Alternating cyclic extrapolation methods for optimization algorithms
Abstract
This article introduces new acceleration methods for fixed-point iterations. Extrapolations are computed using two or three mappings alternately and a new type of step length is proposed with good properties for nonlinear applications. The methods require no problem-specific adaptation and are especially efficient in high-dimensional contexts. Their computation uses few objective function evaluations, no matrix inversion and little extra memory. A convergence analysis is followed by eight applications including gradient descent acceleration for constrained and unconstrained optimization. Performances are on par with or better than competitive alternatives. The algorithm is available as the Julia package SpeedMapping.jl.
Cite
@article{arxiv.2104.04974,
title = {Alternating cyclic extrapolation methods for optimization algorithms},
author = {Nicolas Lepage-Saucier},
journal= {arXiv preprint arXiv:2104.04974},
year = {2021}
}
Comments
30 pages, 10 figures. Changes since previous version: Reorganized intro to improve readability. Made bound checking simpler. Changed the power method application. Added constrained minimization application. Added 9 CUTEst problems. Added interval-censored PH application. Updated software