Forty years ago, Pac-Man first appeared in Japan’s game rooms and went viral around the world. In 1981 alone, Americans threw billions of 25-cent coins into consoles for The Bean Eater, with a total game time of 75,000 hours, equivalent to eight-and-a-half years, the data showed. Now, this classic, which is included in the world’s famous world, and with the help of NVIDIA AI technology, recreate the world!
On the neural network of the NVIDIA DGX system, after up to 50,000 rounds of game training with a total of tens of thousands of frames, nVIDIA Research Institute has created a powerful new AI model, NVIDIA GameGAN, to produce a full version of the Bean Eater game without the need for a basic game engine.
In other words, AI can perfectly reproduce this classic game even if it doesn’t understand the basic rules of the game.
GameGAN is the first neural network model to mimic a computer game engine using a generative anti-network (GAN). The GAN model consists of two competing neural networks, a generator and a discriminator that can learn and create new content that is comparable to the original content.
The trained GameGAN model is able to generate static environmental elements such as uniform maze shapes, beans, reinforced props, ghosts as enemies, and moving elements such as the bean eater itself.
It also learns the simple and complex key rules of the game: bean eaters can’t walk through the maze wall, but need to move around and eat beans; when they eat reinforced props, the ghost turns blue and escapes; the bean eater leaves the maze from one side and is transferred to the other side of the maze; and the bean eater touches the ghost, the screen flashes and the game ends.
When the smart agent tries on a GAME generated by GAN, GameGAN responds to the agent’s behavior, generating a new framework for the game environment in real time.
After training with game scripts at different levels or versions of the game, GameGAN can even generate game levels that have never been there before.
With this feature, game developers can automatically generate new game levels, and AI researchers can more easily develop simulator systems for training autonomous machines.
In fact, regardless of the game, GAN can learn its rules by extracting screen recordings and smart agent’s keystroke tracks from past games, and game developers can use the game script from the original level as training data to design a new level level for the current game.
BandAI NAMCO Research Inc., a research and development company owned by bandAI NAMCO Entertainment, the game publisher, provided data on The Bean Eater used to train GameGAN, and they were shocked to believe that the classic game could be recreated without a game engine.
NVIDIA will release the AI-recreated Bean Eater on AI Playground later this year, which everyone can experience for themselves.