UTRC CDS Seminar: Rachael Tappenden, "Flexible ADMM for Big Data Applications"
Friday, November 20, 2015
301 405 6576
Flexible ADMM for Big Data Applications
School of Applied Math and Statistics
Johns Hopkins University
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.