In the process of training the neural network has determined the basic principles of levels of this game and then they were able to generate new levels without human help.
Researchers from Cornell University made that will radically change the process of developing a new video game. They created a couple of competing neural networks (Generative Adversarial Network, GAN), and trained them in the very first game-Schutter, DOOM.
GAN-nets, exploring the levels of DOOM and made his own card, which was applied not only topographic features of the virtual space, but also the location of the various active objects, including other playable characters, enemies and monsters in this case.
One network was trained only on the basis of the video data stream passed to it, and the second network transmitted the same data, supplemented by further information obtained during the preliminary analysis. After the network is “swallowed” all the levels of DOOM-and they have become able to generate your own levels. Thus, the quality and complexity of the new levels were pretty high, but the artificial intelligence system in seconds of the time did what would have taken many hours of work of a team consisting of designers, artists and programmers.
In conclusion, it should be noted that researchers from Cornwal did not pursue the goal of creating new levels specifically for DOOM legacy or another Schutter from the first person. This technology can be successfully used in relation to a computer game of any other genre, as you can see, by looking at the project page “Video Game Level Corpus”, located on the famous Github.