This is a bugfix release enabling Peacekeeper to actually perform simple mates in a sane manner. There is no significant strength difference between this and v3.00, so testers who use adjudication can safely skip this.

Actually should be this commit, sorry.

This release features the usual improved NNUE and search. The NNUE network is now trained on standard and FRC data generated by an external data source (Lc0). There are also some search patches. It is very sad to say, but Peacekeeper development is paused indefinitely. I partially fried my laptop running SPRT tests, and now have no more working hardware to test patches. However, if any serious bugs are reported, I will make a bugfix release.

In standard chess, this release is 77 elo stronger in selfplay, and I have not yet run a gauntlet against other opponents of similar strength. The version is expected to scale significantly better, and a test at 60+0.6 it gains 105 elo.

Elo   | 77.27 +- 17.00 (95%)
Conf  | 5.0+0.05s Threads=1 Hash=8MB
Games | N: 1010 W: 417 L: 196 D: 397
Penta | [13, 61, 179, 196, 56]

In DFRC, this release is roughly the same strength in selfplay.

Elo   | -10.06 +- 15.85 (95%)
Conf  | 5.0+0.05s Threads=1 Hash=8MB
Games | N: 1002 W: 257 L: 286 D: 459
Penta | [35, 119, 209, 116, 22]

Based on the strength gains, v3.00 should have a blitz elo of about 3630, LTC elo of about 3500, and FRC elo of about 3780 (on CCRL scales).

A note on executables: Different executables for windows have been compiled for different instruction sets. AVX512 is the fastest, followed by AVX2, and POPCNT. Pick the fastest one that doesn't crash on your computer.

EDIT: Binaries added (I suck at releasing)

This release features the usual improved NNUE and search. The NNUE network is now trained on an updated and enlarged 1.88B dataset of DFRC selfplay positions and now uses six king buckets. There are also many search patches.

In standard chess, this release is 40 elo stronger in selfplay, and 61 elo stronger in a gauntlet against other opponents of similar strength.

Elo   | 39.78 +- 15.91 (95%)
Conf  | 5.0+0.05s Threads=1 Hash=8MB
Games | N: 1000 W: 327 L: 213 D: 460
Penta | [11, 75, 225, 167, 22]

In DFRC, this release is 61 elo stronger in selfplay.

Elo   | 60.59 +- 16.01 (95%)
Conf  | 5.0+0.05s Threads=1 Hash=8MB
Games | N: 1002 W: 359 L: 186 D: 457
Penta | [13, 73, 195, 168, 52]

Based on the strength gains, v2.40 should have a blitz elo of about 3560, LTC elo of about 3460, and FRC elo of about 3780 (on CCRL scales).

A note on executables: Different executables for windows have been compiled for different instruction sets. AVX512 is the fastest, followed by AVX2, and POPCNT. Pick the fastest one that doesn't crash on your computer.

EDIT: Binaries re-uploaded to correct naming.

This release features the usual improved NNUE and search. The NNUE network is now trained on an updated and enlarged 1.3B dataset of DFRC selfplay positions. The activation function has been changed to SCReLU and the net now uses four king buckets. Some new search features have gained also significant elo, such as razoring and history pruning.

In standard chess, this release is 96 elo stronger in selfplay, and 90 elo stronger in a gauntlet against other opponents of similar strength.

Elo   | 96.44 +- 16.34 (95%)
Conf  | 5.0+0.05s Threads=1 Hash=8MB
Games | N: 1016 W: 424 L: 149 D: 443
Penta | [4, 52, 176, 217, 59]

In DFRC, this release is 127 elo stronger in selfplay.

Elo   | 127.44 +- 16.75 (95%)
Conf  | 5.0+0.05s Threads=1 Hash=8MB
Games | N: 1008 W: 468 L: 114 D: 426
Penta | [3, 42, 158, 200, 101]

Based on the strength gains, v2.30 should have a blitz elo of about 3530, LTC elo of about 3430, and FRC elo of about 3740 (on CCRL scales).

A note on executables: Different executables for windows have been compiled for different instruction sets. AVX512 is the fastest, followed by AVX2, and POPCNT. Pick the fastest one that doesn't crash on your computer.

Merry Christmas!! 🎄🎁🎅

This (small) release is mostly search patches, such as cutnodes, fail-firm RFP, less futility, and other tweaks, with a new net.

In standard chess, this release is 76 elo stronger in selfplay, and 65 elo stronger in a gauntlet against other opponents of similar strength.

Elo   | 76.45 +- 15.76 (95%)
Conf  | 5.0+0.05s Threads=1 Hash=8MB
Games | N: 1002 W: 372 L: 155 D: 475
Penta | [2, 56, 202, 205, 36]

In DFRC, this release is 84 elo stronger in selfplay.

Elo   | 83.76 +- 15.78 (95%)
Conf  | 5.0+0.05s Threads=1 Hash=8MB
Games | N: 1002 W: 382 L: 145 D: 475
Penta | [9, 60, 179, 191, 62]

Based on the strength gains, v2.20 should have a blitz elo of about 3400, LTC elo of about 3280, and FRC elo of about 3590 (on CCRL scales).

A note on executables: Different executables for windows have been compiled for different instruction sets. AVX512 is the fastest, followed by AVX2, and POPCNT. Pick the fastest one that doesn't crash on your computer.

This release features an improved NNUE, search improvements, and fixing previous mistakes. The network architecture has been enlarged to 768->512x2->1, and trained on 850M DFRC selfplay positions generated from previous nets. Delta pruning which broke quiescence search has been removed with pretty large gains.

LazySMP has also been added.

In standard chess, this release is 141 elo stronger in selfplay, and 138 elo stronger in a gauntlet against other opponents of similar strength.

Elo   | 141.03 +- 18.32 (95%)
Conf  | 5.0+0.05s Threads=1 Hash=8MB
Games | N: 1000 W: 519 L: 134 D: 347

In DFRC, this release is 175 elo stronger in selfplay.

Elo   | 174.55 +- 19.60 (95%)
Conf  | 5.0+0.05s Threads=1 Hash=8MB
Games | N: 1000 W: 582 L: 118 D: 300

Based on the strength gains, v2.10 should have a blitz elo of about 3400, LTC elo of about 3290, and FRC elo of about 3600 (on CCRL scales).

A note on executables: Different executables for windows have been compiled for different instruction sets. AVX512 is the fastest, followed by AVX2, and POPCNT. Pick the fastest one that doesn't crash on your computer.

This release consists mainly of the addition of NNUE. Peacekeeper now uses a 768->256x2->1 neural net with CReLU as the activation function, trained on 197M selfplay positions generated from previous nets and the original HCE. Singular extensions have also been added, with a variety of add-ons.

In standard chess, this release is 220 elo stronger in selfplay, and 177 elo stronger in a gauntlet against other opponents of similar strength.

ELO   | 219.87 +- 23.21 (95%)
CONF  | 5.0+0.05s Threads=1 Hash=8MB
GAMES | N: 1000 W: 691 L: 131 D: 178

In DFRC, this release is 512 elo stronger in selfplay.

ELO   | 511.50 +- 52.03 (95%)
CONF  | 5.0+0.05s Threads=1 Hash=8MB
GAMES | N: 1000 W: 936 L: 36 D: 28

Based on the strength gains, v2.00 should have a blitz elo of about 3280, LTC elo of about 3180, and FRC elo of about 3390 (on CCRL scales).

A note on executables: Different executables for windows have been compiled for different instruction sets. AVX512 is the fastest, followed by AVX2, and POPCNT. Pick the fastest one that doesn't crash on your computer.

Various FRC bugs resulting in crashes and illegal moves were fixed in this version.

There should be an insignificant strength difference from v1.70 in standard chess.

This release includes many small tweaks and bugfixes to perfect HCE (hand-crafted evaluation). Some of the large changes include:

• Updating bestmove on aspiration fail-high
• Adjusting time control
• Making negamax fail-soft (was a bug)
• Adding LMP

FRC (Fischer Random Chess) and DFRC have also been implemented.
EDIT: Using v1.70 for FRC results in crashes and illegal moves, use v1.71 instead.

This release is 119 elo stronger in selfplay, and 104 elo stronger in a gauntlet against other opponents of similar strength.

ELO   | 119.11 +- 12.67 (95%)
CONF  | 5.0+0.05s Threads=1 Hash=8MB
GAMES | N: 2000 W: 977 L: 317 D: 706

Based on the strength gains, v1.70 should have a blitz elo of about 3110, and LTC elo of 3030.

This release includes many changes resulting in considerable strength gains (Otherwise I wouldn't be releasing now). The three main changes from v1.50 are

• Improved evaluation: Added pawn structure, mobility (normal and forward), and bishop pair
• SEE and SEE pruning
• Better time management

This release is 249 elo stronger in selfplay, and 188 elo stronger in a gauntlet against other opponents of similar strength.

ELO   | 249.08 +- 17.31 (95%)
CONF  | 5.0+0.05s Threads=1 Hash=32MB
GAMES | N: 2000 W: 1464 L: 234 D: 302

Based on the strength gains, v1.60 should have a blitz elo of about 2980, and LTC elo at 2920.