The legendary arcade game Pac-Man is turning 40 years old this year, and to celebrate, NVIDIA has trained its GameGAN AI to recreate the iconic title. That might seem like an easy task on the surface, but what’s impressive about NVIDIA’s project is that GameGAN is recreating Pac-Man without the help of a game engine or Pac-Man‘s code. Instead, the GameGAN AI is producing a fully functional version of Pac-Man just by seeing the game in action.
More specifically, GameGAN is a generative adversarial network that leverages two competing neural networks – a generator and a discriminator. Those neural networks worked together through 50,000 episodes of the game, allowing the AI to learn the rules and mechanics of Pac-Man by simply watching the episodes play out. Eventually, the AI was able to create Pac-Man levels that followed the rules of a traditional Pac-Man game.
“This is the first research to emulate a game engine using GAN-based neural networks,” Seung-Wook Kim, an NVIDIA researcher who served as lead author on this project, said in a post to the NVIDIA blog. “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.”
NVIDIA says that “no matter the game,” the GameGAN AI can learn the rules of it “simply by ingesting screen recordings and agent keystrokes from past gameplay.” More complex games might take the GAN more episodes to learn, but it seems that the AI picked up on all of the nuances of Pac-Man with relative ease – in NVIDIA’s version, Pac-Man can’t move through walls, eats the dots as it moves around the field, and the ghosts all turn blue and become consumable when Pac-Man eats a Power Pellet, just like in the normal game.
It’s definitely impressive that NVIDIA was able to train AI to create a fully functional Pac-Man game just by feeding it screen and input recordings, but GameGAN has some real-world uses outside of recreating classic games from the past. NVIDIA explains that GameGAN can be used by developers to automatically generate layouts for new levels, potentially opening the door to an endless amount of content that can be created quickly once the AI has been trained.
Outside of gaming, NVIDIA says that the GameGAN AI could also be used in training autonomous machines, which are typically trained in a simulator to learn the rules of the environment they’ll be working in before they’re put to use. Using GameGAN instead of building a simulator could ultimately save developers a lot of time. Still, such applications seem further off, as Sanja Fidler, director of NVIDIA’s Toronto research lab suggested today.
“We could eventually have an AI that can learn to mimic the rules of driving, the laws of physics, just by watching videos and seeing agents take actions in an environment,” Fidler said. “GameGAN is the first step toward that.” More information about GameGAN can be found over on NVIDIA’s AI Playground, with the company saying that its AI tribute to Pac-Man will be available for everyone to play there later this year.