GameGAN, one of Nvidia’s multiple AIs, managed to recreate Pac-Man from scratch thanks to antagonistic generative networks. The result is impressive.
Pac-Man is partying. The popular video game, one of the most emblematic and benchmarks of all time, is celebrating its 40th anniversary. Many companies in the industry are congratulating the Namco title, as four decades later it is still enjoyed by millions of people. At Nvidia , for example, they have seized the occasion to reveal the latest achievement of their artificial intelligence: recreating Pac-Man from scratch. Yes, you read well.
GameGAN , as Nvidia has named its artificial intelligence, was able to bring Pac-Man to life thanks to machine learning. It is important to make clear, however, that recreation consists solely of generating images, not developing playable code. According to the company, the AI relied on antagonistic gene networks – GAN in English, hence the name of the AI - to make it possible.
Nvidia’s GAN is made up of two neural networks, one generating and the other discriminating . The first aims to study a large number of images until generating something based on what has been learned. The other takes the image created by her “partner” and compares it to the database to determine how close she came to resembling the original images. Both work together to obtain the most accurate result possible.
The recreation of Pac-Man consists of the generation of images, not the development of a playable code.
GameGAN trained by watching 50,000 Pac-Man game sessions . He analyzed the design of the level, how the walls and balls work, the movements of the protagonist and the role of ghosts. At the end, the artificial intelligence could learn the rules of Pac-Man to start generating the correct frames. The result, as you might expect, is surprisingly accurate. You can check it yourself by watching the following video:
The GAN beyond Pac-Man
Nvidia notes that over the past few years, several artificial intelligence solutions have been created that can play games. However, this is the first time that an antagonistic generative network managed to recreate a video game from scratch. For now, it remains an experiment that pays tribute to Pac-Man’s 40 years, but the company anticipates that its technology can benefit, for example, the autonomous driving sector: “Eventually we could have an AI that can learn to imitate driving rules or the laws of physics just by watching videos, ” Engadget was told.
👇 More in NUpgrade