Optimization Online


Accelerating block coordinate descent methods with identification strategies

Ronaldo Lopes (ronaldolps3***at***gmail.com)
Sandra Augusta Santos (sandra***at***ime.unicamp.br)
Paulo J. S. Silva (pjssilva***at***ime.unicamp.br)

Abstract: This work is about active set identification strategies aimed at accelerating block-coordinate descent methods (BCDM) applied to large-scale problems. We start by devising an identification function tailored for bound-constrained composite minimization together with an associated version of the BCDM, called Active BCDM, that is also globally convergent. The identification function gives rise to an efficient practical strategy for Lasso and $\ell_1$-regularized logistic regression. The computational performance of Active BCDM is contextualized using comparative sets of experiments that are based on the solution of problems with data from deterministic instances from the literature. These results have been compared with those of well-established and state-of-the-art methods that are particularly suited for the classes of applications under consideration. Active BCDM has proved useful in achieving fast results due to its dentification strategy. Besides that, an extra second-order step was used, with favorable cost-benefit.

Keywords: Block coordinate descent; active-set identification; large-scale optimization; $\ell_1$ regularization

Category 1: Nonlinear Optimization (Bound-constrained Optimization )


Download: [PDF]

Entry Submitted: 11/22/2017
Entry Accepted: 11/22/2017
Entry Last Modified: 06/20/2018

Modify/Update this entry

  Visitors Authors More about us Links
  Subscribe, Unsubscribe
Digest Archive
Search, Browse the Repository


Coordinator's Board
Classification Scheme
Give us feedback
Optimization Journals, Sites, Societies
Mathematical Optimization Society