Stormphrax 6.0.0

Stormphrax 6.0.0 is significantly stronger than 5.0.0 - a longstanding bug in datagen has been fixed, and the entire dataset has been regenerated, resulting in a massively improved (and much larger) network. The regular search tweaks have also happened, along with improvements to time management, and Stormphrax now supports go searchmoves.

Many thanks again to everyone who has contributed CPU time to the OpenBench instance that Stormphrax is tested on - now too many to list, but as always especially @JouanDeag for contributing hosting together with @noobpwnftw for a constant presence of workers :3

Results against Stormphrax 5.0.0

STC:

Elo   | 87.03 +- 7.22 (95%)
Conf  | 8.0+0.08s Threads=1 Hash=32MB
Games | N: 3024 W: 938 L: 196 D: 1890
Penta | [4, 105, 691, 569, 143]

LTC:

Elo   | 77.25 +- 8.43 (95%)
Conf  | 60.0+0.60s Threads=1 Hash=128MB
Games | N: 1504 W: 373 L: 44 D: 1087
Penta | [0, 24, 414, 275, 39]

FRC STC:

Elo   | 115.83 +- 8.10 (95%)
Conf  | 8.0+0.08s Threads=1 Hash=32MB
Games | N: 3004 W: 1209 L: 243 D: 1552
Penta | [2, 108, 533, 640, 219]

DFRC STC:

Elo   | 136.13 +- 8.64 (95%)
Conf  | 8.0+0.08s Threads=1 Hash=32MB
Games | N: 3022 W: 1409 L: 282 D: 1331
Penta | [6, 100, 439, 693, 273]

Elo testing done with OpenBench's 8moves_v3.epd book for standard chess, 3moves_FRC.epd for FRC, and DFRC.epd for DFRC.
Results are obtained from self-play and will likely not scale perfectly to games against other opponents :)

Full Changelog: v5.0.0...v6.0.0

Stormphrax 5.0.0

Stormphrax 5.0.0 is a pretty major release - SP's search has been completely rewritten from the ground up, and its NNUE implementation has improved drastically, resulting in an engine that is actually faster despite having a network twice the size (stonegardens). Improvements have also been made to UCI compliance - SP can now obey multiple search limits at once (e.g., go movetime 1000 nodes 10000000), and properly follows the specification of go infinite. AVX512 VNNI is now supported, with a new corresponding build vnni512 (this should be the fastest binary on CPUs that support it).

In a breaking change, Stormphrax now requires a specific UCI option (EnableWeirdTCs) to be enabled to play under sudden death (0 increment) or cyclic (movestogo) time controls. SP's time management is (and has always been) designed and tested for normal non-cyclic time controls with increment, and it will not perform to its full potential under other formats. Enabling this option means you are aware of this :3

Many thanks again to everyone who has contributed CPU time to the OpenBench instance that Stormphrax is tested on - now too many to list, but as always especially @JouanDeag for contributing hosting and a constant presence of workers :3

Edit: now with Linux binaries!

Results against Stormphrax 4.1.0

STC:

Elo   | 59.75 +- 6.52 (95%)
Conf  | 8.0+0.08s Threads=1 Hash=32MB
Games | N: 3018 W: 731 L: 217 D: 2070
Penta | [10, 124, 785, 522, 68]

LTC:

Elo   | 44.98 +- 8.65 (95%)
Conf  | 60.0+0.60s Threads=1 Hash=128MB
Games | N: 1002 W: 171 L: 42 D: 789
Penta | [0, 21, 339, 132, 9]

DFRC STC:

Elo   | 88.00 +- 11.76 (95%)
Conf  | 8.0+0.08s Threads=1 Hash=32MB
Games | N: 1504 W: 563 L: 190 D: 751
Penta | [8, 86, 271, 299, 88]

Elo testing done with OpenBench's 8moves_v3 book for standard chess, and 3moves_FRC for FRC.
Results are obtained from self-play and will likely not scale perfectly to games against other opponents :)

Full Changelog: v4.1.0...v5.0.0

Stormphrax 4.1.0

• large tuning run of all params
• various search improvements
• as standard, other miscellaneous tweaks and fixes

Minor release this time, as progress has been somewhat slow - search behaviour has changed fairly dramatically, but the network is the same. Some speedups in NNUE inference, though!

Many thanks again to everyone who has contributed CPU time to the OpenBench instance that Stormphrax is tested on: @Adam-Kulju, @Alex2262, @AndyGrant, @archishou, @cosmobobak, @crippa1337, @dede1751, @GediminasMasaitis, @iudwgerte, @jhonnold, @jw1912, @lucametehau, @PGG106, @raklaptudirm, @spamdrew128, @uwuplant and particularly @JouanDeag :3

Results against Stormphrax 4.0.0

STC:

Elo   | 38.23 +- 6.01 (95%)
Conf  | 8.0+0.08s Threads=1 Hash=32MB
Games | N: 3002 W: 525 L: 196 D: 2281
Penta | [2, 139, 935, 378, 47]

LTC:

Elo   | 26.81 +- 7.79 (95%)
Conf  | 60.0+0.60s Threads=1 Hash=128MB
Games | N: 1000 W: 106 L: 29 D: 865
Penta | [1, 23, 378, 94, 4]

FRC STC:

Elo   | 47.78 +- 10.02 (95%)
Conf  | 8.0+0.08s Threads=1 Hash=32MB
Games | N: 1500 W: 350 L: 145 D: 1005
Penta | [8, 72, 405, 237, 28]

Elo testing done with OpenBench's 8moves_v3 book for standard chess, and 3moves_FRC for FRC.
Results are obtained from self-play and will likely not scale perfectly to games against other opponents :)

Full Changelog: v4.0.0...v4.1.0

Stormphrax 4.0.0

• new network (edgelands)
• huge SMP time management improvements
• NMP verification search
• a few other miscellaneous search improvements
• tuned (almost) everything
• the usual various tweaks and fixes

Stormphrax 4.0.0, no longer using basically no time with multiple threads. This release is a significant strength improvement, most noticeable in SMP games. Puzzle performance should also be improved somewhat, though please note that this is by no means a primary development focus.

Many thanks again to everyone who has contributed CPU time to the OpenBench instance that Stormphrax is tested on: @Adam-Kulju, @Alex2262, @AndyGrant, @archishou, @cosmobobak, @crippa1337, @dede1751, @GediminasMasaitis, @iudwgerte, @jhonnold, @jw1912, @lucametehau, @PGG106, @raklaptudirm, @spamdrew128, @uwuplant and particularly @JouanDeag :3

Results against Stormphrax 3.0.0

STC:

ELO   | 89.48 +- 7.33 (95%)
CONF  | 8.0+0.08s Threads=1 Hash=32MB
GAMES | N: 3024 W: 938 L: 176 D: 1910

LTC:

ELO   | 69.71 +- 10.58 (95%)
CONF  | 60.0+0.60s Threads=1 Hash=128MB
GAMES | N: 1000 W: 229 L: 31 D: 740

SMP (3 threads):

ELO   | 105.63 +- 12.09 (95%)
CONF  | 20.0+0.20s Threads=3 Hash=64MB
GAMES | N: 1000 W: 320 L: 25 D: 655

FRC STC:

ELO   | 108.14 +- 14.31 (95%)
CONF  | 8.0+0.08s Threads=1 Hash=32MB
GAMES | N: 1008 W: 379 L: 75 D: 554

Elo testing done with OpenBench's 8moves_v3 book for standard chess, and 3moves_FRC for FRC.
Results are obtained from self-play and will likely not scale perfectly to games against other opponents :)

Full Changelog: v3.0.0...v4.0.0

Stormphrax 3.0.0

Binaries updated nonfunctionally to correctly name the network.

• new, much bigger network (riverrise)
• significant history improvements
• negative extensions
• fixed (hopefully) reporting false mates
• other small search tweaks
• the usual various tweaks and fixes

Stormphrax 3.0.0, finally with a bucketed network. The dataset now also includes standard chess data, which gives a nice strength boost in standard. The score normalisation method has also changed - a +1.00 evaluation now means an estimated 50% chance to win against an equal opponent, irrespective of move number (previously it was only accurate at move 32).

Many thanks again to everyone who has contributed CPU time to the OpenBench instance that Stormphrax is tested on: @Adam-Kulju, @Alex2262, @AndyGrant, @archishou, @cosmobobak, @crippa1337, @dede1751, @GediminasMasaitis, @iudwgerte, @jhonnold, @jw1912, @lucametehau, @PGG106, @raklaptudirm, @spamdrew128, @uwuplant and particularly @JouanDeag :3

Results against Stormphrax 2.0.0

STC:

ELO   | 56.51 +- 7.54 (95%)
CONF  | 8.0+0.08s Threads=1 Hash=32MB
GAMES | N: 3008 W: 805 L: 320 D: 1883

LTC:

ELO   | 45.42 +- 11.36 (95%)
CONF  | 60.0+0.60s Threads=1 Hash=128MB
GAMES | N: 1000 W: 207 L: 77 D: 716

FRC STC:

ELO   | 22.78 +- 14.57 (95%)
CONF  | 8.0+0.08s Threads=1 Hash=32MB
GAMES | N: 1008 W: 265 L: 199 D: 544

Elo testing done with OpenBench's 8moves_v3 book for standard chess, and 3moves_FRC for FRC.
Results are obtained from self-play and will likely not scale perfectly to games against other opponents :)

Full Changelog: v2.0.0...v3.0.0

Stormphrax 2.0.0

• use slightly more time in non-movestogo TCs
• new network (edgewater)
• implemented singular extensions, along with some related techniques
• implemented late move pruning (yup, this late)
• LMR improvements
• AVX-512 support
• allowed for non-power of 2 TT sizes
• a couple of eval scaling techniques
• significant stability improvements, especially in SMP with higher thread counts
• massive SMP speed improvement with higher thread counts
• various fixes and tweaks

Stormphrax 2.0.0, now with (theoretically) 100% less SMP crashes and hangs! This release brings significant strength improvements, particularly in multi-thread formats and FRC.

Many thanks again to everyone who has contributed CPU time to the OpenBench instance that Stormphrax is tested on: @Adam-Kulju, @Alex2262, @AndyGrant, @archishou, @cosmobobak, @crippa1337, @dede1751, @GediminasMasaitis, @jhonnold, @jw1912, @lucametehau, @PGG106, @raklaptudirm, @spamdrew128, @uwuplant and particularly @JouanDeag :3

Results against Stormphrax 1.0.0

STC:

ELO   | 147.38 +- 8.78 (95%)
CONF  | 8.0+0.08s Threads=1 Hash=32MB
GAMES | N: 3024 W: 1374 L: 163 D: 1487

LTC:

ELO   | 127.76 +- 13.36 (95%)
CONF  | 60.0+0.60s Threads=1 Hash=128MB
GAMES | N: 1000 W: 382 L: 30 D: 588

FRC STC:

ELO   | 190.39 +- 17.54 (95%)
CONF  | 8.0+0.08s Threads=1 Hash=32MB
GAMES | N: 1008 W: 558 L: 55 D: 395

Elo testing done with OpenBench's 8moves_v3 book for standard chess, and 3moves_FRC for FRC.
Results are obtained from self-play and will likely not scale perfectly to games against other opponents :)

Full Changelog: v1.0.0...v2.0.0

Stormphrax 1.0.0

• implemented NNUE
• a few minor tweaks and fixes

Stormphrax is here! Polaris' hand-crafted evaluation function has been replaced with a NNUE, resulting in massive gains in both standard chess and particularly FRC. The current network (deepwoods) was trained on DFRC self-play data from the previous development network - in total there were five previous networks, the first being seeded with random weights and biases, so Stormphrax is a "zero" engine in that it learnt to evaluate positions from zero knowledge, knowing only the rules of chess.

There is still lot more improvement possible for both the NNUE implementation and for search, so development will continue - though training networks is very slow for me currently, so likely not as fast as Polaris development was.

Many thanks again to everyone who has contributed CPU time to the OpenBench instance that Stormphrax is tested on: @Adam-Kulju, @Alex2262, @archishou, @cosmobobak, @crippa1337, @dede1751, @GediminasMasaitis, @jhonnold, @lucametehau, @raklaptudirm, @spamdrew128, @uwuplant and particularly @JouanDeag :3

Results against Polaris 1.8.1

STC:

ELO   | 357.65 +- 16.33 (95%)
CONF  | 8.0+0.08s Threads=1 Hash=32MB
GAMES | N: 3000 W: 2447 L: 126 D: 427

LTC:

ELO   | 357.07 +- 26.69 (95%)
CONF  | 60.0+0.60s Threads=1 Hash=128MB
GAMES | N: 1000 W: 796 L: 23 D: 181

FRC STC:

ELO   | 466.04 +- 42.52 (95%)
CONF  | 8.0+0.08s Threads=1 Hash=32MB
GAMES | N: 1000 W: 907 L: 35 D: 58

Elo testing done with OpenBench's 8moves_v3 book for standard chess, and 3moves_FRC for FRC.
Results are obtained from self-play (well, against Polaris) and will likely not scale perfectly to games against other opponents :)

Full Changelog: ce856d8...v1.0.0