Hier von Experten im Discord Stockfish Channel:

let me try to give just a few pointers, my non-technical, differences between Leela net (NN) and Stockfish net (NNUE); NNUE is rather small, and quick to evaluate, optimised for CPU, integer operations, etc., NN is bigger net best evaluated, as I understand, on a good GPU, but can be also on a CPU, etc., both, NN and NNUE, as I understand, have as the input chess position, presented to the net in some way; the output of NNUE is eval of the position only, the output of NN is some sort of eval of the position, maybe percentages of wins for a side, and draws, etc., but, as I understand, it can also provide a list of good moves in the position, called policy I think; NN is trained with zero knowledge, i.e., some sort of MCTS search playing the games out until the end, NNUE are trained with alpha-beta search knowledge, i.e., evals are assessed by a relatively shallow search using eval function, non NNUE or NNUE

may I ask, if anybody would know, is Leela using MTSC during play, playing games until the end, or is Leela's search similar to alpha-beta, maybe it is still called MCTS, where the positions are evaluated using NN? thanks

https://cdn.discordapp.com/attachments/718853716266188890/862320896357695498/unknown.png

Lothar