Lc0 ratings

By Guenter Stertenbrink Date 2019-03-12 09:14 Edited 2019-03-12 09:46

hier mal meine Zusammenstellung von aktuellen Lc0 ratings , die ich fand

-------------------------------------------------------------------------------
3463(678) Stockfish 10 64-bit 4CPU , CCRL 40/40
3022(458) , Lc0 0.20.1 w36089 64-bit , CCRL 40/40
2842(24) , Lc0 0.20.2 w32930 64-bit , CCRL 40/40
2856(128) , Lc0 0.19.1 w11248 64-bit , CCRL 40/40
2939(142) , Lc0 0.20.1 w11258 64-bit , CCRL 40/40
-----------------------------------------------------------------------------------
3550(1215) , Stockfish 10 64-bit 4CPU , CCRL 40/4
2798(982) , Lc0 0.20.1 w36089 64-bit , CCRL 40/4

http://ccrl.chessdom.com/ccrl/4040/cgi/compare_engines.cgi?family=Leela%20Chess&print=Rating+list&print=Results+table&print=LOS+table&print=Ponder+hit+table&print=Eval+difference+table&print=Comopp+gamenum+table&print=Overlap+table&print=Score+with+common+opponents

http://ccrl.chessdom.com/ccrl/404/cgi/compare_engines.cgi?family=Leela%20Chess&print=Rating+list&print=Results+table&print=LOS+table&print=Ponder+hit+table&print=Eval+difference+table&print=Comopp+gamenum+table&print=Overlap+table&print=Score+with+common+opponents

-------------------------------------------------------------
3555(1300) , Stockfish 10.0 x64 4CPU
3403(1700) , LCZero 0.20.2 Cuda (32930) , CEGT 4/40
2779(1200) , LCZero 0.9 (229) GTX1060 t2 , CEGT 4/40
---------------------------------------------------------------
3422(1861) , Stockfish 10.0 x64 1CPU
2868(1208) , Lc0 0.20.1 11258-112x9 1CPU , CEGT 40/20
http://www.cegt.net/40_40%20Rating%20List/40_40%20All%20Versions/rangliste.html
-----------------------------------------------------
3448(17750) , Stockfish 10 64 , fastgm 60+0.6
2494(17750) , Lc0 CPU 11258-112x9-se , fastgm 60+0.6
------------------------------------------------------
3197(6000) , Stockfish 10 , fastgm 10min
2847(2700) , Lc0 CPU 11258-112x9-se , fastgm 10min:
-------------------------------------------------------------------
3529(5000) , Stockfish 190227 bmi2 , spcc
3419(1200) , LF 0.3 32930/Houdini 6 , spcc
3318(1200) , Lc0 0.20.2 N:32930 , spcc
-----------------------------------------------------------------
3450(1319) , Stockfish 10 64 BMI2 , spcc long
3452(304) , LF 0.3 N:32930 A:Houdini 6 , spcc long
3393(308) , Lc0 v0.20.1 N:32890 , spcc long
3385(246) , Lc0 v0.20.1 N:32567 , spcc long
3384(178) , Lc0 v0.21.0 N:41390 , spcc long
3321(283) , Lc0 v0.20.2 N:40695 , spcc long
--------------------------------------------------------------------
3612(36) , Stockfish 181224 , TCEC14P
3519(36) , LCZero v20rc2-32194 , TCEC14P
---------------------------------------------------------
3588(100) , Stockfish 190203 , TCEC14S
3585(100) , LCZero v20.2-32930 , TCEC14S
-------------------------------------------------------------------
-------------------------------------
mtgostark , test30
3570 , Stockfish 10 4CPU
3448 , 35867
3448 , 36051
3443 , 32335
3423 , 32450
3418 , 11258
3411 , 11248
3400 , 32390
-----------------------------------------
mtgostark , test 40
3506 , LF 32585
3490 , Stockfish 8 4CPU
3436 , 32335
3435 , 32585
3432 , 41450
3430 , 32930
3429 , 41400
3414 , 41350
3411 , 11258

----------------------------------
Elo(best L-version)-Elo(best version) :
-441,-752,-152,-554,-954,-350,-110,+2,-93,-3,-122,-64

best distilled : -350 at fastgm 10min , so that's ~250 worse than undistilled
---------------------------------------

[ ... ipman]

By dkappe Date 2019-03-12 16:11

“Distilled” ist für CPU gedacht. Man kann nicht volle Netze auf gpu damit vergleichen.

By Guenter Stertenbrink Date 2019-03-12 17:57 Edited 2019-03-12 18:01

ist schon klar. Ich haette das vielleicht in 2 Rubriken aufspalten sollen

Cuda --> GPU (TCEC,CEGT 4/40,mtgostark,spcc)

auffallend der 600(!)Elozuwachs bei fastgm
bei Bedenkzeiterhoehung von 60+0,6 auf 600+6

By dkappe Date 2019-03-12 18:59

0,6 ist ziemlich kurz für cpu. Da würde ich ein Netz mit höheren nps wählen, etwa 80x7 oder 64x6. Bei den Netzen ist ein gewisses Minimum an nodes nötig.

By Guenter Stertenbrink Date 2019-03-13 07:29 Edited 2019-03-13 07:48

hier noch der ipman - link :

https://docs.google.com/spreadsheets/d/11vXZe4Dm7nIXWIYD9EoZguR4xOuRWUQHpzSfQW2HaMo/edit#gid=1168769812

3422(50) , Stockfish 10 x64 bmi2 , ipman 60+5
3443(300) , Lc0-cuda v21.0 - 32991 , ipman 60+5

3452(100) , Stockfish 10 x64 bmi2 , ipman 300+5
3459(100) , Lc0-cuda v20.0 - 32579 , ipman 300+5

> right now busy with PR788 + 41501 net

By Guenter Stertenbrink Date 2019-03-13 15:31

41505-41475=+29 (3000 games,2019/03/12)
41500-41400=+44 (3000 games,2019/03/12)
41490-41460=+10 (3000 games,2019/03/12)
41475-41375=+50 (3000 games,2019/03/11)
41475-41445=+30 (3000 games,2019/03/11)
41460-41430=+30 (3000 games,2019/03/10)
41450-41350=+51 (3000 games,2019/03/09)

By Guenter Stertenbrink Date 2019-03-18 13:14 Edited 2019-03-18 13:37

http://ccrl.chessdom.com/ccrl/404/
3551 , Stockfish 10 64-bit 4CPU
3523 , Houdini 6 64-bit 4CPU
3505 , Komodo 11.2 64-bit 4CPU
3458 , Lc0 0.20.2 w32930 GPU
3427 , Fire 7.1 64-bit 4CPU
3411 , Komodo 12.3 MCTS 64-bit 4CPU
3390 , Ethereal 11.25 64-bit 4CPU
3369 , Laser 1.7 64-bit 4CPU

Leela-Ratio = 1.65(1CPU) = 0.5 (4CPU) , also ~3490 mit LR=1 GTX1050

3430 war die Elo bei mtgostark , mit LR=0.58 (4kn/s) macht ca. 3460 fuer LR=1

mit 100kn/s und 40/40 (TCEC,CCCC) ist Lc0-32930 anscheinend ca. 80 Elo besser bei LR=1

By Lothar Jung Date 2019-03-18 14:02

Schau Dir das mal an:

https://groups.google.com/forum/m/#!topic/lczero/Hor6GU9-e4A

Die aktuellen T40-Netze scheinen recht gut zu skalieren.

Lothar