[Paper Review] Co-Creative Level Design via Machine Learning

Matthew Guzdial, Nicholas Liao+1 more|arXiv (Cornell University)|Sep 25, 2018

Artificial Intelligence in Games13 references4 citations

TL;DR

This paper proposes a co-creative level design framework where human designers and machine learning agents collaboratively generate game levels, demonstrating that existing PLGML methods are insufficient for co-creation. It introduces active learning adaptations—'Episodic' and 'Continuous'—showing that episodic fine-tuning improves user satisfaction by 3% on average, while continuous learning fails due to high user variability.

ABSTRACT

Procedural Level Generation via Machine Learning (PLGML), the study of generating game levels with machine learning, has received a large amount of recent academic attention. For certain measures these approaches have shown success at replicating the quality of existing game levels. However, it is unclear the extent to which they might benefit human designers. In this paper we present a framework for co-creative level design with a PLGML agent. In support of this framework we present results from a user study and results from a comparative study of PLGML approaches.

Motivation & Objective

• Address the gap in procedural level generation via machine learning (PLGML) for co-creative design, where human-AI collaboration is needed beyond static content generation.
• Investigate whether existing PLGML approaches can effectively support co-creative level design in practice.
• Demonstrate that co-creative PLGML requires training on co-creative datasets or approximations, not just static game level corpora.
• Evaluate the effectiveness of active learning strategies in adapting PLGML agents to individual human designers' preferences.
• Provide evidence that user design styles vary too widely for generalization, necessitating personalized adaptation in co-creative systems.

Proposed method

• Adapted three existing PLGML models—Markov Chain, Bayesian Network, and LSTM—into iterative, turn-based AI partners for a co-creative level editor interface.
• Implemented a user study using a publicly available Super Mario Bros. level editor, enabling real-time human-AI collaboration through turn-wise additions.
• Designed two active learning variants: 'Episodic' (retraining on each user’s session) and 'Continuous' (incremental weight updates across users).
• Trained agents on a simulated dataset of human-AI interactions, using reward signals based on level completion and design quality metrics.
• Used backend logging to track all human and AI actions, enabling precise evaluation of agent performance and user satisfaction.
• Evaluated agents using a simulated interaction setup where AI partners adapted to user behavior, with performance measured via percentile of maximum possible reward.

Experimental results

Research questions

• RQ1 Can existing procedural level generation via machine learning (PLGML) methods effectively support co-creative level design with human designers?
• RQ2 To what extent do user design styles vary in ways that prevent generalization across individuals in co-creative PLGML systems?
• RQ3 Does active learning improve the ability of PLGML agents to adapt to individual human designers in co-creative level design?
• RQ4 Is episodic retraining (per-user) more effective than continuous learning (cumulative) in adapting to diverse user preferences?
• RQ5 Can approximated co-creative datasets be used to train effective co-creative PLGML agents without requiring full user studies for each new game?

Key findings

• Existing PLGML methods are insufficient for co-creative level design, as they fail to adapt to individual user preferences and design variability.
• The episodic active learning variant improved average user satisfaction by approximately 3% compared to the non-active baseline, measured as the percentile of maximum possible reward.
• The continuous active learning variant underperformed both the episodic and non-active versions for six out of eleven participants, indicating poor generalization across diverse user behaviors.
• Participants 8 and 9 showed the greatest improvement with active learning—over tenfold increase in reward—suggesting that sufficient interaction enables effective personalization.
• The non-active baseline achieved an average reward percentile of 1.45, which improved to 1.47 in the episodic variant and dropped to -4.62 in the continuous variant for the final participant.
• The results suggest that user design patterns are too variable for a single model to generalize across, necessitating personalized adaptation via active learning.