Provably Convergent Two-Timescale Off-Policy Actor-Critic with Function Approximation

TL;DR

This paper introduces COF-PAC, a novel off-policy actor-critic algorithm with proven convergence, utilizing a new emphasis critic trained via Gradient Emphasis Learning, combining ideas from existing temporal difference methods.

Contribution

It presents the first provably convergent two-timescale off-policy actor-critic algorithm with function approximation, featuring a new emphasis critic trained through Gradient Emphasis Learning.

Findings

Proves convergence of COF-PAC with linear critics and nonlinear actor.

Introduces the emphasis critic trained via Gradient Emphasis Learning.

Combines ideas from Gradient and Emphatic Temporal Difference Learning.

Abstract

We present the first provably convergent two-timescale off-policy actor-critic algorithm (COF-PAC) with function approximation. Key to COF-PAC is the introduction of a new critic, the emphasis critic, which is trained via Gradient Emphasis Learning (GEM), a novel combination of the key ideas of Gradient Temporal Difference Learning and Emphatic Temporal Difference Learning. With the help of the emphasis critic and the canonical value function critic, we show convergence for COF-PAC, where the critics are linear and the actor can be nonlinear.