This May, Pac-Man, one of the most classic games in the history of computing, turns 40 years old. And this Friday (22), Nvidia published that it managed to recreate the game from scratch, using its technology based on artificial intelligence, called GameGAN. According to the company, the software was recreated through the “observation” of matches from the original version of the game.
One software watching another …
The approach used by GameGAN is simple: an artificial intelligence is placed to play Pac-Man, while others are watching all the movements of the moves, for several hours, to learn how the game works. After accumulating enough knowledge, in a few days, the process of recreating the game is carried out. In all, 50,000 games of the original Pac-Man were used as “video lessons” for AI.
The company also explained all the training of the software was done from the observation of the pixels generated by the performed plays: there was no code analysis.
It would be as if a person learned to play chess just by watching the movements of other players, but without having access to the rules.
However, this strategy would not enable the apprentice to recreate the pieces of the game, nor the board. Therefore, it is not known exactly how Nvidia’s AI managed to recreate a game just by looking at how it works.
The most logical assumption is that the tool was developed by integrating deep knowledge about the engine of other games. In this way, she managed to create “another game”, adapting the rules observed in Pac-Man, more or less as in a “copy” process. In this process, four workstations equipped with Quadro GV100 GPUs were used.
Nvidia said the company’s goal with GameGAN is not to recreate or copy existing games, but to create complex new games, at the level of GTA 5 or Red Dead Redemption 2.