Many of you may not know this, but today is a very special day for the world of gaming; today marks the 40th anniversary of Pac-Man’s arrival in video game arcades in Japan. In commemoration of this special day, NVIDIA – with the permission of Bandai Namco – recreated the game both in its entirety and the ground up via an AI Model.
The AI model is called the NVIDIA Game Generative Adversarial Network, or GameGAN for short. Officially, and to be absolutely precise, it recreated the entire Pac-Man game with virtually nothing. Not even with the original game engine. Like all NVIDIA AI Model, GameGAN was made to watch up to 50000 episodes of the game being played before ultimately setting off to build the game.
Overall, one of the researchers told us that GameGAN took a total of four days to recreate the whole Pac-Man. Even more impressive is the fact that the AI Model did so without even understanding the game’s fundamental rules.
On that note, GameGAN was also capable of generating new frames, both in-game and in real-time, when pitted against an artificial agent in its game. Moreover, the researchers believe that with a little more time, GameGAN would be able to generate random game layouts that it has never seen before, provided it trains on games with multiple levels.
NVIDIA says that recreating a game like Pac-Man from scratch with GameGAN is just one possibility of using a deep neural network to train an AI model to write and create simulators like this. To that end, NVIDIA’s researchers have expressed hope that sooner or later, there will be an AI Model that can mimic the rules of driving and laws of physics. Simply by watching videos and “seeing agents take action in an environment”.