Not logged inRybka Chess Community Forum
Up Topic Rybka Support & Discussion / Rybka Discussion / Rybka 2.3.2a and Triple Brain ?
- - By Zherkovv (***) Date 2007-12-21 09:23
Long ago one of the features of the "Classic Shredder 6"
interface was the "Triple Brain" analysis features.
So two engines analyse a position and a third engine called
the "Triple Brain" acts as arbiter and uses the analysis of the
other two engines, to determine what line of play is best in that position.
My question (request);

Is "Triple Brain" significant with "Rybka 2.3.2a" and for instance "Zap!Chess"
or any other top engine as the two engines?

Zherkovv
Parent - - By noctiferus (***) Date 2007-12-21 09:31
It's still available under the recent shredder guis
Parent - - By Zherkovv (***) Date 2007-12-21 11:43
Where can i find "Triple Brain" under Shredder 11 GUI ?
Parent - By noctiferus (***) Date 2007-12-21 13:18
here
Parent - - By Uly (Gold) Date 2007-12-21 14:53
But I think that the other 2 engines steal 66% efficiency from the main engine, so you're analyzing the position with a 33% strength Rybka, and it's not cost effective.

Also, when the other two engines agree that Rybka is wrong, but actually Rybka is right (and it's expected to happen often due to the differences in strength) you have a problem.
Parent - - By Uri Blass (*****) Date 2007-12-22 10:36
tripple brain is for 2 engines and not for 3 engines.

I think that this idea can be productive if it is used correctly but the question is if the program that decide which engine to use is smart enough.

Suppose that you have a machine that support smp.
Take 2 different engines with similiar playing strength that do not support smp(for example Delfi5.1 and Slow Chess Blitz WV2.1)

The question is if tripple brain between them is going to have higher rating in ponder off games(note that in this case there is no less of efficiency because the programs can use different processors).

I think that it should give some improvement if the program that choose between delfi and slow chess blitz is a good program.

Uri
Parent - - By Uly (Gold) Date 2007-12-22 15:31

> I think that it should give some improvement if the program that choose between delfi and slow chess blitz is a good program.


What would this third program need to do to be good?
Parent - By Uri Blass (*****) Date 2007-12-22 16:04
The third program need to know to use the pv and the score of both programs to decide which program to choose.

example 1:
If program A shows a positive score that goes down every iteration  when program B show stable score of 0.00 then choose program B because program A seems not to understand the position(justified positive score should improve when you search deeper).

example 2:
If program B show
Move 1 after 5 seconds,Move 2 after 20 seconds,Move 3 after 1 minute when
program A show similiar pv with
Move 1 after 2 seconds.Move 2 after 10 seconds,Move 3 after 30 seconds,Move 4 after 1 minute then
It may be logical to believe that program A simply search deeper than program B in the relevant position and to choose the move of program A.

The third program also need to have smart time management(because the engines are practically used at long time control or at infinite analysis so they are not going to stop by themselves except maybe the case of single legal move) and when one program report fail low and the second program suggest the same move the third program should give more time for both engines and when both engines agree about the best move and use almost no time for the rest of the moves the third program can play faster(playing immediatly when there is a single legal move is obvious).

Uri
Parent - - By richbell (**) Date 2007-12-21 14:54
No, I remember Vas NOT being too positive about this approach in one of the old threads and i guess he is right. Triple brain idea sounds like 'Too many cooks spoil the broth' isn't?
Parent - By Uly (Gold) Date 2007-12-21 15:07
Another way to do it, (if you want to use three engines to select your moves) is to alternate the engines' moves, so if you're using Rybka, Shredder and Zappa, once out of book, the first move will be done with Rybka, then the next with Shredder, the next with Zappa, and then back to Rybka.

One side effect is that hash would be basically negated. Another is that pondering is now impossible. A third one is that since Rybka is stronger than the other two, using Rybka alone will get better results. And lastly, the engines' styles will collide and it could actually get ugly (So, using any engine alone may give better results, but it's unknown since it's expected that Engine X + Rybka would play better than Enigne X alone).

But since it's cost effective (Each engine is playing at 100% on its move) it must still be better than Triple Brain (And you're creating a new, very strong chess playing entity that plays unlike any other, that makes it interesting.)
Parent - By Roland Rösler (****) Date 2007-12-21 15:32
Maybe it´s a question of # cores you have in your machine. If you have less than 8, I think no. With 8 cores or better 16, there is a big chance. But I think, Stefan don´t change the "arbiter" since Shredder 6, although here is big potential for advancement.
Parent - By premraj_n (**) Date 2007-12-22 04:15
Triple brain can be meaning full when all 3 engines have same knowledge i.e. Rybka multiple threads which is nothing but Rybka MP.
I say this because engines have specific style and more than that in chess there are millions of positions and some engines would be able to find a advantage for white and some engine could think its black advantage then how will third brain choose the move? If triple brain plays white and if one engine say is shredder 11 tells its advantage white and simply because of that play shredder's continuation?

we have seen in many games shredder's analysis used to reach 2.50 and still rybka eval say = or opposite advantage and win over shredder because of strategy so triple brain is meaning less unless its centaur where when u have dual load 2 engines, one for each core and manual choose the move with ur chess knowledge (IMs and GMs can do this efficiently).
Parent - - By boo! (**) Date 2007-12-22 21:02
I think the idea of having more than one engine working on a position on one CPU is flawed, except perhaps when the number of cores are very high. In some positions it would help, but in most it would probably make things worse.

Instead, I've been thinking about doing my own little "tripple brain" experiment using the latest rybka and winfinder, where winfinder is running on a separate machine. Basically using winfinder if it's evals suggests tactical shots that is missed by regular rybka. In this setting regular rybka would work as normal and on full steam.

All that is needed is a simple arbiter/client and a uci server on each host. A crude prototype should only take me a night, sadly a night I haven't had lately. Perhaps I can get around to testing this now that it's christmas. We'll see.

Come to think of it, search depth must also be taken into account in the likely scenario that winfinder is running on a slower computer. I wouldn't want winfinder to play tactical moves that regular rybka has already refuted :)
Parent - - By Uly (Gold) Date 2007-12-22 21:32

> Instead, I've been thinking about doing my own little "tripple brain" experiment using the latest rybka and winfinder, where winfinder is running on a separate machine. Basically using winfinder if it's evals suggests tactical shots that is missed by regular rybka. In this setting regular rybka would work as normal and on full steam.


Of course, if the third engine is a human brain, it may actually help tremendously :)
Parent - By BB (****) Date 2007-12-24 01:22
Of course, if the third engine is a human brain, it may actually help tremendously :-)

I was waiting for the name (Herr Professor Doktor) "Ingo Althöfer" (he of the 3-Hirn) to appear in this thread, but I guess this is as close as I can expect.
Parent - - By Uri Blass (*****) Date 2007-12-23 10:50
I think that tripple brain can be productive with 2 engines of similiar strength and different playing style.

winfinder is clearly weaker than rybka2.3.2a so using winfinder is a bad idea.

You should try 2 engines with similiar strength to see if the idea works and it is better not to use rybka because rybka supports smp so using smp can be better for rybka relative to triple brain.

possible candidates to try are ruffian and prodeo that seem to have different playing style based on the following link

http://www.computerchess.org.uk/ccrl/4040.live/cgi/compare_engines.cgi?class=All+engines&print=Ponder+hit+-+most+different+pairs

You can also use  toga1.3.1 and fruit2.3.1

http://www.computerchess.org.uk/ccrl/4040.live/cgi/compare_engines.cgi?class=All+engines&print=Ponder+hit+-+most+different+pairs+%28same+families+only%29

Uri 
Parent - - By boo! (**) Date 2007-12-23 13:21
winfinder is clearly weaker than rybka2.3.2a so using winfinder is a bad idea.

I agree that winfinder is weaker, but that's not a problem. My point is that winfinder has certain qualities that regular rybka does not have. Let me be more rigorous: There exists chess positions where winfinder will play better than rybka, and, equally important, these positions should be fairly easy to detect from the output of the two engines.

You should try 2 engines with similiar strength to see if the idea works and it is better not to use rybka because rybka supports smp so using smp can be better for rybka relative to triple brain.

The whole point of my idea is to run winfinder on a separate machine so that rybka can run on full steam.
Parent - - By Uri Blass (*****) Date 2007-12-23 15:34
The main question is if it is really easy to detect at least part of the positions when winfinder is better.
You can decide to choose winfinder move if the score of winfinder is improving and is clearly higher than the score of rybka but
the question is if winfinder cannot "find" wins that do not exist when rybka's move is practically better.

Uri
Parent - By boo! (**) Date 2007-12-23 16:40
Yes, that is the question
Parent - - By Nelson Hernandez (Gold) Date 2007-12-24 18:59
Personally I am sure that no one engine is always right.  Rybka may be the best engine (i.e. it is right the most often) but it is unproven that a consensus approach wouldn't improve the overall result.  Everything depends on the design and implementation, which could be rather complex and would certainly be unsuitable for anything other than a network of computers.  I can't see something like this being anything other than a rather elaborate (and expensive) custom-built system.

Then again I could be wrong and a consensus approach could weaken results.  You won't know until you try. 
Parent - By Banned for Life (Gold) Date 2007-12-24 20:53
There are well known techniques, for example multiple regression, that are commonly used to combine estimates of varying quality to obtain a better estimate than would be possible from any single input. Given a network of engines (so that the quality of the input estimate is not reduced), this approach could certainly be applied to this problem. If Rybka is much better than the rest and the equivalent of the covariance matrix generated between engine inputs is high, you might end up with moves that are virtually the same as an unaided Rybka. As you indicate, its hard to say if this would be the case without an extensive (and expensive) research project that can't be justified in this field.

Alan
Parent - By Uly (Gold) Date 2007-12-24 20:56

> Everything depends on the design and implementation, which could be rather complex


Uri made it sound simple, maybe he can construct a triple brain engine that works, since he seems to have an idea.

I will try my crazy idea posted above, just for fun :)
Parent - - By NATIONAL12 (Gold) Date 2007-12-24 21:58
in all subjects of life that i can think of,the consensus results in weakness. i believe the same will apply in computer chess, unless you can find someway to weigh the results and then who decides the weighting.
Parent - - By Nelson Hernandez (Gold) Date 2007-12-24 23:01 Edited 2007-12-24 23:03
Hmmm...that's an interesting concept.  "E pluribus unum" is clearly not a phrase you find congenial.  Likewise you may have a problem with "checks and balances" and "separation of powers".  I suppose you are also unfamiliar with the concept of "portfolio diversification".

I think what you meant is that in order to have a chance at spectacular success you have to develop your own vision and innovate.  A great composer does not submit his idea to a committee first. 

I entirely agree with Alan's remarks above, by the way.  There are very smart people out there.  I think some kind of improvement would be likely if a reasonably good mathematician were enlisted in the project.
Parent - - By NATIONAL12 (Gold) Date 2007-12-24 23:11
yes we would not have beethoven, haydn mozart, or bach if you have conceptual opinions, please do not throw latin tags at me although i did latin at school, i am now limited to amo, amas,amant. regards paul.
Parent - - By RFK (Gold) Date 2007-12-25 14:57 Edited 2007-12-25 15:01
Hi National12,

Please note: That everything Nelson says is based in financial cant    "spread your assets"  "where every game is a loser"   "all theory points toward diversifying to reduce the risk of a major... " 

Taken apart there is your slick Nelson! Nothing less then a talking financial spread sheet. :-)
Parent - - By NATIONAL12 (Gold) Date 2007-12-25 16:21
i think you are being a bit hard there robert, i have no disrespect for nelson, i admire people with skins like rhino,s who can take what is thrown at them.but nelson changes the subject halfway through his arguement. we were discussion consensus, when suddenly he starts taking about diversity. then alan joins in and we appear to be talking about something else again.
Parent - By Nelson Hernandez (Gold) Date 2007-12-25 17:45
Diversity very much applies to the case at hand: how do you integrate diverse evaluations of a position into a consensus that performs better than a sole source.

By the way, Vas remarks down below that he thinks this approach is a less than optimal use of resources.  In other words he is saying you could arrange a network of computers in such a way that you could do better than the consensus approach, where each node was building its own private search tree.  Interesting.  So many things will be possible in several years when all of us can buy 80-core machines.
Parent - - By Nelson Hernandez (Gold) Date 2007-12-25 17:39
Well, yes.  I do have an MBA in Finance, and I have spent almost 30 years in the financial field.  Actually I was talking in broader terms about risk management.  I'm not a Wall Street rocket scientist though; I'm just a humble analyst, so don't blame me for the unfolding financial meltdown/depression, OK?
Parent - - By NATIONAL12 (Gold) Date 2007-12-25 17:55
i also in my youth managed to get a diploma in economics from the LSE but what on earth has this got to do with what we are dicussing now.
Parent - - By Nelson Hernandez (Gold) Date 2007-12-25 19:29
Hmmm...OK, let me help you keep score. 

You said consensus always resulted in weakness.
I offered contra examples from the areas of government and finance.
I was mocked as didactic.
I clarified my examples.
I was mocked again for applying financial/risk management concepts.
I affirmed that this was my field of expertise.
You accuse me of changing the subject.
I now accuse you of evading my original points.
Parent - By NATIONAL12 (Gold) Date 2007-12-25 20:00
my arguement was that consesus leads to weakness.i was talking specifically about original thinkers, although i will admit that i did not make this clear in my post.i did not mock you and i will apologise if people misunderstood my post.finance is another thing where original thinking is importent.if you follow the consensus of opinion of wall st or throgmorton st we would all be losing money, all i can say is that i am not losing anything.i also believe that alan & youself have seen which way the wind is blown and protected yourselves. iwill willingly answer any other points you may think i have avoided.
Parent - - By RFK (Gold) Date 2007-12-25 01:27 Edited 2007-12-25 01:58
Oh, for the love God! Nelson, what a pedantic, pretentious didacticism I have ever seen you spill out on a page. Cut the cant,(probably, cut the crap would be more suitable) and talk straight English with National12 for Christ's sake, " portfolio diversification". ahahah! Oh! Brother!

"E pluribus unum" that is hilarious! :-P

sidserious-That's just about a bucks worth for you!
Parent - - By Nelson Hernandez (Gold) Date 2007-12-25 02:32
For goodness sake, I hardly know what to say!  The guy says consensus is weakness, and he can't think of an exception to the rule.  That's utter nonsense, of course.  So I trot out some of the elementary concepts of federalism and constitutionalism that collectively constitute democratic government, i.e. rule by consensus (in contrast to alternatives like despotism, monarchy, dictatorship), and guess who gets bashed?  Me!  Not the guy who fatuously says consensus is always weaker.

Then I turn away from government as an example where consensus is better than a monopoly, and turn toward the world of finance.  Is it better to put all your eggs in one basket or to spread your assets around?  Unless we're talking a Las Vegas casino, where every game is a loser, all theory points toward diversifying to reduce the risk of a major, unrecoverable error.  Again, another example where dispersing rather than consolidating offers better prospects.

Now, will you philistines kindly back off, or shall I start quoting the French guys from "Monty Python and the Holy Grail"?
Parent - By Banned for Life (Gold) Date 2007-12-25 05:49
Unless we're talking a Las Vegas casino, where every game is a loser...

You are only looking at one side of the coin here. Remember, losses in the casino allow me to not pay a state income tax, thus allowing me to spend more of my time in the company of large breasted women. Thank you for your continued support.

Alan
Parent - - By BB (****) Date 2007-12-25 10:53

>Then again I could be wrong and a consensus approach could weaken results. You won't know until you try.


I'd guess that di-worse-ification would mostly hurt (at least in the current computer chess world). The economic analogy might be to the South, where cotton was King, as its marginal return was much higher than anything else. [Portfolio management does not necessarily readily apply to such extremal cases]. However, if the situation is: running another programme on a computer that would otherwise be idle (and thus it doesn't use up any marginal resources), it is difficult to imagine that some gain could not be filched via sufficient ingenuity.

> There are well known techniques, for example multiple regression, that are  commonly used to combine estimates of varying quality to obtain a better estimate than would be possible from any single input.


One concern with blackbox-ing this methodology is that bias due to the positional factors is common in evaluations (one programme does better with king safety, or passed pawns, et cetera), which should be taken into account in the quality scores. [The pessimist could see this as a deterrent - but a maths boffin who wishes to push the boundaries of best current practise might find it to be a positive feature].
Parent - - By Banned for Life (Gold) Date 2007-12-25 18:11
The less correlation there is between the outputs of the engines, the more effective the combination can be. A real-life example of this would be a guidance system which combines the outputs of GPS, an RLG (ring-laser gyro), and Doppler navigation systems to provide outputs far better than any of the individual systems (which each have significant weaknesses on their own).

Alan
Parent - By BB (****) Date 2007-12-26 06:46

>> One concern with blackbox-ing this methodology is that bias ... should be taken into account in the quality scores.
> The less correlation there is between the outputs of the engines, the more effective the combination can be.


Perhaps I should expand on my comment. If the weaknesses are "known" then any adaption of standard methods would not be too novel - my intent was to have the computer system itself determine (via feedback) the relative strengths/weaknesses of the component parts - I think this would be superior, yet more demanding in implementation. The peculiarity of the typical "quality score" is that it is here (in some sense) an average over differing qualities (ability to judge king attacks, ability to judge KP endgames, et cetera), while a more trenchant analytical method would develop feature-specific quality scores (here I use "develop" as connoting the computer action - a useful human could also generate such weightings).
Parent - By Vasik Rajlich (Silver) Date 2007-12-27 12:39
There is another effect here. Often, users confuse being good at something with emphasizing something, which is of course not very accurate.

If you have a program which is truly weak in (for example) king safety, that program should fix its problems.

If you have two programs, one of which emphasizes king safety while the other does not, and they disagree, you should listen to the better one.

Vas
Parent - - By Vasik Rajlich (Silver) Date 2007-12-25 15:43
My personal opinion is that this is not productive - it won't be the most effective use of the processing power, except maybe in some degenerate cases. This should be true regardless of the type of hardware available.

Vas
Parent - - By Roland Rösler (****) Date 2007-12-26 04:35
I think, all the number one(s) like Topalov, Anand, Kramnik and Kamsky will agree. But are you sure, this is your opinion, if Rybka isn´t number one?
Parent - By Vasik Rajlich (Silver) Date 2007-12-27 12:40
Yes - although I may keep it to myself.

Vas
Up Topic Rybka Support & Discussion / Rybka Discussion / Rybka 2.3.2a and Triple Brain ?

Powered by mwForum 2.27.4 © 1999-2012 Markus Wichitill