Harrison Kinsley / Daniel Kukiela
Relying on machine learning and various training methods, AI can already do a lot of amazing things, but what if you don’t use a 3D engine at all and rely on a neural network to calculate a 3D game? YouTuber Harrison Kinsley shared in the video “GAN Theft Auto” produced by him and a group of artificial intelligence enthusiasts such as Daniel Kukiela. The original English name of “Grand Theft Auto V” (Grand Theft Auto V) was replaced by the first word Replaced by the abbreviation of Generative Adversarial Networks. As the name suggests, this is a world of “Grand Theft Auto” that was “simulated” by generating against the Internet.
The generation of the confrontation network is composed of two relatively anti-neuronal networks, one is responsible for generation and the other is responsible for discrimination. In this example, the AI was dropped on a highway in the world of Grand Theft Auto for learning, to understand how the screen should change when the user presses acceleration, braking, turning left, and turning right. Generating the network will produce a screen that it “guess” should be correct, and the discriminating network will be compared with the actual game screen to “guide” to generate the correct screen for the network.
The result is a vague, but roughly discernible world. All the elements in it are generated by the AI neural network based on “experience”, and no 3D drawing or physical operations are used at all. Even so, the AI still learned quite incredible that the angle of the vehicle’s shadow should change with the rotation, and the sun reflection on the car body is also correct. If the car hits an obstacle (for example, when the fence on both sides), AI knows how to stop the picture, and then slide left or right depending on the angle of the impact, and later even add other vehicles on the road, AI can also (roughly Top) React correctly, even distant mountains will change in size with the “distance”.
Training such a GAN is a task that consumes GPU computing power, so NVIDIA loaned a DGX Station A100 with a built-in 64-core AMD CPU and four A100 graphics cards to Kinsley, which can execute 12 AI training models on the machine at the same time. These models continue to run repeatedly on this section of the road, Kinsley also used AI to smooth the picture to make it look less pixelated, and the final result is a scene like driving in a dream.
Unfortunately, due to lack of time, Kinsley and Kukiela were unable to expand the scope of the experiment. The most important thing is that they are not sure how large this “world” can be expanded before the AI will start sending out strange results; or how far they can interact with other vehicles. At present, most of the interaction with the vehicle ends in the opponent disappears after being hit, but there have also been incidents where the opponent split into two when the collision occurred. In a few cases, AI can produce the correct interaction, such as: When the vehicle is blocked on the left, it will make the left turn invalid, but it will take a long time to train if you want a more accurate representation of the interaction with other vehicles.