English

Asymptotically Efficient Off-Policy Evaluation for Tabular Reinforcement Learning

Machine Learning 2020-07-09 v1 Artificial Intelligence Machine Learning

Abstract

We consider the problem of off-policy evaluation for reinforcement learning, where the goal is to estimate the expected reward of a target policy π\pi using offline data collected by running a logging policy μ\mu. Standard importance-sampling based approaches for this problem suffer from a variance that scales exponentially with time horizon HH, which motivates a splurge of recent interest in alternatives that break the "Curse of Horizon" (Liu et al. 2018, Xie et al. 2019). In particular, it was shown that a marginalized importance sampling (MIS) approach can be used to achieve an estimation error of order O(H3/n)O(H^3/ n) in mean square error (MSE) under an episodic Markov Decision Process model with finite states and potentially infinite actions. The MSE bound however is still a factor of HH away from a Cramer-Rao lower bound of order Ω(H2/n)\Omega(H^2/n). In this paper, we prove that with a simple modification to the MIS estimator, we can asymptotically attain the Cramer-Rao lower bound, provided that the action space is finite. We also provide a general method for constructing MIS estimators with high-probability error bounds.

Keywords

Cite

@article{arxiv.2001.10742,
  title  = {Asymptotically Efficient Off-Policy Evaluation for Tabular Reinforcement Learning},
  author = {Ming Yin and Yu-Xiang Wang},
  journal= {arXiv preprint arXiv:2001.10742},
  year   = {2020}
}

Comments

Includes appendix. Accepted for AISTATS 2020

R2 v1 2026-06-23T13:23:46.078Z