Bigger, Regularized, Optimistic: scaling for compute and sample-efficient continuous control

TL;DR

This paper shows that scaling model capacity combined with regularization and optimistic exploration can significantly improve sample efficiency and performance in continuous control reinforcement learning tasks.

Contribution

It introduces the BRO algorithm, which leverages strong regularization and scaling to achieve state-of-the-art results in complex RL benchmarks.

Findings

BRO outperforms existing algorithms on 40 tasks

Achieves near-optimal policies in challenging tasks

Scaling with regularization enhances RL performance

Abstract

Sample efficiency in Reinforcement Learning (RL) has traditionally been driven by algorithmic enhancements. In this work, we demonstrate that scaling can also lead to substantial improvements. We conduct a thorough investigation into the interplay of scaling model capacity and domain-specific RL enhancements. These empirical findings inform the design choices underlying our proposed BRO (Bigger, Regularized, Optimistic) algorithm. The key innovation behind BRO is that strong regularization allows for effective scaling of the critic networks, which, paired with optimistic exploration, leads to superior performance. BRO achieves state-of-the-art results, significantly outperforming the leading model-based and model-free algorithms across 40 complex tasks from the DeepMind Control, MetaWorld, and MyoSuite benchmarks. BRO is the first model-free algorithm to achieve near-optimal policies in the notoriously challenging Dog and Humanoid tasks.