Recently, a computer learned to play Super Mario Bros. better than most humans. Now, in an impressive twist, a group of Georgia Tech researchers has taught a computer to design playable Mario levels just by watching YouTube videos of people playing Mario.
Mark Riedl, the principal investigator for the Mario project, explained how, over time, their AI learned the basic rules of the game. The AI looked at the different “styles” of segments in the level, like sky, or bricks, and made predictions about how likely it was for different styles to be next to each other. Based on these predictions, it was able to create realistic levels. To do this, their computer spent a lot of time examining the parts of YouTube videos which players spent the most time in, as it assumed these areas would have the most data about the game’s structure.
Overall, the AI has built very realistic, successful Mario levels. Sometimes, though, the novice designer still screws up, creating levels with too wide jumps or too high obstacles. Or, Reidl says, it sometimes designs levels which are “supposed to be highly engaging to players but are in fact quite boring because they don’t have any obstacles, monsters, or jumps.” Minimalist Mario, you might call it.
Reidl hopes that their AI will continue to improve its design skills in both Super Mario Bros. and other similar games like Sonic the Hedgehog, eventually using what it learns to design entirely original creations from individual elements of games.