Cluster Rybka

[bob]

Believe what you want. But facts speak louder than hyperbole, IMHO..

By showing data about crafty you prove absolutely nothing about the gain Vasik gets by whatever he does (and you dont know what he does).

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_.

[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?

Nope. It is the result of pure hyperbole, nothing else.

[Eizenhammer]

Believe what you want. But facts speak louder than hyperbole, IMHO..

By showing data about crafty you prove absolutely nothing about the gain Vasik gets by whatever he does (and you dont know what he does).

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_.

It took you several years to notice that fruit is a great program, but by looking at a single output of rybka you know what's going on. People will notice that you are just angry and biased, it is soo obvious ...

The simple fact remains: you can show data from crafty all day, and it proves nothing at all about rybka. But this kind of simple fact usually needs about 200 postings and some insults here in this forum and probably the help of a neutral mathematician, want to make a bet?

[BubbaTough]

For what it is worth, my interpretation of the PV data from Rybka is exactly the same as Bob's. In addition, the Rybka operators confirmed this interpretation is correct. Also, before the tournament I helped the Rybka team test, and we discussed the search in terms of what they should output as their PV, and that discussion was entirely consistent with Bob's interpretation.

How much increase in ELO the cluster provides was not discussed. It seems hard to imagine the Rybka team would have an accurate estimate of the bonus ELO, since the search was clearly much worse at fast time controls (possibly worse than using only a single machine and not a cluster at all for very fast time controls) so they could not really have played massive numbers of games to generate an accurate estimate.

Given all this, I am glad to see Rybka is starting to do work in this area, and look forward to continued progress in this area. I am even more excited about the possibility of Crafty working on this, since open source work on this is more likely to move the state of the art forward (in fact, I would not be surprised if Rybka is improved more by looking at Bob's effort and copying aspects of it than working on it on its own, since parallel search does not seem to be their area of expertise).

-Sam

[bob]

Believe what you want. But facts speak louder than hyperbole, IMHO..

By showing data about crafty you prove absolutely nothing about the gain Vasik gets by whatever he does (and you dont know what he does).

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_.

It took you several years to notice that fruit is a great program, but by looking at a single output of rybka you know what's going on. People will notice that you are just angry and biased, it is soo obvious ...

If you want to join a discussion, have something ready to offer besides something that makes you look idiotic.

(1) it didn't take me "several years" to notice that fruit was a good program. I'd bet I noticed it well before you knew how to spell it, since I play on ICC _every_ day and saw it very soon.

(2) As for looking at a "single output" I played an _entire_ game against cluster rybka in the last ACCA game and had the chance to watch its output for the engire game. yes, that is more than enough to recognize _exactly_ how it splits at the root. If you can not do so, that's not my problem. I know more than enough about parallel search to figuire out that simple approach for what it is, having first done _that_ approach in 1983...

The simple fact remains: you can show data from crafty all day, and it proves nothing at all about rybka. But this kind of simple fact usually needs about 200 postings and some insults here in this forum and probably the help of a neutral mathematician, want to make a bet?

The simple fact remains that the trees for _all_ programs are subject to the _same_ alpha/beta algorithm issues. Whether _you_ understand that or not is irrelevant. I _do_ understand it. As to the "neutral mathematician" I have no idea what you are talking about, and really don't care, either.

[bob]

For what it is worth, my interpretation of the PV data from Rybka is exactly the same as Bob's. In addition, the Rybka operators confirmed this interpretation is correct. Also, before the tournament I helped the Rybka team test, and we discussed the search in terms of what they should output as their PV, and that discussion was entirely consistent with Bob's interpretation.

How much increase in ELO the cluster provides was not discussed. It seems hard to imagine the Rybka team would have an accurate estimate of the bonus ELO, since the search was clearly much worse at fast time controls (possibly worse than using only a single machine and not a cluster at all for very fast time controls) so they could not really have played massive numbers of games to generate an accurate estimate.

Given all this, I am glad to see Rybka is starting to do work in this area, and look forward to continued progress in this area. I am even more excited about the possibility of Crafty working on this, since open source work on this is more likely to move the state of the art forward (in fact, I would not be surprised if Rybka is improved more by looking at Bob's effort and copying aspects of it than working on it on its own, since parallel search does not seem to be their area of expertise).

-Sam

Not worth the trouble to respond. This guy doesn't know what he is talking about, doesn't realize that we had long discussions during the last ACCA event when different people kept asking about the output Rybka was producing. He doesn't know that this was asked on the Crafty mailing list and discussed there. What he does know would not fill a thimble, in fact. There were lots of questions about how strong Cluster Rybka was. I have now explained the answer in detail. It is no stronger than a normal 8-way box and may well be weaker, thanks to the narrow branching factor we see with today's programs. He doesn't know any of that, and trying to explain it is a waste of time because I doubt he _wants_ to understand it, otherwise he would read, and learn, and remain silent until he had something _useful_ to offer.

[frosch]

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.

[frosch]

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

[Eizenhammer]

otherwise he would read, and learn, and remain silent until he had something _useful_ to offer.

You react just as expected, of course.
My useful contribution, repeated for a last time (I won't answer, so you can have the final insult):

You show crafty data, they say nothing about rybka.
It is absoutely useless as a prove.
What you say may very well be true, but you can't prove Vasik wrong by showing your crafty data. Is this understandable? well probably not ...

[bob]

otherwise he would read, and learn, and remain silent until he had something _useful_ to offer.

You react just as expected, of course.
My useful contribution, repeated for a last time (I won't answer, so you can have the final insult):

You show crafty data, they say nothing about rybka.
It is absoutely useless as a prove.
What you say may very well be true, but you can't prove Vasik wrong by showing your crafty data. Is this understandable? well probably not ...

My contribution, ignored by you but not by others, is that the parallel search data I presented says _exactly_ what kind of performance Rybka produces on a cluster. _exactly_. Not "close". _exactly_. No we don't have the same evaluation. No we don't have the same search. But those are absolutely independent of the results a parallel search produces. If you don't get that, you need to study the problem. Once you get the idea, then you will see that for speedup, my test was _exactly_ predicting what Rybka (or any other program that splits in that exact manner) is going to get.

You really should avoid jumping into a conversation where you have absolutely no idea about what is going on. This is a good example. He is not getting 100 Elo. He is not getting 10. Perhaps in Freestyle, which is a different game. But not in normal chess.