Balazs Szorenyi : Convergence Rate Results for Stochastic Approximation Algorithms in Reinforcement Learning

Posted by Gyorgy Turan , part of the Mathematical Computer Science Seminar.

At

April 15, 2019, 3 p.m.

In

427 SEO

Abstract

Stochastic approximation (SA) algorithms are sequential stochastic update rules for finding zeros of a function for which only noisy access is available. Due to their easy applicability and natural fit to optimization problems, SA methods have become a fundamental paradigm in various fields. The underlying theory is particularly important in reinforcement learning, where it can be used to analyze the behavior of an agent's policy. In this talk, I would like to present some recent advancements on showing finite time bounds for the problem, and discuss their relation to some basic reinforcement learning algorithms.