Lc0 NN329 against Houdini 4 in 2 LTC games

[yanquis1972]

what are CLOP settings? is that an automated optimzation tool?

[Laskos]

what are CLOP settings? is that an automated optimzation tool?

Yes, I saw them somewhere for recent nets. But I don't trust so much them, first they are derived from self-games, second, one is ought to use short time control for optimization (many games). The issues of Lc0 NN-MCTS engine against a standard AB engine with time control are very intricate. If we had no standard engines to compare with, maybe Lc0 would be more "understandable" as strength goes comparing just to humans. It is more "comparable" strength-wise to humans than to standard engines.

[yanquis1972]

ah, thanks. if i'm thinking of the same thing i've been using those settings as a baseline for awhile. i'm trying your settings atm. even with the slightly negative FPU leela seems to have her mind made up within a second most of the time.

[jp]

What is the default setting re. opening book?

I'm using Sedat's 2017 .abk book.
It's around 7-8 Moves deep.

Thanks for that info.

[whereagles]

From LC0 games at TCEC vs. these ones just reported, I seem to see an enormous difference between running leela on GPU vs CPU, about ~400 ELO.

Is this right, or am I just late to the party?

[shrapnel]

Is this right, or am I just late to the party?

You are right AND also Late to the party .
Maybe not that late, as it looks like it will still be a while before it can match upto the latest Stockfish/Komodo.

[shrapnel]

Scale thinking time=2.87
Cpuct MCTS option=9
First Play Urgency Reduction=-0.25

GREAT Settings, Mr Laskos, but only with 20x256 Leela !
Using these Settings and 20x256 weight file, lco beat its old nemesis Komodo 11.2.2 quite impressively in a 10'+10" Game.
Btw, I noticed that this new Network weight file slows down even my Asus 1080 Ti quite a bit.
I know its only one Game, but its all I have Time for right now....
https://lichess.org/g2ebFdCr
More impressively, Analysis shows that lc0 made 0 inaccuracies, mistakes and blunders.
20x256 Leela...getting somewhere now....

[Laskos]

Scale thinking time=2.87
Cpuct MCTS option=9
First Play Urgency Reduction=-0.25

GREAT Settings, Mr Laskos, but only with 20x256 Leela !
Using these Settings and 20x256 weight file, lco beat its old nemesis Komodo 11.2.2 quite impressively in a 10'+10" Game.
Btw, I noticed that this new Network weight file slows down even my Asus 1080 Ti quite a bit.
I know its only one Game, but its all I have Time for right now....
https://lichess.org/g2ebFdCr
More impressively, Analysis shows that lc0 made 0 inaccuracies, mistakes and blunders.
20x256 Leela...getting somewhere now....

I had a win over Komodo 10.2 of the 192x15 net too, but Komodo was on one core. How many cores Komodo used in your case? Btw, your game finishes a bit too soon, Lc0 is winning, but it's not 100% it would have converted, endgames are not her strong point.

[shrapnel]

. How many cores Komodo used in your case? Btw, your game finishes a bit too soon, Lc0 is winning, but it's not 100% it would have converted, endgames are not her strong point.

6 Cores. I will put Komodo on 8 Cores and play again later.
Yes, I know Score looks odd in the lichess analysis, only +1, but while playing in Arena GUI, Score was being seen as +3 and as I adjusted Arena Settings for win at +3, Arena GUI adjudicated it as a Leela CUDA win.
I'll change Arena GUI settings too.