NVIDIA Recreates Pac-Man on 40th Anniversary with AI

[Guest_Jim_*]

PAC-MAN is 40 years old now, so naturally there is some celebration for this anniversary and NVIDIA decided its tribute to the classic would be to use an example of fairly new technology. By leveraging the power of generative adversarial networks, or GAN, a form of artificial intelligence, NVIDIA has recreated PAC-MAN just by having its new GameGAN system watch as it is played.

Generative adversarial networks are a kind of neural network that can be applied to accelerate the machine learning process. Typically the process for training a network would require providing it with reference information, and potentially that information has been first filtered by humans to make sure it is all accurate. That human component can slow the process down because if a neural network may require thousands of images to finally learn to identify whether a specific object is contained in an arbitrary image, each image would need to first be looked at by a human. By using a GAN this process is sped up because it gives a pair of neural networks the task of defeating each other. One network tries to generate information accurate enough to fool the second, so it is called the generator, and the second network tries to find the flaw to expose the fake, so it is called the discriminator. Though they are adversaries, some information is shared between the networks so both can improve until you eventually have the ability to generate information that cannot be distinguished from actual reference information.

NVIDIA has previously deployed GANs with GauGAN, a technology that can take in rough images from a user and generate detailed scenes that will appear realistic. The potential here would be to ease the amount of effort required by those who need to supply synthetic yet realistic images, such as architects, landscape designers, and game developers. Now the company has put together GameGAN with PAC-MAN as one of its first examples.

What GameGAN has done is taken in about 50,000 PAC-MAN episodes to learn from so it could then imitate the original. It watched both the screenplay of the game as well as keyboard actions for its training. From this information, GameGAN was able to determine the rules of the game as well as its mechanics, such as how consuming a capsule causes the ghosts to change color and run from Pac-Man, instead of chasing him. It also had to learn something that might seem obvious, but as it was not starting from a game engine even realizing how keyboard input controls Pac-Man was important to achieve.

While just imitating the original game via an AI method is important on its own, GameGAN did something else with this process that was important, and that was disentangling the static and dynamic components visible in the image. One reason this is important is that it allows the system to create a more interpretable model that can also be better applied to tasks later on that explicitly relate to dynamic and not static elements. In PAC-MAN the map would be one of the static elements and by separating it from the dynamic elements, like the ghosts, the map could be more easily replaced. For training the researchers behind GameGAN used a modified version of the game to work with random mazes, so that ability to distinguish between static and dynamic undoubtedly helped it.

In addition to working with PAC-MAN the researchers also worked with ViZDoom, an AI research platform actually developed to teach AI to play DOOM using just visual information. Thanks to the disentangling of components, GameGAN can replace elements from both, such as allowing one to play as Mario in either game or changing the environment of ViZDoom without compromising the gameplay.

Though this research shows GameGAN being used to effectively reverse-engineer a classic game, its potential uses include automated generation of game levels as well as easier creation of simulators for training autonomous machines. For those wondering about playing the GameGAN version of PAC-MAN, it will be made available later this year in NVIDIA's AI Playground

 

 

Source: NVIDIA and NVIDIA GameGAN Paper

Back to original news post