nnue-pytorch icon indicating copy to clipboard operation
nnue-pytorch copied to clipboard

Published 20 hours ago • verified publisher icon official-stockfish

[WIP] Attempt to optimize the position skipping to fit a distribution.

Open Sopel97 opened this issue 3 years ago • 1 comments

The old algorithm work online and therefore might discard excessive amounts of positions. The new algorithm filters the buffers that are filled in each worker thread (currently each buffer is around 1M elements). This offline algorithm allows more exact distribution match and less wasted positions. It is however tied to the implementation of the buffering, which is not done for .bin files (so it currently is not implemented for .bin files), and the buffer size will now have (a small) effect. Similarly FEN stream filtering no longer uses this type of skipping.

Sopel97 avatar Jun 21 '22 11:06 Sopel97

Appears to be bugged currently in some way. Testing shows networks degrade during training which is wrong.

Sopel97 avatar Jul 02 '22 09:07 Sopel97