UTRC CDS Seminar: Rachael Tappenden, "Flexible ADMM for Big Data Applications"

Friday, November 20, 2015
11:00 a.m.
2460 AVW

Flexible ADMM for Big Data Applications

Rachael Tappenden
School of Applied Math and Statistics
Johns Hopkins University

Abstract
In this talk we present a flexible Alternating Direction Method of Multipliers (F-ADMM) algorithm for solving optimization problems involving a strongly convex objective function that is separable into n blocks, subject to linear equality constraints. The F-ADMM algorithm updates the blocks of variables in a Gauss-Seidel fashion, and the subproblems arising within F-ADMM include a regularization term so that they can be solved efficiently. The algorithm is globally convergent. We also introduce a hybrid variant called H-ADMM that is partially parallelizable, which is important in a big data setting. Convergence of H-ADMM follows directly from the convergence properties of F-ADMM. We present numerical experiments to demonstrate the practical performance of this algorithm.