[Paper Review] Designer Modeling through Design Style Clustering

Alberto Alvarez, José Font+1 more|arXiv (Cornell University)|Apr 3, 2020

Artificial Intelligence in Games36 references4 citations

TL;DR

This paper proposes modeling designer behavior in mixed-initiative game design through clustering of room design styles and frequent sequence mining to identify archetypical design trajectories, or 'Designer Personas'. Using data from 48 users in the Evolutionary Dungeon Designer, the method identifies 12 cohesive room style clusters and four dominant design personas, enabling AI to anticipate designer intent and guide procedural content generation with low computational cost.

ABSTRACT

We propose modeling designer style in mixed-initiative game content creation tools as archetypical design traces. These design traces are formulated as transitions between design styles; these design styles are in turn found through clustering all intermediate designs along the way to making a complete design. This method is implemented in the Evolutionary Dungeon Designer, a research platform for mixed-initiative systems to create adventure and dungeon crawler games. We present results both in the form of design styles for rooms, which can be analyzed to better understand the kind of rooms designed by users, and in the form of archetypical sequences between these rooms, i.e., Designer Personas.

Motivation & Objective

• To address the challenge of enabling seamless human-AI collaboration in procedural game content generation by modeling designer behavior beyond individual actions.
• To overcome limitations of supervised models trained on individual edits by shifting abstraction to design styles and trajectories.
• To identify archetypical design paths—'Designer Personas'—that reflect recurring patterns in how designers progress through level design.
• To enable AI systems to anticipate designer preferences and suggest content aligned with observed design styles and progression patterns.
• To provide a scalable, data-driven method for modeling designers using unsupervised clustering and frequent sequence mining on real design session data.

Proposed method

• Clustering intermediate design states from 180 unique dungeon rooms into 12 cohesive room style clusters based on design step sequences.
• Using unsupervised learning to identify representative design styles from the sequence of edits made during level creation.
• Applying frequent sequence mining to design session traces to discover recurring archetypical trajectories through style space.
• Mapping each designer’s session to the identified style clusters to determine alignment with specific Designer Personas.
• Visualizing design progress across clusters to validate the emergence of distinct, recurring design paths.
• Using the resulting model to guide procedural content generation by predicting the next likely design style based on current position in style space.

Experimental results

Research questions

• RQ1 How can designer behavior in mixed-initiative game design be modeled beyond individual edits, using higher-level abstractions?
• RQ2 What are the dominant archetypical design trajectories that emerge from real user design sessions in a dungeon-crawler creation tool?
• RQ3 Can clustering of intermediate design states reveal coherent and meaningful room design styles that reflect designer intent?
• RQ4 To what extent do designers follow predictable, recurring paths through style space, and how can these be used to guide AI assistance?
• RQ5 How can the resulting Designer Personas improve the alignment between procedural content generation and designer preferences?

Key findings

• The method successfully identified 12 cohesive and representative room design styles through clustering of intermediate design steps from 180 unique rooms.
• Four dominant archetypical design trajectories—Split Central-Focus, Goal-Oriented, Architectural Focus, and Early Design—were discovered through frequent sequence mining of design sessions.
• Design sessions were found to align with multiple personas depending on progression stage, with some designs initially deviating but later converging on a clear path.
• The approach enables low-cost anticipation of designer intent by identifying current style and likely next steps within the style space map.
• The model successfully captured design uncertainty and adaptation, as seen in early-stage designs that later align with known personas.
• The resulting Designer Personas provide a foundation for guiding search-based or reinforcement learning-based content generation with higher stylistic fidelity.