Convergence analysis of an accelerated stochastic admm with larger stepsizes

JIANCHAO BAI,DEREN HAN,HONGCHAO ZHANG,HAO SUN
2020-01-01
Abstract:In this paper, we develop an accelerated stochastic Alternating Direction Method of Multipliers (ADMM) for solving the structured convex optimization problem whose objective is the sum of smooth and nonsmooth terms. The proposed algorithm uses the ideas of both ADMM and accelerated stochastic gradient method, and the involved dual variables are updated twice with a larger stepsize region than the so-called (0, √ 5+1 2 ). To simplify computation of the nonsmooth subproblem, we add a nonnegative proximal term to transform the subproblem into a proximal mapping. By a variational analysis, we show that the proposed stochastic algorithm converges in expectation with the worst-case O(1/T ) convergence rate, where T denotes the number of outer iterations. Convergence rates of a special stochastic augmented Lagrange method and 3-block extensions of the proposed algorithm are also investigated. Numerical experiments on testing a big-data problem in machine learning verify that the proposed method is effective and promising.
What problem does this paper attempt to address?