TalkChess.com

jonkr wrote: ↑Sat Nov 07, 2020 9:14 pm I probably should have changed the name of slow64.exe to slow64-avx.exe : it now uses avx2 instructions for the endgame neural nets.
(My only 2 computers support avx2 so haven't tested on older ones, but perhaps there's a better way to handle failure.)

The slow64-noAvx.exe should work on everything the version 2.3 worked on, so that's the one to use if the Slow 2.3 worked, it will be about 20% slower in late endgame but otherwise the same in early/midgame.
slow64-noPop.exe has no hardware pop count instructions and no avx instructions, so is for very old cpus, and will have slower midgame play too due to no popcount.

• I estimate it's about 40 elo stronger in standard chess than 2.4
• Can beat Stockfish 8 in 2-moves book matches by a few elo on my computer.
• The play based on win percentages does feel more human-like to me.

• requires AVX2 for strongest play
• struggles in Chess960 variant (can't recommend it over 2.4)
• can take longer to close out games. Probably from valuing sure wins so high. (Some day I may see if I can train it try to close out as aggressively as possible.)

jonkr wrote: ↑Mon Jan 11, 2021 5:05 am SlowChess Blitz 2.5 is now available at
https://www.3dkingdoms.com/chess/slow.htm

On the minus side

• requires AVX2 for strongest play
• struggles in Chess960 variant (can't recommend it over 2.4)

Modern Times wrote: ↑Mon Jan 11, 2021 5:53 am My usual hardware is not AVX2, and that combined with the comment above means I should probably run 2.4 ?

jonkr wrote: ↑Mon Jan 11, 2021 5:05 am SlowChess Blitz 2.5 is now available at
https://www.3dkingdoms.com/chess/slow.htm

You should use slow64-avx if your computer supports AVX2.
The alternative is slow64-noPop. It is plain SSE2. Hardware popcount doesn't really matter since not used in neural net eval, so I didn't compile any more versions.

Slow 2.5 now uses neural nets for every stage of the game. I decided it was time to drop Classic from the name. The endgame nets are the same as 2.4, except with more training and the addition of a Q v R(s) network. The main addition is a network to replace the general evaluations. It's the same structure as the endgame nets, but with a more neurons on first later ( 320x32x32x1 ) Inputs are the same too, just piece/sq, with horizontal symmetry based on white king.

Playing style is similar to 2.4 but more different than any of the hard-coded evals were to each other. Replacing a hard-coded eval with a neural net is a big change, even one based partially on the same training positions and partially on search with same eval. Older hce versions with just certain terms added and factors tweaked showed an extra high draw % playing against previous version, and this one doesn't show that. I think eval definitely main factor in how an engine plays.

On the plus side

• I estimate it's about 40 elo stronger in standard chess than 2.4
• Can beat Stockfish 8 in 2-moves book matches by a few elo on my computer.
• The play based on win percentages does feel more human-like to me.

On the minus side

• requires AVX2 for strongest play
• struggles in Chess960 variant (can't recommend it over 2.4)
• can take longer to close out games. Probably from valuing sure wins so high. (Some day I may see if I can train it try to close out as aggressively as possible.)

I think it's of some interest to have a neural net program that is not standard NNUE and not based on it in any way. (Although it's clearly much weaker than any stockfish NNUEs, and I'm not sure how much is based on my lack of computer resources for training, and how much is from not being the best possible architecture or training process.)

At this point the idea of working more to make it stronger sounds too tedious so I'm releasing it, but if the past is a guide I expect eventually I will start experimenting again then make enough progress to make another version.