Watch This Neural Network Learn How To Play 'Mario Kart'

From Uproxx - November 7, 2017

Artificial intelligence just keeps getting better at beating us at games, and what game more lends itself to a computer being a cheating, blue-shelling bastard than Mario Kart? Thats the idea behind MariFlow, a Mario Kart-playing neural network programmed by Seth Bling based upon a neural program called TensorFlow.

The AI has already earned gold medals in the 50cc, Mushroom, and Flower Cups, but has only been able to silver in Star Cup. The video above shows four screens: three with MariFlow running and one with Seths dad playing. It was easy to guess which player was Seths dad because he was the only one driving like somebodys dad whos never played Mario Kart before.

The neural network is modeled after Seths play style and works by predicting what button presses Seth would make and mimicking his style of play at a rate of 15 predictions per second. It isnt a normal feed forward neural network; its a recurrent neural network, meaning it retains some of the information it had in previous steps, allowing each neuron to act as a memory cell while simultaneously performing computations.

Seth trained the recurrent neural network through backpropagation using 15 hours of his own gameplay footage, in a similar human-mimicking method used by the nunchuck robot. (Okay, maybe the nunchuck robot wasnt trained via backpropagation, but we still wanted to mention nunchuck robot, the robot that knows nunchaku.)


Continue reading at Uproxx »