Last month, you might have seen that a group of researchers created a machine learning system that can change the presentation of the Grand Theft Auto V into something that looks almost photorealistic. Apparently, at the same time, another AI fan group was working on something even more impressive involving the title of the Rockstar Open World. On Friday, Youtuber Harrison Kinsley shares videos that show off auto theft, a neural network that can produce a stretch that can be played from the world of Grand Theft Auto V by itself.
Kinsley and Collaborator Daniel Kukieła made Gan Theft Auto with Gamegeg, which last year re-created Pac-Man by watching AI playing through the game. Gamegan, as suggested as its name, is a network of generative hostility. Each bro consists of two competing neural networks: generators and discriminators. The generator was trained on the sample dataset and then was told to produce content based on what he saw. The discriminators, meanwhile, will compare the output generator with the original dataset, in the “train” process of colleagues to issue closer content and closer to the source material.
“Every pixel you see here is generated from neural networks when I play,” Kinsley said in the video. “Neural networks are the whole game. There are no rules written here by us or the [Rage] machine.”
Training bro is a solid assignment of the GPU. Nvidia lends the Kinsley A DGX Station A100 computer to make the project come true. This system is equipped with four GPU A100 GPUs and a 64-core AMD server CPU. Kinsley and Kukieła use all the computing forces to run 12 AI based on rules simultaneously. The programs will encourage the same highway stretch, collect neural network data needed to start producing its own world of gam. Both also develop AI Supersampling to clean the output of neural networks so it won’t look so pixelated.
As you can see from videos, the network model the surprising number of systems of the game. When the car moves, as well as the shadow below it and the sun reflection on the windshield. The mountains in the distance are even closer. It wasn’t something Kinsley expected AI would do when he and Kukieła first started training AI.
There are also qualities like dreams to gameplay and it’s a part of a portion to the fact that neural networks don’t replicate every aspect of GTA V perfectly. For one, the collision gives a problem. Kinsley said there was one example where he saw a police explorer who will come into two just like it will hit his car.
If you want to try auto theft for yourself, Kinsley and Kukieła have uploaded the project to GitHub. They say most computers must be able to run the demo.