On the Performance of Temporal Difference Learning With Neural Networks

Abstract

Neural Temporal Difference (TD) Learning is an approximate temporal difference method for policy evaluation that uses a neural network for function approximation. Analysis of Neural TD Learning has proven to be challenging. In this paper we provide a convergence analysis of Neural TD Learning with a projection onto $B({\theta }_0,\omega )$, a ball of fixed radius $\omega$ around the initial point ${\theta }_0$. We show an approximation bound of $O(\epsilon )+\tilde{O}(1/\sqrt{m})$ where $\epsilon$ is the approximation quality of the best neural network in $B({\theta }_0,\omega )$ and $m$ is the width of all hidden layers in the network.