Leela Chess Zero how large is the probability vector in the output layer

Question

Recently, Alpha Zero and Leela Chess Zero are based on the idea that a given chess position s can be given both an evaluation number, as well as a probability vector for the next move. This number and vector are the output neurons of a CNN. My question is: given that different chess positions s and s' have a different amount of legal chess moves, how is this probability vector structured if it doesn't have a fixed size?

Answer 1

The next move's probability vector (called the policy) is dependent on the entire number of possible moves from each chess situation, which is a fixed quantity. AlphaZero had 4672 potential moves (as mentioned on page 13 of the paper), while Leela Chess Zero has roughly 1800. (you can see them all listed in the bitboard.cc code).