Machine-learning system recreates Super Mario game based on just a few clips
Image credit: Dreamstime
Researchers at the Georgia Institute of Technology have demonstrated an artificial intelligence (AI) system capable of recreating the basic software of a video game using less than two minutes of Super Mario Bros footage.
The basic software framework of any video game – the game engine – controls the framework of the game, from rendering graphics and sound to the mechanics of the characters, and their interactions with objects and the setting.
The researchers created a machine-learning system: the system “learns” to detect patterns based on the supervised or unsupervised processing of large datasets. In this case, the researchers trained the system on video game footage of an avatar racing towards a goal. The technique in this case relied on a simple search algorithm which tests possible sets of rules for the “best fit” for the footage.
After consuming less than two minutes of new gameplay frames, the system was able to build its own model of the game engine and make predictions about future events in the game, such as how enemies might react to certain actions.
The researchers found that the game engine predicted video games more accurate to the original game when compared to an alternative test, which used a convolutional neural network (typically used to process photographs or video footage) to predict game mechanics. This gave the team a more general model of the game based on a small sample of footage.
“Our AI creates the predictive model without ever accessing the game’s code and makes significantly more accurate future event predictions than those of convolutional neural networks,” said Matthew Guzdial, a PhD student and lead researcher.
“A single video won’t produce a perfect clone of the game engine, but by training the AI on just a few additional videos you get something that’s pretty close.”
The researchers were able to recreate the game engines of the platform game Super Mario Bros and are currently experimenting with other simple games in which all the action appears onscreen, including Sega’s Sonic the Hedgehog and Nintendo’s Mega Man.
Having created a cloned game engine for Super Mario Bros, the Georgia Institute of Technology team tested the framework by employing a second AI to play the game, avoiding traps and escaping enemies. They found that the AI performed just as well playing with the AI-generated game engine as with the original game engine.
“Intelligent agents need to be able to make predictions about their environment if they are to deliver on the promise of advancing different technology applications,” said Guzdial. “Our model can be used for a variety of tasks in training or education scenarios, and we think it will scale to many types of games as we move forward.”