When something is as popular, revered, and utterly inescapable as Super Mario Bros., there's a tendency for it to become a bit… boring. Thanks to a new AI, though, Mario will be trouncing through new and trippier environs, like an underwater castle.
The AI is an updated version of a system designed by computer scientists Mark Riedl and Matthew Guzdial last year, which taught itself to design more traditional Super Mario Bros. levels after being trained on YouTube videos of people playing through the game. The new AI does the same thing, but with an added twist: it can recognize different "styles" of level design in abstract, and blend their elements together based on the structural similarities between the elements in the models of the different styles.
Riedl and Guzdial's paper describing the system is viewable on the Arxiv preprint server, and has not yet been peer reviewed.
But mashing up Mario levels isn't just fun and games. Super Mario Bros. is actually a preferred platform for some artificial intelligence researchers to test their theories, because of the balance the game strikes between being simple enough to work with, but complex enough to be a challenge for computers. Riedl's idea, he wrote me in an email, is to test the bounds of machine creativity.
"Conceptual blending is a general theory of creativity and not limited to games," Riedl wrote. "We see blending in a lot of human creativity. Star Trek is a Western but set in space. We can imagine what a flying car might look like because we have seen cars and we have seen airplanes.
The result of their AI's work is something 10-year-old me would have died to play, and actually looks a bit like the mashed-up "fantasy world" stages in 1985's Super Mario Bros.: The Lost Levels. And although they haven't generated that underwater castle level yet, Riedl said, he and Guzdial plan on it. Chalk up one more win for the robots.