Emphatic Temporal-Difference Learning

TL;DR

This paper reviews and unifies recent advances in emphatic temporal-difference learning algorithms, highlighting their stability, convergence, and empirical advantages in off-policy reinforcement learning scenarios.

Contribution

It provides a unified summary of key theoretical results and demonstrates the practical benefits of emphatic algorithms with flexible features.

Findings

Algorithms are stable and convergent under off-policy training.

Empirical benefits include improved stability and flexibility.

State-dependent discounting and resource allocation enhance learning.

Abstract

Emphatic algorithms are temporal-difference learning algorithms that change their effective state distribution by selectively emphasizing and de-emphasizing their updates on different time steps. Recent works by Sutton, Mahmood and White (2015), and Yu (2015) show that by varying the emphasis in a particular way, these algorithms become stable and convergent under off-policy training with linear function approximation. This paper serves as a unified summary of the available results from both works. In addition, we demonstrate the empirical benefits from the flexibility of emphatic algorithms, including state-dependent discounting, state-dependent bootstrapping, and the user-specified allocation of function approximation resources.