Representing and Generating Levels Over Time through Playtrace Reconstructive Partitioning

The cake representation and Playtrace Reconstructive Partitioning (PRP) algorithm offer a novel, domain-independent framework for procedural content generation by encoding game levels as a series of temporal slices. By reconstructing levels from existing playtraces rather than relying on hand-authored constraints, this approach captures dynamic gameplay information to ensure high-quality, playable output. The primary thesis posits that partitioning temporal entities derived from player behavior allows for the generation of diverse, valid game environments that maintain structural integrity across various grid-based systems.

Evaluations within the Sokoban domain demonstrate that PRP achieves 100% playability, consistently outperforming or matching six state-of-the-art procedural generation baselines in terms of tile diversity and duplication rates. Ablation studies reveal that while providing an empty partial specification yields the highest overall diversity, more constrained inputs allow for greater designer control. Furthermore, successful case studies in OvercookedAI and SkeleWalker confirm that the methodology generalizes effectively across different game mechanics with minimal domain-specific adjustments.

Despite these successes, the approach faces limitations, including potential information loss during the conversion of gameplay data and the necessity for specific domain-processing logic. Current research covers grid-based environments, but future developments aim to refine the reconstruction process and extend the application of PRP into more complex, three-dimensional spaces such as Minecraft. By bridging the gap between raw player data and structured level design, this methodology provides a robust foundation for future co-creative design tools that prioritize both algorithmic efficiency and human-centric gameplay experiences.