I am currently trying to understand the Stockfish NNUE HalfKP 256x2-32-32 neural network, which is listed on chessprogramming.org.
I understand the way how the board is interpreted. What I don't fully understand is how the evaluation actually works.
On chessprogramming.org is a picture of the ANN's structure. The first "layer" is divided into two halves; top one is black, bottom one is white. They represent the player colors. To the right are two halves; one pink-ish and the other one green-ish. The color represents the side to move. Green-ish means to move, Pink-ish means not to move.
Q1: Is my understanding correct?
On top and on the bottom of the structure are things like "2x41024 (...)" and "256 x 41024 (...)" written. As far as I understand it, the things written on top represent the way the neurons are connected to each other or rather how and what information is transmitted from neuron to neuron. The things written beneath represent the neurons and their description.
Q2: Again, is my understanding correct?
Beneath the ANN there is a total of four descriptions. There are three descriptions that say some number and than "calculated parameters".
Q3: What does "calculated parameters" mean? Are these more or less the same as principles used to evaluate a position in hand-crafted function, i. e. Material, Pawn structur and so on?
Q4: What are the numbers written above "calculated parameters"? Are these the total number of parameters, i. e. neurons in one layer?
Q5: Explained shortly, how would you describe the process of evaluating a position using this method? In hand-crafted
Since I am asking you for a full explanation of the structure I will stop with individual questions now. Since you probably already could tell, I have little-to-none prior knowledge concerning ANN's and programming.
The structure I am referring to can be found here.
