I think that's the epsilon parameter. That's the chance that, instead of playing the move Mario thinks has the highest value, he plays a random one instead. Tha

I think that's the epsilon parameter. That's the chance that, instead of playing the move Mario thinks has the highest value, he plays a random one instead. That in theory should help prevent him from falling into a local minimum. This probability starts at 100% at the beginning of training and over the course of 1 million steps, ramps down to 1%