Self-Supervised Discovering of Interpretable Features for Reinforcement Learning.

Area of Science:

• Artificial Intelligence
• Computer Vision
• Machine Learning

Background:

• Deep reinforcement learning (RL) excels in complex control but lacks transparency.
• Interpretability is crucial for safety-critical RL applications.
• Existing methods for vision-based RL offer limited explanations.

Purpose of the Study:

• To develop a self-supervised framework for interpretable deep reinforcement learning.
• To enable non-experts to understand RL agent decision-making processes.
• To generate fine-grained attention masks highlighting task-relevant information.

Main Methods:

• Proposed a self-supervised interpretable network (SSINet).
• SSINet generates attention masks to visualize decision-making evidence.
• Evaluated on Atari 2600 games and the Duckietown self-driving simulator.

Main Results:

• The framework provides empirical evidence of agent decision-making.
• It explains performance variations and failures, especially in novel scenarios.
• Demonstrated effective mask generation in a self-supervised, label-free manner.

Conclusions:

• The method offers valuable insights into vision-based RL decision processes.
• Highlights the potential of self-supervised learning for label-free vision tasks.
• Suggests new paradigms for self-supervised segmentation and detection.