An optimal method for stochastic composite optimization (Q431018): Difference between revisions

The stochastic composite optimization problem under consideration consists in minimizing the sum \(f+h\) of two convex functions \(f,h\) defined on a compact convex set \(X\subset \mathbb{R}^{n}\), assuming that \(\nabla f\) and \(h\) are globally Lipschitz. The author proposes two subgradient-type methods, namely, a modified version of the mirror-descent SA method due to \textit{A. Nemirovski, A. Juditsky, G. Lan} and \textit{A. Shapiro} [SIAM J. Optim. 19, No. 4, 1574--1609 (2009; Zbl 1189.90109)], which substantially improves the rate of convergence, and an accelerated stochastic approximation method, which can achieve an optimal rate of convergence. He illustrates the advantages of the latter method over other existing methods by discussing its performance for a particular class of stochastic optimization problems.