Question for Bob Hyatt

[lkaufman]

Doch also does a lot of the sorts of pruning you mention, but does not display the characteristic Rybka behavior of rarely changing her mind and of missing things that are immediately seen on the next move. This behavior did not start with Rybka 3, it was certainly the case with Rybka 2.xx. I believe it was also the case with Rybka 1, but I never spent much time with that version so I'm not certain. Others who have commented seem to feel that already with Rybka 1 the main-line concentration was obvious and unprecedented. Is it possible you simply did not spend enough time with the Strelka code to see what was responsible?

[Zach Wegner]

I agree with you that the extreme focus on the PV is counter-intuitive, and clearly the programming world agrees with you as no one else appears to be doing it aside from the Rybka-derivatives.

It's not counter intuitive to me, at least anymore. And it's not only Rybka that does it--AFAIK this all began with Fruit, and Vas later took it to an extreme. Fabien (and Tord too, for that matter) deserves a lot of credit in that regard.

[Zach Wegner]

"The parameter tuning of material seems a neural network to me. the parameter tuning of the other parameters seems more thoroughly tested". Wow, I've never been called a "neural network" before, but I guess it's a compliment.

Sort of, most people I know have one of those.

[BubbaTough]

I believe it was also the case with Rybka 1, but I never spent much time with that version so I'm not certain. Others who have commented seem to feel that already with Rybka 1 the main-line concentration was obvious and unprecedented.

I think unprecedented is a significant exaggeration, particularly once you discount the extra/hidden depth, which naturally acts to stabilize the pv.

Is it possible you simply did not spend enough time with the Strelka code to see what was responsible?

Possible I guess, but unlikely. I must admit I did not study it line by line, day after day...but the search was pretty simple...and pretty standard. Most of the fancy tricks were for efficiency, and the few non-pv "aggressive pruning" tricks were things I had already tried before, and assumably that is true for a number of others too. I doubt there was something major I missed. I assume subsequent versions are quite different of course.

-Sam

[lkaufman]

sure, it's a matter of degree, but I think Rybka behaves quite differently than all other programs (derivatives aside) because she does this to a vastly greater degree.

[lkaufman]

"once you discount the extra/hidden depth, which naturally acts to stabilize the pv." Could you explain this comment a bit? I'm not exactly sure what you refer to here with extra/hidden depth. Thanks.

[BubbaTough]

sure, it's a matter of degree, but I think Rybka behaves quite differently than all other programs (derivatives aside) because she does this to a vastly greater degree.

I am sure you are right...but perhaps that was not so much the case for Rybka 1. Of course, I only say this based on having looked at Strelka.

-Sam

[BubbaTough]

"once you discount the extra/hidden depth, which naturally acts to stabilize the pv." Could you explain this comment a bit? I'm not exactly sure what you refer to here with extra/hidden depth. Thanks.

My understanding is that Rybka subtracted a couple ply before outputing its depth number. A lot of people (not me) were pretty annoyed when this was revealed. That means it was searching a bit deeper than it appears, and I believe the deeper you search the less likely you are to switch pv.

-Sam

[lkaufman]

Oh, you're just referring to the fact that Rybka did not count the last three plies, on the grounds that they are highly selective. I don't think that this affects anything; the question is, at a given time limit, why is Rybka much less likely to change her mind than other programs? It doesn't matter what depth is reported.

[BubbaTough]

Oh, you're just referring to the fact that Rybka did not count the last three plies, on the grounds that they are highly selective. I don't think that this affects anything

Well, I would disagree with this. A number of programs do very aggressive pruning here, and they count the these plies. Therefore, Rybka is outputing less depth than others while doing the same thing. Keep in mind, I am not bothered at all by this.

the question is, at a given time limit, why is Rybka much less likely to change her mind than other programs? It doesn't matter what depth is reported.

Well, being a very fast, very well coded engine...better than all others by a large margin, it was getting much deeper in the same time than other programs (though it may not have looked that way because of the reporting). If my claim that going deeper stabilizes the pv, then in part the stabler pv can be explained by the extra depth.

Again, I am only talking about Strelka really, which I am confident does not have a super special search besides being really efficient. Rybka 2 and 3 I am sure are quite different.

-Sam