Player Style Clustering without Game Variables

This research addresses the challenge of identifying player styles in video games without relying on internal game variables, which typically require source code access and manual selection. The primary thesis is that a hybrid approach—combining unsupervised state representation learning with a reference-based clustering algorithm—can effectively extract player styles from raw gameplay video data. This methodology eliminates the need for domain-specific variable selection and reduces the risk of bias inherent in manually defined metrics.

The study proposes a hybrid encoder architecture, the STDIM-VAE, which integrates visual frame information with high-level game variables to create robust representations. To process these representations, a novel reference-based clustering algorithm is introduced, which functions without the need for hyperparameter tuning. The effectiveness of these methods was evaluated using two distinct environments: Black Smoke (a 2D fully observable environment) and VizDoom (a 3D partially observable environment). The researchers utilized a mixture of human and AI-generated gameplay data to establish ground truth labels for performance comparison.

Key findings indicate that while traditional clustering based on expert-selected game variables remains superior when play styles are well-defined and known, the proposed encoding-based method significantly outperforms variable-based approaches when play styles are unknown or unanticipated. In scenarios involving both expected and unexpected play styles, the proposed method achieved an Adjusted Mutual Information (AMI) score of 0.868, compared to 0.603 for traditional methods. These results suggest that the proposed framework is a highly effective tool for game analytics, particularly for analyzing external data sources like streaming content or "Let’s Play" videos where internal game variables are inaccessible.