New Aspects of Black Box Conditional Gradient: Variance Reduction and One Point Feedback

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Veprikov, Andrey, Bogdanov, Aleksandr, Minashkin, Vladislav, Beznosikov, Aleksandr
Format: Preprint
Published: 2024
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866915043443998720
author Veprikov, Andrey
Bogdanov, Aleksandr
Minashkin, Vladislav
Beznosikov, Aleksandr
author_facet Veprikov, Andrey
Bogdanov, Aleksandr
Minashkin, Vladislav
Beznosikov, Aleksandr
contents This paper deals with the black-box optimization problem. In this setup, we do not have access to the gradient of the objective function, therefore, we need to estimate it somehow. We propose a new type of approximation JAGUAR, that memorizes information from previous iterations and requires $\mathcal{O}(1)$ oracle calls. We implement this approximation in the Frank-Wolfe and Gradient Descent algorithms and prove the convergence of these methods with different types of zero-order oracle. Our theoretical analysis covers scenarios of non-convex, convex and PL-condition cases. Also in this paper, we consider the stochastic minimization problem on the set $Q$ with noise in the zero-order oracle; this setup is quite unpopular in the literature, but we prove that the JAGUAR approximation is robust not only in deterministic minimization problems, but also in the stochastic case. We perform experiments to compare our gradient estimator with those already known in the literature and confirm the dominance of our methods.
format Preprint
id arxiv_https___arxiv_org_abs_2409_10442
institution arXiv
publishDate 2024
record_format arxiv
spellingShingle New Aspects of Black Box Conditional Gradient: Variance Reduction and One Point Feedback
Veprikov, Andrey
Bogdanov, Aleksandr
Minashkin, Vladislav
Beznosikov, Aleksandr
Optimization and Control
This paper deals with the black-box optimization problem. In this setup, we do not have access to the gradient of the objective function, therefore, we need to estimate it somehow. We propose a new type of approximation JAGUAR, that memorizes information from previous iterations and requires $\mathcal{O}(1)$ oracle calls. We implement this approximation in the Frank-Wolfe and Gradient Descent algorithms and prove the convergence of these methods with different types of zero-order oracle. Our theoretical analysis covers scenarios of non-convex, convex and PL-condition cases. Also in this paper, we consider the stochastic minimization problem on the set $Q$ with noise in the zero-order oracle; this setup is quite unpopular in the literature, but we prove that the JAGUAR approximation is robust not only in deterministic minimization problems, but also in the stochastic case. We perform experiments to compare our gradient estimator with those already known in the literature and confirm the dominance of our methods.
title New Aspects of Black Box Conditional Gradient: Variance Reduction and One Point Feedback
topic Optimization and Control
url https://arxiv.org/abs/2409.10442