Lc0 NN329 against Houdini 4 in 2 LTC games

[Lion]

Perfect, thanks!

[shrapnel]

Well, this is embarrassing !
Komodo 11.2.2 x64 with Black, beat lc0 within 30 Moves
All Komodo settings were at default.
Well, I suppose its back to the Drawing board to try and find the best settings for lc0.
I've noticed that lco almost never utilizes the full 60 seconds when it is its turn to move.
It didn't seem to matter so much against Stockfish, but here this flaw is perhaps making it pay a heavy price.
lco seems very uncomfortable against Komodo.
Odd.
Anyway, here's the Game.
https://lichess.org/F5qUw9Z2

[Milos]

Well, this is embarrassing !
Komodo 11.2.2 x64 with Black, beat lc0 within 30 Moves
All Komodo settings were at default.
Well, I suppose its back to the Drawing board to try and find the best settings for lc0.
I've noticed that lco almost never utilizes the full 60 seconds when it is its turn to move.
It didn't seem to matter so much against Stockfish, but here this flaw is perhaps making it pay a heavy price.
lco seems very uncomfortable against Komodo.
Odd.
Anyway, here's the Game.
https://lichess.org/F5qUw9Z2

If you indeed gave all 8 cores of your CPU to Komodo it would be at least 400 Elo stronger than Lc0-cudnn on your precious 1080Ti even with best possible settings. So being killed in 30 moves while playing white is nothing out of ordinary.
If you were more informed you'd know what is the actual strength on Lc0 on particular hardware and wouldn't daydream of "putting Komodo, SF and H to dust" .

[Lion]

Well, this is embarrassing !
Komodo 11.2.2 x64 with Black, beat lc0 within 30 Moves
All Komodo settings were at default.
Well, I suppose its back to the Drawing board to try and find the best settings for lc0.
I've noticed that lco almost never utilizes the full 60 seconds when it is its turn to move.
It didn't seem to matter so much against Stockfish, but here this flaw is perhaps making it pay a heavy price.
lco seems very uncomfortable against Komodo.
Odd.
Anyway, here's the Game.
https://lichess.org/F5qUw9Z2

I am not surprised.
I had several years ago had my own computer play on playchess.com the same variation with black and had many wins in a similar way.
So I am not surprised that LCO with no knowledge of the opening/variation falls the same way top engines loosed in the same opening/variation.

[Laskos]

With my more tactical settings for Lc0, which seemed to perform badly in short time control games, I played one game at 10 minutes + 10 seconds increment of Lc0 NN335 on GTX 1060 as White against latest SF dev on 4 i7 cores as Black, from standard opening position. It was a draw in Queen's Indian Defence.
https://lichess.org/3w2TWU3Z

Maybe, after all, the feeling that improving tactics by fiddling with parameters of Lc0 will be beneficial at LTC is not that far off. And this game was not even that long time control. Only one game played, will maybe play to four games later.

[jp]

Well, this is embarrassing !
Komodo 11.2.2 x64 with Black, beat lc0 within 30 Moves
All Komodo settings were at default.

What is the default setting re. opening book?

[yanquis1972]

With my more tactical settings for Lc0, which seemed to perform badly in short time control games, I played one game at 10 minutes + 10 seconds increment of Lc0 NN335 on GTX 1060 as White against latest SF dev on 4 i7 cores as Black, from standard opening position. It was a draw in Queen's Indian Defence.
https://lichess.org/3w2TWU3Z

Maybe, after all, the feeling that improving tactics by fiddling with parameters of Lc0 will be beneficial at LTC is not that far off. And this game was not even that long time control. Only one game played, will maybe play to four games later.

unless they're quite a radical departure from what i've seen suggested, or the TC is classical, i doubt it. messing around yesterdat with the WAC revised 200, i couldn't improve on a limited set that give leela the most difficulty, going from 5s to 60s+ with various extremes. imo an essential step in her core evolution simply hasn't taken place yet. but i'd love to be wrong!

[shrapnel]

What is the default setting re. opening book?

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

[shrapnel]

With my more tactical settings for Lc0...

What Settings are you using, Mr Laskos ?

[Laskos]

With my more tactical settings for Lc0...

What Settings are you using, Mr Laskos ?

I tried my extreme settings, which are performing much better in tactical test-suites, but worse in positional ones. They are:

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

All games are from the initial standard opening position. These settings achieved one draw and 3 losses against SF dev on 4 cores at 10'+ 10''. Then I started to use CLOP settings and default settings, the results were similar, totally in 16 games at 10'+ 10'' with NN325, 6 draws and 10 losses against SF dev on 4 cores. I am unable to discern which settings are the best at this TC. The estimated rating of SF dev on 4 cores is 3600 CCRL 40/4' Elo. That of Lc0 NN335 on GTX 1060 about 3300 in these conditions. I will maybe wait until bigger nets (256x20) will start to train on new self-games, hope they will improve quickly, at least at their beginning. My impression is that 192x15 architecture was not fully exploited, they seem to run into over-fitting and some specific issues I don't understand.

Here is a game where Lc0 NN335 held a draw as Black from standard opening position:
https://lichess.org/XfEJUjWe