How Building ‘Super Mario Bros.’ Levels Is Pushing the Boundaries of AI

AI researcher and YouTuber Tommy Thompson lays out the latest AI research using 'Super Mario Bros.' in a new video.
AI researchers use Super Mario Bros.
Screengrab: YouTube

What makes a great Super Mario Bros. level? Is it the aesthetic mix of pleasant music, bustable question marks, colored bricks, and Goombas? Is it the slowly ramping difficulty curve that leads to a cathartic finish? Is it simply the fact that Mario is present? Once we know what the magic traits are, can we teach a machine to create levels on its own?

For years, artificial intelligence researchers have been using Super Mario Bros. to train machines to generate their own video game levels and even whole video games. Now, AI researcher and YouTuber Tommy Thompson has put together a video delving into the state of the current research on Super Mario Bros. level generation at his channel, AI and Games.

His video, “Building Mario Levels with Machine Learning,” focuses on four different methods of level generation from four different researchers and tries to answer the existential question: what is the defining characteristic of a Super Mario Bros. level?

It’s a hard question to answer, one confounded by the thousands of strange and unique levels generated by humans using Super Mario Maker an official game that allows players to make their own Super Mario levels and machines who learned how to make Super Mario Bros. levels from humans.

Listen to CYBER, Motherboard’s new weekly podcast about hacking and cybersecurity.