If you watched 50,000 games of Pac-Man, you’d probably be pretty good at drawing out what you saw. Nvidia’s GameGAN artificial intelligence model, however, was able to recreate the classic dot-chomping game for its 40th birthday from scratch without a traditional game engine in a matter of days.
Generative adversarial networks, or GANs, are made up of two competing neural networks. In the case of GameGAN, the model mimics a computer game engine, which is a first, Nvidia said in its blog post about the project.
Researchers from the chipmaker’s AI Research Lab in Toronto used the company’s DGX systems to train the GameGan neural networks on 50,000 Pac-Man episodes combined with keystroke data of an AI agent playing the game, as well as research data from game developer Bandai Namco.
Once trained, the GameGAN model generated static elements such as maze shapes and dots in addition to moving elements like the infamous Inky, Pinky, Blinky and Clyde ghosts and, of course, Pac-Man himself. It also learned all the key game rules like what happens when Pac-Man gobbles a Power Pellet or teleports from one side of the screen to the other.
The Pac-Man AI tribute will be available later this year on Nvidia’s AI Playground, where you’ll be able to test out the research demos firsthand, Nvidia said.
Although the recreation of the game is certainly cool, there is a bigger picture. Autonomous robots are typically trained in a simulator, and the creation of these simulators is a time-consuming process for developers, according to Nvidia. Training a neural network like GameGAN could possibly replace the need for developers to write these simulators.