HPNAS-RL jointly searches RL architectures and hyperparameters for driving
Original title: Hyperparameter and Neural Architecture Search via Evolutionary Optimization for Reinforcement Learning: HPNAS-RL
In a driving simulator the joint search beat standalone NAS, producing higher-reward policies with fewer parameters.
The paper presents HPNAS-RL, a dual-branch evolutionary method that searches neural architectures and hyperparameters together for reinforcement-learning policies in autonomous driving. Architecture search targets convolutional layers used for visual features, while hyperparameter search tunes RL training settings, building on earlier Evolutionary Multi-Objective NAS work. Parallel evaluation and teacher-student policy transfer are used to speed search and improvegeneralization. In a high-fidelity driving simulator, the joint search consistently beat standalone NAS, with higher rewards, fewer parameters, and more stable policies.
Unconfirmed: Gains are reported only for a driving simulator, with no field or public-road results.
- NAS covers convolutional layers of the RL policy; HPO covers RL hyperparameters.
- The method extends prior Evolutionary Multi-Objective NAS work.
- Teacher-student policy transfer and parallel evaluation speed the search.
- The paper reports no numeric reward or parameter counts.
From the original article
1 min read at arxiv.orgIn this study, we propose a dual-branch evolutionary framework that simultaneously performs Neural Architecture Search (NAS) and Hyperparameter Optimization (HPO) for Reinforcement Learning (RL) in Autonomous Driving (AD) tasks.
By Nihal Acharya Adde · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.