Cluster Rybka

[Rolf]

But in the meantime you cant just wave hands.

One of these days in the not too distant future, cluster crafty will play cluster rybka. Then we'll see who's been waving their hands.

I hope so. But I have doubts. The communication between commercials and open sourcerers cant function because the masses listening in the background who only wait to copy. Then Bob still has a life as an academic but Vas future kids are starving. So difficult to understand?

[bob]

That was a pretty trollish post.

No

Even if bob were a little bit biased against Rybka (and I'm not saying whether he is), the business of speedup from parallel algorithms has a lot more to do with how AlphaBeta search fundamentally works, than the specific details of any particular engine.

Dr. Hyatt been working on parallel search techniques for decades, so he has lots of experience in this area. If he looks at the output of Cluster Rybka and it clearly indicates to him that Rybka is splitting only at the root, then he is very probably correct (and someone who believes otherwise, should have to bring forward some convincing interpretation of the same data to prove their view).

And if Rybka is splitting only at the root, there is no possible way (regardless of whatever details went into its implementation) for the speedup not to suck. Splitting only at the root just doesn't work.

I understand how the chain of arguments is intended to work and just wanted to hint at the fact that it is flawed, one either understands this or not.
In the end there is the simple argument: Look, what I do does not work with crafty, so what he does with rybka cannot work either. By summing up unproved assumptions you can prove everything, and this statement is ridiculous as an argument.
Rybkas gain by the cluster may suck or not, I don't know, but crafty's results are totally independent from it: this is really easy and should not be argued about at all.

Your statement is simply wrong, and more than one has tried to explain why. Alpha/beta is _well_ understood. You can look at the Knuth/Moore paper to understand exactly how it works and why. Then you can find the only really mathematical analysis I have ever done and published in a paper in the Journal of Parallel Computing, which took the Knuth/Moore analysis and stretched it into a parallel search analysis. And splitting alpha/beta trees at the root is bad for _all_ alpha/beta tree searchers. Now if you want to offer proof that Rybka doesn't use alpha/beta, then you have a leg to stand on. But we already know it does alpha/beta, and it doesn't matter whether it is an alpha/beta tree for chess, checkers, go, othello, or you-name-it, the parallel algorithm has to conform to the restrictions imposed by alpha/beta.

Splitting at the root provides little if any speedup. I posted recent numbers. I posted such numbers in 1983. Monty Newborn posted such numbers back in the late 1970's. This is not new. It is well-understood. It just doesn't work very well. Certainly not to the +100 Elo claimed which means a factor of 4 using 5 cluster nodes. Not going to happen. Getting a factor of 2 is almost impossible with today's narrow branching factor.

So you may as well drop your misinformed argument about what works or doesn't work in Crafty doesn't mean a thing about Rybka. If you were talking about evaluation, or about search methodology, I would agree. But parallel search is parallel search, there is ample math to support this.

Bob perhaps I can explain what is going on although I dont even know what splitting at the root and above this in acluster - really is!

You are talking about your only really mathematical analysis 25 years ago and you obviously claim that even today that holds water and is believed in CC like sort of natural constants.

Is it absolutely impossible that actually such constants are different? How can you prove that?

The reasoning by Peter E. sounds better and besides personal feelings I would also say that you believe in something that cant be the same after 25 years and for different hardware.

What you are doing as a teacher that is ok, you ask "if thzere is something new then please he should step forward and explain."

But in the meantime you cant just wave hands.

Let me please repeat what I would be able to imagine, but it's really my last idea. Please explain this to me.

How sure can you be that the output of such a cluster shows today in 2009 what he's really doing? If I were Vas I would know what you know and how I could play so that you see exactly what you believe and expect to see.

Why dont you work in a more opened style where even your data cant be taken for the njudgement of new stuff?

What exactly brought you to the point that you act like those astronoms who showed what's absolutely impossible to exist?

I write that after having read the thread backwards to page 8.

I am not going to waste time on this. Alpha/beta has _not_ changed since it was defined in the 1950's. The math still works.

As far as "being sure" he has already admitted that this is what is happening, so why do we need to have a discussion abut why this is right, when Vas has already said it was right???

[bob]

want to make a bet? I saw his output. There is one, and only one way that output can be produced. By splitting only at the root. That is a poor parallel search approach, and it is poor regardless of the program using it. It's been discussed for 30 years, and dismissed for 30 years. The data I showed explains why.

But there is absolutely _zero_ doubt about how his cluster search works. And I do mean _zero_.

you offer a bet, although you didn't want to waste more time with this issue? you also do tests with crafty to back up your opinion, though you didn't want to waste more time.
BUT you don't take a bet about the 100+ elo improvement for going from 1 node to 5 nodes, because you don't want to waste more time?

I try, but that's hard to understand.

I didn't do _any_ tests with Crafty, and didn't waste any time.

We now know my conjecture was correct because Vas has _said_ it was correct. So we are _still_ wasting time when you don't keep up with current discussions _here_ to know about this...

[bob]

Could the increase in Elo be due to evaluation rather than speed increase? By somehow making good (very good) use of the additional information from splitting at the root?

Could be from cosmic rays, inter-dimensional communication, or extra vitamin-D. Or from incorrect measurement. No idea. But no +100 Elo either.

[bob]

if I were as confident as you on this subject, I would take that bet from Vas.

isn't it easy money? you also could bet for some insight in rybka if you win

what bet???