Differentiable Architecture Search for Reinforcement Learning
This work addresses the problem of automating architecture design for RL practitioners, showing that DARTS can be surprisingly effective, though it is incremental as it adapts an existing NAS method to RL.
The paper tackled the applicability of gradient-based neural architecture search (NAS) to reinforcement learning (RL), finding that discrete architectures discovered using DARTS achieved up to 250% performance compared to manual designs across various RL environments and algorithms, with only 3x more computation time.
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.