Efficient Parameter-Free First-Order Methods for Nonsmooth Composite Minimax Optimization
In the authors' words
In this paper we propose first-order methods for a class of nonsmooth composite strongly convex--strongly concave and nonconvex--concave minimax optimization. We first develop an inexact proximal method and an accumulative regularizated method for strongly convex--strongly concave problems. The latter achieves the optimal dependence on the curvature parameters and places the smaller curvature inside the accuracy logarithm. Using this method as a subsolver, we propose a proximal point method for nonconvex--concave problems. Under suitable assumptions, it finds an -stationary point with an operation complexity of , which improves the best-known bounds by removing the logarithmic factor. We further develop parameter-free variants for both problem classes, and achieve the same complexity without knowledge of any problem constants. All proposed methods are equipped with verifiable termination criteria.
Appeared: Wednesday, September 23. arXiv. Preprint, not yet peer-reviewed.