LC0 graph has gone impressive

[Dann Corbit]

https://lczero.org/

It seemed before that LC0 would top out and then fall down like Malthus' dismal theory or something.
But the current trajectory is very impressive indeed.
I know the 6200 figure does not mean 6200 Elo against SF and Komodo.
But I suspect that the current iteration of LC0 on dual 2080 cards would be very, very strong.
Maybe even give the original Alpha Zero a run for its money.

[Guenther]

https://lczero.org/

It seemed before that LC0 would top out and then fall down like Malthus' dismal theory or something.
But the current trajectory is very impressive indeed.
I know the 6200 figure does not mean 6200 Elo against SF and Komodo.
But I suspect that the current iteration of LC0 on dual 2080 cards would be very, very strong.
Maybe even give the original Alpha Zero a run for its money.

Obviously selfplay rating is completely meaningless meanwhile. Something has changed since previous NNs'.
The selfplay rating is calculated by matches, but if one looks at current 30 net matches they really play
more or less the same 3-4 French lines for often more than 25 plies and may be in 2-3% out of the total match games,
there are other openings.

I haven't followed what they did in the past weeks/months, but this was not the case e.g. for the last 10 nets.
Temperature decay is there for variation during the learning games, but during the rating matches there still should
be some more randomness which is far too low currently.
Right now we can only say it improves in French play. I wonder when the last 30 net was matched with below 95% French games?

[Guenther]

Ok my 'guts' numbers were a bit off, but the trend is obvious.
There is much less variety in matchplay than before.
The other trend is a shifting from 'not e4' towards 'e4' openings (currently almost French).

Thanks to the LC0 team for the match games download (sql file ~ 2GB extracted by now)
http://lczero.org/old_networks/matches.sql.gz

Code: Select all

Match-ID:5256   NN31478 vs. NN31477
---------------------------------------------
Count   Moves           Opening           %
389     1.e4 e6         French          80,87
9       1.e4 c5         Sicilian         1,87
21      1.e4 e5         KPs              4,37
3       1.c4            English          0,62
4       1.Nf3           Reti             0,83
3       1.d4            QP               0,62
13      1.b3/b4/e3/g3                    2,70
37      not 1.e4                         7,69
---------------------------------------------
481     total                          100,00

Code: Select all

Match-ID:5220   NN31442 vs. NN31441
---------------------------------------------
Count   Moves           Opening           %
345     1.e4 e6         French          72,18
4       1.e4 c5         Sicilian         0,84
75      1.e4 e5         KPs             15,69
5       1.c4            English          1,05
3       1.Nf3           Reti             0,63
2       1.d4            QP               0,42
7       1.b3/b4/e3/g3                    1,46
28      not 1.e4                         5,86
---------------------------------------------
478     total                          100,00

Code: Select all

Match-ID:4940   NN31174 vs. NN31173
---------------------------------------------
Count   Moves           Opening           %
339     1.e4 e6         French          73,70
32      1.e4 c5         Sicilian         6,96
9       1.e4 e5         KPs              1,96
5       1.c4            English          1,09
6       1.Nf3           Reti             1,30
8       1.d4            QP               1,74
9       1.b3/b4/e3/g3                    1,96
40      not 1.e4                         8,70
---------------------------------------------
460     total                          100,00

Code: Select all

Match-ID:2235   NN20590 vs. NN20589
---------------------------------------------
Count   Moves           Opening           %
8       1.e4 e6         French           1,59
38      1.e4 c5         Sicilian         7,54
6       1.e4 e5         KPs              1,19
46      1.c4            English          9,13
37      1.Nf3           Reti             7,34
69      1.d4            QP              13,69
71      1.b3/b4/e3/g3                   14,09
398     not 1.e4                        78,97
---------------------------------------------
504     total                          100,00

Code: Select all

Match-ID:1513   NN11198 vs. NN11197
---------------------------------------------
Count   Moves           Opening           %
0       1.e4 e6         French           0,00
37      1.e4 c5         Sicilian         7,96
0       1.e4 e5         KPs              0,00
67      1.c4            English         14,41
145     1.Nf3           Reti            31,18
48      1.d4            QP              10,32
123     1.b3/b4/e3/g3                   26,45
423     not 1.e4                        90,97
---------------------------------------------
465     total                          100,00

[Damir]

Hi Guenther

Sorry for asking such a stupid question, but is it possible to view this match as PGN ?

[Guenther]

Hi Guenther

Sorry for asking such a stupid question, but is it possible to view this match as PGN ?

Could you precise your question? (Thousands of such matches have been done already and my previous post showed stats about five of them)
I guess you know, that you can view single games per match id online? (select a match id then select a game id with a result)
http://lczero.org/matches/

[Damir]

Sorry I meant in the link you posted. http://lczero.org/old_networks/matches.sql.gz
Is it possible to view this file in PGN format in Chessbase GUI ?

[Guenther]

Sorry I meant in the link you posted. http://lczero.org/old_networks/matches.sql.gz
Is it possible to view this file in PGN format in Chessbase GUI ?

No, but the moves etc., could be extracted and pgn files could be reconstructed, if desired.

[Damir]

This chesstempo website is crap. It wants you to register and than pay for a membership in order to see the games or positions...

[Damir]

Leela is advertising that website. It seems can not see the games or position whitout registering it, and than have to enter the ID to see a certain game or position.. as usual with limitation if are not a payed member. grrrr

[Damir]

Sorry I meant in the link you posted. http://lczero.org/old_networks/matches.sql.gz
Is it possible to view this file in PGN format in Chessbase GUI ?

No, but the moves etc., could be extracted and pgn files could be reconstructed, if desired.

How can I do that ?