When Pac-Man appeared on May 22, 1980, it set a record for development time, requiring a whopping 17 months for development, coding, and completion. Now, 40 years later, it took NVIDIA just four days to train its new gaming AI, to fully recreate Pac-Man, based only on observing the game of another AI. This was reported in the company’s blog.
Called GameGAN, it is a generative adversarial network similar to those used to create (and detect) photorealistic images of people who in reality do not exist. In general, GANs work by pairing two neural networks, a generator, and a discriminator. The generator learns from a large sample of the data set, and then receives instructions for creating an image based on what it saw. The discriminator then compares the generated image with the sample dataset to determine how similar they are to each other. Switching between these networks, artificial intelligence will gradually create more and more realistic images.
In the case of GameGAN, the generative network was trained to use 50,000 game sessions, and then she was ordered to recreate it as a whole – from static walls and granules to ghosts, Pac-Man himself, and the rules governing their interaction. The whole process took place on the four GP100s. GameGAN, however, was not provided with any basic code or access to the game engine. Instead, the AI watched the game of another AI and recreated the engine itself and all the components of the game.
“In recent years, many AIs have been created that can play games. But this is the first GAN created that can actually reproduce the game itself”.
Roar Lebaredian, NVIDIA Vice President of Modeling Technologies
This is the same creation process as the methods of procedural generation that have existed since the late 70s, but a much more efficient method. This method could also improve the development time of real autonomous machines. Since robots operating in warehouses and assembly lines can pose a threat to the safety of their fellow humans, these machines are usually trained in the first place, so if they make a mistake, then no real harm is done. The problem is that developing these digital learning scenarios is a time-consuming and time-consuming task. Now, you can simply train a deep learning model that can predict the consequences of your actions, and use it instead.
In the end, artificial intelligence may appear that can learn to imitate the rules of driving, the laws of physics, just by watching a video and watching the actions of agents in the environment. GameGAN is the first step to this.
NVIDIA’s GameGAN Pac-Man is a fully functional game that both people and AI can play when the company releases it online later this summer.