MaskCtrl: Training mask networks as self-explainable and performant controllers via deep reinforcement learning.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Control Systems

Background:

• Explainability is crucial for safe and accountable control in deep reinforcement learning (DRL).
• Current explanation methods often train models offline, leading to feature importance misalignment and overlooking critical state features.
• This disconnection from environmental feedback hinders the accuracy of explanations.

Purpose of the Study:

• To propose MaskCtrl, a novel deep reinforcement learning framework for training self-explainable controllers.
• To develop controllers that achieve both high decision-making performance and accurate identification of critical state features.
• To address the feature importance misalignment problem in DRL explainability.

Main Methods:

• MaskCtrl trains mask networks as controllers using a DRL-based mechanism.
• The framework leverages system rewards as environmental feedback to correct feature misalignment during training.
• This approach integrates decision-making and explanation model training.

Main Results:

• Self-explainable controllers demonstrate decision-making performance comparable to original neural controllers.
• MaskCtrl achieves up to 50.58% higher fidelity in critical feature identification compared to offline methods.
• Enhanced feature alignment improves adversarial attack effectiveness by 25.2% and robustness against perturbations.

Conclusions:

• MaskCtrl offers a DRL-based solution for creating self-explainable controllers with accurate feature identification.
• The proposed method significantly improves the reliability and robustness of DRL systems.
• Self-explainable controllers provide a more trustworthy and accountable alternative to traditional explanation models.