|GAN Theft Auto - The Neural Network Is The Game|
|Written by Nikos Vaggalis|
|Friday, 16 July 2021|
GAN Theft Auto is a fork of the Nvidia's GameGAN emulating dynamic game environments neural network but applied to Grand Theft Auto.
A couple of researchers replicated a "simple" scene of Grand Theft Auto;a car driving around a highway.The difference is that the car as well as the surrounding environment obeys to rules not pre-written by a game engine but the neural network itself.
Whenever you press the left and right keys to control the car, the NN does the rest.That is turning the wheels,moving the car, and handling the physics alongside the collisions with other cars or items on the level,characteristically described as
Every pixel you see is generated from a neural network.We're used to seeing ai playing within environments but here the ai is the environment it's just a model that we're running with python that takes key presses from the player and outputs pixel values
In other words, we are playing inside of a neural network.
The work is based on Nvidia's GameGan,
a generative model that learns to visually imitate a desired game by ingesting screenplay and keyboard actions during training. Given a key pressed by the agent, GameGAN "renders" the next screen using a carefully designed generative adversarial network.
This was officially showcased when a GameGan network was trained on 50,000 episodes of the timeless PAC-MAN game and by then it could generate a fully functional version of the game without an underlying game engine.
Harrison Kinsley and Daniel Kukiela took the idea a step further in having a GameGAN network generate a playable demo version of a highway stretch of GTA.The attempt even attracted interest from Nvdia itself who loaned them a dgx station beefed up computer sporting four of the 80 gigabyte A100 GPU cards for a total of 320 gigabytes of VRAM, because as we know training neural networks is really expensive business.
And to train a model they needed data,loads of it. Of course manually playing the original GTA game was to expensive time wise so they've obtained the necessary data by having 12 AIs playing the game for them instead.Needless to say the first versions of the models were crude as such colliding with police cars would split them in half or if you drove into a wall the model would just get very confused.
But the final attempts improved that much that even had unforeseeable consequences like the network making the background appear closer as you drove towards it!
Nicely done research wise, but what does that practically mean for the future? It could mean that you can create game engines that learn for example real world physics without being instructed to do so. That could apply to not just games but to other simulations and tools.
However in looking at the future of building game engines based on AI,isn't that a chicken and egg situation in that in the first place you have to play the original game in order to extract the data/model needed to train the NN to generate the game? In any case this is just the beginning.
Of course the work is open source with the code and instructions on how to play demo available on GitHub.
or email your comment to: firstname.lastname@example.org
|Last Updated ( Friday, 16 July 2021 )|