Nvidia Research used AI to recreate a full-functioning version of the computer game PAC-MAN, which turned 40 on Friday after it wasted/enriched the lives of millions around the globe.
That means Nvidia had to train 50,000 episodes of the game on something Nvidia created called GameGAN and did so without an underlying game engine. GameGAN is an AI model which was used to recreate the game without understanding the game’s fundamental rules.
If that sounds confusing, the AI process involved may be a little like plopping an inexperienced 10-year-old in front of an old PAC-MAN console with a roll of quarters and saying, “Have fun!” before walking away for several hours. Neural networks try to imitate some of what the human brain does so that analogy isn’t totally off, but Nvidia could use GameGAN to imitate moves in piloting future autonomic vehicles, among other machines.
(They really had PAC-MAN standing consoles in arcades, bars and fine restaurants everywhere in the U.S. in the 1980s and there really were these things called quarter rolls that you had to purchase at a bank on your own.)
GAN stands for generative adversarial networks. A GAN-based model is comprised of two competing neural networks, a generator and a discriminator. In a blog, Nvidia said the PAC-MAN content created with GameGAN is “convincing enough to pass for the original.” (Except for creating the sleazy crowd at the bar, cigarette smoke and 50-cent beers.)
Nvidia researcher Seung-Wook Kim said the work is the first research to emulate a game engine using GAN-based neural networks. “We wanted to see whether the AI could learn the rules of an environment just by looking at the screenplay of an agent moving through the game, and it did,” he said.
This process would probably be like a 10-year-old watching a 12-year-old play PAC-MAN on a single quarter for over an hour and then trying for herself by wasting 10 quarters in 10 minutes on the game. After 50,000 tries, the 10-year-old would get better for sure, and maybe even after 10 tries—the power of the human brain. But what Nvidia did goes farther than playing the game.
The lessons learned from Nvidia’s process could help AI researchers more easily develop simulator systems for training autonomous machines used in many settings, like robots that carry parts around a factory.
The GameGAN model could be used to disentangle the background in a game from moving characters, so it would be possible to recast the game to take place in outdoor hedge or swap out PAC-MAN for an emoji. Game developers can use the GameGAN concept to experiment with new ideas, Nvidia said.
AI could eventually be used to mimic the rules of driving and physics just by watching videos and GameGAN is the first step in that direction, said Sanja Fidler, director of Nvidia’s Toronto research lab.
GameGAN learns game rules by ingesting screen recordings and agent keystrokes from past games. Kim and his collaborators used data from Bandai Namco Research to train the neural networks on PAC-MAN episodes, totaling a few million frames in total. Bandai Namco Entertainment published the game first in Japan and then brought it to the U.S. in 1981.
The classic maze chase game was a hit in gaming arcades and moved into versions for PCs, gaming consoles and cell phones. To play the game, the 8-bit hero PAC-MAN moved through a maze and gobbled up up dots, avoiding ghosts named Inky, Blinky and Clyde. Nvidia’s work on the AI tribute to PAC-MAN will be available on AI Playground later in 2020, Nvidia said.