16
0

Convergence of Batch Asynchronous Stochastic Approximation With Applications to Reinforcement Learning

Abstract

Ever since its introduction in the classic paper of Robbins and Monro in 1951, Stochastic Approximation (SA) has become a standard tool for finding a solution of an equation of the form f(θ)=0f(\theta) = 0, when only noisy measurements of f()f(\cdot) are available. In most situations, \textit{every component} of the putative solution θt\theta_t is updated at each step tt. In some applications such as QQ-learning, a key technique in Reinforcement Learning (RL), \textit{only one component} of θt\theta_t is updated at each tt. This is known as \textbf{asynchronous} SA. The topic of study in the present paper is to study \textbf{Block Asynchronous SA (BASA)}, in which, at each step tt, \textit{some but not necessarily all} components of θt\theta_t are updated. The theory presented here embraces both conventional (synchronous) SA as well as asynchronous SA, and all in-between possibilities. We also prove bounds on the \textit{rate} of convergence of θt\theta_t to the solutions. As a prelude to the new results, we also briefly survey some results on the convergence of the Stochastic Gradient method, proved in a companion paper by the present authors.

View on arXiv
Comments on this paper