A Projection-free Incremental Gradient Method for Large-scale Constrained Optimization
Lists
Only you see this
Abstract
The problem of minimizing finite sums where each component function is associated with a block of dataset, is very popular in machine learning. Among well-known avenues to address this class of problems is the incremental gradient (IG) methods where a single component function is selected at each iteration in a cyclic (or randomized) manner. When the problem is constrained, the existing IG schemes often require a projection step onto the feasible set at each iteration. This step is computationally costly, in particular, in large-scale applications. To the best of our knowledge, the only variant of the IG method that does not require a projection-type step is the incremental augmented Lagrangian method which can only address the separable equality-constrained optimization problems. Motivated by this gap, we consider the problem of minimizing a strongly convex and smooth finite sum function, subject to nonseparable linear constraints, associated with each component function. By introducing a novel regularization-based relaxation technique, we develop an iteratively regularized IG algorithm that requires no projections. With a proper choice of update rules for the stepsize and the regularization parameters, we derive a sublinear rate of convergence for the objective function and establish an asymptotic convergence for the infeasibility of the generated iterates. We compare the performance of the proposed scheme with that of the classical projected IG schemes and SAGA. The effectiveness of our method is validated by increasing both the dimensionality of the solution space and the number of constraints.
BibTeX
@misc{kaushik2020projection,
author = {Harshal D. Kaushik and Farzad Yousefian},
title = {{A Projection-free Incremental Gradient Method for Large-scale Constrained Optimization}},
year = {2020},
eprint = {2006.07956},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2006.07956},
}Loading related content...