Sample-Efficient Deep Reinforcement Learning via Episodic Backward Update

TL;DR

This paper introduces Episodic Backward Update (EBU), a new deep reinforcement learning algorithm that efficiently propagates value information through entire episodes, improving learning speed and performance in various environments.

Contribution

The paper presents a novel episodic backward update method with a theoretical convergence proof and demonstrates superior sample efficiency in Atari games.

Findings

Achieves human-normalized performance using only 5-10% of samples

Converges theoretically in both deterministic and stochastic settings

Outperforms traditional methods in sample efficiency

Abstract

We propose Episodic Backward Update (EBU) - a novel deep reinforcement learning algorithm with a direct value propagation. In contrast to the conventional use of the experience replay with uniform random sampling, our agent samples a whole episode and successively propagates the value of a state to its previous states. Our computationally efficient recursive algorithm allows sparse and delayed rewards to propagate directly through all transitions of the sampled episode. We theoretically prove the convergence of the EBU method and experimentally demonstrate its performance in both deterministic and stochastic environments. Especially in 49 games of Atari 2600 domain, EBU achieves the same mean and median human normalized performance of DQN by using only 5% and 10% of samples, respectively.