W4SMHuQA)
Abstract
When solving unbalanced minimax optimization problems, single-loop algorithms such as GDA or extragradient method could suffer from inferior convergence. On the other hand, recent algorithms with best-known complexities require sophisticated multi-loop schemes, which are hard to implement and unrealistic in game applications. In this talk, we show that for several interesting minimax settings, we can design single-loop algorithms that achieve complexities as good as or even better than state-of-the-art multi-loop algorithms. We discuss how acceleration can be achieved through the idea of primal-dual lifting.