AI learns how to build Super Mario levels by watching YouTube

First, we created artificial intelligence that can learn how to play and beat Super Mario World by finding the fastest route. Now, we have AI that can build its own Super Mario Bros. levels. Next thing you know, AI will master Tetris and learn how to build human prisons out of Tetris blocks.

This new piece of gaming AI comes from researchers at the Georgia Institute of Technology.

The most amazing part of their AI system is how it learns the rules and gameplay of the 1985 platform game Super Mario Bros. The AI simply watches gameplay videos on YouTube and Twitch. It observes and interprets the rules of the game, as well as players' behavior (where they spend more time collecting bonus items, for example) and the placement of various parts of the terrain such as clouds and pipes.

"For example, pipes in the Mario games tend to stick out of the ground, so the system learns this and prevents any pipes from being flush with grassy surfaces," according to a blog post on Georgia Tech's College of Computing website. Then the algorithm applies what it's learned from human design choices and behavior to creating playable new sections of game levels.

Ph.D. computer science student Matthew Guzdial and associate professor of Interactive Computing Mark Riedl presented their findings at the International Conference on the Foundations of Digital Games (FDG) that started Monday in Pacific Grove, Calif.

Their AI eventually created 151 unique level sections from just 17 game samples, and its output increased to 334 level sections as the system lessened the constraints.

"Our system creates a model or template, and it's able to produce level sections that have never been seen before, do not appear random and can be traversed by the player," Riedl said. "One could say that the system 'studies' the design of Mario levels until it is able to create new playable areas."

So they've basically taught a program how to be a gamer and a game designer that doesn't need sleep, sunlight or water (assuming that human game designers actually get all three of those things in their daily lives). Riedl and Guzdial noted in their research paper (PDF) presented to the FDG that this technology could be adapted to other games "as gameplay video becomes more accessible and as open machine vision toolkits become more advanced."

However, if AI learns how to obtain a real fire flower and shoot fireballs in the real world, someone needs to pull the plug on it -- and fast.