Integral v6

This update should let Integral breach into the top 15, maybe even top 10 if it scales well enough. It's genuinely crazy seeing where this project is now.

Notable things in this update are:

• Support for Multi-PV
• Many search improvements and an SPSA tune
• Bug fixes for wrong scores returned when using Syzygy tablebases

And of course, special thanks to everyone in my OpenBench instance and the kind, helpful people in the Stockfish discord.

Progression Tests

STC

Elo   | 48.98 +- 6.48 (95%)
Conf  | 8.0+0.08s Threads=1 Hash=16MB
Games | N: 3006 W: 946 L: 525 D: 1535
Penta | [5, 197, 700, 574, 27]
https://chess.aronpetkovski.com/test/6574/

LTC

Elo   | 52.95 +- 6.17 (95%)
Conf  | 40.0+0.40s Threads=1 Hash=32MB
Games | N: 3002 W: 963 L: 509 D: 1530
Penta | [1, 185, 679, 631, 5]
https://chess.aronpetkovski.com/test/6601/

Integral v5

This update should let Integral breach into the top 20!

Notable things in this update are:

• Added NNUE for evaluation (trained on positions from self-play games)
• Many search improvements and an SPSA tune
• New time management
• Pretty print output

Progression Tests

STC

Elo   | 315.08 +- 14.93 (95%)
Conf  | 8.0+0.08s Threads=1 Hash=16MB
Games | N: 2008 W: 1523 L: 78 D: 407
Penta | [0, 1, 85, 390, 528]
https://chess.aronpetkovski.com/test/5552/

LTC

Elo   | 310.76 +- 19.37 (95%)
Conf  | 40.0+0.40s Threads=1 Hash=32MB
Games | N: 1002 W: 734 L: 19 D: 249
Penta | [0, 2, 23, 235, 241]
https://chess.aronpetkovski.com/test/5555/

Special thanks to everyone in my OpenBench instance and the kind, helpful people in the Stockfish discord.

Integral v4

This is the final "HCE" version of Integral. Notable things in this update are:

• Multiple search improvements.
• Self-play data generation system.
• New evaluation dataset of 40 million positions (combined from lichess-big3-resolved, Ethereal, and self-generated data).
• Pawn-relative piece square tables.

Progression Test

+158 elo at STC over Integral v3 on UHO_Lichess_4852_v1.epd

Elo   | 158.51 +- 16.36 (95%)
Conf  | 8.0+0.08s Threads=1 Hash=16MB
Games | N: 1000 W: 568 L: 141 D: 291
Penta | [2, 24, 143, 207, 124]
https://chess.aronpetkovski.com/test/3783/

Special thanks to everyone in my OpenBench instance and the kind, helpful people in the Stockfish discord.

Integral v3

A lot of improvements have been made to the search as well as the evaluation. Syzygy tablebase support has been added, and Integral now has multithreaded search using the Lazy SMP technique.

Progression Test

+381 elo at STC over Integral v2

Elo   | 381.51 +- 27.65 (95%)
Conf  | 8.0+0.08s Threads=1 Hash=16MB
Games | N: 1014 W: 825 L: 14 D: 175
Penta | [0, 2, 30, 137, 338]
https://chess.aronpetkovski.com/test/2268/

Special thanks to everyone in my OpenBench instance and the kind, helpful people in the Stockfish discord.

Integral v2

Focused on improving the evaluation by adding several evaluation terms and tuning. Time management has been reworked as well.

Estimated Elo Rating

+207 elo at STC over Integral v1

Score of ./integral.exe vs ./integral_x86_64_bmi2.exe: 157 - 26 - 62  [0.767] 245
...      ./integral.exe playing White: 84 - 10 - 29  [0.801] 123
...      ./integral.exe playing Black: 73 - 16 - 33  [0.734] 122
...      White vs Black: 100 - 83 - 62  [0.535] 245
Elo difference: 207.3 +/- 41.7, LOS: 100.0 %, DrawRatio: 25.3 %
SPRT: llr 2.95 (100.1%), lbound -2.94, ubound 2.94 - H1 was accepted

Credits

• Everyone in the Stockfish Discord who were extremely helpful and kind
• Andrew Grant's paper on Texel Tuning and Weiss' implementation of it
• Alexandria for the time management scaling factors

Integral v1

The first public release of Integral 🎉

Estimated Elo Rating

Using Stash v22 (2770 CCRL Blitz) as a baseline for comparison, Integral is estimated to be around 2830 CCRL Blitz

Score of ./integral.exe vs ./stash-22.0-windows-x86_64-modern.exe: 317 - 166 - 179  [0.614] 662
...      ./integral.exe playing White: 168 - 87 - 75  [0.623] 330
...      ./integral.exe playing Black: 149 - 79 - 104  [0.605] 332
...      White vs Black: 247 - 236 - 179  [0.508] 662
Elo difference: 80.7 +/- 23.0, LOS: 100.0 %, DrawRatio: 27.0 %
SPRT: llr 2.95 (100.2%), lbound -2.94, ubound 2.94 - H1 was accepted

Also thank you to the people in the Stockfish Discord who were extremely helpful and kind!