Fat Fritz works with a neural network which it has itself created or “learned” on the basis of chess games. That is the evaluation function, which finally evolves from a huge number of weightings which individually have no firm significance.

 

So the network teaches (trains) a large black box, which at the end spits out probabilities for the moves. It also does not give a traditional evaluation, the result being thus

 

e4 54%

d4 55%

c4 51%

 

etc., which could again be calculated as an evaluation.