Terrain-adaptive PCGML in Minecraft
Arthur van der Staaij, Mike Preuß, Christoph Salge · 2024
We make a first step towards terrain-adaptive PCGML in Minecraft by introducing an automated system to create “volume-to-volume” datasets suitable for machine learning by leveraging handwritten black-box Minecraft settlement generation algorithms. Using this system, we create ten terrain-adaptive Minecraft ML datasets - including ones based on the currently best-performing algorithm submitted to the Generative Design in Minecraft (GDMC) competition. Finally, we train and qualitatively evaluate various GAN-based volume-to-volume models on all ten of our datasets. Although we do not obtain good results in all cases, we demonstrate that terrain-adaptive PCGML in Minecraft is indeed feasible.