[2108.00473] By-product-free Alternating Projection Algorithms for Basic Nonconvex-Concave Minimax Issues

Obtain a PDF of the paper titled By-product-free Alternating Projection Algorithms for Basic Nonconvex-Concave Minimax Issues, by Zi Xu and three different authors

Obtain PDF

Summary:On this paper, we examine zeroth-order algorithms for nonconvex-concave minimax issues, which have attracted broadly consideration in machine studying, sign processing and plenty of different fields in recent times. We suggest a zeroth-order alternating randomized gradient projection (ZO-AGP) algorithm for clean nonconvex-concave minimax issues, and its iteration complexity to acquire an $varepsilon$-stationary level is bounded by $mathcal{O}(varepsilon^{-4})$, and the variety of operate worth estimation is bounded by $mathcal{O}(d_{x}+d_{y})$ per iteration. Furthermore, we suggest a zeroth-order block alternating randomized proximal gradient algorithm (ZO-BAPG) for fixing block-wise nonsmooth nonconvex-concave minimax optimization issues, and the iteration complexity to acquire an $varepsilon$-stationary level is bounded by $mathcal{O}(varepsilon^{-4})$ and the variety of operate worth estimation per iteration is bounded by $mathcal{O}(Okay d_{x}+d_{y})$. To the very best of our data, that is the primary time that zeroth-order algorithms with iteration complexity gurantee are developed for fixing each basic clean and block-wise nonsmooth nonconvex-concave minimax issues. Numerical outcomes on knowledge poisoning assault downside and distributed nonconvex sparse principal part evaluation downside validate the effectivity of the proposed algorithms.