As gamers who picked up the SNES Classic Edition can attest, Super Mario Kart is still a masterpiece 25 years after its release. Later games in the series have refined the gameplay and graphics, but the original version can still present a challenge for anyone who attempts to master it…even when the challenger is a machine. As you can see in the video below, one man has attempted to train a neural network to play Super Mario Kart. However, the network, known as “MariFlow,” is still somewhat limited in what it can accomplish within the game.
Via Polygon, YouTube user SethBling posted a video about MariFlow, the recurrent neural network he used to play Super Mario Kart. MariFlow has a form of artificial intelligence that allows it to learn by example and act like a memory cell. To prepare MariFlow for Super Mario Kart, Seth programmed it to examine 15 hours of game footage and analyze his choices to emulate his style of play. So far, it’s been able to get gold medals in the mushroom and flower cups, but it has only been able to land a silver in the star cup.
MariFlow can and does make some mistakes along the way, but the network was able to “forget” the errors and it eventually became a fairly competent Mario Kart player. We’re curious to see how it would fare against a human opponent, although the video appears to humorously suggest it could easily defeat SethBling’s father.
SethBling has previously unveiled MarI/O, a neural network that appeared to achieve mastery of an early level in Super Mario World. In this particular game, MarI/O is clearly superior to many human players, as its jumps and movements easily allow it to quickly advance to the end of the level.
What did you think about MariFlow? Unleash the red koopa shell and share your thoughts in the comment section below!
Images: Nintendo, Seth Bling
Press pause and read more Nintendo stories!
- Hear the Super Mario Bros. theme played on calculator.
- Why Nintendo’s Famicom cartridges have little holes.
- Could Nintendo’s next classic be the Game Boy?
