Last month, you may have found that a group of researchers has created a machine learning system that could transform the presentation of the great self-flight V into something that seems almost photorealistic. It turns out at about the same time, another group of AI enthusiasts worked on something a little more impressive involving the title of the open world of Rockstar. Friday, YouTuber Harrison Kinsley shared a video showing Gan Theft Auto, a network of neurons that can generate a large automotive game play set.
Kinsley and Collaborator Daniel Kukieła rendered the Gan Thft Auto with Gamegan, who recreated last year Pac-Man, looking at another I play through the game. Gamegan, as his name suggests, is a contradictory generator network. Each Gan consists of two networks of competing neurons: a generator and a discriminator. The generator is formed on an example of a dataset and then produce content depending on what it has seen. The discriminator, meanwhile, will compare the generator output with the original dataset, in the “coaching” process its counterparty at the output of content which is closer and closer to the source material.
“Every pixel you see here is generated from a neuron network while I play,” Kinsley said in the video. “The neuron network is the entire game. There is no rule written here by us or the motor [rage]. “
Training A GAN is an intense task of GPU. Nvidia lent Kinsley a DGX A100 computer station to make the project a reality. The system comes with four of the company’s GPU A100 and a 64-core AMD server CPU. Kinsley and Kukieła used all this computer power to perform simultaneously 12 IA based on rules. These programs would lead the same highway stretch, collecting the data that the neuronal network had to start generating its own world of play. Both have also developed a superampling AI to clean the output of the neuron network, so that It would not seem so pixeled.
As you can see from the video, the network models a surprising number of game systems. As the car moves, the shadows under it and the reflection of the sun on its rear windshield. The mountains in the distance are also getting closer. It’s not something that Kinsley was necessarily expected to do when she and Kukieła began training IAs.
There is also a dream quality of the gameplay and it is partly thanks that the neuronal network does not perfectly reproduce all aspects of GTA V. For one, collisions give it problems. Kinsley says there was a case where he saw a police cruise coming to divide into two as it was crash with his car.
If you want to try Gan Theft Auto for yourself, Kinsley and Kukieła downloaded the project to Github. They say that most computers should be able to execute the demo.