Differentiable Architecture Search for Reinforcement Learning

TL;DR

This paper explores the application of gradient-based neural architecture search, specifically DARTS, to reinforcement learning, demonstrating significant performance improvements with efficient search compared to random methods.

Contribution

It adapts DARTS for RL, showing it can find architectures that outperform manual designs with less computational cost.

Findings

Discovered architectures achieve up to 250% performance gains.

DARTS improves architectures up to 30x more efficiently than random search.

Gradient-based NAS is effective for RL architecture optimization.

Abstract

In this paper, we investigate the fundamental question: To what extent are gradient-based neural architecture search (NAS) techniques applicable to RL? Using the original DARTS as a convenient baseline, we discover that the discrete architectures found can achieve up to 250% performance compared to manual architecture designs on both discrete and continuous action space environments across off-policy and on-policy RL algorithms, at only 3x more computation time. Furthermore, through numerous ablation studies, we systematically verify that not only does DARTS correctly upweight operations during its supernet phrase, but also gradually improves resulting discrete cells up to 30x more efficiently than random search, suggesting DARTS is surprisingly an effective tool for improving architectures in RL.