GenDICE: Generalized Offline Estimation of Stationary Values

TL;DR

GenDICE introduces a novel, consistent method for estimating stationary distribution-based quantities in reinforcement learning from fixed data, using divergence minimization techniques, with strong theoretical guarantees and empirical results.

Contribution

It proposes a new algorithm, GenDICE, for offline estimation of stationary values that is simple, consistent, and effective, extending the applicability of stationary distribution estimation.

Findings

GenDICE is consistent under general conditions.

It performs well on benchmark problems like off-line PageRank.

The method provides a new approach for offline stationary distribution estimation.

Abstract

An important problem that arises in reinforcement learning and Monte Carlo methods is estimating quantities defined by the stationary distribution of a Markov chain. In many real-world applications, access to the underlying transition operator is limited to a fixed set of data that has already been collected, without additional interaction with the environment being available. We show that consistent estimation remains possible in this challenging scenario, and that effective estimation can still be achieved in important applications. Our approach is based on estimating a ratio that corrects for the discrepancy between the stationary and empirical distributions, derived from fundamental properties of the stationary distribution, and exploiting constraint reformulations based on variational divergence minimization. The resulting algorithm, GenDICE, is straightforward and effective. We prove its consistency under general conditions, provide an error analysis, and demonstrate strong empirical performance on benchmark problems, including off-line PageRank and off-policy policy evaluation.