Error Analysis of Fitted Q-iteration with ReLU-activated Deep Neural Networks

Lican Kang, Han Yuan, Chang Zhu

Keywords: Reinforcement learning, Fitted Q-iteration, Deep neural networks, Non-asymptotic error bound

TL;DR: In this paper, we provide an error analysis for deep-fitted $Q$-iteration applying ReLU-activated FNN for value function approximation.

Abstract: Deep reinforcement learning (RL) has grown rapidly with the development of backbone feedforward neural networks (FNNs). However, there remains a theoretical gap when researchers conduct error analysis of the FNNs-based RL process. In this work, we provide an error analysis for deep-fitted $Q$-iteration applying ReLU-activated FNNs for value function approximation.