Computer Goes From Super Mario Noob to Master in 24 Hours
You, too, once figured out how to optimize Mario.
Software developer and Mario speedrunner Seth Bling has created a computer algorithm that’s conquered Super Mario World. What makes MarI/O tick is machine learning — a method that cribs from biology, relying on so-called neural networks and evolution through iteration. Mathematical equations act as artificial neurons, and these building blocks cobbled together form a network. The bot starts off just the way you did when you handled your first controller — sort of dumbly standing in place, until you make the connection that the right button means go right.
Unlike you, the program can numerically assign fitness values as it avoids pits or fireballs (higher numbers for networks that go farther). And after a few dozen generations, a champ is born out of the fittest networks.
By putting the process to video, Bling has, in a sense, given you a glimpse into what your own brain was processing the first time you suited up as an Italian plumber to stomp turtles and hoard floating coins. Want to create your own side-scrolling computer champ? Bling has posted the source code here.