Procedural Content Generator For Platformer Levels

Abstract

Procedural Content Generation (PCG) are algorithms that can generate game content or levels with little to no human intervention. As discussed by PCGML, datasets, and benchmarks in the field of game generation are very limited and lack video gameplay data. Furthermore, there is no unified clear framework for the evaluation of GAN-based PCG algorithms. Therefore, in this thesis, we provide a new clean video gameplay dataset composed of two games Super Mario Bros and Super Mario Bros Lost Levels. We show that learning from language in Platformer PCG outperforms learning from video frames. Moreover, we discuss three approaches to extract meaningful data from the two games to perform learning from language. The approach generates a variety of levels learned from different sources (one level, multiple levels, multiple games). Thus, we show that learning from multiple games is possible with GANs learning from langauge. Furthermore, we categorize several evaluation approaches used in the literature into style difference, playability, and fun and provide an evaluation framework for each to compare different GAN architectures (Simple GAN, DCGAN and WGAN) over different datasets and we show that in most cases WGAN learning outperforms GAN and DCGAN.