Would the Triple-Brain Concept be a match for the new Rybka?

[Eraserheads]

I was fascinated with SMK's Triple-Brain concept when it was first introduced some years ago. The idea of two engines suggesting moves, and being adjudicated by a third, deciding entity was revolutionary. But with the kind of hardware available, I got to feel that back then this concept was way ahead of its time.

Now the availability of 8core machines is even more within the reach of the average computer chess fan.

Perhaps an octa-brain opponent for Rybka 3 would not be so much of a pushover? Getting a wide variety of 'heads' with varying styles would be ideal - like getting one bean-cruncher, one speculator, one attacking engine, one defensive engine etc... to completely fit the facets of this 8-headed beast. This would hopefully make it avoid missing unusual moves.

The adjudicating engine could be Rybka 2.3.2a

What do you guys think?

[CRoberson]

It wasn't a new idea. It falls under the category of Multi-Agent Systems.
It sounds good at first but has a fundamental problem: how do you
decide when they all disagree? Of course, how do you decide at
any point? A simple voting system could be functional and then a
weighted system could be good. There is a lot of research going on
in that field and its applications to many areas outside of computer
chess or even computer games.

Given the scaling problems we see beyond 4 procs. The question
becomes: would Rybka on an oct box be outperformed by 2 or more
engines on quad boxes and then using the "triple-brain"?

I think it is interesting that throughout chess history it has been
demonstrated that one very good player beats teams (Kasparov vs
the world, Morphy vs a Count and a Duke). I don't see why one should think this would be different
with computers.

[Uri Blass]

<snipped>

Now the availability of 8core machines is even more within the reach of the average computer chess fan.

I think that the average computer chess fan has only one or 2 cores.
I also think that the triple brain is a failure without a smart program that
know which move to choose.

A smart program that knows which move to choose does not have to be a chess engine.

If people do not develop programs only for the task of choosing the engine then programmers do not take the idea seriously so the idea is going to fail.

If people develop programs only for the task of choosing the engine then the first thing is to find the best chooser program and it is possible to do it by games like shredder and Fritz and chooser X against shredder and Fritz and chooser Y.

Uri

[M ANSARI]

I think triple brain is a good idea if you use multiple independent machines connected by LAN on a local network and you assign an IP address to each engine. You have the MAIN engine playing and say two other computers assigned with different engine parameters or even totally different engines. This is already possible with Chessnet software but is used for analysis. I am not sure if you can configure it to play as one combined entity ... obviously that would make things more difficult as you would have to have some sort of arbiter who would choose a move in case they all disagree. You could ofcourse easily code an engine to act as the main arbiter and use a trigger evaluation difference to allow the variation to be checked by the MAIN engine if trigger has been activated. Only then would it overrule its original choice.

[parrish]

Too many cooks spoil the soup.

[ozziejoe]

It seems like a highly restricted version of triple brain might work. For example, you could have the triple brain go with rybka's choices almost always, except in instances where rybka has evaluated the top two lines to be about equal.(say within 5 centipawns). In this instance, , say, deep fritz and zappa could break the deadlock. (or if they don't have a clear, and agreeing opinion, you coudl stay with rybka).

With rybka likely to be 200 points better than the next engine, than it does not make sense to go against her judgemnts very often.


best
J

[michiguel]

I was fascinated with SMK's Triple-Brain concept when it was first introduced some years ago. The idea of two engines suggesting moves, and being adjudicated by a third, deciding entity was revolutionary. But with the kind of hardware available, I got to feel that back then this concept was way ahead of its time.

Now the availability of 8core machines is even more within the reach of the average computer chess fan.

Perhaps an octa-brain opponent for Rybka 3 would not be so much of a pushover? Getting a wide variety of 'heads' with varying styles would be ideal - like getting one bean-cruncher, one speculator, one attacking engine, one defensive engine etc... to completely fit the facets of this 8-headed beast. This would hopefully make it avoid missing unusual moves.

The adjudicating engine could be Rybka 2.3.2a

What do you guys think?

The concept was introduced by Ingo Althofer as Three-Hirn, not SMK

The third brain could a human (not necessarily a GM). I believe that in that case 3-Hirn will beat any other program, including Rybka. The engines should be as different in style as possible.

Miguel

[bob]

I was fascinated with SMK's Triple-Brain concept when it was first introduced some years ago. The idea of two engines suggesting moves, and being adjudicated by a third, deciding entity was revolutionary. But with the kind of hardware available, I got to feel that back then this concept was way ahead of its time.

Now the availability of 8core machines is even more within the reach of the average computer chess fan.

Perhaps an octa-brain opponent for Rybka 3 would not be so much of a pushover? Getting a wide variety of 'heads' with varying styles would be ideal - like getting one bean-cruncher, one speculator, one attacking engine, one defensive engine etc... to completely fit the facets of this 8-headed beast. This would hopefully make it avoid missing unusual moves.

The adjudicating engine could be Rybka 2.3.2a

What do you guys think?

That was not a "revolutionary idea" at all. It was simply a rehash of what had been done multiple times in the past.

For example, the last version of MacHack had the normal engine, and then the "cheops" engine that was a tactical searcher. The tactical search was one input, the normal positional search was a second, and the result was then subject to arbitration.

Sun Phoenix did the same thing, with a deeper tactical search done in parallel with the normal positional search, and again the results were arbitrated.

Humans have done the same thing using two or more programs and trying to select the best suggestion from the available choices.

This isn't new at all. And it is not particularly effective. Just notice how often it has been used in tournaments since Sun Phoenix retired...

Shredder has never played with it in a real tournament either.

[bob]

I think triple brain is a good idea if you use multiple independent machines connected by LAN on a local network and you assign an IP address to each engine. You have the MAIN engine playing and say two other computers assigned with different engine parameters or even totally different engines. This is already possible with Chessnet software but is used for analysis. I am not sure if you can configure it to play as one combined entity ... obviously that would make things more difficult as you would have to have some sort of arbiter who would choose a move in case they all disagree. You could ofcourse easily code an engine to act as the main arbiter and use a trigger evaluation difference to allow the variation to be checked by the MAIN engine if trigger has been activated. Only then would it overrule its original choice.

But the issue is exactly as it has been since this idea was first used about 20 years ago... "what do you do when the two engines disagree?" Some (Phoenix) used the "second brain" to simply tactically confirm that the primary move was not bad, if it was, the primary move was discarded and another one proposed. It is a very difficult problem, with no good answer to date. You can't even reasonably choose between a shallow-searched positional move and a deeply-searched tactical move, unless the deep search finds a clear win of material or mate...

[bob]

I was fascinated with SMK's Triple-Brain concept when it was first introduced some years ago. The idea of two engines suggesting moves, and being adjudicated by a third, deciding entity was revolutionary. But with the kind of hardware available, I got to feel that back then this concept was way ahead of its time.

Now the availability of 8core machines is even more within the reach of the average computer chess fan.

Perhaps an octa-brain opponent for Rybka 3 would not be so much of a pushover? Getting a wide variety of 'heads' with varying styles would be ideal - like getting one bean-cruncher, one speculator, one attacking engine, one defensive engine etc... to completely fit the facets of this 8-headed beast. This would hopefully make it avoid missing unusual moves.

The adjudicating engine could be Rybka 2.3.2a

What do you guys think?

The concept was introduced by Ingo Althofer as Three-Hirn, not SMK

The third brain could a human (not necessarily a GM). I believe that in that case 3-Hirn will beat any other program, including Rybka. The engines should be as different in style as possible.

Miguel

I think there would have to be one more assumption for that to have even a small chance of happening. That assumption is that you can find an engine that is tactically better than Rybka even if it is positionally weaker. Two engines are not going to beat a program that plays better positionally and tactically than either of the two engines being used... That is a problem this won't overcome.