v9

Bit-Genie 9

ELO   | 232.62 +- 13.67 (95%)
CONF  | 32.0+0.20s Threads=1 Hash=8MB
GAMES | N: 3000 W: 2142 L: 388 D: 470

A number of search patches and cleanups, but mainly the addition of a neural network for position evaluation. The network has been trained on 50 million fens from self-play Bit-Genie games generated using a custom fen generator branch. The trainer is an original C++ implementation and is available here.

Special thanks to Finn Eggers for guiding me through the process of creating the neural network, and
to Andrew Grant for allowing Bit-Genie's testing on OpenBench, off-the-record contributors to Bit-Genie!

Note:
The self-play games generated for this network were although from Bit-Genie's search, evaluation was from an older network that was trained on Koivisto dateset. Later networks will have been trained on data generated by Bit-Genie's HCE for originality concerns.

v8

Bit-Genie 8

ELO   | 185.28 +- 14.05 (95%)
CONF  | 32.0+0.20s Threads=1 Hash=64MB
GAMES | N: 1816 W: 1052 L: 166 D: 598

Change log

• Source cleanups [#116] [#104] [#99][#88]
• Speed-ups [#78][#77][#129][#93]
• Add tuner to repo [#84]
• Improve TT replacement [#95][#119]
• LMR tune [#82]
• Low score capture pruning [#83]
• Reverse futility pruning [#89][#91][#100]
• Check-extension [#97]
• Aspiration-window tune [#98][#125][#126]
• Capture history-heuristic [#102][#105]
• Other move-ordering improvements [#108][#109][#110]
• Pawn-support bonus [#113]
• Pawn-shield [#114]
• Tune NMP [#121]
• Improve LMP [#122][#123][#124][#128]

v7

Bit-Genie 7

Changes

• Fix PV node detection in search [#55]
• Polyglot book support [#57]
• LMR improvement [#61]
• IID [#64][#65]
• Aspiration windows [#67][#68]
• Null move pruning improvements [#71][#75]
• Enhance late move pruning [#76]

---

40+0.4s regression test

ELO   | 194.25 +- 18.25 (95%)
SPRT  | 40+0.4s Threads=1 Hash=16MB
LLR   | 0.00 (-2.94, 2.94) [0.00, 0.00]
Games | N: 1027 W: 591 L: 70 D: 366

v6

Bit-Genie 6

Changes

• Better evaluation of passed pawns [#23][#25][#46]
• Control evaluation [#29]
• Re-tune of all evaluation parameters using lichess' quiet-labeled data-set [#32]
• Improved TT usage [#35]
• Add late-move pruning [#44]
• Improve move-ordering [#49]
• Add ucinewgame command [#45]
• LMR tweak [#51]
• Fix TT size allocation [#52]
• Fix node counting in search, and now provide a more accurate NPS [#50]

Self-play shows v6 to be about 100 elo stronger than v5

Special thanks to Koivisto authors for allowing Bit-Genie into Koivisto's OpenBench instance for rapid progress

ELO   | 190.85 +- 36.99 (95%)
SPRT  | 60.0+0.6s Threads=1 Hash=64MB
LLR   | 3.00 (-2.94, 2.94) [0.00, 5.00]
Games | N: 292 W: 173 L: 27 D: 92

v5

Bit-Genie 5

Changes

• Improve late-move reductions [#11]
• Re-tune of all evaluation parameters with a bigger data set [#10 ][#13]
• Evaluation of king-safety [#21 ][#20 ] (Thanks to Luca Metehau)
• Improvements in history-heuristic for move-ordering [#12]
• Other miscellaneous changes in evaluation [#14]
• Fix for bug which very rarely caused illegal moves [#18 ]

Elo-Test

Test tournament played at 60+0.6s with 32 mb hash

Rank Name                          Elo     +/-   Games   Score    Draw
   1 Loki-3.0.0                    124      38     292   67.1%   18.5%
   2 Bit-Genie-v5                   75      36     292   60.6%   22.6%
   3 GooglePlex-1.6                 48      35     291   56.9%   23.0%
   4 FoxSEE-7.2                     21      35     293   53.1%   21.5%
   5 Fornax-1-2                    -99      37     291   36.1%   19.9%
   6 Bit-Genie-v4                 -179      40     293   26.3%   17.1%

Finished match

Elo test v5 against v4, 60+0.6s

Score of Bit-Genie-5 vs Bit-Genie-4: 14 - 3 - 5

Elo difference: 190.8 +/- 156.8

v4

Bit-Genie 4

Added

• Improve evaluation

  • Tuning of some parameters
  • Improve mobility evaluation

• Improve late move reductions to allow engine to search deeper (higher depths)

• Fix for score cp display issue which appeared again in v3

• Update make-file to work with OpenBench for testing

• Format all source files to fix in-consistent indents and other small cleanups in source code

• Fix for issue which caused engine to not terminate immediately when testing with cutechess

Elo test against v3

Time control: 10 + 0.1 s

ELO   | 186.66 +- 34.54 (95%)
SPRT  | 10.0+0.1s Threads=1 Hash=8MB
LLR   | 0.08 (-2.94, 2.94) [0.00, 0.10]
Games | N: 440 W: 291 L: 75 D: 74

v3

Bit-Genie 3

Added

• Tune material values
• 50-move rule draw detection
• Null-move pruning
• Late move reductions
• Reduce transposition entry size, making transposition tables more efficient

Elo-test

Time control: 4+0.2s, Opening book upto 4 moves

Score of Bit-Genie 3 vs Bit-Genie 2: 180 - 77 - 51  [0.667] 308
...      Bit-Genie 3 playing White: 97 - 33 - 25  [0.706] 155
...      Bit-Genie 3 playing Black: 83 - 44 - 26  [0.627] 153
...      White vs Black: 141 - 116 - 51  [0.541] 308
Elo difference: 120.8 +/- 37.3, LOS: 100.0 %, DrawRatio: 16.6 %

v2

Happy to release version 2 of Bit-Genie.

Changes

• Improved use of transposition tables
• Bug fix in move legality checks
• Add support for 'stop' and 'quit' commands during search

Elo test against v1.1

• Time Control: 2+0.1

Score of Bit-Genie v2 vs Bit-Genie v1.1: 255 - 154 - 95  [0.600] 504
...      Bit-Genie v2 playing White: 129 - 73 - 50  [0.611] 252
...      Bit-Genie v2 playing Black: 126 - 81 - 45  [0.589] 252
...      White vs Black: 210 - 199 - 95  [0.511] 504
Elo difference: 70.6 +/- 27.8, LOS: 100.0 %, DrawRatio: 18.8 %

• Time Control: 4+0.4

Score of Bit-Genie v2 vs Bit-Genie v1.1: 239 - 160 - 105  [0.578] 504
...      Bit-Genie v2 playing White: 117 - 81 - 54  [0.571] 252
...      Bit-Genie v2 playing Black: 122 - 79 - 51  [0.585] 252
...      White vs Black: 196 - 203 - 105  [0.493] 504
Elo difference: 54.9 +/- 27.3, LOS: 100.0 %, DrawRatio: 20.8 %

v1.1

Version 1.1 fixes some bugs present in v1 related to time-management and move-legality which caused random forfeits and (sometimes) crashes

Bit-Genie 1

The first release of Bit-Genie. Estimated elo: 1800-1850