Structure and Reduction of MCTS for Explainable-AI

MCTS data structures contain valuable reasoning information: The paper demonstrates that Monte Carlo Tree Search—the algorithm component that simulates future scenarios in advanced AI systems like AlphaZero—naturally contains the "reasoning" behind decisions, making it a promising source for generating explanations without requiring separate analysis.

Explanations can be extracted with minimal computational overhead: The authors show that basic explainability metrics can be calculated as an integrated part of the algorithm's normal operation, meaning organizations don't need to choose between having powerful AI systems and having explainable ones—explanations can be generated at relatively low additional cost.

Information theory provides tools for simplifying complex decision trees: The paper uses information-theoretic methods to reduce and simplify the complex tree structures that these algorithms create, which is essential for translating the AI's internal reasoning into formats that humans can actually understand.

Explainability is feasible for real-world planning problems: By focusing on sequential decision-making systems that solve complex, real-life planning challenges (the type of problems AlphaZero tackles), the research suggests that explainability solutions can be practical for deployed AI systems, not just theoretical exercises.

Both offline and online explanations are possible: The paper indicates that explanations of AI decisions can be generated either after-the-fact (offline) or in real-time (online), providing flexibility for different use cases where stakeholders need to understand AI reasoning.