A major approach to saddle point optimization \(\min_x\max_y f(x, y)\) is a gradient based approach as is popularized by generative adversarial networks (GANs). In contrast, we analyze an alternative approach relying only on an oracle that solves a minimization problem approximately. Our approach locates approximate solutions \(x'\) and \(y'\) to \...

Saddle Point Optimization with Approximate Minimization Oracle

10.48550/arxiv.2103.15985