Very roughly, it works as follows: 1/ Initialise a random neural network. 2/ For any given position, use the neural network to estimate the probability of win

Very roughly, it works as follows: 1/ Initialise a random neural network. 2/ For any given position, use the neural network to estimate the probability of win (or draw). 3/ Play multiple branches out to the end, to get an estimate of the 'real' probability of win (for the current skill level) 4/ Modify the neural network so its estimate of the probability of win more closely matches the actual probability of win 5/ Goto 2 :)