v3.0.1

This version removes Syzygy TB support for release executables, due to crashes on certain machines. Syzygy tablebases are still available through make syzygy, though the issue causing the crash is yet to be determined.

Check out the v3.0.0 patch notes for the changes from v2.0.0

Full Changelog: v3.0.0...v3.0.1

v3.0.0

The cooking is complete!
Carp 3.0 irons out a lot of the oddities in the previous release, though still maintaining a lot of
its weird quirks (it is, quite possibly, the strongest Fail Hard engine around). The progtests should
speak for themselves:

STC (8+0.08)

ELO   | 438.76 +- 18.13 (95%)
CONF  | 8.0+0.08s Threads=1 Hash=16MB
GAMES | N: 3024 W: 2609 L: 33 D: 382

LTC (60+0.6)

ELO   | 364.09 +- 24.95 (95%)
CONF  | 60.0+0.60s Threads=1 Hash=128MB
GAMES | N: 1000 W: 787 L: 6 D: 207

Changes Overview

• Network improvements: moved to a (768->1024)x2->1 network trained on 3.2B positions
• Countless search improvements and bug fixes.
• Support for Syzygy endgame tablebases.
• Major project structural refactor.
• Massively improved chess backend.

PR List

• Carp0013 by @dede1751 in #9
• Compact moves by @dede1751 in #10
• Use 32b eval internally by @dede1751 in #11
• Nodetm by @dede1751 in #12
• Use more of the time increment by @dede1751 in #13
• TT null move by @dede1751 in #14
• TT static eval by @dede1751 in #15
• Update best_move on Fail High by @dede1751 in #16
• Proper handling of possible mates in QS by @dede1751 in #17
• Remove Delta Pruning in QS by @dede1751 in #18
• Refactor carp::chess crate by @dede1751 in #19
• Better threat generation for legal movegen by @dede1751 in #20
• Fix SEE rejecting equal exchanges. by @dede1751 in #22
• QS TT move by @dede1751 in #23
• Material Scale by @dede1751 in #24
• Singular Extensions by @dede1751 in #25
• Refactor search by @dede1751 in #26
• PGO by @dede1751 in #27
• SEE pruning by @dede1751 in #28
• Fix conthist by @dede1751 in #29
• Fix hist score by @dede1751 in #30
• Implement a MovePicker for staged sorting by @dede1751 in #31
• Fix MovePick scores by @dede1751 in #32
• Thread refactor by @dede1751 in #42
• Keep histories between moves by @dede1751 in #43
• Shift killers back between moves by @dede1751 in #44
• PV Table by @dede1751 in #45
• TT Replacement bug fix + policy change by @dede1751 in #46
• always set global stop flag after main thread stops by @Ciekce in #47
• Fix repetition detection by @dede1751 in #48
• Use a smaller max value for standard histories by @dede1751 in #49
• Refactor codebase with workspaces by @dede1751 in #50
• QS Draws by @dede1751 in #51
• Release CI by @dede1751 in #52
• Build file by @dede1751 in #53
• Use 8 BBs in Board by @dede1751 in #54
• Stop saving temporary static eval entries by @dede1751 in #55
• Syzygy by @dede1751 in #56
• Refactor Movegen by @dede1751 in #57
• QS promos by @dede1751 in #58
• Caphist by @dede1751 in #59
• Sane Replacement by @dede1751 in #61
• Remove killers by @dede1751 in #62
• Fix evals by @dede1751 in #63
• Carp0019 by @dede1751 in #64

Binaries

Binaries are released following the microarchitecture levels. It is recommended to at least use V3. As for MacOS, I was unable to build a V3 binary using Github CI, so it is recommended to build locally, as is explained in the ReadME. If possible, a native build is always recommended.

Special Thanks

Carp development wouldn't be possible without the hardware contributed by the amazing people in the
SWEHosting OpenBench instance: @Adam-Kulju, @Alex2262, @archishou, @Ciekce, @cosmobobak, @crippa1337, @GediminasMasaitis, @jhonnold, @jw1912, @lucametehau, @raklaptudirm, @PGG106, @spamdrew128, @uwuplant and most importantly @JouanDeag.

New Contributors

• @Ciekce made their first contribution in #47

Full Changelog: v.2.0.0...v3.0.0

v2.0

Carp 2.0 introduces NNUE, likely pushing past 3000 CCRL ELO. I'm finally confident enough to start submitting Carp to engine tournaments.
Starting from 2.0 I will also start focusing more on adding tooling/functionality around the engine, transitioning to a private OpenBench instance for testing.

Score of Carp v2.0 vs Carp v1.3: 101 - 8 - 4  [0.912] 113
...      Carp v2.0 playing White: 55 - 1 - 0  [0.982] 56
...      Carp v2.0 playing Black: 46 - 7 - 4  [0.842] 57
...      White vs Black: 62 - 47 - 4  [0.566] 113
Elo difference: 405.1 +/- 118.7, LOS: 100.0 %, DrawRatio: 3.5 %
SPRT: llr 2.96 (100.5%), lbound -2.94, ubound 2.94 - H1 was accepted

v1.3

Version 1.3 adds a lot of search selectivity improvements, pushing Carp to an estimated 2700-2750 elo on 4 cores.

v1.2

Carp v1.2 brings much more aggressive stockfish-like pruning while also improving overall performance. It is without a doubt superior to v1.1 (it just takes 70 games of sprt to get +10 elo at 5% confidence) but more extensive testing needs doing to determine actual playing strength.

v1.1.0

Version 1.1.0 adds fully legal move generation, along with a few search improvements. Excuse the release rollback but a few tentative changes which turned out to be not so good got pushed to release.

v1.0.0

Initial release: the engine has a solid search function and, while the evaluation is still quite rudimentary, it's capable of playing relatively "human" moves. Without any proper testing, I would estimate it to be rated in the low 2000s.