Quantifying Feature Importance of Games and Strategies via Shapley Values

Shapley values can make AI game strategies understandable to humans: The paper demonstrates that Shapley values—a mathematical tool from cooperative game theory—can explain why AI systems make specific moves in games, moving beyond "black box" decision-making that humans cannot interpret.

Two complementary explanation approaches: The authors propose methods to explain both the game structure itself (which features matter most) and individual AI strategies (why a particular AI plays a certain way), providing explanations at different levels of analysis.

Explanations align with human intuition: The paper reports that their Shapley value-based explanations "resonate with and augment human understanding," suggesting this approach produces interpretations that make sense to people rather than generating technically correct but meaningless outputs.

Interpretability matters for strategy evaluation: By making strong AI strategies interpretable, researchers and players can better understand why strategies work, which has practical value for game design, AI development, and human-AI collaboration rather than just knowing that they work.

Bridges game AI and responsible AI goals: This work connects advances in game-playing AI with the broader responsible AI objective of explainability, showing how XAI techniques can be applied to make complex AI systems more transparent and trustworthy.