An analysis of DOOM level generation using Generative Adversarial Networks

Abstract

Generative Adversarial Networks (GANs) learn models of data distributions that can be employed to generate synthetic data with similar characteristics. In this paper, we analyze how GANs can create levels for the iconic first-person-shooter Doom. We designed a framework to train GANs to extract regularities from human-designed levels and trained them using more than a thousand levels, taken from the most extensive online library of Doom content. We trained two GAN models: an unconditional one using only visual information about the levels; a conditional one integrating the same visual information with features capturing high-level structures of the levels. We evaluated the two models by comparing the levels they generated against the human-designed levels used for training. First, we compared the levels using topological metrics inspired by the ones used in robotics showing that the conditional model produces levels more similar to the human-designed ones. Next, we compared the levels using the high-level structural features used for the conditional network, showing that the generated levels are similar to human-designed ones when considering features describing the spatial layout. Finally, we analyzed how much the generation of levels in the conditional network can be controlled using the input features. Our results show that some input features (like the ones related to the number of rooms and the size of the walkable area) influence the generation process. In contrast, the remaining features appear to be ineffective in this respect.