It might be rudimentary, but AI can recreate a game just from watching a game now. Skynet can't be far off, right?
Pac-Man turns 40 years old today, and to celebrate the occasion some researchers at gaming hardware company Nvidia has been hard at work training an AI based around the game. This AI wasn't created to play Pac-Man, however - it was created to build it.
The concept is simple, though the execution is so brutally complicated we won't get into explaining it in detail. Nvidia created a learning AI and sat it in front of footage of Pac-Man being played - 50,000 matches worth. That AI, GameGAN, extrapolated an understanding of the rules, visuals and mechanics of Pac-Man from the footage and then set to its job - to create a fully-functional version of the game derived from its learnings.
GameGAN is named for 'generative adversarial networks', and Nvidia says it is the first neural network built to ape video game engines.
The result here is pretty shocking. While crude and a little sluggish compared to the real thing, the AI has managed to recreate Pac-Man in a playable form - and in partnership with Bandai Namco, Nvidia plans to make the AI's take on Pac-Man available on its AI Playground website later in the year.
"This is the first research to emulate a game engine using GAN-based neural networks," said Seung-Wook Kim, an NVIDIA researcher and lead on the project. "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."
While this is a neat experiment just for showing off the power of AI and the technology available to the researchers at Nvidia, it does have greater potential implications, too: these tools could eventually be used by game developers to have AI help generate game level layouts, for instance. It could also - in classic skynet fashion - be used to help train other AI for other purposes.
Nvidia's official blog post on the AI's Pac-Man clone puts forward a striking example: of a camera on a car, with an AI monitoring both the environment in front of the camera and the driver's inputs to the vehicle. In theory, the AI could learn the rules of driving and of how the car moves on the road just by watching this sort of footage, just in the same way that GameGAN was able to recreate Pac-Man.
If you want to learn more about this striking experiment, Nvidia has a detailed blog post on it and a full, lengthy research paper that gets into the nitty gritty of how it was accomplished. It's impressive stuff, and is the perfect way to celebrate Pac-Man's birthday - at the cutting edge of video games in 1980, and again now.