Patricia 5.1

[Whiskers]

https://github.com/Adam-Kulju/Patricia/releases/tag/5.1

I think I've come close to reaching the limit for Patricia's current aggressiveness methodology. She has the search of a top 20 engine and it gets harder and harder to make gains there, so eval is really the only thing left to work on.

Future plans:
Patricia's current net arch is very small and very simple. I recently secured a career in compsci (sidenote: I hadn't even started college when I began engine dev. Time flies!) so now I can focus on getting/renting GPUs and figuring out multilayer inference + datagen for good. I expect I could probably get 50-100 Elo out of that.
The retraining on filtered data scheme I have is probably near optimal at this point - significant improvement requires an entirely different approach. My current idea is to train a net to try and predict the "aggressiveness" of a position, either through manual classification or by running EAS calculations on individual games. That'd let me get an aggression score separate from the objective eval score and then I could figure out how to best combine them.

Note: I am moving across the country next week so any needed bug fixes (there are usually bug fixes after I release an engine) will likely have to wait until I am somewhat settled in.

Thank you and have a blessed day.

[peter]

Code: Select all

# Name                          Points     %     Elo    +/-   W-L-D         1     2     
----------------------------------------------------------------------------------------
1 Chess System Tal 2.07 E1019   310.5/500  62.1  +85.8  20.5  +175-54=271         62.1% 
2 Patricia 5.1                  189.5/500  37.9     --    --  +54-175=271   37.9%       

500/500 total games, 54% draws, time control:180+1, book:UHO_2024_6mvs_+085_+094.pgn

Code: Select all

                         early           bad  avg.win 
Rank  EAS-Score  sacs    sacs   shorts  draws  moves  Engine/player 
---------------------------------------------------------------------------
   1    867464  57.41%  48.39%  68.52%  08.49%   56   Patricia 5.1  
   2     77228  08.00%  14.29%  15.43%  50.55%   78   Chess System Tal 2.07 E1019  

[bastiball]

Just added https://computerchess.org.uk/404FRC/cgi ... tricia_5_1 on the frc list