Advantage-Weighted Regression: Simple and Scalable Off-Policy Reinforcement Learning

TL;DR

Advantage-Weighted Regression (AWR) is a simple, scalable off-policy reinforcement learning algorithm that uses standard supervised learning techniques to effectively leverage off-policy data for both continuous and discrete actions.

Contribution

We introduce AWR, a novel off-policy RL method that employs simple supervised learning steps, enabling effective learning from static datasets and complex tasks with minimal implementation complexity.

Findings

AWR achieves competitive performance on OpenAI Gym benchmarks.

AWR outperforms many off-policy algorithms on static datasets.

AWR successfully handles complex continuous control tasks.

Abstract

In this paper, we aim to develop a simple and scalable reinforcement learning algorithm that uses standard supervised learning methods as subroutines. Our goal is an algorithm that utilizes only simple and convergent maximum likelihood loss functions, while also being able to leverage off-policy data. Our proposed approach, which we refer to as advantage-weighted regression (AWR), consists of two standard supervised learning steps: one to regress onto target values for a value function, and another to regress onto weighted target actions for the policy. The method is simple and general, can accommodate continuous and discrete actions, and can be implemented in just a few lines of code on top of standard supervised learning methods. We provide a theoretical motivation for AWR and analyze its properties when incorporating off-policy data from experience replay. We evaluate AWR on a suite of standard OpenAI Gym benchmark tasks, and show that it achieves competitive performance compared to a number of well-established state-of-the-art RL algorithms. AWR is also able to acquire more effective policies than most off-policy algorithms when learning from purely static datasets with no additional environmental interactions. Furthermore, we demonstrate our algorithm on challenging continuous control tasks with highly complex simulated characters.