A programmer has actually taught a neural network to play Mario Kart.
After showing a neural network 15 hours of in-game footage and refining its behaviour, programmer SethBling managed to get the network to win the gold trophy, by itself, in Mario Kart‘s 50cc Mushroom Cup.
SethBling also created MarI/O, another neural network experiment designed to play Super Mario World. MariFlow, the name of the Mario Kart neural network, learns from player information, with SethBling stating that he actually aimed to make the robot play like his father.
The neural network features four layers of computation and creates a set of predictions for buttons it thinks SethBlink is likely to input at specific points on the track. Unlike traditional AI though, MariFlow is able to remember information, though the experience needed to be weighted by humans so that the neural network is able to know which commands are more important.
For example, over the course of the 15 hours, SethBling often took control of the kart manually to teach the network what it should actually be doing.
In total, MariFlow has managed to grab medals in Mario Kart‘s Mushroom Cup, Flower Cup and a silver medal in the Star Cup. While the entire experiment is a little silly, it’s fascinating to see the concept of a neural network broken down to such a basic level.
Since the original Mario Kart for the Super Nintendo doesn’t feature the dreaded Blue Shell, the neural network likely won’t ever have to experience the pain of going from first to last in only a few seconds.