We have previously heard about AI systems being adept at learning to play complex games that humans love, such as Go, Space Invaders, or even StarCraft. But what if an AI was tasked at creating or re-creating a game? Researchers at the Georgia Institute of Technology are doing something similar; they're getting AI to learn how video games work. In tests, an AI has been able to re-create the game engine for titles such as Super Mario Bros., just by watching it being played for a few minutes.
Georgia Institute of Technology researchers Matthew Guzdial, Boyang Li, Mark O. Riedl recently published a paper called Game Engine Learning from Video (PDF). From the paper abstract we learn that the team isn't training AIs to (re)create computer games for the sake of the computer entertainment industry. Rather the point is that "Intelligent agents need to be able to make predictions about their environment. In this work we present a novel approach to learn a forward simulation model via simple search over pixel input." And for this task they chose Super Mario Bros., "as an initial test of our approach as it represents a physics system that is significantly less complex than reality".
The AIs analyse thousands of frames of Mario gameplay to see what happens when certain on-screen evens occur. It watches and tries to determine cause and effect - what happens when Mario picks up a coin or jumps on a mushroom, for example. The AI has a visual dictionary of character and object sprites used in the game and knows of basic concepts such as position and velocity.
Apparently the re-created video games developed using this method are "glitchy, but passable," says The Verge. Georgina Tech researchers think that isn't very important, as they train the AI on a number of 2D platformers such as the aforementioned Super Mario Bros. and Mega Man. It is hoped that what is learned can be used with AI's trying to learn how to make sense of and operate in the real world, starting with smaller, limited domains to suit this budding technology.
