Saddle Point Optimization with Approximate Minimization Oracle

A major approach to saddle point optimization 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 and to and at a given point and updates toward these approximate solutions with a learning rate . On locally strong convex--concave smooth functions, we derive conditions on to exhibit linear convergence to a local saddle point, which reveals a possible shortcoming of recently developed robust adversarial reinforcement learning algorithms. We develop a heuristic approach to adapt derivative-free and implement zero-order and first-order minimization algorithms. Numerical experiments are conducted to show the tightness of the theoretical results as well as the usefulness of the adaptation mechanism.
View on arXiv