This patch ports the efficiently updatable neural network (NNUE) evaluation to Stockfish.
Both the NNUE and the classical evaluations are available, and can be used to
assign a value to a position that is later used in alpha-beta (PVS) search to find the
best move. The classical evaluation computes this value as a function of various chess
concepts, handcrafted by experts, tested and tuned using fishtest. The NNUE evaluation
computes this value with a neural network based on basic inputs. The network is optimized
and trained on the evalutions of millions of positions at moderate search depth.
The NNUE evaluation was first introduced in shogi, and ported to Stockfish afterward.
It can be evaluated efficiently on CPUs, and exploits the fact that only parts
of the neural network need to be updated after a typical chess move.
[The nodchip repository](https://github.com/nodchip/Stockfish) provides additional
tools to train and develop the NNUE networks.
This patch is the result of contributions of various authors, from various communities,
including: nodchip, ynasu87, yaneurao (initial port and NNUE authors), domschl, FireFather,
rqs, xXH4CKST3RXx, tttak, zz4032, joergoster, mstembera, nguyenpham, erbsenzaehler,
dorzechowski, and vondele.
This new evaluation needed various changes to fishtest and the corresponding infrastructure,
for which tomtor, ppigazzini, noobpwnftw, daylen, and vondele are gratefully acknowledged.
The first networks have been provided by gekkehenker and sergiovieri, with the latter
net (nn-97f742aaefcd.nnue) being the current default.
The evaluation function can be selected at run time with the `Use NNUE` (true/false) UCI option,
provided the `EvalFile` option points the the network file (depending on the GUI, with full path).
The performance of the NNUE evaluation relative to the classical evaluation depends somewhat on
the hardware, and is expected to improve quickly, but is currently on > 80 Elo on fishtest:
60000 @ 10+0.1 th 1
https://tests.stockfishchess.org/tests/view/5f28fe6ea5abc164f05e4c4c
ELO: 92.77 +-2.1 (95%) LOS: 100.0%
Total: 60000 W: 24193 L: 8543 D: 27264
Ptnml(0-2): 609, 3850, 9708, 10948, 4885
40000 @ 20+0.2 th 8
https://tests.stockfishchess.org/tests/view/5f290229a5abc164f05e4c58
ELO: 89.47 +-2.0 (95%) LOS: 100.0%
Total: 40000 W: 12756 L: 2677 D: 24567
Ptnml(0-2): 74, 1583, 8550, 7776, 2017
At the same time, the impact on the classical evaluation remains minimal, causing no significant
regression:
sprt @ 10+0.1 th 1
https://tests.stockfishchess.org/tests/view/5f2906a2a5abc164f05e4c5b
LLR: 2.94 (-2.94,2.94) {-6.00,-4.00}
Total: 34936 W: 6502 L: 6825 D: 21609
Ptnml(0-2): 571, 4082, 8434, 3861, 520
sprt @ 60+0.6 th 1
https://tests.stockfishchess.org/tests/view/5f2906cfa5abc164f05e4c5d
LLR: 2.93 (-2.94,2.94) {-6.00,-4.00}
Total: 10088 W: 1232 L: 1265 D: 7591
Ptnml(0-2): 49, 914, 3170, 843, 68
The needed networks can be found at https://tests.stockfishchess.org/nns
It is recommended to use the default one as indicated by the `EvalFile` UCI option.
Guidelines for testing new nets can be found at
https://github.com/glinscott/fishtest/wiki/Creating-my-first-test#nnue-net-tests
Integration has been discussed in various issues:
https://github.com/official-stockfish/Stockfish/issues/2823https://github.com/official-stockfish/Stockfish/issues/2728
The integration branch will be closed after the merge:
https://github.com/official-stockfish/Stockfish/pull/2825https://github.com/official-stockfish/Stockfish/tree/nnue-player-wip
closes https://github.com/official-stockfish/Stockfish/pull/2912
This will be an exciting time for computer chess, looking forward to seeing the evolution of
this approach.
Bench: 4746616
This corrects a functional change in
ddcbacd04d
changing evaluation of KPPvK. Bench remains unchanged at low depth
With this patch, 8/8/5k1p/8/7p/7K/8/8 b - - 1 11 is again correctly evaluated as a draw.
closes https://github.com/official-stockfish/Stockfish/pull/2807
Bench: 4366686
In some French games, Stockfish likes to bring the Knight to a bad outpost spot. This is evident in TCEC S18 Superfinal Game 63, where there is a Knight outpost on the queenside but is actually useless. Stockfish is effectively playing a piece down while holding ground against Leela's break on the kingside.
This patch turns the +56 mg bonus for a Knight outpost into a -7 mg penalty if it satisfies the following conditions:
* The outpost square is not on the CenterFiles (i.e. not on files C,D,E and F)
* The knight is not attacking non pawn enemies.
* The side where the outpost is located contains only few enemies, with a particular conditional_more_than_two() implementation
Thank you to apospa...@gmail.com for bringing this to our attention and for providing insights.
See https://groups.google.com/forum/?fromgroups=#!topic/fishcooking/dEXNzSIBgZU
Reference game: https://tcec-chess.com/#div=sf&game=63&season=18
Passed STC:
LLR: 2.93 (-2.94,2.94) {-0.50,1.50}
Total: 6960 W: 1454 L: 1247 D: 4259
Ptnml(0-2): 115, 739, 1610, 856, 160
https://tests.stockfishchess.org/tests/view/5f08221059f6f0353289477e
Passed LTC:
LLR: 2.98 (-2.94,2.94) {0.25,1.75}
Total: 21440 W: 2767 L: 2543 D: 16130
Ptnml(0-2): 122, 1904, 6462, 2092, 140
https://tests.stockfishchess.org/tests/view/5f0838ed59f6f035328947a2
various related tests show strong test results, but so far no generalizations or simplifications of conditional_more_than_two() are found. See PR for details.
closes https://github.com/official-stockfish/Stockfish/pull/2803
Bench: 4366686
Probcut is a heuristic that wasn't changed a lot in past years,
all attempts to change it using information / writing info to transposition table failed.
This patch has a number of differences that can be summarized as follows:
* For TT write/read we use depth - 3. Because probcut search is depth - 4 but we actually do the move prior to it so effectively we do depth - 3 search;
* In any case of depth of eval from transposition table being >= depth - 3 we either produce cutoff or refuse to even do probcut search, this is allowing us to write info of probcut to transposition table because we know that we wouldn't be overwriting some deeper data with our depth - 3 search - this is an important aspect of this patch;
* For some not really known reason this patch completely ignores tte->bound() - which was the case for previous patch that made probcut interact with TT, maybe 2) is the reason, although it's unproven.
A first version of this patch passed STC and LTC
passed STC
https://tests.stockfishchess.org/tests/view/5f05908a59f6f03532894613
LLR: 2.95 (-2.94,2.94) {-0.50,1.50}
Total: 95776 W: 18300 L: 17973 D: 59503
Ptnml(0-2): 1646, 10944, 22377, 11279, 1642
passed LTC
https://tests.stockfishchess.org/tests/view/5f06b54059f6f035328946bb
LLR: 2.94 (-2.94,2.94) {0.25,1.75}
Total: 57128 W: 7266 L: 6938 D: 42924
Ptnml(0-2): 372, 5163, 17217, 5389, 423
However, an additional bugfix was needed to avoid checking a condition on ttMove if was not available. This passed non-regression bounds on top of the first version:
at STC
https://tests.stockfishchess.org/tests/view/5f080e5059f6f03532894766
LLR: 2.94 (-2.94,2.94) {-1.50,0.50}
Total: 14096 W: 2800 L: 2628 D: 8668
Ptnml(0-2): 225, 1620, 3238, 1688, 277
at LTC
https://tests.stockfishchess.org/tests/view/5f0836a559f6f0353289479c
LLR: 2.95 (-2.94,2.94) {-1.50,0.50}
Total: 25352 W: 3228 L: 3139 D: 18985
Ptnml(0-2): 175, 2350, 7549, 2415, 187
closes https://github.com/official-stockfish/Stockfish/pull/2804
Bench 4540940
A number of engines, GUIs and tournaments start to report WDL estimates
along or instead of scores. This patch enables reporting of those stats
in a more or less standard way (http://www.talkchess.com/forum3/viewtopic.php?t=72140)
The model this reporting uses is based on data derived from a few million fishtest LTC games,
given a score and a game ply, a win rate is provided that matches rather closely,
especially in the intermediate range [0.05, 0.95] that data. Some data is shown at
https://github.com/glinscott/fishtest/wiki/UsefulData#win-loss-draw-statistics-of-ltc-games-on-fishtest
Making the conversion game ply dependent is important for a good fit, and is in line
with experience that a +1 score in the early midgame is more likely a win than in the late endgame.
Even when enabled, the printing of the info causes no significant overhead.
Passed STC:
LLR: 2.94 (-2.94,2.94) {-1.50,0.50}
Total: 197112 W: 37226 L: 37347 D: 122539
Ptnml(0-2): 2591, 21025, 51464, 20866, 2610
https://tests.stockfishchess.org/tests/view/5ef79ef4f993893290cc146b
closes https://github.com/official-stockfish/Stockfish/pull/2778
No functional change
We lower the endgame value of the evaluation when we detect that there
is only one queen left on the board (more precisely, we use a scale
factor of 37/64, or about 0.58, for the endgame part of the evaluation).
Hopefully this helps a little bit for the assessment of positions with
queen imbalance, which are one of the well-known Stockfish weaknesses.
STC:
LLR: 2.94 (-2.94,2.94) {-0.50,1.50}
Total: 21600 W: 4176 L: 3955 D: 13469
Ptnml(0-2): 351, 2457, 5003, 2598, 391
https://tests.stockfishchess.org/tests/view/5ef871b6020eec13834a94e8
LTC:
LLR: 2.97 (-2.94,2.94) {0.25,1.75}
Total: 248328 W: 30596 L: 29720 D: 188012
Ptnml(0-2): 1544, 22345, 75665, 22911, 1699
https://tests.stockfishchess.org/tests/view/5ef87aec020eec13834a94fe
Closes https://github.com/official-stockfish/Stockfish/pull/2781
Bench: 4441323
* Supports popcnt (thanks @daylen)
* bits = 64 is now the default
Tested with g++ (Ubuntu/Linaro 7.5.0-3ubuntu1~18.04) 7.5.0 on ThunderX CN8890,
yields about 9% speedup.
Also tested with clang version 6.0.0-1ubuntu2 (tags/RELEASE_600/final).
closes https://github.com/official-stockfish/Stockfish/pull/2770
No functional change.
the option was, since at least 2014, not correctly implemented,
ignoring all dynamic adjustments to optimum time in search.
Instead of fixing it, remove it, no need to expose an option that
will influence time management negatively.
closes https://github.com/official-stockfish/Stockfish/pull/2765
No functional change.
The idea of this patch is that if capture can be described as
"less valuable piece takes more valuable piece" it's not really correct
to add only piece value of captured piece to static evaluation
since there can be more threats in other places and opponent can't really
do much but recapture our capturing piece which leaves us space for
more captures thus winning more material and increasing static eval.
passed STC
https://tests.stockfishchess.org/tests/view/5ef0167b122d6514328d760f
LLR: 2.96 (-2.94,2.94) {-0.50,1.50}
Total: 24736 W: 4838 L: 4607 D: 15291
Ptnml(0-2): 438, 2812, 5648, 3021, 449
passed LTC
https://tests.stockfishchess.org/tests/view/5ef073bc122d6514328d7693
LLR: 2.93 (-2.94,2.94) {0.25,1.75}
Total: 46152 W: 5865 L: 5567 D: 34720
Ptnml(0-2): 312, 4160, 13886, 4354, 364
closes https://github.com/official-stockfish/Stockfish/pull/2761
bench 4789930
We increase a little bit the midgame value of pawns on a4, h4, a6 and h6.
Original idea by Malcolm Campbell, who tried the version restricted to the
pawns on the H column a couple of weeks ago and got a patch which almost
passed LTC. The current pull request just adds the same idea for pawns on
the A column.
Possible follow-ups: maybe tweak the a5/h5 pawn values, and/or add a malus
for very low king mobility in midgame?
STC:
LLR: 2.94 (-2.94,2.94) {-0.50,1.50}
Total: 33416 W: 6516 L: 6275 D: 20625
Ptnml(0-2): 575, 3847, 7659, 4016, 611
https://tests.stockfishchess.org/tests/view/5ee6c4e687586124bc2c10d4
LTC:
LLR: 2.95 (-2.94,2.94) {0.25,1.75}
Total: 134368 W: 16869 L: 16319 D: 101180
Ptnml(0-2): 908, 12083, 40708, 12521, 964
https://tests.stockfishchess.org/tests/view/5ee74e60aae8aec816ab756a
closes https://github.com/official-stockfish/Stockfish/pull/2747
Bench: 5299456
Tuned search constants after many search patches since the last
successful tune.
1st LTC @ 60+0.6 th 1 :
LLR: 2.97 (-2.94,2.94) {0.25,1.75}
Total: 57656 W: 7369 L: 7036 D: 43251
Ptnml(0-2): 393, 5214, 17336, 5437, 448
https://tests.stockfishchess.org/tests/view/5ee1e074f29b40b0fc95af19
SMP LTC @ 20+0.2 th 8 :
LLR: 2.95 (-2.94,2.94) {0.25,1.75}
Total: 83576 W: 9731 L: 9341 D: 64504
Ptnml(0-2): 464, 7062, 26369, 7406, 487
https://tests.stockfishchess.org/tests/view/5ee35a21f29b40b0fc95b008
The changes were rebased on top of a successful patch by Viz (see #2734)
and two different ways of doing this were tested. The successful test
modified the constants in the patch by Viz in a similar manner to the
tuning run:
LTC (rebased) @ 60+0.6 th 1 :
LLR: 2.94 (-2.94,2.94) {0.25,1.75}
Total: 193384 W: 24241 L: 23521 D: 145622
Ptnml(0-2): 1309, 17497, 58472, 17993, 1421
https://tests.stockfishchess.org/tests/view/5ee43319ca6c451633a995f9
Further work: the recent patch to quantize eval #2733 affects search quit
quite a bit, so doing another tune in, say, three months time might be a
good idea.
closes https://github.com/official-stockfish/Stockfish/pull/2735
Bench 4246971
We replace the current decrease of the complexity term in initiative
when shuffling by a direct damping of the evaluation. This scheme may
have two benefits over the initiative approach:
a) the damping effect is more brutal for fortresses with heavy pieces
on the board, because the initiative term is almost an endgame term;
b) the initiative implementation had a funny side effect, almost a bug,
in the rare positions where mg > 0, eg < 0 and the tampered eval
returned a positive value (ie with heavy pieces still on the board):
sending eg to zero via shuffling would **increase** the tampered
eval instead of decreasing it, which is somewhat illogical. This
patch avoids this phenomenon.
STC:
LLR: 2.94 (-2.94,2.94) {-0.50,1.50}
Total: 43072 W: 8373 L: 8121 D: 26578
Ptnml(0-2): 729, 4954, 9940, 5162, 751
https://tests.stockfishchess.org/tests/view/5ee008ebf29b40b0fc95ade2
LTC:
LLR: 2.94 (-2.94,2.94) {0.25,1.75}
Total: 37376 W: 4816 L: 4543 D: 28017
Ptnml(0-2): 259, 3329, 11286, 3508, 306
https://tests.stockfishchess.org/tests/view/5ee03b06f29b40b0fc95ae0c
Closes https://github.com/official-stockfish/Stockfish/pull/2727
Bench: 4757174
Conceptually group hash clusters into super clusters of 256 clusters.
This scheme allows us to use hash sizes up to 32 TB
(= 2^32 super clusters = 2^40 clusters).
Use 48 bits of the Zobrist key to choose the cluster index. We use 8
extra bits to mitigate the quantization error for very large hashes when
scaling the hash key to cluster index.
The hash index computation is organized to be compatible with the existing
scheme for power-of-two hash sizes up to 128 GB.
Fixes https://github.com/official-stockfish/Stockfish/issues/1349
closes https://github.com/official-stockfish/Stockfish/pull/2722
Passed non-regression STC:
LLR: 2.93 (-2.94,2.94) {-1.50,0.50}
Total: 37976 W: 7336 L: 7211 D: 23429
Ptnml(0-2): 578, 4295, 9149, 4356, 610
https://tests.stockfishchess.org/tests/view/5edcbaaef29b40b0fc95abc5
No functional change.
This is a non-functional code style change which provides a safe access handler for LineBB.
Also includes an assert in debug mode to verify square correctness.
closes https://github.com/official-stockfish/Stockfish/pull/2719
No functional change
