DOOM Level Generation using Generative Adversarial Networks

[Najork]

I have just come across a reference to this publication available on arXiv DOOM Level Generation using Generative Adversarial Networks. The authors have trained artificial neural networks on a sample of 1088 Doom levels from idgames and then used the trained networks to create new levels. Unfortunately they provide very limited details of the new levels in the paper in the form of a few tiny pictures. Evaluation of the levels uses metrics which compare somewhat abstract characteristics of the generated levels with human made levels. There is no mention of anyone actually playing the levels, and I could not find any links for downloading a sample on the authors' web pages. This seems to be a purely theoretical exercise at the moment, the authors say in their introduction:

"Content creation is nowadays one of the most expensive and time consuming tasks in the game development process. ... In this context, levels are of paramount importance, especially in first person shooter and platform games, as they greatly affect the player experience. Unfortunately, level design usually heavily relies on domain expertise, good practices, and an extensive playtesting. To deal with these issues, several game researchers are spending considerable effort on studying and designing procedural content generation systems that, exploiting machine learning and search algorithms, can model the level design process and assist human designer."

[Reisal]

It would be nice if they can apply some of this science to randomly generated Doom levels, something that would succeed OBLIGE.

[Najork]

As far as I can understand from the paper the levels generated by the network are produced by feeding random variations to the abstracted feature set in the network which characterises human made levels (Figure 3). The authors say that their eventual goal is to generate level designs usable in practice, but at present the metrics for evaluating levels seem to be based only on geometric and abstract characteristics of the levels so it's a moot point whether current generated levels would be interesting or enjoyable to play.

 

As an aside it should be possible to compare OBLIGE levels with the set of levels from idgames to examine their similarity or difference in terms of the authors' measures. However, at present computational restrictions appear to limit the size of levels that they can work with.

Edited May 3, 2018 by Najork : On reflection saw that I missed Glaice's point