I saw some games.
In the third Leela refused to draw (repetition of moves) in a superior position and lost.
It seems to me that such a good result vs. komodo (8 +3 -6 =11) happened because Komdo 8 is a very sophisticated program in positional understanding.
In my opinion, you should test against Komodo and an equivalent version of Stockfish who has strong tactical understanding:
example
10 games against Komodo 8
10 games against Stockfish 6
and consider the overall output in order to estimate Elo of your settings.
Kanizsa wrote: ↑Wed Jun 06, 2018 4:01 pm
I saw some games.
In the third Leela refused to draw (repetition of moves) in a superior position and lost.
It seems to me that such a good result vs. komodo (8 +3 -6 =11) happened because Komdo 8 is a very sophisticated program in positional understanding.
In my opinion, you should test against Komodo and an equivalent version of Stockfish who has strong tactical understanding:
example
10 games against Komodo 8
10 games against Stockfish 6
and consider the overall output in order to estimate Elo of your settings.
Bye
It was Komodo 10.2, so, a pretty recent top engine.
Komodo is not a tactical slouch, and in the earlier matches, those ended 3.5/20, Komodo 10.2 was punishing Lc0 both positionally and tactically, especially in endgames. I might leave overnight a gauntlet of Lc0 against some Komodo and some Houdini Tactical, to see how it fares comparatively.
Well these new setting have created a monster. LCO now makes Hiarcs 14 look like a novice. In this game Hiarcs is crushed tactically and by positional play. My favourite game of their match.
Kanizsa wrote: ↑Wed Jun 06, 2018 4:01 pm
I saw some games.
In the third Leela refused to draw (repetition of moves) in a superior position and lost.
It seems to me that such a good result vs. komodo (8 +3 -6 =11) happened because Komdo 8 is a very sophisticated program in positional understanding.
In my opinion, you should test against Komodo and an equivalent version of Stockfish who has strong tactical understanding:
example
10 games against Komodo 8
10 games against Stockfish 6
and consider the overall output in order to estimate Elo of your settings.
Bye
It was Komodo 10.2, so, a pretty recent top engine.
Komodo is not a tactical slouch, and in the earlier matches, those ended 3.5/20, Komodo 10.2 was punishing Lc0 both positionally and tactically, especially in endgames. I might leave overnight a gauntlet of Lc0 against some Komodo and some Houdini Tactical, to see how it fares comparatively.
LC0 on GTX1060 6GB, standard engines on 1 CPU core.
Gauntlet 1' +1'' LC0 cuDNN NN373 against Komodo 8 (abot 3240 CCRl) and Houdini 4 Tactical (about 3200 CCRL), total 100 games:
Rank Name Elo +/- Games Score Draws
0 LC0_exp_NN373 -28 53 100 46.0% 40.0%
1 Komodo_8 42 73 50 56.0% 44.0%
2 Houdini_4 T 14 78 50 52.0% 36.0%
100 of 100 games finished.
Although Komodo 8 is much more positional than Houdini 4 Tactical, their performance against LC0 is according to their ratings, so, no there are no much variation in this. They both are anyway very much superior to LC0 in tactics, so small differences in style don't matter. Again, LC0 cuDNN NN373 in these conditions performs at about 3200 CCRl level in these conditions (1' +1'').
To me this is a convincing demonstration. You have rejected the hypothesis that LC0 plays well especially against Komodo because positionally strong and badly against Stockfish because tactically strong. These style differences certainly affect Elo, but appear numerically marginal and negligible
