After watching 50000 hours of Pac-Man, Nvidia's AI generated a playable clone - Ars Technica

If you've seen other farms of computers trained on existing games, this has usually come in the form of them learning how to play the game in question.

After watching thousands of hours of a particular game and tracking the most successful moves and reactions in the course of a versus match, these AI routines can then control games, repeat and juggle thousands of strategies, and battle humans.

Nvidia's latest experiment starts in similar fashion, as its AI research team trained a farm of four computers—each equipped with a Quadro GV100 workstation-grade GPU—on 50,000 hours of Pac-Man gameplay.

"Our AI didn't see any of [Pac-Man's] code, just pixels coming out of the game engine," Nvidia representative Hector Marinez said to Ars Technica.

"Any of us could watch hours of people playing Pac-Man, and from that, you could potentially write your own Pac-Man game, just by observing the rules," Marinez said.

Marinez and the rest of Nvidia's available researchers did not clarify exactly what coding subroutine drove this AI's ability to write executable code of its own, nor whether it leaned on existing game engines.

(After all, if simply watching hours of Pac-Man could teach me how to code video games by hand, I would have a very different career.) Eventually, they handed us a copy of the associated paper that explains how Nvidia's research team pulled it off.

In our interview, Nvidia suggested how this technology might some day revolutionize the work pipeline of a video game studio. The representatives didn't do much to back up their bold claim that "there's a straight line from Pac-Man to having GameGAN produce games and simulations" of the same quality as modern 3D mega-hits.

"We've created an AI agent that can learn the rules of the game just by observing it," one Nvidia representative said.

It needed 50,000 hours to interpret a single Pac-Man maze, and again, that was choked up in terms of fidelity; one researcher admitted that higher resolutions are "still an open challenge for these kinds of networks." How many equivalent processing hours will be needed to recreate a dated 3D game like Grand Theft Auto: San Andreas, let alone best-in-class 3D worlds.