Importance Sampling Placement in Off-Policy Temporal-Difference Methods

TL;DR

This paper investigates the impact of importance sampling placement in off-policy TD methods, revealing that a subtle change in update rules acts as a control variate, reducing variance and enhancing learning performance.

Contribution

It demonstrates that correcting the TD target instead of the entire TD error can be viewed as a control variate, leading to improved off-policy learning algorithms.

Findings

Variance reduction improves learning stability

Modified update rules outperform traditional methods

Enhanced performance across various algorithms

Abstract

A central challenge to applying many off-policy reinforcement learning algorithms to real world problems is the variance introduced by importance sampling. In off-policy learning, the agent learns about a different policy than the one being executed. To account for the difference importance sampling ratios are often used, but can increase variance in the algorithms and reduce the rate of learning. Several variations of importance sampling have been proposed to reduce variance, with per-decision importance sampling being the most popular. However, the update rules for most off-policy algorithms in the literature depart from per-decision importance sampling in a subtle way; they correct the entire TD error instead of just the TD target. In this work, we show how this slight change can be interpreted as a control variate for the TD target, reducing variance and improving performance. Experiments over a wide range of algorithms show this subtle modification results in improved performance.