Abstract
Procedural content generation via machine learning (PCGML) is a growing field in procedural content generation (PCG) in which models are trained on existing content to generate new game content with minimal human intervention. PCGML can allow game developers to generate a large amount of complex playable levels without the need to manually design each level. Our research involves generating new maps for the platform game Mega Man using two proposed ensembles of models, multi-layer multidimensional Markov chains (Multi-ML-MDMC) and multi-layer Markov random fields (Multi-ML-MRF), respectively. We evaluate our models with the following metrics: map playability, map novelty, room regenerations, and map generation time. Extensive experiments show that our Multi-ML-MDMC and Multi-ML-MRF approaches yield a superior performance on Mega Man map generation compared to the existing Multi-GAN and Multi-MDMC baseline approaches.