Infinite-horizon Off-Policy Policy Evaluation with Multiple Behavior Policies

TL;DR

This paper introduces EMP, a new method for off-policy policy evaluation that effectively estimates stationary distribution corrections using data from multiple behavior policies, resulting in improved accuracy.

Contribution

The paper proposes the estimated mixture policy (EMP), a novel, partially policy-agnostic approach that reduces variance and improves estimation of stationary distribution corrections in off-policy evaluation.

Findings

EMP achieves significantly better accuracy than existing methods.

EMP reduces variance in estimating state stationary distributions.

EMP performs well in both continuous and discrete environments.

Abstract

We consider off-policy policy evaluation when the trajectory data are generated by multiple behavior policies. Recent work has shown the key role played by the state or state-action stationary distribution corrections in the infinite horizon context for off-policy policy evaluation. We propose estimated mixture policy (EMP), a novel class of partially policy-agnostic methods to accurately estimate those quantities. With careful analysis, we show that EMP gives rise to estimates with reduced variance for estimating the state stationary distribution correction while it also offers a useful induction bias for estimating the state-action stationary distribution correction. In extensive experiments with both continuous and discrete environments, we demonstrate that our algorithm offers significantly improved accuracy compared to the state-of-the-art methods.