Hello Andrea, could you implement the new network size in Brainlearn (it doesn't have to be in Shashchess)? It will be a long time until the next official SF release. I implemented Dimension 3072 for myself and did some testing. I also carried out tests with MCTS, first in analysis mode and then also on Lichess while playing. The MCTS search is better for analysis. The larger network scores better, especially against weaker engines on Lichess.
A test at Lichess with D3072 and MCTS (1 thread). Matchmaking mode, 6 cores used:
Very good Andrea. The Engine plays very strong with small network. Who needs bigger nnue, when small one are just as good..
The engine runs more than 50% faster on my machine than the current Stockfish and other Stockfish based engines.
ShashChess seems to "scale" better than Crystal. In my testsuite was 21 positions unsolved in 1 minute. When I increase time to 3 min Shash solves 9 more, but Crystal only 5. One nice example:
[d]1k6/3p4/1B6/4Pp1p/1p5R/1p4p1/pP3n2/K6n w - - 0 1
Thanks for your reply but I don't see these options in Fritz gui.
I have MCTS not cheked by default with
MCTSThreads 1
MCTS Multi Strategy 20
MCTS Multi MinVisits 5