Achieving efficient interpretability of reinforcement learning via policy distillation and selective input gradient...

Area of Science:

• Artificial Intelligence
• Machine Learning
• Deep Learning

Background:

• Deep Reinforcement Learning (RL) excels in various tasks but lacks interpretability for real-world applications.
• Saliency maps are common for deep neural network interpretability, but existing RL methods are computationally expensive or ineffective.
• Current saliency map techniques for RL struggle with real-time performance and policy interpretability.

Purpose of the Study:

• To develop an interpretable and computationally efficient method for generating saliency maps in deep RL.
• To enhance the robustness of RL policies against adversarial attacks.
• To validate the proposed approach across diverse RL tasks.

Main Methods:

• Proposed Distillation with selective Input Gradient Regularization (DIGR) approach.
• Utilized policy distillation and input gradient regularization for enhanced interpretability.
• Evaluated DIGR on MiniGrid, Atari (Breakout), and CARLA Autonomous Driving.

Main Results:

• DIGR produces interpretable saliency maps efficiently.
• The approach improves the robustness of RL policies against adversarial attacks.
• Experimental results demonstrate the effectiveness of DIGR across multiple tasks.

Conclusions:

• DIGR offers a computationally efficient and interpretable solution for deep RL.
• The method enhances RL policy robustness, making it suitable for real-world deployment.
• DIGR represents a significant advancement in the interpretability and reliability of RL systems.