Prediction, Knowledge, and Explainability: Examining the Use of General Value Functions in Machine Knowledge.

Area of Science:

• Computational Reinforcement Learning
• Artificial Intelligence
• Explainable AI (XAI)

Background:

• Growing research in computational reinforcement learning focuses on agents encoding world knowledge via predictions.
• General Value Functions (GVFs) are a prominent method for representing these predictions, with significant theoretical and applied advancements.
• The explainability of GVF-based systems remains an underexplored area.

Purpose of the Study:

• To explore the potential of General Value Functions (GVFs) as a framework for explainable AI.
• To propose a subjective, agent-centric approach to explainability in sequential decision-making.
• To investigate the role of self-explanation in an agent's decision-making process.

Main Methods:

• Articulating a subjective, agent-centric perspective on explainability for sequential decision-making.
• Reviewing existing applications of GVFs, particularly those involving human-agent collaboration.
• Analyzing how agents can introspectively explain their own predictions.

Main Results:

• GVFs can serve as a mechanism for achieving explainable AI.
• An agent's ability to self-explain its predictions is crucial for external explainability.
• Making subjective explanations public enhances operational clarity in collaborative scenarios.

Conclusions:

• General Value Functions offer a promising avenue for developing explainable AI systems.
• Explainability in AI can be framed through an agent's internal, subjective understanding of its predictions.
• Publicly sharing these subjective explanations can significantly improve human-agent collaboration and system transparency.