Dungeon level generation using generative adversarial network: an experimental study for top-down view games

Abstract

Manual designing levels for games is a complex task, often demanding time and effort from the game designer. An option for this is using algorithms to generate such levels, improving its scalability. Currently, such procedural content generation methods can be guided by hand-crafted rules or, as in more recent approaches, by learning from existing data. In deep learning, generative adversarial networks have been proven to effectively generate samples for a given distribution in an unsupervised manner. Unfortunately, the training and usage of GANs show many challenges, such as convergence problems, instability, and mode collapse. In addition to these problems, we show that learning to generate valid levels is not trivial. In this paper, we demonstrate an empirical analysis of the use of GANs, pointing out the best state-of-the-art practices and highlighting future directions in this research field.