Programmer lets neural network imitate own Mario Kart playstyle

Programmer and YouTube personality SethBling has trained a neural network to imitate his own style of play in Super Mario Kart. For this he used recordings of races driven by himself.

In the accompanying video, he explains that he used 15 hours of video footage to train the neural network. Using the material, the network had to predict which button SethBling would press in certain situations. The programmer used a so-called recurrent neural network, or rnn.

If the rnn makes a wrong prediction, the ‘weight’ of a certain input is increased, so that the network makes a different trade-off in the future and comes closer to the gameplay of SethBling, the programmer explains. He found that if he only used his recordings, the network would get into situations it couldn’t get out of. That is why he also held play sessions in which he let the neural network play and intervene only in certain cases.

SethBling used Google’s open source library TensorFlow to build its network and has made instructions available online. Super Mario Kart recently celebrated its 25th anniversary.