SPCC: Testrun of Slow Chess 2.6 finished

[pohl4711]

AB-testrun of Slow Chess 2.6 finished.

https://www.sp-cc.de

(Perhaps you have to clear your browsercache or reload the website)

[jonkr]

Thanks for testing, nice to see Slow consistently scoring better than Komodo 14. My goal for the next version is to at least sometimes beat Stockfish 10, which was the strongest engine when I restarted chess programming. (Depending on how it goes and how much time I spend, Slow might stay near the top of the rating lists, but I still think overall days are numbered.)

[AndrewGrant]

Jesus. Elo gains are massive. Here I was thinking this +40 LTC 8moves_v3 net for TCEC was a massive revolution. And now days later you match the Ethereal 13.00 NNUE release it appears. I lost so much time in the last half year not working on this stuff.

[jonkr]

Jesus. Elo gains are massive. Here I was thinking this +40 LTC 8moves_v3 net for TCEC was a massive revolution. And now days later you match the Ethereal 13.00 NNUE release it appears. I lost so much time in the last half year not working on this stuff.

I was a bit surprised that the PawnKingMaterial neural net hash idea that I'd been putting off fully implementing and testing for months turned out to be an instant +15 Elo, then another instant +15 Elo adding the dropout layer to the first part of tuning, then with more training and net generating even more gains. (Technically it was an instant -250 elo, then -60 elo, but I won't count the bugged implementation attempts.) I think I had saturated the amount of knowledge my old net could hold and was just passively running my trainer, so a fairly efficient increase in net-size helped a lot, but also I've been getting more consistent nets as well as stronger nets, so something about the structure is better too, I assume it generalizes better.

(Slow does always score about +10 elo on Stefan's list compared to my own tests, I think maybe its good at the HERT opening book and also the gambit book in Ed's test.)

I still think NNUE (or whatever variation / improvements stockfish is doing now) will be clearly better for reaching higher levels of play, latest Stockfish scores like +270 elo against Slow 2.6 in fast bullet games.