Not logged inRybka Chess Community Forum
Up Topic Rybka Support & Discussion / Aquarium / Leela chess
- - By Boombaard (**) [nl] Date 2018-10-18 17:37
I wonder if it is possible to run Leela (Lc0) in Aquarium?
Parent - By Ghengis-Kann (***) [us] Date 2018-10-18 18:31
Should be no problem, but you might be better off asking for advice on the LCZero discord.

I just checked and there are currently 529 users online.
Parent - - By Dadi Jonsson (Silver) [is] Date 2018-10-18 20:36
It works like any other UCI compatible engine in Aquarium.
Parent - - By pawnslinger (****) [us] Date 2018-10-19 06:16
Indeed.  I have used it too.  Works fine.
Parent - - By Boombaard (**) [nl] Date 2018-10-19 08:02
In an IDEA project it is possible to run 2 enigines at the same time. Wouldn't it be great to use Stockfish and Leela to correct each others ideas and end up with the best of two worlds?
Parent - - By pawnslinger (****) [us] Date 2018-10-19 16:02
Way ahead of you.  I do this with multiple engines ALL the time.  Nothing new here.
Parent - - By Boombaard (**) [nl] Date 2018-10-19 19:15
But I can see you in the distance! ;-)
Some people say you get an unreliable mix of evaluations in 1 project. Normally I make 2 projects and manually change best lines between them. Or can that be done more efficient?
Parent - By pawnslinger (****) [us] Date 2018-10-21 18:04
Well, yes, but the trick is this... save your best engine for the IDeA evaluations.... let's say Stockfish.  It is free and probably one of the best, if not THE best engine at this time.  Now use different engines to feed IA into the IDeA queue.  In this way, you can use lc0 to feed positions to SF for its considerations.  All in the one project.  I actually like to use several engines for this... in some of my old posts you can see what my engine line-up is.  Recently I have added 2 new ones to the list: Ethereal and Pedone.  Makes about a dozen engines that I currently use to feed into SF.  Most are older engines that are no longer in favor - I like quirky moves to consider <grin>.
Parent - - By ChiefPushesWood (**) [us] Date 2018-10-19 18:10
'Slinger, what card are you using? I don't have an Nvidia card at the moment so I'm definitely looking at one for this computer or my next build.

CPW
Parent - - By pawnslinger (****) [us] Date 2018-10-19 19:12
I am not using a good card for lc0.  I am using a GTX 1060 3gb model.  My wife gave it to me for Christmas last year.  The rule for lc0 is more CUDA cores, more memory, the better.  Let your budget be your guide.  I think they were using 2x GTX 1080 Ti cards in the TCEC trournament... the TCEC Cup is underway now and lc0 has a match coming up with Stockfish --- it will be the one to watch.  But I do not know what hardware lc0 is using.  I do know it plays a LOT better with many CUDA cores and video ram.

On my dinky card it doesn't play up to SF's level, but my card is really puny compared to what they run it on in tournys.
Parent - - By Ghengis-Kann (***) [us] Date 2018-10-19 22:35
LC0 is a fascinating experiment to me because I have a theory about why human players hit a rating plateau that, if correct, will cause LC0 to plateau as well.

I believe our intuition is an accumulated memory of which moves worked well in certain types of positions and which did not.
These moves can come from our own games or from study, but in either case a person's intuition "hardens" after awhile in a way that makes them tend to find certain types of moves and miss others.

I won a game awhile back where my opponent could have crushed me by putting his queen where my pawn could take it. Neither one of us saw it at the time, but the computer found it instantly because it does not have an intuition.

The way LC0 is training itself seems strongly analogous to my hypothesis about how humans acquire intuition, so I predict it will hit a rating plateau caused by missing certain types of moves that are either usually bad, but good in a particular position, or usually good but bad in a particular position.
Parent - - By Carl Bicknell (*****) [gb] Date 2018-10-20 17:42

> The way LC0 is training itself seems strongly analogous to my hypothesis about how humans acquire intuition, so I predict it will hit a rating plateau caused by missing certain types of moves that are either usually bad, but good in a particular position, or usually good but bad in a particular position.


The big difference being that Lc0 isn't JUST an evaluation function, it also does roll outs of games before deciding on a move. Humans don't do this. When you are playing human-human chess, everything happens in your head and then you play the move on the board. You don't get to try 20,000 games with that move to see if it's any good.
So, in many ways, humans may have an EVALUATION FUNCTION similar to LC0, but their search is closer to ALPHA BETA engines, like Stockfish.

We just don't know how good Lc0 will get.
Parent - - By cma6 (****) Date 2018-10-26 04:04
Carl,
  If Lc0 is quite a bit weaker than SF, and if in i.a. one can get to depth of only 10-11 ply with Lc0 vs. 40 ply on SF, what is the point of the Lc0 exercise in competitive chess?
Parent - By pawnslinger (****) [us] Date 2018-10-27 19:25
The problem is that lc0 depth is not the same as depth in an A/B engine.  Especially in view of the Monte Carlo methods often used in these types of programs.  In other words, they are actually looking much deeper into the position than it reports... but how to judge the equivalency???  It is hard to do, and lc0 is definitely reporting a depth much lower than its true effective depth.
Parent - - By Carl Bicknell (*****) [gb] Date 2018-10-27 20:48

> Carl,
>   If Lc0 is quite a bit weaker than SF, and if in i.a. one can get to depth of only 10-11 ply with Lc0 vs. 40 ply on SF, what is the point of the Lc0 exercise in competitive chess?


So there are two big differences between Lc0 and SF. First, the search is different, Lc0 does some kind of self-play game roll-out behind the scenes and SF does a classical search. So "depth 10" will mean very different things to both engines.

Secondly, the evaluation function is very different. SF has a good, relatively light weight, set of bonuses and penalties for all kinds of positional factors (knight in the corner = penalty; big pawn centre = bonus). Lc0 has learnt its own and we don't really know what it is. But the big thing is: it's self learning and will get better.

SF has improved faster than any other classical engine I know of in recent years, but Lc0 has improved much faster even than Stockfish. So the real issue is: maybe in 2-3 months Lc0 will catch up SF...we don't know. But we have to try to find out.
Parent - By cma6 (****) Date 2018-10-28 02:29
Thanks, Carl, for providing the big picture.
Parent - - By pawnslinger (****) [us] Date 2018-10-20 18:22 Edited 2018-10-21 08:39
If you dig into the lc0 forum, you will see that THIS does happen.  lc0 routinely plateaus, and then the "teachers" fret over what to do next to take lc0 to the next level.

It is fascinating to read thru that forum.  To see how they "bake bread".

lc0 forum - https://groups.google.com/forum/#!forum/lczero

I am a "lurker" there, don't know enough to even understand half of what they talk about.  But it is interesting anyway.
Parent - By Ghengis-Kann (***) [us] Date 2018-10-22 17:46
The rollouts are something I would like to understand better.

Suppose one possible next move leads to 200,000 rollouts where  199,999 are wins and one that loses by force.
It seems to me if the move is rejected then we are back to alpha/beta search, and if it is not then it is a blind spot that will be exploited by traditional engines.

To make it more difficult, what if the forced loss wasn't within the horizon, like the move leads to an exchange down ending that is a forced loss after 80 more moves?
Parent - - By cma6 (****) Date 2018-10-26 04:01
Pawnslinger,
What is your source for the latest LC0 engine and what parameters do you set when adding it into Aquarium?
Or is this idea a waste of time--to use LC0 engine for i.a. analysis--unless one has a decent video card in the system?
Parent - - By pawnslinger (****) [us] Date 2018-10-27 19:33 Edited 2018-10-27 19:35
I have been using this url: https://github.com/LeelaChessZero/lc0/wiki/Getting-Started

I tried using a development page to download from, but I stopped using it because it was too buggy (not quite the right word, but kind of).  I no longer have that url.  Perhaps someone else has it.  I think it is in another thread here on this board (someone gave it to me here).

Personally, I don't think lc0 is quite ready for prime time yet... and I certainly need a better graphics card.  From what I understand, the graphics card power is VERY important.  When the card was wonky during the TCEC tournament, lc0 had a spate of bad games until it was fixed.  So a very big deal.

And I think they were using dual GT 1080 Ti for lc0 in that event.
Parent - By cma6 (****) Date 2018-10-28 02:32
Pawnslinger:
  Thanks much for the link and recap of your experience with lc0 so far.
- - By zaarcis (***) [lv] Date 2018-10-23 12:36
I tried to make the aqt tree for Lc0 to see its evaluations in another tree, but it doesn't work. (But with Stockfish I don't have such problem.)
How to do that? (Thanks.)
Parent - - By Ghengis-Kann (***) [us] Date 2018-10-23 16:25 Upvotes 1
Hello Zaacris.

I have not tried Leela specifically, but my general method is to assign the IA tree in the engine properties section (change it from default to something specific) and then add that tree to a tree configuration in the Sandbox.
Parent - - By zaarcis (***) [lv] Date 2018-10-23 16:31
Hmm, then I should retry this process. I did it in the same way (I believe) twice, without success for now.
Parent - - By zaarcis (***) [lv] Date 2018-10-23 16:37 Edited 2018-10-23 16:42
One hypothesis is that Lc0 reaches even depth = 10 too slowly, and Aquarium is still waiting for results of some reasonable depth (and then I stop IA disappointed). :D
Not sure if it's correct, of course.
Parent - By pawnslinger (****) [us] Date 2018-10-23 17:24 Upvotes 1
Well, yes, you must adjust your depth parameters when using lc0.  The depth reported by lc0 is much too shallow to be using the same values as you would for an A/B engine like Stockfish.
Parent - By zaarcis (***) [lv] Date 2018-10-24 09:04
Yes, that was it. If I wait until depth 10 or 11, then the results show up.
Up Topic Rybka Support & Discussion / Aquarium / Leela chess

Powered by mwForum 2.27.4 © 1999-2012 Markus Wichitill