SF 11 vs T60 Lc0

[mwyoung]

ops i forgot i meant vs stockfish!,how many cores amd+sf to catch lco with 1 2080ti

Probably in the range 16-32 cores, with the best SV or LS nets and new Lc0 parameters.

Lc0's strength is not subject to the "new" parameters. They are only new to you. People that know things. Have known for a long time that the default settings are not the best. And don't use the default settings to test Lc0.

[dkappe]

Rather than trying to pick a pointless fight, the “new” parameters can be found here: http://talkchess.com/forum3/viewtopic.php?f=2&t=72854

Of course the people that “know”, know that a big grid search for parameters at different time controls is expensive, long winded and not fun. That’s why these “special settings” only come along every once in a while, probably aren’t best or even good at most time controls, and are passed around like an illicit drug or secret handshake.

[mwyoung]

Rather than trying to pick a pointless fight, the “new” parameters can be found here: http://talkchess.com/forum3/viewtopic.php?f=2&t=72854

Of course the people that “know”, know that a big grid search for parameters at different time controls is expensive, long winded and not fun. That’s why these “special settings” only come along every once in a while, probably aren’t best or even good at most time controls, and are passed around like an illicit drug or secret handshake.

There is no pointless fight. But do not act like people did not know before today. That the default settings are garbage. Many many people have know this before today. And have posted this here, and on other sites, and with better settings. And if you think you know things, and have no clue what you are talking about. Then you need to say so...and stop acting like the "new" settings is a sudden revelation. This denigrates the hard work as you have described. By many people that have put time into testing the neural networks correctly.

In Lasko's own words.

"In the past I tried to fiddle with Lc0 engine parameters, it was always within even 1 SD error margins, it is hard to have many games at acceptable TC with Leela. Besides that, I do not understand what many of these parameters do precisely, and just blindly trying here and there is useless."

Enough said....

[Ovyron]

Of course the people that “know”, know that a big grid search for parameters at different time controls is expensive, long winded and not fun.

Really? This marks the first time that people test chess engines without having fun in the process.That's as bad as playing chess without enjoying the game, just for some of the wins, where I'd give the advice to play some other game (that is enjoyed despite the losses); here I'd tell them to just switch to a hobby that is fun to them, and leave these parameter testing to people that enjoy the process. You don't want to wake up years later and wonder where did all the time go.

[Nay Lin Tun]

Hardware

https://ibb.co/Bz6RRFb

Do you know if Lc0 engine (0.23.2) overcame 70-80k nps threshold in the opening and midgame? Also, the strength scaling from 2 GPUs to 4 GPUs has improved? It seemed bad to me until months ago.

Typical speeds in the openings are of the order 100M nps for Stockfish and some 50k nps for T60 larger net?

Lc0 speed in opening and middle game is approx 50k ( between 30 k and 70k).

This larger net get benefits from 4 GPU whereas T40 cant use more than 3 ( practically 2, the third one has minimal benefit).

SF is around 100 to 110 Mnps.

[BrendanJNorman]

There is no pointless fight. But do not act like people did not know before today. That the default settings are garbage. Many many people have know this before today. And have posted this here, and on other sites, and with better settings. And if you think you know things, and have no clue what you are talking about. Then you need to say so...and stop acting like the "new" settings is a sudden revelation. This denigrates the hard work as you have described. By many people that have put time into testing the neural networks correctly.

Come on man, please learn how to structure a sentence.

You do not start a new sentence with "And".

You do not start a new sentence with "But".

Whether the point you make is right or wrong, people do judge you (and the quality of your arguments) by your level of literacy in your native tongue.

If you want to brush up your prose, I can recommend you start with the fantastic book Elements of Style.

https://www.amazon.com/Elements-Style-C ... BYCQ5MVDQA

[Laskos]

ops i forgot i meant vs stockfish!,how many cores amd+sf to catch lco with 1 2080ti

Probably in the range 16-32 cores, with the best SV or LS nets and new Lc0 parameters.

Lc0's strength is not subject to the "new" parameters. They are only new to you. People that know things. Have known for a long time that the default settings are not the best. And don't use the default settings to test Lc0.

I hope you are fine.

[Kiudee]

Rather than trying to pick a pointless fight, the “new” parameters can be found here: http://talkchess.com/forum3/viewtopic.php?f=2&t=72854

Of course the people that “know”, know that a big grid search for parameters at different time controls is expensive, long winded and not fun. That’s why these “special settings” only come along every once in a while, probably aren’t best or even good at most time controls, and are passed around like an illicit drug or secret handshake.

Tuning a set of parameters in a big grid search is very inefficient and I would agree that it’s not fun at all.
Still, any method used for tuning at some point runs into the curse of dimensionality - some sooner than others.
If you are fine with only finding a local optimum (maybe you did good work in restricting all the parameter ranges), then local search methods like CLOP can work, but usually the optimization landscapes become more and more chaotic, the more parameters you add.

Here I disagree with dkappe in that designing optimization techniques which work in "higher" dimensions with the signal-to-noise ratio of chess is certainly fun .

[pohl4711]

Tuning a set of parameters in a big grid search is very inefficient and I would agree that it’s not fun at all.
Still, any method used for tuning at some point runs into the curse of dimensionality - some sooner than others.
If you are fine with only finding a local optimum (maybe you did good work in restricting all the parameter ranges), then local search methods like CLOP can work, but usually the optimization landscapes become more and more chaotic, the more parameters you add.

Here I disagree with dkappe in that designing optimization techniques which work in "higher" dimensions with the signal-to-noise ratio of chess is certainly fun .

In the meantime, there a plenty of testing-results, that are showing a clear Elo-gain, when Lc0 plays with your Kiudee-setting. Congrats for this!
My comparable testrun of Lc0 0.23.1 LS 12.2 with Kiudee-setting (Lc0 default result is already in my longtime-testrun list(https://www.sp-cc.de/nn-longtime-testing.htm)) is running and will take 4-5 more days. But looks very promising at this early stage, too.

[Ovyron]

Here I disagree with dkappe in that designing optimization techniques which work in "higher" dimensions with the signal-to-noise ratio of chess is certainly fun .

And I'm glad to hear it! I appreciate your work