On May 22, 1980, Pac-Man was launched, the legendary video game of the ice cream cone that brought about a real revolution first in arcades and later on computers and consoles.
His legacy has been enormous, and now NVIDIA has wanted to pay a unique tribute to the Bandai Namco title. For this they have created an artificial intelligence system that simply playing 50,000 games of Pac-Man has recreated the game from scratch, by itself, respecting the mechanics and rules of the game.
No code is generated, images are generated
The system created by NVIDIA has been dubbed GameGAN, and as its name suggests, makes use of Generative Adversarial Networks (GAN ), a type of algorithm that takes advantage of the philosophy of unsupervised learning.
This approach to this type of problem is already familiar in NVIDIA, which in the past has used it in other areas such as the artistic with GauGAN or the generation of animal images from stuffed animals with GANimal . The idea is that the system learns the rules of the game simply observing it again and again.
According to NVIDIA GameGAN it is the first neural network model that emulates the behavior of a video game engine. It makes use of two neural networks that compete with each other (one generating and the other discriminating), and ends up creating new content that emulates the original content that they were observing.
The idea is relatively simple: as the artificial agent plays the game generated by this unique engine, GameGAN responds to those actions by generating the frames in real-time. This engine does not generate code that is then executed, but basically it generates frames in a coherent way as the different actions that are happening in the game are recognized and evaluated.
GameGAN wants to go beyond Pac-Man
By observing the development of these games, the general model of the static elements of the environment, but also respects its rules. For example, that Pac-man cannot cross the walls or that when he passes over a small ball, he eats it.
With this system it would, therefore, be possible to offer developers a tool to automatically create new levels for their games, but it also allows AI researchers to have a tool to develop simulation systems to train autonomous machines.
The creators of this system stressed that with GameGAN they could “have artificial intelligence that learns to emulate the rules of driving and the laws of physics simply by watching videos and watching how agents make certain decisions in their environment.”
The Pac-Man game created by this artificial intelligence will be available soon on the NVIDIA AI Playground website thanks to the agreement with Bandai Namco for anyone to enjoy.