Pikafish 2024-03-10

Comparing with Pikafish 2023-12-03:

Book: 3mvs 151-350_1097123

TC: 10+0.1
Total/Win/Draw/Lose: 4036 / 1165 / 1979 / 892
PTNML: 33 / 382 / 942 / 601 / 60
WinRate: 53.38%
ELO: 23.28[16.64, 30.04]
LOS: 100.00
LLR: 3.35[-2.94, 2.94]

TC: 60+0.6
Total/Win/Draw/Lose: 2124 / 597 / 1133 / 394
PTNML: 2 / 162 / 544 / 339 / 15
WinRate: 54.78%
ELO: 33.06[25.41, 41.01]
LOS: 100.00
LLR: 3.02[-2.94, 2.94]

Book: 3mvs 151-350_1097123
TC: 180+1.8
Total/Win/Draw/Lose: 1414 / 432 / 706 / 276
PTNML: 0 / 76 / 402 / 226 / 3
WinRate: 55.52%
ELO: 38.23[29.85, 47.12]
LOS: 100.00
LLR: 2.97[-2.94, 2.94]

Shows very good scaling.

Pikafish 2023-12-03

What's New:

1. Rule refinement and now totally comply with https://pikafish.org/rule.pdf. since Pikafish is only focused on computer rule, all options related to old rule have been removed.
2. Network architecture and weights updated, engine parameter tuned via SMP MTC.
3. Enhanced zugzwang resilience, more tough positions can now be solved.

Pikafish 2023-03-05

This release heavily increases playing strength at STC via parameter tweaks compared to Pikafish 2023-02-18 on https://pikafish.org :

Slightly better at STC
Book: 3mvs 151-350_1097123
TC: 10+0.1
Total/Win/Draw/Lose: 24868 / 6340 / 12462 / 6066
PTNML: 174 / 2944 / 6034 / 2998 / 284
WinRate: 50.55%
ELO: 3.58[1.06, 6.07]
LOS: 99.72
LLR: 3.04[-2.94, 2.94]

Horribly behind at LTC (within anti-plagiarism safe margin)
Book: 3mvs 151-350_1097123
TC: 60+0.6
Total/Win/Draw/Lose: 3652 / 732 / 1921 / 999
PTNML: 12 / 527 / 1009 / 272 / 6
WinRate: 46.34%
ELO: -24.84[-36.57, -13.65]
LOS: 0.00
LLR: -12.11[-2.94, 2.94]

Compared with the last release on Github:

Scales well at STC:
Book: 3mvs 151-350_1097123
TC: 10+0.1
Total/Win/Draw/Lose: 5220 / 1291 / 2997 / 932
PTNML: 18 / 452 / 1322 / 789 / 29
WinRate: 53.44%
ELO: 23.37[13.56, 33.47]
LOS: 100.00
LLR: 11.55[-2.94, 2.94]

Scales well at LTC:
Book: 3mvs 151-350_1097123
TC: 60+0.6
Total/Win/Draw/Lose: 3142 / 790 / 1858 / 494
PTNML: 2 / 191 / 888 / 489 / 1
WinRate: 54.71%
ELO: 32.28[21.04, 44.26]
LOS: 100.00
LLR: 11.47[-2.94, 2.94]

What's New:

1. Introduced a new lossless feature merging architecture to reduce the size of the weights file while remains the same playing strength.
2. Allow repetition fold changing.
3. Add dot product instruction set support for Android devices, around 10% speed up measured.

Pikafish 2023-02-16

This version is extremely powerful compared to 2022-12-26:
59.73 ELO stronger than 2022-12-26 in STC:
Book: 3mvs 151-350_1097123
TC: 10+0.1
Total/Win/Draw/Lose: 2302 / 771 / 1156 / 375
PTNML: 5 / 139 / 523 / 423 / 61
WinRate: 58.60%
ELO: 59.73[43.51, 78.12]
LOS: 100.00
LLR: 10.69[-2.94, 2.94]

41.84 ELO stronger than 2022-12-26 in LTC:
Book: 3mvs 151-350_1097123
TC: 60+0.6
Total/Win/Draw/Lose: 2378 / 689 / 1289 / 400
PTNML: 0 / 166 / 598 / 395 / 30
WinRate: 56.08%
ELO: 41.84[28.53, 56.26]
LOS: 100.00
LLR: 10.33[-2.94, 2.94]

1. We switched to advisor and bishop missing bucketing NNUE architecture, an idea by mualanhlung017.
2. Part of the Chinese Rule is also supported in this release. You can enable it by setting the Repetition Rule options to ChineseRule. In Chinese Rule, detection for the illegal sequence x check y chase z mate threat is enabled where x, y, and z do not equal zero simultaneously.
3. For speed concerns, you can tweak the Mate Threat Depth option, it defines how many checks one side needs to make at most to form a checkmate when the opponent gives up moving this round. The larger the value, the slower the speed. Default to 1.
4. When Mate Threat Depth is set to 0, it means that the engine does not detect Mate Threat, and the illegal sequence becomes x check y chase where x and y do not equal zero simultaneously

Pikafish 2022-12-26

Compared to version 2022-10-22, we are having 60 move rule and 3-fold repetition at root nodes that almost fully comply with 亚规 now.
In addition to that, we also use the WDL model for score outputs and add all of these as tweakable UCI options for convenience.
Strict threefold even in search is also an option disabled by default but will be great to enable when analyzing positions.
The new compression architecture has been applied and a new arch net has also been released.

A small test confirmed at least 26 ELO gains against Pikafish version 2022-10-22:
Book: 3mvs 140-200_150560
TC: 10+0.1
Total/Win/Draw/Lose: 1548 / 359 / 949 / 240
PTNML: 14 / 109 / 404 / 238 / 9
WinRate: 53.84%
ELO: 26.20[15.87, 36.90]
LOS: 100.00
LLR: 3.29[-1.56, 2.77]

Pikafish 2022-10-22

This build still uses the nnue file trained using the same data as the last release, but with some adjustments, and gets a relatively good enhancement in LTC games. As for high score miseval in the endgame, wait for later releases.

versus 2022.10.15
Book: 3 Moves 140-200 Red Advantage
TC: 10+0.1
Total/Win/Draw/Lose: 1974 / 497 / 1112 / 365
PTNML: 13 / 172 / 488 / 298 / 16
WinRate: 53.34%
ELO: 22.69[12.93, 32.71]
LOS: 100.00
LLR: 3.94[-2.94, 2.94]

Book: 3 Moves 140-200 Red Advantage
TC: 60+0.6
Total/Win/Draw/Lose: 1954 / 448 / 1260 / 246
PTNML: 2 / 111 / 552 / 307 / 5
WinRate: 55.17%
ELO: 35.56[24.43, 47.48]
LOS: 100.00
LLR: 6.63[-2.94, 2.94]

Pikafish 2022-10-15

This build plays at the level of the latest version of the commercial engine AlphaCat aka AlphaCatFriendly 2022-09-26.

Pikafish 2022-10-07

Pikafish-2022-10-07

tune piece value and update see_ge (~97000 iter)

Pikafish 2022-10-01

Pikafish-2022-10-01

speed up for legal

Pikafish 2022-09-19

move prefetch

passed STC:
Total: 21758 W: 4130 L: 4024 D: 13604
Ptnml(0-2): 131, 2238, 6061, 2292, 157

passed LTC:
Total: 30804 W: 4449 L: 4332 D: 22023
Ptnml(0-2): 50, 2309, 10556, 2448, 39