This research presents a novel method for procedural maze generation that incorporates human behavioral tendencies to calibrate difficulty levels. While traditional procedural content generation often relies on algorithmic complexity, this approach utilizes supervised learning to model how human players navigate mazes, specifically focusing on path selection at branch points. By training a model on human navigation data, the authors developed a test player capable of simulating human-like decision-making, which is then used to evaluate and categorize the difficulty of generated mazes.
The methodology involved two primary phases. First, researchers conducted subject experiments with 20 participants who navigated 21 mazes of varying sizes and visibility ranges. Using LightGBM, the team analyzed 23 features—such as distance to the goal and straight-line movement tendencies—to predict player path selection. The resulting model demonstrated a moderate to high positive correlation with actual human behavior. Second, a test player was programmed to navigate mazes based on these learned probabilities. The difficulty was defined by the number of extra steps taken beyond the shortest possible path.
Experimental validation involved 10 additional participants navigating 35 generated mazes categorized into easy, moderate, and difficult groups. Statistical analysis, including Student’s t-tests, confirmed that the difficulty levels predicted by the test player aligned significantly with the actual performance of human players. The study concludes that integrating human-centric behavioral models into procedural generation effectively bridges the gap between algorithmic map creation and player experience. Future research aims to refine the test player’s simulation of human hesitation and expand the scope to include more complex game elements like enemies and items, particularly for role-playing game environments.
