Q-learning with function approximation could diverge in the off-policy setting and the target network is a powerful technique to address this issue. In this manuscript, we examine the sample complexity of the associated target Q-learning algorithm in the tabular case with a generative oracle. We point out a misleading claim in [Lee and He, 2020] and establish a tight analysis. In particular, we demonstrate that the sample complexity of the target Q-learning algorithm in [Lee and He, 2020] is O(∣S∣2∣A∣2(1−γ)−5ε−2). Furthermore, we show that this sample complexity is improved to O(∣S∣∣A∣(1−γ)−5ε−2) if we can sequentially update all state-action pairs and O(∣S∣∣A∣(1−γ)−4ε−2) if γ is further in (1/2,1). Compared with the vanilla Q-learning, our results conclude that the introduction of a periodically-frozen target Q-function does not sacrifice the sample complexity.
@article{arxiv.2203.11489,
title = {A Note on Target Q-learning For Solving Finite MDPs with A Generative Oracle},
author = {Ziniu Li and Tian Xu and Yang Yu},
journal= {arXiv preprint arXiv:2203.11489},
year = {2022}
}