ATA: An Abstract-Train-Abstract approach for explanation-friendly deep reinforcement learning.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Deep Reinforcement Learning

Background:

• Explaining decision-making in deep reinforcement learning (DRL) neural networks is difficult.
• Abstract policy graphs (APGs) are effective for model explanation but face challenges in achieving high fidelity and explainability.
• Larger cluster sizes in APGs correlate with higher fidelity.

Purpose of the Study:

• To introduce a novel approach, Abstract-Train-Abstract (ATA), for constructing high-fidelity and explainable APGs.
• To improve the accuracy of predicting actions associated with abstract states in DRL models.
• To enhance user understanding and prediction accuracy of DRL decision-making processes.

Main Methods:

• Developed the Abstract-Train-Abstract (ATA) method, integrating abstraction-based training and abstraction-oriented clustering.
• Abstraction-based training expands the scope of abstract state clusters.
• Abstraction-oriented clustering ensures states within a cluster map to the same action.

Main Results:

• ATA achieved up to 26.63% higher fidelity compared to state-of-the-art methods.
• The approach maintained competitive reward levels.
• A user study showed ATA improved user prediction accuracy by an average of 35.7%.

Conclusions:

• ATA offers a significant advancement in creating explainable and high-fidelity APGs for DRL.
• The method enhances the accuracy of action prediction by effectively clustering abstract states.
• ATA demonstrably improves both model performance and human understanding of DRL systems.