A dual-based stochastic inexact algorithm for a class of stochastic nonsmooth convex composite problems

In this paper, a dual-based stochastic inexact algorithm is developed to solve a class of stochastic nonsmooth convex problems with underlying structure. This algorithm can be regarded as an integration of a deterministic augmented Lagrangian method and some stochastic approximation techniques. By utilizing the sparsity of the second order information, each subproblem is efficiently solved by a superlinearly convergent semismooth Newton method. We derive some almost surely convergence properties and convergence rate of objective values. Furthermore, we present some results related to convergence rate of distance between iteration points and solution set under error bound conditions. Numerical results demonstrate favorable comparison of the proposed algorithm with some existing methods.