Information Structures for Causally Explainable Decisions.

Area of Science:

• Artificial Intelligence
• Decision Science
• Cognitive Science

Background:

• AI agents require explainability for trustworthy decision-making under uncertainty.
• Normative decision theories inform AI by specifying actions, outcomes, probabilities, risks, and trade-offs.
• AI systems must adapt to changing conditions by integrating rapid, learned responses with deliberative causal inference.

Purpose of the Study:

• To review how causal models and related concepts enable AI agents to provide trustworthy, explainable decision recommendations.
• To outline a framework for AI systems to recognize and respond to changes affecting user goals.
• To explore the integration of fast, intuitive AI responses with slower, deliberative reasoning for adaptive decision-making.

Main Methods:

• Review of concepts including conditional independence, causal models, heuristic search, uncertainty reduction, and value of information.
• Identification of probabilistic causal dependencies among variables.
• Development of methods for detecting relevant changes, representing them efficiently, and updating causal models and plans.

Main Results:

• Causal models link actions to outcome probabilities, providing rationales for AI decisions.
• AI systems can integrate learned patterns (System 1) with causal inference and planning (System 2) for adaptive responses.
• Efficient representation and updating of causal models support goal achievement in uncertain environments.

Conclusions:

• Causally explainable AI decisions enhance trustworthiness by clarifying the reasoning behind recommendations.
• AI agents can achieve goals more effectively in uncertain environments by efficiently using information via causal models.
• The reviewed concepts provide a principled framework for building adaptive, explainable AI decision support systems.