Re-trainable Procedural Level Generation via Machine Learning (RT-PLGML) as Game Mechanic

Re-trainable Procedural Level Generation via Machine Learning (RT-PLGML) introduces a novel game design paradigm where players actively train and refine procedural generation models as a core mechanic. Unlike traditional procedural content generation, which typically relies on pre-trained or static rules, RT-PLGML empowers players to provide in-game examples that directly influence how the system generates future content. This approach shifts the player's role from a passive consumer of generated levels to an active participant in the machine learning process, offering new opportunities for emergent gameplay and interactive design.

The implementation of RT-PLGML presents several technical and design challenges. Models must be capable of rapid, few-shot learning to accommodate small, player-provided datasets while remaining computationally efficient enough for real-time interaction. Furthermore, the system requires explainable machine learning interfaces to help players—who may lack technical expertise—understand how their inputs affect the generator's output. Balancing player agency with the designer's need for structural coherence remains a primary hurdle, as the system must prevent unbalanced or uninteresting gameplay while allowing for creative freedom.

To illustrate these concepts, the research proposes two 2D, tile-based game prototypes: DungeonCollapse and PatternCollapse. Both utilize the WaveFunctionCollapse algorithm to allow players to define training examples through in-game actions, such as placing specific tiles or patterns to influence the generation of subsequent areas. These concepts serve as a foundation for future research into human-computer interaction and machine learning education within gaming environments. By treating the model as a dynamic, user-controllable tool, RT-PLGML provides a framework for investigating how players develop mental models of AI and how developers can create more transparent, interactive, and engaging procedural experiences.