Разные матчи и турниры движков

[Владимир001]

2m+1s 2 core frıtz17 gui syzgyz 3-4-5 i7 3820 4ghz play 144 games

CEVDET-PC, Blitz 2.0min+1.0sec 0


1 CF EXT 291120 x64 SSE41 NB +12 +9/=131/-4 51.74% 74.5/144
2 CF EXT 291120 x64 SSE41 NA -12 +4/=131/-9 48.26% 69.5/144

1 CF EXT 291120 x64 SSE41 NB MANTIS-XO.bin
2 CF EXT 291120 x64 SSE41 NA book2.bin

games
https://pixeldrain.com/u/bUTnThEw

[Владимир001]

SF14 vs Glaurung, 10k games.
[Match "Stockfish 14 vs Glaurung; TC=180+1.80; startpos; no EGTB; single-core; results = WDL=+9999=1-0"]
Dual Xeon Platinum 8176. 56-cores/112T.
command is:

./c-chess-cli -each tc=180+1.8 option.Threads=1 option.Hash=16 -engine cmd=./stockfish_14_x64_bmi2 "name=Stockfish 14" -engine cmd=glaurung "name=Glaurung" -games 10000 -concurrency 50 -repeat -draw number=40 count=8 score=10 -pgn out.pgn

+1720 ELO rating of SF14 vs Glaurung, according to ELO calculator:
https://www.3dkingdoms.com/chess/elo.htm
PGN:
https://www.dropbox.com/s/miu675o5om...g-out.pgn?dl=0

[Владимир001]

CCWC 2020 Quarter Finals Match 4 BrainLearn vs Lc0 Jthorthos RESULT = BrainLearn Win - W=18 L=14 D=68

https://pixeldrain.com/u/TqDpido1

[Владимир001]

S21 Суфийская партия 92, что, если черные сыграют 19 .. h4 ?:
Stockfish 14_202107131735 - Церера 0.92, TC 3h + 15s, закончилась ничьей, как и 30min + 5s.
Это Stockfish с более 1,5-кратным TC и должен иметь в среднем больше узлов / ходов с менее быстрыми потоками в аналогичной ситуации, чем в реальной игре.

https://pixeldrain.com/u/mpXi9kC9

[Владимир001]

Результаты матча 5 полуфинала CCWC 2020 = Поздравляем Stockfish с выходом в финал CCWC

Победитель матча 1 Cfish
13 Dev + nn-9e3c6298299a.nnue + Cerebellum3Merge
против
Победитель матча 2 Stockfish
15 разработчиков 310721 + nn-56a5f1c4173a.nnue

РЕЗУЛЬТАТ = Выигрыш Stockfish - W = 27 L = 4 D = 69

https://pixeldrain.com/u/EJghramF

[Владимир001]

TCEC S21 Sufi game 34 black move 34, playout by Stockfish 14_202107131735 - Ceres 0.91d_69626 15min+1s 1/2-1/2 in kibitz machine

https://pixeldrain.com/u/W7kyZ6fg

[Владимир001]

CCWC 2021 Quarter Finals Match

https://pixeldrain.com/l/fiAjiitv#item=0

[Владимир001]

Цитата:

Хорошо, просто из любопытства, вот игра, в которой я тестирую свой Stockfish, белый, против Stockfish 14, мой Stockfish когда-либо так немного изменял код, так что он использовал чистый NNUE eval, и когда-либо так слегка изменял поиск, чтобы исключить некоторые настройки, которые мне показались ненаучными, и т. д., мой Stockfish мог использовать более новый NNUE, оба движка использовали архитектуру HalfKAv2 512x2-16-32-1 NNUE, игра закончилась ничьей

http://www.jurjevic.org.uk/chess/gam...kfish_001.html

--

[Владимир001]

SF14 vs Glaurung, 1000 игр. TC = 10 мин. + 6 с.
---
Цель этого нового теста:
* Выясните, сможет ли Глаурунг защитить линию с более высоким контролем времени.
* Многопоточность должна привнести больше случайности в игры. Больше разнообразия.
* увеличен хэш с 16 МБ до 512 МБ для увеличения мощности обоих движков.
---
CLI:
root @ rampage-104: ~ # ./cutechess-cli -each tc = 600 + 6 proto = uci option.Threads = 2 option.Hash = 512 -engine cmd = / root / stockfish_14_x64_bmi2 "name = Stockfish 14" -engine cmd = $ (which glaurung) "name = Glaurung" -games 1000 -concurrency 56 -pgnout out-sf14-glaurung-thread2.pgn
---
Результаты:
Полное уничтожение Глаурунга без всякой надежды на защиту.
Завершенная партия 1000 (Глаурунг - Стокфиш 14): 0: 1
Счет Stockfish 14 vs Glaurung: 1000 - 0 - 0 [1.000] 1000
ELO: + бесконечность.
Глаурунг не смог защитить себя и был полностью уничтожен. Больше времени на часах не помогло ему выдержать линию.
PGN грязный (из-за ранних попыток тестирования; поэтому учитываются только последние 1000 игр). PGN:

https://pixeldrain.com/u/eiCWmzst

[Владимир001]

Многие спрашивали, сильнее ли Stockfish 14, чем Alpha-Zero? Наконец-то я дал ответ: да, и вот почему:
Игра-симулятор A0 с точно такими же условиями, установленными Google DeepMind Alpha Zero против Stockfish 8 (в качестве нашего базового «якоря»), но на этот раз против SF8 играет SF14 вместо A0.
TC = 60 секунд / ход, CPU = Dual Xeon Platinum 8176 (56 ядер / 112 потоков), мне требуется 12 дней, чтобы запустить это моделирование! Никакой книги (что означает «startpos») и 1 Гб хеша (как в оригинальном тесте, проведенном командой AlphaZero).
Завершенная партия 100 (Stockfish 8 vs Stockfish 14): 1 / 2-1 / 2 {3-кратное повторение}
Счет Stockfish 14 vs Stockfish 8: 48 - 0 - 52 [0,740] 100
---
AlphaZero против Stockfish 8:
как белый: + 25-0 = 25,
как черный: +3-0 = 47.
A0 против SF8 было +100 ELO
https://www.3dkingdoms.com/chess/elo.htm
---
Stockfish 14 против Stockfish 8:
как белый: + 42-0 = 8,
черными: +6-0 = 44.
SF14 против SF8 +182; Так что Stockfish14 на +82 ELO сильнее AlphaZero !!!
SF14 смог достичь одной огромной цели, которая сделала Alpha-Zero и стала большим шоком для меня лично, а именно сыграть 100 игр против SF8 с НУЛЕВЫМИ поражениями, ура !!!
---
CLI:
root @ rampage-102: ~ / c-chess-cli # ./c-chess-cli -each tc = 60 + 60 option.Threads = 112 option.Hash = 1024 -engine cmd =. / stockfish_14_x64_bmi2 "name = Stockfish 14 "-engine cmd =. / stockfish_8_x64_bmi2" name = Stockfish 8 "-games 100 -concurrency 1 -repeat -pgn sf14-a0-simul.pgn
---
PGN:
https://www.dropbox.com/s/3qa3jsnl3x...simul.pgn?dl=0

[Владимир001]

Хеш 1 ГБ (большой) против 16 МБ (стандартный):
---
TC = 20 минут + 10 секунд, book = bjbraams_chessdb_1207. Количество потоков ЦП = 112; Двойной Xeon Platinum 8176.
---
CLI:
./cutechess-cli -each tc = 1200 + 10 proto = uci option.Threads = 112 -engine cmd = / root / stockfish_14_x64_bmi2 "name = Stockfish 14-1G-hash" option.Hash = 1024 -engine cmd = / root / stockfish_14_x64_bmi2 "name = Stockfish-14-16MB-hash" option.Hash = 16 -games 50 -concurrency 1 -repeat -openings file =. / bjbraams_chessdb_1207_lines.pgn format = pgn order = random plies = 16 -pgnout out-sf14-1g -hash-test-book-run01.pgn
---
Завершенная игра 50 (Stockfish-14-16MB-hash vs Stockfish 14-1G-hash): 1 / 2-1 / 2 {3-кратное повторение}
Оценка Stockfish 14-1G-hash vs Stockfish-14-16MB-hash: 10 - 3 - 37 [0,570] 50
---
PGN = https://www.dropbox.com/s/goi3synhzr...run01.pgn?dl=0
---
+48 ЭЛО

[Владимир001]

На fastgm Руби обновился, и на этот раз SF не нужно было защищать проваленный дебют. В топ-10 SF только 1 проигрыш: http://www.fastgm.de/60min.html

[Владимир001]

https://pixeldrain.com/l/MfvugBh4#item=0
SF 14 VS SF 13

[Владимир001]

https://pixeldrain.com/u/pS1LHQHU

[Viktor]

Цитата: Сообщение от Владимир001

Многие спрашивали, сильнее ли Stockfish 14, чем Alpha-Zero?

Еще на Immortal-е помню проводил тест с последней версией Стока на тот период. Это, видимо, был Сток 9 или даже с аброка 8-я. Так вот она тогда показала такой же результат с 8 Стоком!
Еще говорили, почему Гугл взял старую версию? Да потому и взял, чтобы обыграть.

[Владимир001]

http://talkchess.com/forum3/viewtopi...rt=230#p901299

[Владимир001]

https://pixeldrain.com/l/N1x3mR5j#item=0

[Владимир001]

DarkToga 1.1 wins JCER - Arena Amateur Engines Tournament, 2021.08.04 - 2021.08.10

Temp 3'+3" Lenovo Ideapad 330 i3-8130U 12GB GUI Arena 3.5.1
Book Perfect 2021 All JCER games 01.08.2021 Table: Scid vs PC
JCER Test Tournament: 132 games

[Владимир001]

lc0game.pgn

https://pixeldrain.com/u/rXHBbJDX

[Владимир001]

https://pixeldrain.com/l/S1qLcnD8#item=0

[Владимир001]

Ceres.pgn

https://pixeldrain.com/u/uYozf1Vn

[Владимир001]

Gauntlet: sf14 vs smooth & alphazero
LC0 version (both): pr1596 (efe6bf23)
LC0 options (smooth): threads=1, minibatch-size=48, multi-gather=false, time-manager=smooth(midpoint=45.2,steepness=5.93,init-tree-reuse=0.52,max-tree-reuse=0.73,tree-reuse-update-rate=3.39,nps-update-period=20.0,init-timeuse=0.7,min-timeuse=0.34,timeuse-update-rate=5.51,max-move-budget=0.42,init-piggybank=0.09,per-move-piggybank=0.12,max-piggybank-use=0.94,max-piggybank-moves=36.5)
LC0 options (alphazero): threads=1, minibatch-size=48, multigather=false, time-manager=alphazero(alphazero-time-pct=5.0)
SF options: defaults
LC0 Time control: 5m+5s
SF Time control: 1m+1s
LC0 Net: 744501
SF Net: nn-9e3c6298299a
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:

   # PLAYER                :  RATING  ERROR  POINTS  PLAYED   (%)  CFS(%)    W    D    L  D(%)
   1 sf14                  :     0.0   ----    56.0     100    56      97   14   84    2    84
   2 lc0-smooth            :   -35.2   38.0    22.5      50    45      70    0   45    5    90
   3 lc0-alphazero_5pct    :   -49.4   37.9    21.5      50    43     ---    2   39    9    78

[Владимир001]

Match: Gauntlet of Lc0 SF 14 1CPU (single core) vs Lc0 69471/69525/69626/69680
Lc0-version: v0.28.0-rc1
Lc0-options : -Backend=cudnn -Threads=1 -NNCacheSize=1000000 -MiniBatchSize=256
SF-version: Stockfish 14 x64 bmi2 (compile by @Ipmanchess)
SF-options: threads=1, hash=256 MB, NNUE=true, net=nn-3475407dc199.nnue
Time control: 2min + 1s
Hardware: notebook with mobile i7-7700HQ 4x2.8GHz (8 threads with HT on), mobile Geforce GTX 1050 TI 4GB
Book: Balsa_v500.pgn, 10 plies, sequentially, colors reversed
Tablebases: Syzygy 5-men on SSD
Speed: Average around 4k nodes/move for T60 ; T60: 1.23k nps, SF 14 1CPU: 1097k nps => Leela Ratio (1:875) = 0.97
Adjudication: Draw at 50 moves rule, 3-fold repetition, after 200 moves, at 40 moves with eval = 0.00 ; 5-men Gaviota tablebase adjudication
Software: Arena 3.5.1, ponder = off
Elo rating: Ordo 1.2.6
Comments: TC as in CCRL Blitz.

   # PLAYER            :  RATING  ERROR  POINTS  PLAYED   (%)  CFS(%)    W    D    L  D(%)
   1 Stockfish_14_1CPU :     0.0   ----   249.5     400  62.4     100  105  289    6  72.3
   2 Lc0.net.69680     :   -82.4   33.7    38.5     100  38.5      62    1   75   24  75.0
   3 Lc0.net.69471     :   -89.9   34.6    37.5     100  37.5      50    3   69   28  69.0
   4 Lc0.net.69626     :   -89.9   34.5    37.5     100  37.5      56    0   75   25  75.0
   5 Lc0.net.69525     :   -93.7   34.9    37.0     100  37.0     ---    2   70   28  70.0

https://pixeldrain.com/u/956sauGJ

[Владимир001]

Match: sf14 vs lc0 tm configs

Развернуть для просмотра

LC0 version (smooth_v1): pr1596 (efe6bf23)
LC0 version (smooth_v2): pr1596 (efe6bf23)
LC0 version (alphazero_v1): pr1596 (efe6bf23)
LC0 version (alphazero_v2): pr1596 (efe6bf23)
LC0 version (hyperpruning_zz): pr1601 (90a534f0)
LC0 options (all): threads=1, minibatch-size=48, multi-gather=false
LC0 options (smooth_v1): smart-pruning-factor=1.33, time-manager=smooth(midpoint=45.2,steepness=5.93,init-tree-reuse=0.52,max-tree-reuse=0.73,tree-reuse-update-rate=3.39,nps-update-period=20.0,init-timeuse=0.7,min-timeuse=0.34,timeuse-update-rate=5.51,max-move-budget=0.42,init-piggybank=0.09,per-move-piggybank=0.12,max-piggybank-use=0.94,max-piggybank-moves=36.5)
LC0 options (smooth_v2): smart-pruning-factor=2.66, time-manager=smooth(midpoint=45.2,steepness=5.93,init-tree-reuse=0.52,max-tree-reuse=0.73,tree-reuse-update-rate=3.39,nps-update-period=20.0,init-timeuse=0.7,min-timeuse=0.34,timeuse-update-rate=5.51,max-move-budget=0.42,init-piggybank=0.09,per-move-piggybank=0.12,max-piggybank-use=0.94,max-piggybank-moves=36.5)
LC0 options (alphazero_v1): smart-pruning-factor=1.33, time-manager=alphazero(alphazero-time-pct=15.0)
LC0 options (alphazero_v2): smart-pruning-factor=2.66, time-manager=alphazero(alphazero-time-pct=15.0)
LC0 options (hyperpruning_zz): smart-pruning-factor=1.0, hyperpruning-threshold=15.32, time-manager=alphazero(alphazero-time-pct=21.4)
SF options: defaults
LC0 Time control: 18s+0.3s
SF Time control: 3.6s+0.06s
LC0 Net: 744501
SF Net: nn-9e3c6298299a
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: adding more soon

[свернуть]

   # PLAYER                 :  RATING  ERROR  POINTS  PLAYED   (%)  CFS(%)    W     D    L  D(%)
   1 lc0_smooth_v2          :    28.6   15.2   405.5     750    54      92  171   469  110    63
   2 lc0_alphazero_v2       :    13.1   14.6   389.0     750    52      55  159   460  131    61
   3 lc0_smooth_v1          :    11.7   14.9   387.5     750    52      94  152   471  127    63
   4 sf                     :     0.0   ----  1857.0    3750    50      79  674  2366  710    63
   5 lc0_alphazero_v1       :    -6.1   15.0   368.5     750    49      99  133   471  146    63
   6 lc0_hyperpruning_zz    :   -30.5   15.5   342.5     750    46     ---   95   495  160    66

https://pixeldrain.com/u/WV4Yj4PP

[Владимир001]

Match: sf14 vs lc0 legacy tm
LC0 version: pr1534 (c6562bfa)
LC0 options: threads=1, minibatch-size=48, multi-gather=false, smart-pruning-factor=2.648, smart-pruning-q-effect=0.5214, time-manager=legacy(midpoint-move=30.8205,steepness=7.1756,immediate-use=0.7637,book-ply-bonus=0.3299,slowmover=0.5024)
SF options: defaults
LC0 Time control: 18s+0.3s
SF Time control: 3.6s+0.06s
LC0 Net: 744501
SF Net: nn-9e3c6298299a
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:

   # PLAYER           :  RATING  ERROR  POINTS  PLAYED   (%)  CFS(%)    W    D    L  D(%)
   1 lc0_legacy_v1    :    31.9   14.6   409.0     750    55     100  168  482  100    64
   2 sf               :     0.0   ----   341.0     750    45     ---  100  482  168    64


TM Comparison:
   # PLAYER                 :  RATING  ERROR  POINTS  PLAYED   (%)  CFS(%)    W     D    L  D(%)
   1 lc0_legacy_v1          :    31.9   15.4   409.0     750    55      62  168   482  100    64
   2 lc0_smooth_v2          :    28.6   15.4   405.5     750    54      92  171   469  110    63
   3 lc0_alphazero_v2       :    13.1   14.7   389.0     750    52      55  159   460  131    61
   4 lc0_smooth_v1          :    11.7   15.3   387.5     750    52      93  152   471  127    63
   5 sf                     :     0.0   ----  2198.0    4500    49      79  774  2848  878    63
   6 lc0_alphazero_v1       :    -6.1   14.9   368.5     750    49      99  133   471  146    63
   7 lc0_hyperpruning_zz    :   -30.5   14.5   342.5     750    46     ---   95   495  160    66

[Владимир001]

Match: sf14 vs lc0 alphazero zz-ish tm
LC0 version: pr1534 (c6562bfa)
LC0 options: threads=1, minibatch-size=48, multi-gather=false, smart-pruning-factor=4.11, smart-pruning-q-effect=0.21, time-manager=alphazero(alphazero-time-pct=31.8)
SF options: defaults
LC0 Time control: 18s+0.3s
SF Time control: 3.6s+0.06s
LC0 Net: 744501
SF Net: nn-9e3c6298299a
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: zz tune still in progress but looks like these values will be close to final

   # PLAYER                  :  RATING  ERROR  POINTS  PLAYED   (%)  CFS(%)    W    D    L  D(%)
   1 lc0_alphazero_zz-ish    :     1.9   14.3   377.0     750    50      60  130  494  126    66
   2 sf                      :     0.0   ----   373.0     750    50     ---  126  494  130    66


TM Comparison:
   # PLAYER                      :  RATING  ERROR  POINTS  PLAYED   (%)  CFS(%)    W     D     L  D(%)
   1 lc0_legacy_v1               :    31.9   14.5   409.0     750    55      62  168   482   100    64
   2 lc0_smooth_v2               :    28.6   14.3   405.5     750    54      92  171   469   110    63
   3 lc0_alphazero_v2            :    13.1   15.4   389.0     750    52      55  159   460   131    61
   4 lc0_smooth_v1               :    11.7   14.8   387.5     750    52      82  152   471   127    63
   5 lc0_alphazero_zz-ish        :     1.9   15.3   377.0     750    50      59  130   494   126    66
   6 sf                          :     0.0   ----  2571.0    5250    49      77  900  3342  1008    64
   7 lc0_alphazero_v1            :    -6.1   15.9   368.5     750    49      99  133   471   146    63
   8 lc0_hyperpruning_zz         :   -30.5   15.1   342.5     750    46     ---   95   495   160    66

[Владимир001]

Match: sf14 vs lc0_smooth_v3
LC0 version: pr1605 (4343441d)
LC0 options: threads=1, minibatch-size=48, multi-gather=false, smart-pruning-factor=2.66, smart-pruning-q-effect=0.5, time-manager=smooth(midpoint=45.2,steepness=5.93,init-tree-reuse=0.52,max-tree-reuse=0.73,tree-reuse-update-rate=3.39,nps-update-period=20.0,init-timeuse=0.7,min-timeuse=0.34,timeuse-update-rate=5.51,max-move-budget=0.42,init-piggybank=0.09,per-move-piggybank=0.12,max-piggybank-use=0.94,max-piggybank-moves=36.5)
SF options: defaults
LC0 Time control: 18s+0.3s
SF Time control: 3.6s+0.06s
LC0 Net: 744501
SF Net: nn-9e3c6298299a
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: settings same as smooth_v2 except added smart pruning q effect

   # PLAYER           :  RATING  ERROR  POINTS  PLAYED   (%)  CFS(%)    W    D    L  D(%)
   1 lc0_smooth_v3    :     9.8   14.4   385.5     750    51      91  139  493  118    66
   2 sf               :     0.0   ----   364.5     750    49     ---  118  493  139

TM Comparison:
   # PLAYER                      :  RATING  ERROR  POINTS  PLAYED   (%)  CFS(%)     W     D     L  D(%)
   1 lc0_legacy_v1               :    31.9   15.3   409.0     750    55      62   168   482   100    64
   2 lc0_smooth_v2               :    28.6   15.4   405.5     750    54      92   171   469   110    63
   3 lc0_alphazero_v2            :    13.1   15.3   389.0     750    52      55   159   460   131    61
   4 lc0_smooth_v1               :    11.7   14.9   387.5     750    52      57   152   471   127    63
   5 lc0_smooth_v3               :     9.8   15.0   385.5     750    51      77   139   493   118    66
   6 lc0_alphazero_zz-ish        :     1.9   14.4   377.0     750    50      60   130   494   126    66
   7 sf                          :     0.0   ----  2935.5    6000    49      78  1018  3835  1147    64
   8 lc0_alphazero_v1            :    -6.1   15.2   368.5     750    49      99   133   471   146    63
   9 lc0_hyperpruning_zz         :   -30.5   14.6   342.5     750    46     ---    95   495   160    66

[Владимир001]

Match: (Gauntlet) Fixed node w/ SF13 vs. LCZero nets 69693/69680

Развернуть для просмотра

LC0 version: v0.28.0-rc1
SF version: SF13 (nn-62ef826d1a6d.nnue)
LC0 options: nodes=1500 Backend=cuda-fp16, NNCacheSize=40000, Minibatch-Size=32, CPuct=2.0521, CPuctBase=100000, CPuctFactor=0.872, FpuValue=0.1831, PolicyTemp=1.0322
MLH Options: MovesLeftMaxEffect=0.1, MovesLeftSlope=0.0002, MovesLeftScaledFactor=1.0
SF Options: Threads=4 , Hash=512 , nodes=800000
Hardware: CPU: i7-10700K , GPU: RTX 2080 Super
TC: Inf. ; Lc0 Speed: 1,500 n/move , SF Speed: 800,000 n/move (1:533)
Book: Openings-dd-16ply-15000.pgn (Chads 16-ply openings - random)
Tablebases: Syzygy 5-man
Adjudication: Draw movenumber=35 movecount=6 score=18 Resign movecount=5 score=600
Software: cutechess-cli , Ordo -q -D -W -s 2000 -a 0 -A "Stockfish_13" -p SF13vsLCZero0280.pgn -U "0,1,2,3,4,5,7,9,8,10,6"
Comment: 1,000 games each for LCZero nets 69693 & 69680 at 1,500 npm against Stockfish 13 at 800,000 npm ... I was going to just add all of these games with my games from the test above, but thatd give net 69680 2,000 games, while the other 2 nets sit at only 1,000. These openings were far more balanced. **

[свернуть]

  # PLAYER                      :  RATING  ERROR  POINTS  PLAYED   (%)    W    L     D  D(%)  CFS(%)
   1 Stockfish_13                :     0.0   ----  1065.0    2000    53  426  296  1278    64     100
   2 LCZero_v0.28.0_T60.69680    :   -21.5   12.7   470.0    1000    47  156  216   628    63      65
   3 LCZero_v0.28.0_T60.69693    :   -25.0   13.0   465.0    1000    47  140  210   650    65     ---

https://pixeldrain.com/u/mixYYZs8

[Владимир001]

Hardware: i5-8400 CPU @ 2.80GHz, 6 cores, 32GB, gpu RTX 2060, 6GB
Gauntlet: SF-dev vs. LC0 384x30b nets
Network IDs: J94-100, ..., 69146, ..., 69471, 69524, 69677
LC0-version: v0.27.0
LC0-options: minibatch-size=90
SF-version: SF-dev 20120516
SF-options: 4 Threads, 512MB Hash
Software: Cutechess-CLI
Time control: LC0: 1,500 nodes/move, SF-dev: 500,000 nodes/move
Book: Chad's 8-ply book with 10000 positions, order=sequential
Tablebases: 6-men
Adjudication: -draw movenumber=50 movecount=5 score=8 -resign movecount=5 score=1000
Ordo: -a 0 -A "Stockfish.dev" -W -D -U "0,1,2,3,4,5,6,7,8,9,10" -s 2000 -n 1
Comment: previous nets
https://forchess.ru/showpost.php?p=1...postcount=1570

   # PLAYER             :  RATING  ERROR    POINTS  PLAYED   (%)  CFS(%)       W       D       L  D(%)
+  1 lc0.net.69677      :    37.2    3.0   11031.0   20000    55      60    4893   12276    2831    61
   2 lc0.net.69524      :    36.6    3.0   11016.0   20000    55      61    4837   12358    2805    62
-  3 lc0.net.69146      :    36.0    3.0   10999.0   20000    55      53    4795   12408    2797    62
   4 lc0.net.69471      :    35.9    3.0   10994.5   20000    55      65    4806   12377    2817    62
     ...
- 22 lc0.net.J94-100    :    29.5    3.0   10817.5   20000    54      51    4562   12511    2927    63
     ...
- 41 Stockfish.dev      :     0.0   ----  366855.5  800000    46     ---  120310  493091  186599    62

White advantage = 55.85 +/- 0.23
Draw rate (equal opponents) = 64.41 % +/- 0.06

PGN:
https://pixeldrain.com/u/Kut1gzB1

Развернуть для просмотра

[свернуть]

[Владимир001]

Match: sf14 vs lc0 time management

Развернуть для просмотра

LC0 version: pr1605 (4343441d)
LC0 options (a0_15pct): threads=1, minibatch-size=48, multi-gather=false, smart-pruning-q-effect=0.0, time-manager=alphazero(alphazero-time-pct=15.0)
LC0 options (smooth_sufi): threads=1, minibatch-size=48, multi-gather=false, smart-pruning-q-effect=0.2, time-manager=smooth(midpoint=45.2,steepness=5.93,init-tree-reuse=0.52,max-tree-reuse=0.73,tree-reuse-update-rate=3.39,init-timeuse=0.7,min-timeuse=0.34,timeuse-update-rate=5.51,max-move-budget=0.42,init-piggybank=0.09,per-move-piggybank=0.12,max-piggybank-use=0.94,max-piggybank-moves=36.5)
LC0 options (smooth_default): threads=1, minibatch-size=48, multi-gather=false, smart-pruning-q-effect=0.2, time-manager=smooth
SF options: defaults
LC0 Time control: 60s+1s
SF Time control: 8.5s+0.14s
LC0 Net: 744501
SF Net: nn-9e3c6298299a
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: a0_15pct performing better than smooth_sufi here, while in previous tests at shorter time control a0_15pct performed worse than smooth_sufi. could it be that the a0 tm scales a bit better? ltc test by @Navs also shows a0 doing better (with high error bars)

[свернуть]

   # PLAYER                :  RATING  ERROR  POINTS  PLAYED   (%)  CFS(%)    W     D    L  D(%)
   1 lc0-a0_15pct          :    12.6   15.4   259.0     500    52      76   76   366   58    73
   2 lc0-smooth_sufi       :     4.9   15.6   253.5     500    51      73   72   363   65    73
   3 sf                    :     0.0   ----   743.5    1500    50      86  191  1105  204    74
   4 lc0-smooth_default    :    -8.4   15.5   244.0     500    49     ---   56   376   68    75