In a video he uploaded yesterday, he shows an algorithm that he wrote that actually learns how to play Super Mario Bros. As he describes in the video, the program knows absolutely nothing about the game at the start. It doesn't even known that pushing the "right button" will make the character go right. And yet, the computer has learned to get through the level as fast as even the best human player could.
The program learns through a process called 'neuro evolution', which mimics the way the human brain actually learns to make decisions. The potential for a program like this reaches far beyond the scope of robot gamers, however. A complex enough version of this program could one day allow computers to make super efficient decisions in fields ranging from manufacturing, to health, and even to military applications.
Nathan Hutchcraft is the creator and co-host of the YouTube web series, NathanAndRose. He also writes for several online sources, as well as having a nearly insatiable love for jalapeno pizza.