本研究提出一套分析框架，利用智能體探索所產生的行為軌跡來評估遊戲中的策略多樣性。我們的方法從軌跡中擷取關鍵決策狀態，並將問題轉化為圖論模型，透過多樣性近似最短路徑演算法於狀態轉移圖中找出多條具策略差異性的通關路徑。此方法可自動且量化地分析遊戲的策略彈性，無需人工規則或回饋，即可提供客觀的設計品質評估依據。 ;In modern game design, strategic diversity—the ability for players to complete a game through multiple viable strategies—is a key indicator of design quality, contributing to creativity, engagement, and replayability. Traditional evaluation methods rely on human testers, which are costly, time-consuming, and prone to subjective bias. Recent advances in reinforcement learning have introduced agents capable of autonomously exploring game environments and exhibiting diverse behaviors. However, while much attention has been devoted to increasing agent diversity, the problem of analyzing the diversity of the game environment itself remains underexplored.
In this work, we propose a framework to assess gameplay diversity by analyzing agent-generated behavior trajectories. Our approach identifies key decision states from the agent’s exploration and formulates a graph-based problem to compute diverse paths toward the goal. By applying diverse approximate shortest path algorithms on a state-transition graph, we uncover multiple distinct solutions within the game. This provides a quantitative, automated means to evaluate the strategic flexibility of a game, offering objective insights into its design quality without relying on handcrafted rules or human feedback.
