DAG-NAS: Explainable Neural Architecture Search\\for Reinforcement Learning via Scalar-level DAG Modeling
Keywords: automated machine learning, neural architecture search, directed acyclic graphs, reinforcement learning
Abstract: We present an explainable and effective Neural Architecture Search (NAS) framework for Reinforcement Learning (RL). We model a feed-forward neural network as a Directed Acyclic Graph (DAG) that consists of scalar-level operations and their interconnections. We train the model for RL tasks using a differentiable search method, followed by pruning the search outcomes. This process results in a compact neural architecture that achieves high performance and enhances explainability by emphasizing crucial information for solving the RL problem. This process results in a compact and efficient neural architecture that enhances explainability by emphasizing crucial information for solving the RL problem. We apply our NAS framework to the Actor-Critic PPO algorithm, targeting both actor and critic networks. We evaluate its performance across various RL tasks. Extensive experiments demonstrate that our architectures achieve comparable performance with significantly fewer parameters while also enhancing explainability by highlighting key features.
Primary Area: other topics in machine learning (i.e., none of the above)
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 148
Loading