Toward Co-creative Dungeon Generation via Transfer Learning

Co-creative procedural content generation via machine learning often faces significant barriers due to the high cost of collecting human interaction data for every new game title. To address this, a transfer learning framework allows agents to leverage knowledge from established domains, such as Super Mario Bros., to assist in the design of new environments like Zelda dungeon rooms. By synthesizing interaction data from existing, non-interactive level layouts, this approach bypasses the need for extensive, game-specific user studies, providing a scalable pathway for developing collaborative design tools.

The research evaluates both reinforcement learning and supervised learning models to determine their efficacy in approximating human-like design behaviors. Findings indicate that reinforcement learning is better suited for the iterative, step-by-step nature of co-creative design, whereas supervised learning excels at predicting final room objectives. When compared to agents trained from scratch, the transfer learning models demonstrate a consistent, though modest, improvement in performance. This suggests that while domain disparity and the simplicity of target environments can limit the magnitude of gains, the underlying methodology remains a viable strategy for cross-domain knowledge application.

Ultimately, the study establishes that transfer learning serves as a functional bridge for co-creative systems, particularly when data scarcity is a concern. While the performance advantages are currently incremental, the ability to repurpose existing design intelligence across distinct game genres offers a promising direction for future development. The research concludes that selecting the appropriate learning architecture—prioritizing reinforcement learning for interactive sessions and supervised learning for autonomous generation—is critical for optimizing the quality and utility of procedural content generation tools.