Generalizable Episodic Memory for Deep Reinforcement Learning

TL;DR

This paper introduces GEM, a novel episodic memory method for deep reinforcement learning that improves sample efficiency and performance in continuous and discrete domains by organizing experience for implicit planning.

Contribution

GEM is a new episodic memory approach that effectively organizes state-action values for better experience aggregation and planning in continuous and discrete RL tasks.

Findings

GEM outperforms existing methods on MuJoCo tasks.

GEM shows significant improvements on Atari games.

The method reduces overestimation bias in value propagation.

Abstract

Episodic memory-based methods can rapidly latch onto past successful strategies by a non-parametric memory and improve sample efficiency of traditional reinforcement learning. However, little effort is put into the continuous domain, where a state is never visited twice, and previous episodic methods fail to efficiently aggregate experience across trajectories. To address this problem, we propose Generalizable Episodic Memory (GEM), which effectively organizes the state-action values of episodic memory in a generalizable manner and supports implicit planning on memorized trajectories. GEM utilizes a double estimator to reduce the overestimation bias induced by value propagation in the planning process. Empirical evaluation shows that our method significantly outperforms existing trajectory-based methods on various MuJoCo continuous control tasks. To further show the general applicability, we evaluate our method on Atari games with discrete action space, which also shows a significant improvement over baseline algorithms.