DAG-NAS : An explainable neural architecture search framework for reinforcement learning

Area of Science:

• Artificial Intelligence
• Machine Learning
• Deep Learning

Background:

• Neural Architecture Search (NAS) is crucial for optimizing deep learning models.
• Explainability in Reinforcement Learning (RL) remains a significant challenge.
• Current NAS methods often produce complex, black-box models.

Purpose of the Study:

• To introduce an explainable NAS framework for Reinforcement Learning (RL).
• To enhance the interpretability of neural network architectures in RL.
• To develop compact and high-performance RL models.

Main Methods:

• Modeling feed-forward neural networks as Directed Acyclic Graphs (DAGs) at the scalar level.
• Employing a differentiable search method for training DAGs.
• Implementing a pruning strategy on search results to refine architectures.

Main Results:

• The NAS framework successfully identified compact neural architectures.
• Achieved comparable performance to existing methods with significantly fewer parameters.
• Demonstrated enhanced explainability by highlighting critical information pathways.

Conclusions:

• The proposed NAS framework effectively balances performance, parameter efficiency, and explainability in RL.
• Scalar-level DAG representation offers a promising direction for interpretable AI.
• The method is validated across various RL tasks, including Actor-Critic PPO.