Commit Graph

5723 Commits

Author SHA1 Message Date
Tomasz Sobczyk
3041adb080 Cleanup layers. 2020-10-18 19:32:15 +09:00
Tomasz Sobczyk
0d4c3014ca Cleanup features. 2020-10-17 23:19:16 +09:00
Tomasz Sobczyk
ca760c3a5b Cleanup architecture files. 2020-10-17 20:01:09 +09:00
Tomasz Sobczyk
c93f8732bf Force Use NNUE to pure when learning. 2020-10-17 08:44:38 +09:00
Tomasz Sobczyk
3cf193a90e Properly handle cases in verify and init when SkipLoadingEval is set. 2020-10-17 08:44:38 +09:00
Tomasz Sobczyk
5db46d0c82 Verify whether there is a network being used during training. 2020-10-17 08:44:38 +09:00
Tomasz Sobczyk
e503cc4ea8 Add one more empty line between progress reports. 2020-10-17 00:13:50 +09:00
Tomasz Sobczyk
5856237e3f Rename hirate to startpos 2020-10-16 09:07:02 +09:00
Tomasz Sobczyk
904adb9a32 Indentation consistency in learn folder 2020-10-15 22:11:31 +09:00
Tomasz Sobczyk
880d23af1c Move sfen input/output streams to sfen_stream.h 2020-10-15 20:37:03 +09:00
Tomasz Sobczyk
14f83ad7b9 Move public search/qsearch interface from namespace Learner to namespace Search 2020-10-15 20:37:03 +09:00
Tomasz Sobczyk
0494adeb2c Move nnue evaluation stuff from evaluate.h to nnue/evaluate_nnue.h 2020-10-15 20:37:03 +09:00
Unai Corzo
288a604411 Scale factor tweak
Add !pawnsOnBothFlanks heuristic to scale factor.

STC https://tests.stockfishchess.org/tests/view/5f8080575b3847b5d41f9134
LLR: 2.94 (-2.94,2.94) {-0.25,1.25}
Total: 250960 W: 49779 L: 49168 D: 152013
Ptnml(0-2): 4224, 28822, 58802, 29383, 4249

LTC https://tests.stockfishchess.org/tests/view/5f832f498ea73fb8ddf83ddb
LLR: 2.95 (-2.94,2.94) {0.25,1.25}
Total: 88584 W: 11827 L: 11388 D: 65369
Ptnml(0-2): 585, 8079, 26578, 8412, 638

closes https://github.com/official-stockfish/Stockfish/pull/3179

bench: 3834252
2020-10-14 19:32:12 +02:00
FauziAkram
4a5cc1365f RookOnQueenFile Removal
Removing Rook On Queen File looks beneficial, and it might even bring some ELO.
I will try to reintroduce it with a different method later on.

Passed STC:
https://tests.stockfishchess.org/tests/view/5f7cea204389873867eb10cb
LLR: 2.94 (-2.94,2.94) {-1.25,0.25}
Total: 18624 W: 3800 L: 3568 D: 11256
Ptnml(0-2): 308, 2131, 4257, 2253, 363

Passed LTC:
https://tests.stockfishchess.org/tests/view/5f7d76a4e936c6892bf50598
LLR: 2.95 (-2.94,2.94) {-0.75,0.25}
Total: 117864 W: 15515 L: 15340 D: 87009
Ptnml(0-2): 926, 11127, 34671, 11262, 946

closes https://github.com/official-stockfish/Stockfish/pull/3176

Bench: 3756191
2020-10-14 19:29:22 +02:00
Joost VandeVondele
ba73f8ce0d Update default net to nn-04cf2b4ed1da.nnue
Further tune the net parameters, now the last but one layer (32x32).
To limit the number of parameters optimized, the network layer was
decomposed using SVD, and the singular values were treated
as parameters and tuned.

Tuning branch: https://github.com/vondele/Stockfish/tree/svdTune
Tuner: https://github.com/vondele/nevergrad4sf

passed STC:
https://tests.stockfishchess.org/tests/view/5f83e82f8ea73fb8ddf83e4e
LLR: 2.94 (-2.94,2.94) {-0.25,1.25}
Total: 8488 W: 944 L: 795 D: 6749
Ptnml(0-2): 39, 609, 2811, 734, 51

passed LTC:
https://tests.stockfishchess.org/tests/view/5f83f4118ea73fb8ddf83e66
LLR: 2.94 (-2.94,2.94) {0.25,1.25}
Total: 169016 W: 8043 L: 7589 D: 153384
Ptnml(0-2): 133, 6623, 70538, 7085, 129

closes https://github.com/official-stockfish/Stockfish/pull/3181

Bench: 3945198
2020-10-14 13:28:21 +02:00
Tomasz Sobczyk
4a340ad3b2 Add docs for auto_lr_drop 2020-10-12 23:56:26 +09:00
Tomasz Sobczyk
4a2bf16b30 Add option "auto_lr_drop" that specifies the amount of positions from previous lr drop after which to reduce lr by newbob_decay. 2020-10-12 23:56:26 +09:00
Tomasz Sobczyk
7d62b3f799 Store additional bits for fullmove clock and 50 more rule halfmove clock at the end of the bit stream. This change keeps backwards compatibility. 2020-10-11 20:59:27 +09:00
Tomasz Sobczyk
de20887e11 Update readme. Link to docs. 2020-10-10 14:37:51 +09:00
Tomasz Sobczyk
2af4bf7eac Move the docs folder one above, it was in src by mistake. 2020-10-10 14:37:51 +09:00
Tomasz Sobczyk
ef57ac78a3 Print gensfen speed when outputting status. 2020-10-09 08:14:48 +09:00
Joost VandeVondele
adddf339bb Output sfens/second in the trainer, to track performance more easily 2020-10-08 08:29:42 +09:00
Tomasz Sobczyk
3f55b3af42 Change some gensfen parameter defaults. 2020-10-07 23:09:33 +09:00
Tomasz Sobczyk
8830209125 Change some learn parameter defaults. 2020-10-07 23:09:33 +09:00
Tomasz Sobczyk
2e57f3fa22 Switch to set recommended learn UCI options 2020-10-07 23:09:33 +09:00
Tomasz Sobczyk
d1c44dca04 Switch to set recommended gensfen UCI options 2020-10-07 23:09:33 +09:00
Tomasz Sobczyk
5fa28b12fa Allow setting UCI options programmatically. 2020-10-07 23:09:33 +09:00
Tomasz Sobczyk
80cbc3ffee Fix grammar and spelling. Add recommendations for UCI options. 2020-10-07 16:08:26 +09:00
Tomasz Sobczyk
31f9d66f12 Initial documentation for learn, gensfen, convert, and binpack. 2020-10-07 16:08:26 +09:00
FauziAkram
767b4f4fbe Pawn Tuning
Tuning of pawns, for classical evaluation:

Passed STC:
https://tests.stockfishchess.org/tests/view/5f771f0e52560f5fc78559ec
LLR: 2.96 (-2.94,2.94) {-0.25,1.25}
Total: 252696 W: 50321 L: 49692 D: 152683
Ptnml(0-2): 4614, 29845, 57049, 29978, 4862

Passed LTC:
https://tests.stockfishchess.org/tests/view/5f77cfef090dcf9aaa16d38b
LLR: 2.94 (-2.94,2.94) {0.25,1.25}
Total: 48184 W: 6556 L: 6193 D: 35435
Ptnml(0-2): 335, 4516, 14100, 4733, 408

closes https://github.com/official-stockfish/Stockfish/pull/3169

bench: 4016121
2020-10-05 19:01:46 +02:00
Unai Corzo
17fb3a8ce0 Simplify away futility pruning for captures
Remove futility pruning for captures.

STC https://tests.stockfishchess.org/tests/view/5f749bfed930428c36d34c56
LLR: 2.94 (-2.94,2.94) {-1.25,0.25}
Total: 38064 W: 4011 L: 3929 D: 30124
Ptnml(0-2): 192, 3004, 12567, 3068, 201

LTC https://tests.stockfishchess.org/tests/view/5f74d99bf18675b1ce2f7412
LLR: 2.94 (-2.94,2.94) {-0.75,0.25}
Total: 184984 W: 8567 L: 8610 D: 167807
Ptnml(0-2): 146, 7593, 77058, 7548, 147

closes https://github.com/official-stockfish/Stockfish/pull/3166

bench: 3890648
2020-10-05 18:59:02 +02:00
Joost VandeVondele
9382f854b3 Schedule threads fairly under valgrind
fixes a rare case that can cause CI to fail when running multithreaded under valgrind.

closes https://github.com/official-stockfish/Stockfish/pull/3165

No functional change.
2020-10-05 18:56:49 +02:00
noobpwnftw
91cb4a6770 Skip eval dampening in Use NNUE = pure case 2020-10-03 19:31:21 +09:00
noobpwnftw
6f7a228707 Minor cleanups
Remove unused code and magic numbers
2020-10-01 16:52:20 +09:00
Joost VandeVondele
f848d67341 Use fair scheduling of threads under valgrind
fixes some rare case where the master search thread makes no progress,
observed in CI.
2020-10-01 09:36:44 +09:00
nodchip
b889debfb5 Merge pull request #171 from noobpwnftw/sf_merge
Merge SFdev
2020-09-30 10:18:41 +09:00
noobpwnftw
d865159bd6 Fix variable initialization in test commands 2020-09-29 17:30:08 +08:00
noobpwnftw
a8b502a975 Merge remote-tracking branch 'remotes/origin/master'
Bench: 3618595
2020-09-29 17:09:14 +08:00
noobpwnftw
b44d539c94 Fix a bug that LR is not correctly scaled when initial LR is not 1.0 2020-09-29 16:18:42 +09:00
Stéphane Nicolet
5af09cfda5 Include pawns in NNUE scaling
We now include the total pawn count in the scaling factor for the output
of the NNUE evaluation network. This should have the effect of trying to
keep more pawns when SF has the advantage, but exchange them when she
is defending.

Thanks to Alexander Pagel (Lolligerhans) for the idea of using the
value of pawns to ease the comparison with the rest of the material
estimation.

Passed STC:
LLR: 2.93 (-2.94,2.94) {-0.25,1.25}
Total: 15072 W: 1700 L: 1539 D: 11833
Ptnml(0-2): 65, 1202, 4845, 1355, 69
https://tests.stockfishchess.org/tests/view/5f7235a63b22d6afa50699b3

Passed LTC:
LLR: 2.93 (-2.94,2.94) {0.25,1.25}
Total: 25880 W: 1270 L: 1124 D: 23486
Ptnml(0-2): 23, 980, 10788, 1126, 23
https://tests.stockfishchess.org/tests/view/5f723b483b22d6afa5069a99

closes https://github.com/official-stockfish/Stockfish/pull/3164

Bench: 3776081
2020-09-28 22:42:26 +02:00
Stefan Geschwentner
6f0aa186d8 Tweak reduction formula.
Replace log(i) with log(i + 0.25 * log(i)). This increases especially for low values the reductions. But for bigger values there are nearly no changes.

STC:
LLR: 2.94 (-2.94,2.94) {-0.25,1.25}
Total: 49640 W: 5505 L: 5289 D: 38846
Ptnml(0-2): 270, 4074, 15924, 4274, 278
https://tests.stockfishchess.org/tests/view/5f71f04d3b22d6afa5069478

LTC:
LLR: 2.94 (-2.94,2.94) {0.25,1.25}
Total: 43856 W: 2209 L: 2021 D: 39626
Ptnml(0-2): 32, 1776, 18128, 1956, 36
https://tests.stockfishchess.org/tests/view/5f7232ee3b22d6afa50699a2

closes https://github.com/official-stockfish/Stockfish/pull/3163

Bench: 3555769
2020-09-28 22:34:25 +02:00
SFisGOD
5efbaaba77 Update default net to nn-baeb9ef2d183.nnue
Further optimization of Sergio's nn-03744f8d56d8.nnue
This patch is the result of collaboration with Joost VandeVondele.

STC:
LLR: 2.96 (-2.94,2.94) {-0.25,1.25}
Total: 37000 W: 4145 L: 3947 D: 28908
Ptnml(0-2): 191, 3016, 11912, 3166, 215
https://tests.stockfishchess.org/tests/view/5f71e7983b22d6afa5069475

LTC:
LLR: 2.96 (-2.94,2.94) {0.25,1.25}
Total: 60224 W: 2992 L: 2769 D: 54463
Ptnml(0-2): 48, 2420, 24956, 2637, 51
https://tests.stockfishchess.org/tests/view/5f722bb83b22d6afa506998f

closes https://github.com/official-stockfish/Stockfish/pull/3161

Bench: 3720073
2020-09-28 22:29:31 +02:00
FauziAkram
ba46599aa2 Tweaking Mobility and Safe Check
Passed STC:
https://tests.stockfishchess.org/tests/view/5f70d86d3b22d6afa50693b9
LLR: 2.94 (-2.94,2.94) {-0.25,1.25}
Total: 100368 W: 20323 L: 19914 D: 60131
Ptnml(0-2): 1927, 11641, 22605, 12118, 1893

Passed LTC:
https://tests.stockfishchess.org/tests/view/5f71bb553b22d6afa5069457
LLR: 2.94 (-2.94,2.94) {0.25,1.25}
Total: 77648 W: 10613 L: 10181 D: 56854
Ptnml(0-2): 634, 7280, 22594, 7652, 664

closes https://github.com/official-stockfish/Stockfish/pull/3160

Bench: 3861984
2020-09-28 22:26:37 +02:00
Vizvezdenec
a5e68d9b25 Adjust null move pruning constants
Idea is that division by fraction of 2 is slightly faster than by other numbers so parameters are adjusted in a way that division in null move pruning depth reduction features dividing by 256 instead of dividing by 213.
Other than this patch is almost non-functional - difference starts to exist by depth 133.

passed STC
https://tests.stockfishchess.org/tests/view/5f70dd943b22d6afa50693c5
LLR: 2.95 (-2.94,2.94) {-0.25,1.25}
Total: 57048 W: 6616 L: 6392 D: 44040
Ptnml(0-2): 304, 4583, 18531, 4797, 309

passed LTC
https://tests.stockfishchess.org/tests/view/5f7180db3b22d6afa506941f
LLR: 2.95 (-2.94,2.94) {0.25,1.25}
Total: 45960 W: 2419 L: 2229 D: 41312
Ptnml(0-2): 43, 1779, 19137, 1987, 34

closes https://github.com/official-stockfish/Stockfish/pull/3159

bench 3789924
2020-09-28 22:22:54 +02:00
Joost VandeVondele
36c2886302 Update default net to nn-04a843f8932e.nnue
an optimization of Sergio's nn-03744f8d56d8.nnue tuning the output layer (33 parameters) on game play.

WIP code to make layer parameters tunable is https://github.com/vondele/Stockfish/tree/optionOutput
Optimization itself is using https://github.com/vondele/nevergrad4sf
Writing of the modified net using WIP code based on the learner code https://github.com/vondele/Stockfish/tree/evalWrite

Most parameters in the output layer are changed only little (~5 for int8_t).

passed STC:
https://tests.stockfishchess.org/tests/view/5f716f6b3b22d6afa506941a
LLR: 2.94 (-2.94,2.94) {-0.25,1.25}
Total: 15488 W: 1859 L: 1689 D: 11940
Ptnml(0-2): 79, 1260, 4917, 1388, 100

passed LTC:
https://tests.stockfishchess.org/tests/view/5f71908e3b22d6afa506942e
LLR: 2.93 (-2.94,2.94) {0.25,1.25}
Total: 8728 W: 518 L: 400 D: 7810
Ptnml(0-2): 7, 338, 3556, 456, 7

closes https://github.com/official-stockfish/Stockfish/pull/3158

Bench: 3789924
2020-09-28 16:55:40 +02:00
noobpwnftw
c065abdcaf Use incremental updates more often
Use incremental updates for accumulators for up to 2 plies.
Do not copy accumulator. About 2% speedup.

Passed STC:
LLR: 2.95 (-2.94,2.94) {-0.25,1.25}
Total: 21752 W: 2583 L: 2403 D: 16766
Ptnml(0-2): 128, 1761, 6923, 1931, 133
https://tests.stockfishchess.org/tests/view/5f7150cf3b22d6afa5069412

closes https://github.com/official-stockfish/Stockfish/pull/3157

No functional change
2020-09-28 16:54:35 +02:00
Stéphane Nicolet
1dbd2a1ad5 Tweak nnue scaling to keep more material
Current master uses a constant scale factor of 5/4 = 1.25 for the output
of the NNUE network, for compatibility with search and classical evaluation.
We modify this scale factor to make it dependent on the phase of the game,
going from about 1.5 in the opening to 1.0 for pure pawn endgames.

This helps Stockfish to avoid exchanges of pieces (heavy pieces in particular)
when she has the advantage, keeping more material on the board when attacking.

Passed STC:
LLR: 2.95 (-2.94,2.94) {-0.25,1.25}
Total: 14744 W: 1771 L: 1599 D: 11374
Ptnml(0-2): 87, 1184, 4664, 1344, 93
https://tests.stockfishchess.org/tests/view/5f6fb0a63b22d6afa506904f

Passed LTC:
LLR: 2.95 (-2.94,2.94) {0.25,1.25}
Total: 8912 W: 512 L: 393 D: 8007
Ptnml(0-2): 7, 344, 3637, 459, 9
https://tests.stockfishchess.org/tests/view/5f6fcf533b22d6afa5069066

closes https://github.com/official-stockfish/Stockfish/pull/3154

Bench: 3943952
2020-09-27 08:24:50 +02:00
noobpwnftw
9d4bf4fe0c Optimize accumulators for null move. 2020-09-27 09:39:16 +09:00
noobpwnftw
96a3180770 Update instrumented_learn.sh
Fix typo.
2020-09-27 09:32:26 +09:00
noobpwnftw
5e8a49f7f2 Restore lambda and gradient function post-merge and minor fixes.
bench: 3788313
2020-09-26 12:55:02 +09:00