Quantifying Feature Importance of Games and Strategies via Shapley Values

This analysis introduces two novel methodologies for quantifying feature importance in two-player, zero-sum games using Shapley values. The primary objective is to bridge the gap between high-performing, opaque game AI and human interpretability. By applying Shapley-based explanations, the research aims to provide insights into both the fundamental nature of a game and the specific decision-making logic of individual AI agents, thereby facilitating better human-AI collaboration and strategic learning.

The first method, Shapley Game Feature Importance (SGFI), provides a global interpretation of a game by measuring the impact of specific features on the expected return of a Nash equilibrium strategy. By systematically obscuring features through game abstraction, the model identifies which elements are most critical for optimal play. The second method, Shapley Strategy Feature Importance (SSFI), is a model-agnostic approach that quantifies the contribution of specific features to an AI’s action selection at the local level, without requiring additional training.

Empirical validation was conducted using the card game Goofspiel. In the SGFI analysis, the "Center" and "Deck" cards were identified as the most significant features, aligning with human strategic intuition. The SSFI analysis demonstrated that the importance of features can fluctuate significantly across different game states, providing a granular view of how an AI weighs variables like opponent hand composition and point differentials.

The study concludes that these Shapley-based frameworks successfully produce intuitive explanations that augment human understanding of complex game environments. While acknowledging that feature importance alone does not fully explain the underlying mechanics of a strategy, the research posits that these tools offer a robust foundation for developing more transparent AI, ultimately supporting more effective human learning and decision-making in competitive domains.