Universal Gradient Methods for Stochastic Convex Optimization

Fuente: arXiv
Guardado en:
Detalles Bibliográficos
Autores principales: Rodomanov, Anton, Kavis, Ali, Wu, Yongtao, Antonakopoulos, Kimon, Cevher, Volkan
Formato: Preprint
Publicado: 2024
Materias:
Acceso en línea:
Etiquetas: Agregar Etiqueta
Sin Etiquetas, Sea el primero en etiquetar este registro!
_version_ 1866916319646973952
author Rodomanov, Anton
Kavis, Ali
Wu, Yongtao
Antonakopoulos, Kimon
Cevher, Volkan
author_facet Rodomanov, Anton
Kavis, Ali
Wu, Yongtao
Antonakopoulos, Kimon
Cevher, Volkan
contents We develop universal gradient methods for Stochastic Convex Optimization (SCO). Our algorithms automatically adapt not only to the oracle's noise but also to the Hölder smoothness of the objective function without a priori knowledge of the particular setting. The key ingredient is a novel strategy for adjusting step-size coefficients in the Stochastic Gradient Method (SGD). Unlike AdaGrad, which accumulates gradient norms, our Universal Gradient Method accumulates appropriate combinations of gradient- and iterate differences. The resulting algorithm has state-of-the-art worst-case convergence rate guarantees for the entire Hölder class including, in particular, both nonsmooth functions and those with Lipschitz continuous gradient. We also present the Universal Fast Gradient Method for SCO enjoying optimal efficiency estimates.
format Preprint
id arxiv_https___arxiv_org_abs_2402_03210
institution arXiv
publishDate 2024
record_format arxiv
spellingShingle Universal Gradient Methods for Stochastic Convex Optimization
Rodomanov, Anton
Kavis, Ali
Wu, Yongtao
Antonakopoulos, Kimon
Cevher, Volkan
Optimization and Control
We develop universal gradient methods for Stochastic Convex Optimization (SCO). Our algorithms automatically adapt not only to the oracle's noise but also to the Hölder smoothness of the objective function without a priori knowledge of the particular setting. The key ingredient is a novel strategy for adjusting step-size coefficients in the Stochastic Gradient Method (SGD). Unlike AdaGrad, which accumulates gradient norms, our Universal Gradient Method accumulates appropriate combinations of gradient- and iterate differences. The resulting algorithm has state-of-the-art worst-case convergence rate guarantees for the entire Hölder class including, in particular, both nonsmooth functions and those with Lipschitz continuous gradient. We also present the Universal Fast Gradient Method for SCO enjoying optimal efficiency estimates.
title Universal Gradient Methods for Stochastic Convex Optimization
topic Optimization and Control
url https://arxiv.org/abs/2402.03210