LCZero

[Владимир001]

Tune of: CPuct, CPuctAtRoot, FpuValue
LC0 version: 0.26.3
SF version: 12
LC0 options: threads=1, minibatch-size=96, root-has-own-cpuct-params=false, moves-left-max-effect=0.2, moves-left-threshold=0, moves-left-slope=0.004, moves-left-quadratic-factor=0, moves-left-scaled-factor=1, move-overhead-ms=0
SF options: contempt=0, moveoverhead=0
Tuning ranges: CPuct: Real(0.5, 2.5), FpuValue: Real(-0.25, 1.0)
Tuning configuration: acq_function: vr/ei, rounds: 10
Time control (LC0): 0.07s per move
Time control (SF): 0.04s per move
Net (LC0): 703810
Net (SF): nn-eba324f53044.nnue
Book: custom-3k (1-3 ply)
Hardware: gtx970 + i5-4570
Tablebases: syzygy 5-man
Adjudication: -draw movenumber=50 movecount=5 score=10 -resign movecount=5 score=1000
Software: chess-tuning-tools
Comment: stc tune of these params. full config json is attached. @zz4032 409 iterations and median lc0 npm should be ~1100

Optimums found:

Cpuct            1.5674
CPuctAtRoot      1.5674
FpuValue         0.242

90.0% confidence intervals of the parameters:

Parameter  Lower bound  Upper bound
-----------------------------------
CPuct             1.41         2.28
FpuValue        -0.071        0.324

Развернуть для просмотра

[свернуть]

[Владимир001]

Развернуть для просмотра

Gauntlet: Lc0 26.3 cuda + J92-270 vs SF-12 and SF Classical
Time control: 300s + 3s, Ryzen 9 4900H, RTX 2060 mobile
Openings: TCEC-19 (50 pos.)
Stockfish: threads=8, hash=1024
Lc0 26.3: --nncache=1000000 --threads=2 --minibatch-size=112 --max-prefetch=50 --max-collision-events=50 --moves-left-max-effect=0.2 --moves-left-threshold=0.0 --moves-left-slope=0.004 --moves-left-constant-factor=0.0 --moves-left-scaled-factor=1.0 --moves-left-quadratic-factor=0.0
Bench startpos (15s): J92-270 (cuda-fp16) = 6.4 knps, SF-12 (avx2) = 8100 knps, SF Classical (stockfish_20073110_x64_modern) = 13 000 knps.
TBs and adj.: Syzygy 5-men, resign (two-sided) movecount=5 score=600, draw movenumber=35 movecount=5 score=5
Comments: Results looks good for J92-270. +45, =55, -0 with the white pieces!

Score of lc0-26.3-cuda-J92-270 vs Stockfish 12 8CPU: 16 - 18 - 66 [0.490]
...      lc0-26.3-cuda-J92-270 playing White: 15 - 0 - 35  [0.650] 50
...      lc0-26.3-cuda-J92-270 playing Black: 1 - 18 - 31  [0.330] 50
...      White vs Black: 33 - 1 - 66  [0.660] 100
Elo difference: -6.9 +/- 39.8, LOS: 36.6 %, DrawRatio: 66.0 %
100 of 100 games finished.

Score of lc0-26.3-cuda-J92-270 vs Stockfish Classical 8CPU: 34 - 12 - 54 [0.610]
...      lc0-26.3-cuda-J92-270 playing White: 30 - 0 - 20  [0.800] 50
...      lc0-26.3-cuda-J92-270 playing Black: 4 - 12 - 34  [0.420] 50
...      White vs Black: 42 - 4 - 54  [0.690] 100
Elo difference: 77.7 +/- 46.2, LOS: 99.9 %, DrawRatio: 54.0 %
100 of 100 games finished.

[свернуть]

Games: https://gofile.io/d/EwxnSI

[SergeyMZ]

Подскажите пожалуйста, какая сейчас нейросетка предпочтительна для анализа шахматных позиций на 30-180 минут движком Лилка 26.3?
А также, какой параметр NNCache для карты Nvidia RTX 2080 с указанным выше движком и ОЗУ=32 Гб является оптимальным?

[Maratka]

Я бы сказал, что лучше всего анализировать последними версиями SF'ша.

А так вообще - ну вот тесты сетей:

Цитата: Сообщение от Владимир001

Match: combined results (314 positions) of ERET test + Vesely 203 positions testset
time control: analysis with time limit of 60s per move/position
AB engine options: threads=4, hash=4096 MB
Lc0-version: v0.26.2 + v0.26.3
LC0-options: -Backend=cudnn -NNcache=3000000 -MaxCcollisionEvents=256
MLH-settings S19 SuFi: -MovesLeftMaxEffect=0.2 -MovesLeftSlope=0.004 -MovesLeftThreshold=0.0 -MovesQuadraticFactor=0.0 -MovesScaledFactor=1.0
hardware: notebook with mobile i7-7700HQ 4x2.8GHz, mobile Geforce GTX 1050 TI 4GB
software: Arena GUI, "Automatic Analysis" feature with EPD file (Vesely_testsuiteLC0.epd).
Tablebases: Syzygy 5-men on SSD
Comments: For J92 so far those results fit nicely to the results of the test gauntet against SF Dev at 5min+5s TC (280 not tested there yet).

A) CPU engines (mobile i7-7700HQ 4x2.8GHz):
Cfish 161020 NNUE 4CPU : 280/314 = 89.2% <== new, *net: nn-04cf2b4ed1da.nnue*

B) GPU engines (mobile Geforce GTX 1050 TI 4GB):
Lc0.net.JH.J92-280 (30x384) : 220/314 = 70.0% <== new
Lc0.net.JH.J92-260 (30x384) : 218/314 = 69.4% <== new
Lc0.net.JH.J92-270 (30x384) : 217/314 = 69.1% <== new
Lc0.net.JH.J92-210 (30x384) : 216/314 = 68.8%
Lc0.net.JH.J92-240 (30x384) : 215/314 = 68.5%
Lc0-0.22.0-dev.Leelenstein 14.3 : 215/314 = 68.5%
Lc0.net.JH.J92-190 (30x384) : 214/314 = 68.2% ** TCEC S19 SuFi net **
Lc0.net.65162 : 213/314 = 67.8%
Lc0.net.65411 : 205/314 = 65.3% <== new

[SergeyMZ]

С сетями понял, спасибо. У меня самого стоит 190-ая. Попробую 280-ую. Только с параметром NNCache так и неясно: у Владимира карта намного слабее моей.

[Maratka]

Я думаю, Вы сами быстрее проверите, чем ответ получите.
Начните с настройки по умолчанию.

[Владимир001]

Развернуть для просмотра

Match Gauntlet
LC0-version: lc0-v0.26.3 cudnn-version
LC0 options: Standard option except Threads=1, NNCacheSize=0, MinibatchSize=1.
Time control: 1 Node/move
Hardware: i7-4790 with 16 GByte RAM + GTX1650
Speed: Not important for this type of testing.
Book: Chad-opening book with 100 000 different openings with the most popular human openings.
Book options: Opening order: Sequential. Play each opening twice. Play all opening in the book both for White and Black.
Tablebases: No Tablebases.
Adjudication: Only 5-men syzygy
Software: CuteChess 1.2.0 https://github.com/cutechess/cutechess/releases / Graphic tools is programmed by glbchess64, in Python programming tools.
Comment: This is result out of Ordo 1.2.6 https://github.com/michiguel/Ordo/releases/tag/v1.2.6
Ordo string: ordo.exe -Q -N 0 -D -a 0 -A "J92-145" -W -n4 -s500 -G -V -U "0,1,2,3,4,5,6,7,8,9,10" -j stats.txt -o rating.txt -p games.pgn
Results: pgn-file: https://ufile.io/3n3m9l9h
Explanation: Trying to find the best net in clean policy Elo.

   # PLAYER     :  RATING  ERROR    POINTS  PLAYED   (%)  CFS(%)      W      D      L  D(%)
   1 J92-280    :      11      1  103100.0  200000  51.5     100  66752  72696  60552  36.3
   2 J92-145    :       0      1   96900.0  200000  48.5     ---  60552  72696  66752  36.3

White advantage = 23.98 +/- 0.57
Draw rate (equal opponents) = 36.51 % +/- 0.10

My list until now (Elo against J92-145) 1 Node/move:

White advantage = 23.98 +/- 0.57
Draw rate (equal opponents) = 36.51 % +/- 0.10
J92-280:             11     Elo    (Test with 200k games)
J92-260:             11     Elo    (Test with 200k games)
J92-240:              8     Elo    (Test with 200k games)
J92-210:              5     Elo    (Test with 200k games)
J92-220:              3     Elo    (Test with 200k games)
J92-205:              3     Elo    (Test with 200k games)
J92-230:              2     Elo    (Test with 200k games)
J92-190:              1     Elo    (Test with 200k games)
J92-145:              0     Elo    (Reference)
J92-100:             -2     Elo    (Test with 200k games)

[свернуть]

[Владимир001]

Развернуть для просмотра

Match: gauntlet of SF Dev 20092822 4CPU vs Lc0 J92-280/270/260/240/220/210/190 (S19 SuFi1 MLH) + Cfish NNUE + SF11
Lc0-version: v0.26.2 + v0.26.3 (J92-270/260)
Lc0-options : -Backend=cudnn -NNCacheSize=2000000 -MaxOutOfOrderEvalsFactor=2.4 -MaxCollisionEvents=256 -TwoFoldDraws=true
MLH-settings S19 SuFi: -MovesLeftMaxEffect=0.2 -MovesLeftSlope=0.004 -MovesLeftThreshold=0.0 -MovesQuadraticFactor=0.0 -MovesScaledFactor=1.0
SF-version: Stockfish 20092822 x64 bmi2 (date: 28.09.2020)
SF/Cfish-options: threads=4 hash=2048 MB, contempt=24, NNUE=true, net=nn-baeb9ef2d183.nnue (default)
Time control: 5min + 5s
Hardware: notebook with mobile i7-7700HQ 4x2.8GHz (8 threads with HT on), mobile Geforce GTX 1050 TI 4GB
Book: Balsa_Special.pgn, 10 plies, sequentially, colors reversed
Tablebases: Syzygy 5-men on SSD
Speed: J92: 961 nps, SF Dev NNUE 4CPU: 3680k nps, Cfish NNUE 4CPU: 4198k nps
Adjudication: Draw at 50 moves rule, 3-fold repetition, after 200 moves, at 50 moves with eval = 0.00 ; 5-men Gaviota tablebase adjudication
Software: Arena 3.5.1, ponder = off
Elo rating: Ordo 1.2.6
Comments: J92-280 added. The table includes the ERET test results (60s time limit).

  # PLAYER                  :  RATING  ERROR  POINTS  PLAYED   (%)  CFS(%)    W    D    L  D(%)  ERET test
   1 Cfish-NNUE 071020_4CPU  :     7.1   21.5    51.0     100  51.0      74    2   98    0  98.0  96/111
   2 Stockfish_20092822_4CPU :     0.0   ----   520.0     900  57.8     100  151  738   11  82.0  95/111
   3 Lc0.net.JH.J92-280      :   -46.1   25.1    43.5     100  43.5      58    3   81   16  81.0  80/111    <= NEW
   4 Lc0.net.JH.J92-260      :   -49.7   25.6    43.0     100  43.0      58    1   84   15  84.0  81/111
   5 Lc0.net.JH.J92-210      :   -53.3   26.3    42.5     100  42.5      50    1   83   16  83.0  81/111
   6 Lc0.net.JH.J92-270      :   -53.3   25.7    42.5     100  42.5      77    1   83   16  83.0  81/111
   7 Lc0.net.JH.J92-220      :   -67.9   28.7    40.5     100  40.5      57    2   77   21  77.0  79/111
   8 Lc0.net.JH.J92-190      :   -71.5   29.4    40.0     100  40.0      57    1   78   21  78.0  79/111
   9 Lc0.net.JH.J92-240      :   -75.2   29.1    39.5     100  39.5      75    0   79   21  79.0  80/111
  10 Stockfish_11_4CPU       :   -90.1   32.4    37.5     100  37.5     ---    0   75   25  75.0  91/111

https://dropmefiles.com/aUSZe

[свернуть]

[Владимир001]

[Владимир001]

[Владимир001]

Championship Chess Match after 60/100 games

   # PLAYER               :  RATING  ERROR  POINTS  PLAYED   (%)  CFS(%)    W    D    L  D(%)
   1 Lc0-Sufi             :       3     10    30.5      60  50.8      72    2   57    1  95.0
   2 stockfish-Sufi 19    :      -3     10    29.5      60  49.2     ---    1   57    2  95.0

White advantage = 0.00 +/- 0.00
Draw rate (equal opponents) = 95.27 % +/- 3.34


Medians
Name                     Games    Depth       NPS           Nodes
-------------------------------------------------------------------------------------
stockfish-Sufi 19    :      60     54        34829194   3006149734
Lc0-Sufi             :      60     17           31537      7364415

[Владимир001]

Подача заявки на ICLR 2021 | «Лямбда-сети» обеспечивают точность SOTA, экономят большую память

https://syncedreview.com/2020/10/21/...ve-memory/amp/

[Владимир001]

[Владимир001]

Цитата: Сообщение от Владимир001

https://www.twitch.tv/VendableBee420

[Владимир001]

Цитата: Сообщение от Владимир001

Championship Chess Match after 60/100 games

   # PLAYER               :  RATING  ERROR  POINTS  PLAYED   (%)  CFS(%)    W    D    L  D(%)
   1 Lc0-Sufi             :       3     10    30.5      60  50.8      72    2   57    1  95.0
   2 stockfish-Sufi 19    :      -3     10    29.5      60  49.2     ---    1   57    2  95.0

White advantage = 0.00 +/- 0.00
Draw rate (equal opponents) = 95.27 % +/- 3.34


Medians
Name                     Games    Depth       NPS           Nodes
-------------------------------------------------------------------------------------
stockfish-Sufi 19    :      60     54        34829194   3006149734
Lc0-Sufi             :      60     17           31537      7364415

Championship Chess Match 74/100 games

   # PLAYER               :  RATING  ERROR  POINTS  PLAYED   (%)  CFS(%)    W    D    L  D(%)
   1 Lc0-Sufi             :       2      8    37.5      74  50.7      72    2   71    1  95.9
   2 stockfish-Sufi 19    :      -2      8    36.5      74  49.3     ---    1   71    2  95.9

White advantage = 0.00 +/- 0.00
Draw rate (equal opponents) = 96.17 % +/- 2.73

Medians
Name                     Games    Depth       NPS           Nodes
-------------------------------------------------------------------------------------
stockfish-Sufi 19    :      74     55        34623844   3018981086
Lc0-Sufi             :      74     17           31396      7528856

[Владимир001]

Развернуть для просмотра

Match: ERET test (Eigenmann Rapid Engine Test) with 111 test positions
time control: analysis with time limit of 60s per move/position
LC0-version: v0.21.2 up to v0.26.3
LC0-options: -Backend=cudnn -NNcache=3000000 -MaxCcollisionEvents=256
MLH-settings S19 SuFi: -MovesLeftMaxEffect=0.2 -MovesLeftSlope=0.004 -MovesLeftThreshold=0.0 -MovesQuadraticFactor=0.0 -MovesScaledFactor=1.0
AB engine options: threads=4, hash=4096 MB
hardware: mobile Geforce GTX 1050 TI 4GB, mobile i7-7700HQ 4x2.8GHz (8 threads with HT)
software: Arena GUI, "Automatic Analysis" feature with EPD file.
Tablebases: Syzygy 5-men on SSD
Results (solutions, percent):

A) CPU engines (mobile i7-7700HQ 4x2.8GHz):
Cfish 221020 NNUE 4CPU          :  97 = 87.4%   <== new, *net: nn-eba324f53044.nnue*
CorChess NNUE 1.3_201021 4CPU   :  97 = 87.4%   <== new, *net: nn-eba324f53044.nnue*
Stockfish 20102220 NNUE 4CPU    :  95 = 85.6%   <== new, *net: nn-eba324f53044.nnue*
Stockfish 20092822 NNUE 4CPU    :  95 = 85.6%   ** TCEC S19 Champion ** (net: nn-baeb9ef2d183.nnue)
Stockfish 12 4CPU (NNUE)        :  94 = 84.7%   *net: nn-82215d0fd0df.nnue*
BrainFish_200724_4CPU           :  92 = 82.9%   *last version before NNUE support is added*
Stockfish 20073110 4CPU         :  92 = 82.9%   *last version before NNUE support was added*
Stockfish 11 4CPU               :  91 = 82.9%
Stockfish 10 4CPU               :  90 = 81.1%
Komodo-13.2.5_4CPU              :  88 = 79.3%
Komodo 11.01 4CPU (freeware)    :  84 = 75.7%
Nemorino 6.00 NNUE 4CPU         :  77 = 69.4%   *net: nemorino_net_6.nnue*
SlowChess Classic 2.2 4CPU      :  77 = 69.4%
Xiphos 0.6 4CPU                 :  76 = 68.4%
Ethereal 12.75 pext-avx2 4CPU   :  73 = 65.8%
Igel 2.8.0 NNUE 4CPU            :  72 = 64.9%   *default net (night nurse v0.2.5b)*
Laser 1.7 4CPU                  :  72 = 64.9%
Komodo-MCTS-13.2.5_4CPU         :  62 = 55.9%
Lc0.net.703350-blas_4CPU        :  60 = 54.1%
Lc0.net.591003-blas_kiudee_4CPU :  60 = 54.1%
Fire 7.1 4CPU                   :  60 = 54.1%
Lc0.net.11258-112x9se_blas_4CPU :  59 = 53.2%
Lc0.net.DK.Bad_Gyal_7_blas_4CPU :  57 = 51.4%
Lc0.net.LD2_blas_4CPU  (10x128) :  56 = 50.4%

B) GPU engines (mobile Geforce GTX 1050 TI 4GB):
Lc0.net.65162  (24x320)         :  83 = 74.8%   *current record for T60 nets*
Lc0-0.22.0-dev.Leelenstein 14.3 :  83 = 74.8%
Lc0.net.JH.J92-270  (30x384)    :  81 = 73.0%   *record for 384x30 nets*
Lc0.net.JH.J92-260  (30x384)    :  81 = 73.0%   *record for 384x30 nets*
Lc0.net.JH.J92-210  (30x384)    :  81 = 73.0%   *record for 384x30 nets*
Lc0.net.JH.J92-280  (30x384)    :  80 = 72.1%
Lc0.net.JH.J92-240  (30x384)    :  80 = 72.1%
Lc0.net.65500 (24x320)          :  80 = 72.1%   <== new
Lc0.net.JH.J92-250  (30x384)    :  79 = 71.2%
Lc0.net.JH.J92-230  (30x384)    :  79 = 71.2%
Lc0.net.JH.J92-220  (30x384)    :  79 = 71.2%
Lc0.net.JH.J92-205  (30x384)    :  79 = 71.2%
Lc0.net.JH.J92-190  (30x384)    :  79 = 71.2%   ** TCEC S19 SuFi net **
Lc0.net.JH.J92-180  (30x384)    :  78 = 70.3%
Lc0.net.64460                   :  77 = 69.4%
Allie-v0.7.Leelenstein 14.3     :  77 = 69.4%
Lc0.net.JH.J92-290  (30x384)    :  76 = 68.4%   <== new
Lc0.net.65411                   :  76 = 68.4%
Lc0.net.65297                   :  76 = 68.4%
Lc0.net.42810                   :  76 = 68.4%
Lc0.net.JH.J92-100  (30x384)    :  75 = 67.6%   ** TCEC S19 DivP net **
Lc0.net.SV.384x30-T60-4619      :  75 = 67.6%
Lc0.net.SV.384x30-T60-4229      :  75 = 67.6%   ** TCEC Cup 6 net **
Lc0.net.JH.T40B.4-160           :  75 = 67.6%   ** S16 DivP net **
Lc0.net.JH.T40B.4-260           :  75 = 67.6%
Lc0.net.703810  (10x128)        :  75 = 67.6%
Lc0.net.JH.J13B.2-136  (24x320) :  74 = 66.7%   *Terminator 2 net*
Lc0.net.SV.256x20-T40-1541      :  74 = 66.7%
Lc0.net.591003_kiudee  (10x128) :  74 = 66.7%
Lc0.net.42850                   :  73 = 65.8%
Lc0.net.11248                   :  69 = 62.2%
Lc0.net.LD2          (10x128)   :  67 = 60.4%
Lc0.net.32930                   :  66 = 59.4%
Lc0.net.DK.Bad_Gyal_9  (10x128) :  65 = 58.6%

[свернуть]

Comment: EPD file:
https://glarean-magazin.ch/wp-conten...gine-Chess.zip

[Владимир001]

jhorthos:

Развернуть для просмотра

Match: T73 vs T70 best network
Lc0-version: 26.1
Lc0 options: FpuStrategy="reduction" FpuValue=0.23 FpuStrategyAtRoot="absolute" FpuValueAtRoot=1.0 CPuct=1.3 CPuctAtRoot=1.9 PolicyTemperature=1.4 MovesLeftMaxEffect=0.2 MovesLeftThreshold=0.0 MovesLeftSlope=0.0007 MovesLeftQuadraticFactor=0.85 MovesLeftScaledFactor=0.15 MovesLeftConstantFactor=0.0 MoveOverheadMs=0
Time/Node control: 2+0.04s
Hardware: RTX 2080, AMD 2950X
Book: book-ply12-unifen-Q-0.0-0.4.pgn, random with reverse.
Tablebases: 6-man Syzygy
Adjudication: 6-man Syzygy and -draw movenumber=50 movecount=6 score=10 -resign movecount=5 score=500 twosided=true
Chess Interface: cutechess-cli, time forfeit disabled
Comments: Ongoing test to see if the added blocks in T73 (128x14b) improve or degrade performance relative to T70 (128x10b). [Not properly comparable, but FYI T70 gained 127 self-Elo after the current T73 LR (0.004).]
Results:

# PLAYER : RATING ERROR POINTS PLAYED (%) W L D D(%) CFS(%)
1 lc0.net.703810 : 0.0 ---- 11033.0 18000 61 5384 1318 11298 63 100
2 lc0.net.730218.tc2 : -58.5 8.6 836.5 2000 42 174 501 1325 66 61
3 lc0.net.730248.tc2 : -60.3 8.9 831.5 2000 42 168 505 1327 66 98 << latest
4 lc0.net.730196.tc2 : -74.3 9.2 793.5 2000 40 160 573 1267 63 82
5 lc0.net.730173.tc2 : -80.4 9.4 777.0 2000 39 161 607 1232 62 51
6 lc0.net.730153.tc2 : -80.6 9.0 776.5 2000 39 155 602 1243 62 98
7 lc0.net.730131.tc2 : -91.9 6.5 1492.5 4000 37 258 1273 2469 62 61
8 lc0.net.730112.tc2 : -93.5 9.6 742.0 2000 37 113 629 1258 63 91
9 lc0.net.730096.tc2 : -102.9 9.4 717.5 2000 36 129 694 1177 59 ---

(games too big to attach)

[свернуть]

[Alvir]

Встречаем следующую сеть J92 384x30b от jhorthos, J92-290 :
https://storage.lczero.org/files/net...ontrib/J92-290

[Владимир001]

Match: PR963 (betamcts) vs baseline
LC0 version (both): deeabd7
LC0 options (both): threads=1, minibatch-size=192, cpuct=1.32, cpuct-at-root=1.9, cpuct-factor=2.815, cpuct-factor-at-root=2.815, fpu-value=0.23, fpu-strategy-at-root=absolute, policy-softmax-temp=1.4
LC0 options (betamcts): betamcts-level=2, betamcts-trust=4.2, betamcts-prior=0, lcb-percentile=0.2, betamcts-update-interval=10, april-factor=0, april-factor-parent=0
LC0 options (baseline): betamcts-level=0, betamcts-trust=0, betamcts-prior=0, lcb-percentile=0, betamcts-update-interval=10, april-factor=0, april-factor-parent=0
MLH options (both): moves-left-max-effect=0.2, moves-left-threshold=0, moves-left-slope=0.007, moves-left-scaled-factor=0.15, moves-left-quadratic-factor=0.85
Time control: 800 nodes/move
Net: 703810
Book: custom-3k (1-3 ply)
Hardware: gtx970 + i5-4570
Tablebases: syzygy 5-man
Adjudication: -draw movenumber=50 movecount=5 score=10 -resign movecount=5 score=700 twosided=true
Software: cutechess-cli
Comment: best result so far at fixed npm similar to training environment @Naphthalin

# PLAYER : RATING ERROR POINTS PLAYED (%) CFS(%) W D L D(%)
1 lc0-703810-betamcts : 26.9 9.6 969.0 1800 54 100 386 1166 248 65
2 lc0-703810 : 0.0 ---- 831.0 1800 46 --- 248 1166 386 65

White advantage = -2.35 +/- 5.01
Draw rate (equal opponents) = 65.33 % +/- 1.09

https://pixeldrain.com/u/mNv2iULm

[Владимир001]

Цитата: Сообщение от Alvir

Встречаем следующую сеть J92 384x30b от jhorthos, J92-290 :
https://storage.lczero.org/files/net...ontrib/J92-290

https://youtu.be/cqPSHH-YWMk

[Alvir]

Win10 x64, Ryzen 3700x, RTX 2060super, BanksiaGui, book-TCEC19_NoomenSelect.pgn, option default, 128Mb, 1'+1".

Result:
-----------------------------------------------------------------------------
  #  name             games    wins   draws  losses  score%    los%  elo+/-
  1. lc0-v0.26.3-290    100      16      71      13    51.5    71.1    10.4
  2. lc0-v0.26.3-190    100      13      71      16    48.5    28.9   -10.4

Cross table:
-----------------------------------------------------------------------------
  #  name               score%   games         1         2
  1. lc0-v0.26.3-290      51.5     100         x      51.5
  2. lc0-v0.26.3-190      48.5     100      48.5         x

Tech:
-----------------------------------------------------------------------------

Tech (average nodes, depths, time/m per move, others per game), counted for computing moves only, ignored moves with zero nodes:
  #  name               nodes/m         NPS  depth/m   time/m    moves     time
  1. lc0-v0.26.3-290        56K       30981     10.5      1.8     65.8    118.6
  2. lc0-v0.26.3-190        63K       34644     10.4      1.8     65.7    118.8
     all ---                57K       32814     10.4      1.8     65.8    118.7

Возможно новая сеть получше сети, игравшей в SuFi 19.

[svoitsl]

Цитата: Сообщение от Alvir

Возможно новая сеть получше сети, игравшей в SuFi 19.

Это их уже для TCEC 20 тренируют?
Вряд ли до конца года удасться выжать из них хотя бы 20 Эло.

[Владимир001]

Развернуть для просмотра

Match: combined results (314 positions) of ERET test + Vesely 203 positions testset
time control: analysis with time limit of 60s per move/position
AB engine options: threads=4, hash=4096 MB
Lc0-version: v0.26.2 + v0.26.3
LC0-options: -Backend=cudnn -NNcache=3000000 -MaxCcollisionEvents=256
MLH-settings S19 SuFi: -MovesLeftMaxEffect=0.2 -MovesLeftSlope=0.004 -MovesLeftThreshold=0.0 -MovesQuadraticFactor=0.0 -MovesScaledFactor=1.0
hardware: notebook with mobile i7-7700HQ 4x2.8GHz, mobile Geforce GTX 1050 TI 4GB
software: Arena GUI, "Automatic Analysis" feature with EPD files.
Tablebases: Syzygy 5-men on SSD
Comments: For J92 so far those results fit nicely to the results of the test gauntet against SF Dev at 5min+5s TC (290 not tested there yet).

A) CPU engines (mobile i7-7700HQ 4x2.8GHz):
Cfish 221020 NNUE 4CPU          :  280/314 = 89.2%   <== NEW, *net: nn-eba324f53044.nnue*
CorChess NNUE 1.3_211020 4CPU   :  279/314 = 88.9%   <== NEW, *net: nn-eba324f53044.nnue*

B) GPU engines (mobile Geforce GTX 1050 TI 4GB):
Lc0.net.JH.J92-280  (30x384)    :  220/314 = 70.0%
Lc0.net.JH.J92-260  (30x384)    :  218/314 = 69.4%
Lc0.net.JH.J92-270  (30x384)    :  217/314 = 69.1%
Lc0.net.JH.J92-290  (30x384)    :  216/314 = 68.8%   <== NEW
Lc0.net.JH.J92-210  (30x384)    :  216/314 = 68.8%
Lc0.net.JH.J92-240  (30x384)    :  215/314 = 68.5%
Lc0-0.22.0-dev.Leelenstein 14.3 :  215/314 = 68.5%
Lc0.net.JH.J92-190  (30x384)    :  214/314 = 68.2%   ** TCEC S19 SuFi net **
Lc0.net.65536                   :  213/314 = 67.8%   <== NEW
Lc0.net.65162                   :  213/314 = 67.8%
Lc0.net.65411                   :  205/314 = 65.3%

[свернуть]

[Владимир001]

Развернуть для просмотра

Match: J92-290 vs stockfish 12 in fixed nodes
LC0 version: 0.26.2, built with cuda11 + cudnn7
LC0 options: cudnn-fp16, 1 thread, default for the rest
SF options: 1 thread, 1GB Hash.
Time control: Leela: 1 kn/move, SF 12: 400Kn/move
Hardware: RTX 2070, i7-7700
Book: Balsa 2724 openings
Tablebases: 6-man TB for both.
Adjudication: 6-man TB, -draw movenumber=50 movecount=5 score=8 -resign movecount=5 score=1000
Software:cutechess-cli
Comments: J92-290 got -2 elo vs SF 12.

Leela: 1 thread, 1Kn/move
SF 12: 1 thread, 400Kn/move

Balsa 2724

   # PLAYER                :  RATING  ERROR   POINTS  PLAYED   (%)  CFS(%)     W      D     L
-  1 Stockfish_12.400k     :       0   ----  24875.5   49032  50.7      57  9722  30307  9003
   2 lc0.net.J92-270.1k    :      -1      6   2720.0    5448  49.9      67  1063   3314  1071
+  3 lc0.net.J92-290.1k    :      -2      6   2705.5    5448  49.7      53  1012   3387  1049
   4 lc0.net.J92-260.1k    :      -3      6   2703.0    5448  49.6      55  1004   3398  1046
   5 lc0.net.J92-280.1k    :      -3      6   2699.5    5448  49.6      54  1045   3309  1094
   6 lc0.net.J92-190.1k    :      -4      6   2696.0    5448  49.5      54   996   3400  1052
   7 lc0.net.J92-250.1k    :      -4      6   2693.0    5448  49.4      61  1015   3356  1077
   8 lc0.net.J92-230.1k    :      -5      6   2684.0    5448  49.3      81   980   3408  1060
   9 lc0.net.J92-220.1k    :      -9      6   2656.0    5448  48.8      96   978   3356  1114
  10 lc0.net.J92-100.1k    :     -16      6   2599.5    5448  47.7     ---   910   3379  1159

[свернуть]

https://pixeldrain.com/u/C46LJNKD

[Владимир001]

Развернуть для просмотра

[свернуть]

[Владимир001]

Match Gauntlet
LC0-version: lc0-v0.26.3 cudnn-version
LC0 options: Standard option except Threads=1, NNCacheSize=0, MinibatchSize=1.
Time control: 1 Node/move
Hardware: i7-4790 with 16 GByte RAM + GTX1650
Speed: Not important for this type of testing.
Book: Chad-opening book with 100 000 different openings with the most popular human openings.
Book options: Opening order: Sequential. Play each opening twice. Play all opening in the book both for White and Black.
Tablebases: No Tablebases.
Adjudication: Only 5-men syzygy
Software: CuteChess 1.2.0 https://github.com/cutechess/cutechess/releases / Graphic tools is programmed by glbchess64, in Python programming tools.
Comment: This is result out of Ordo 1.2.6 https://github.com/michiguel/Ordo/releases/tag/v1.2.6
Ordo string: ordo.exe -Q -N 0 -D -a 0 -A "J92-145" -W -n4 -s500 -G -V -U "0,1,2,3,4,5,6,7,8,9,10" -j stats.txt -o rating.txt -p games.pgn
Results: pgn-file: https://ufile.io/7peqd2vr
Explanation: Trying to find the best net in clean policy Elo.

   # PLAYER     :  RATING  ERROR    POINTS  PLAYED   (%)  CFS(%)      W      D      L  D(%)
   1 J92-270    :      17      1  104779.0  200000  52.4     100  68788  71982  59230  36.0
   2 J92-145    :       0      1   95221.0  200000  47.6     ---  59230  71982  68788  36.0

White advantage = 26.34 +/- 0.57
Draw rate (equal opponents) = 36.22 % +/- 0.10

My list until now (Elo against J92-145) 1 Node/move:

White advantage = 23.98 +/- 0.57
Draw rate (equal opponents) = 36.51 % +/- 0.10
J92-270:             17     Elo    (Test with 200k games)
J92-280:             11     Elo    (Test with 200k games)
J92-260:             11     Elo    (Test with 200k games)
J92-240:              8     Elo    (Test with 200k games)
J92-210:              5     Elo    (Test with 200k games)
J92-220:              3     Elo    (Test with 200k games)
J92-205:              3     Elo    (Test with 200k games)
J92-230:              2     Elo    (Test with 200k games)
J92-190:              1     Elo    (Test with 200k games)
J92-145:              0     Elo    (Reference)
J92-100:             -2     Elo    (Test with 200k games)

[Alvir]

Цитата: Сообщение от Владимир001

Match Gauntlet
LC0-version: lc0-v0.26.3 cudnn-version
LC0 options: Standard option except Threads=1, NNCacheSize=0, MinibatchSize=1.
Time control: 1 Node/move

Никак не пойму, кому нужны эти тесты с 1 Node/move, может кто-нибудь рассказать, для чего это делается? Или просто, ради фана, -бестолковая трата времени? Эти результаты не имеют ничего общего с тестами с нормальными контролями или как? Похоже на выбор автомобиля по повороту ключа зажигания...

[svoitsl]

Цитата: Сообщение от Alvir

кому нужны эти тесты с 1 Node/move

"если звезды зажигают — значит — это кому-нибудь нужно? "
Вы где нибудь видели доказательство того,что при одном и том же движке и при сети одинаковой размерности более сильный движок при скорости N был бы хуже при N+1? Я нет.
на практике такое может быть, но в теории при большом числе игр в это сложно поверить

[Владимир001]

Производительность Leela на различных конфигурациях GPU/CPU:

https://docs.google.com/spreadsheets...it?usp=sharing

[Владимир001]

Развернуть для просмотра

Tune of: BetaMctsTrust, BetaMctsPrior
LC0 version (both): deeabd7
LC0 options (both): threads=1, minibatch-size=192, cpuct=1.32, cpuct-at-root=1.9, cpuct-factor=2.815, cpuct-factor-at-root=2.815, fpu-value=0.23, fpu-strategy-at-root=absolute, policy-softmax-temp=1.4, april-factor=0, april-factor-parent=0, moves-left-max-effect=0.2, moves-left-threshold=0, moves-left-slope=0.004, moves-left-quadratic-factor=0, moves-left-scaled-factor=1
LC0 options (engine1): betamcts-level=2, lcb-percentile=0.2, betamcts-update-interval=10
LC0 options (engine2): betamcts-level=0, betamcts-trust=0, betamcts-prior=0, lcb-percentile=0, betamcts-update-interval=10
Tuning ranges: BetaMctsTrust: Real(0.0, 20), BetaMctsPrior: Real(0.0, 0.1)
Tuning configuration: acq_function: vr/ei, rounds: 10
Time control: 800 nodes/move
Net: 703810
Book: custom-3k (1-3 ply)
Hardware: gtx970 + i5-4570
Tablebases: syzygy 5-man
Adjudication: -draw movenumber=50 movecount=5 score=10 -resign movecount=5 score=1000
Software: chess-tuning-tools
Comment: tuned for training

Optimums found:
BetaMctsTrust 4.2
BetaMctsPrior 0.04

90.0% confidence intervals of the parameters:
Parameter Lower bound Upper bound
---------------------------------------
BetaMctsTrust 0.1 16.7
BetaMctsPrior 0.001 0.095

[свернуть]