Hyperspherical Normalization for Scalable Deep Reinforcement Learning

TL;DR

This paper introduces SimbaV2, a novel reinforcement learning architecture that stabilizes training with hyperspherical normalization and reward scaling, enabling scalable, high-performance learning on complex continuous control tasks.

Contribution

The paper proposes SimbaV2, a new RL architecture that stabilizes training for large models using hyperspherical normalization and distributional value estimation.

Findings

Achieves state-of-the-art results on 57 continuous control tasks

Effectively scales with larger models and more compute

Stabilizes optimization in RL with novel normalization techniques

Abstract

Scaling up the model size and computation has brought consistent performance improvements in supervised learning. However, this lesson often fails to apply to reinforcement learning (RL) because training the model on non-stationary data easily leads to overfitting and unstable optimization. In response, we introduce SimbaV2, a novel RL architecture designed to stabilize optimization by (i) constraining the growth of weight and feature norm by hyperspherical normalization; and (ii) using a distributional value estimation with reward scaling to maintain stable gradients under varying reward magnitudes. Using the soft actor-critic as a base algorithm, SimbaV2 scales up effectively with larger models and greater compute, achieving state-of-the-art performance on 57 continuous control tasks across 4 domains. The code is available at https://dojeon-ai.github.io/SimbaV2.