IMPALA: Scalable Distributed Deep-RL with Importance Weighted Actor-Learner Architectures

TL;DR

IMPALA introduces a scalable distributed reinforcement learning architecture that efficiently handles large data and training times, achieving superior multi-task performance through a novel off-policy correction method and resource-efficient design.

Contribution

The paper presents IMPALA, a new distributed RL agent that scales to thousands of machines, improves data efficiency, and enables effective multi-task learning with a novel off-policy correction.

Findings

IMPALA outperforms previous agents with less data.

It achieves stable, high-throughput learning.

Demonstrates positive transfer in multi-task settings.

Abstract

In this work we aim to solve a large collection of tasks using a single reinforcement learning agent with a single set of parameters. A key challenge is to handle the increased amount of data and extended training time. We have developed a new distributed agent IMPALA (Importance Weighted Actor-Learner Architecture) that not only uses resources more efficiently in single-machine training but also scales to thousands of machines without sacrificing data efficiency or resource utilisation. We achieve stable learning at high throughput by combining decoupled acting and learning with a novel off-policy correction method called V-trace. We demonstrate the effectiveness of IMPALA for multi-task reinforcement learning on DMLab-30 (a set of 30 tasks from the DeepMind Lab environment (Beattie et al., 2016)) and Atari-57 (all available Atari games in Arcade Learning Environment (Bellemare et al., 2013a)). Our results show that IMPALA is able to achieve better performance than previous agents with less data, and crucially exhibits positive transfer between tasks as a result of its multi-task approach.