Deep learning novelty exploration for minecraft building generation

Abstract

Computational creativity (CC) refers to the study of computational systems that exhibit behaviors that an unbiased observed would consider creative. CC systems have shifted focus toward intrinsic motivation (IM) and open-endedness (OE), which are at the heart of creative behavior in biological systems. In this project, we apply these concepts to a procedural content generator that autonomously creates Minecraft buildings according to its own evolving definition of novelty. This work addresses a research gap in PCG, specifically in Minecraft, which currently focuses on generating adaptive settlement layouts without prioritizing creativity. This system follows the fundamentals of the DeLeNoX algorithm. An autoencoder identifies the high-level features of buildings, compressing them into onedimensional latent vectors. The system alternates between phases of exploration and transformation. In exploration, CPPN-NEAT is used to evolve populations of buildings using constrained novelty search. We calculate an individual’s novelty as the average euclidean distance to the nearest K neighbors in the latent space. In transformation, the autoencoder is retrained with a dataset of the most novel individuals created in the previous exploration phase/s. We experiment with different approaches to the retraining of the autoencoder and observe their impact on the diversity and complexity of the content generated. We assess the results quantitatively by comparing population diversities across experiments, and by visualizing their expressive range using a set of building properties. Finally, we compare the structures qualitatively and observe the effective change in complexity in the structures over time. Our results show that the transformation phase is most effective when it uses larger training sets and includes examples from all previous iterations of the algorithm. This allows the system to more effectively scale in effective complexity of building features, which become more similar to examples of realistic buildings over time.