EUROPT 2024
Abstract Submission

148. On Almost Sure Convergence Rates for Stochastic Gradient Methods

Invited abstract in session WF-6: Stochastic Gradient Methods: Bridging Theory and Practice, stream Challenges in nonlinear programming.

Wednesday, 16:20 - 18:00
Room: M:H

Authors (first author is the speaker)

1. Sara Klein
Mathematics Institute, University of Mannheim

Abstract

Stochastic gradient methods are among the most important algorithms in training machine learning problems. While classical assumptions such as strong convexity allow for simple analysis, they are often not satisfied in applications. In recent years, global and local gradient domination properties have been shown to be a sufficient relaxation of strong convexity. They have been proven to hold in diverse settings, such as policy gradient methods. In this talk, we will discuss almost sure convergence rates for SGD (with and without momentum) under global and local gradient domination assumptions. Afterwards, we will apply the results to reinforcement learning, more precisely to softmax parameterized stochastic policy gradient methods.This is joint work with Simon Weissmann and Leif Döring.

Keywords

Status: accepted


Back to the list of papers