Author: Robert Hyatt
Date: 16:56:10 11/24/98
Go up one level in this thread
On November 24, 1998 at 18:07:00, Amir Ban wrote: >On November 24, 1998 at 10:59:13, Ernst A. Heinz wrote: > >>On November 24, 1998 at 09:59:02, Amir Ban wrote: >>> >>> [...] >>> >>> Junior is more aggressively extended than most PC programs, and therefore >>> more than DB. >> >>Amir, >> >>I do not think that your above chain of reasoning is valid because we really >>do not know how much "DB" extends. I am convinced that "DB" extends far more >>than most PC programs as well. Whether "Junior" exceeds "DB" in this respect >>is open to speculation but not to conclusive argument. >> >>BTW, I happen to remember a post on RGCC shortly before the Kasparov rematch. >>The poster had apparently talked to Murray Campbell during a conference and >>reported that Murray said something about the search trees of "DB" being very >>broad at the top, then stringy because heavily extended in the middle, and >>quite bushy again at the bottom where the special-purpose chess processors >>resorted to 4 plies or 5 plies of non-selective (i.e. also non-extended?) >>full-width search again. Especially the final full-width part of the chess >>processors would result in a completely different tree-structure than those of >>all others if my memory serves me right and the original report was true. >> >>=Ernst= > >I don't really believe this. First, because Murray is likely to have said the >tree is "heavily extended" if what they did was not more than check and >recapture extensions. As every programmer knows, that's already heavy, but every >does it, and for Junior that's really only the beginning. I doubt Murray meant >"in comparison to other chess programs", because his audience would not care >about that, and as a Deep Blue team member he probably didn't care either. > Murray doesn't report like that. I know him quite well, and he knew *exactly* what we did in Cray Blitz, for example, and in HiTech. And many discussions centered around what they were doing in DT (and then deep blue later) that they couldn't do in HiTech "because it didn't have nearly enough horsepower to make the extensions effective." Again, take the math. Cray Blitz did 10 ply searches (minimum) with the normal recapture extension, in check extension, full (pv and fail-high) extensions, plus following checks and some non-checking threats in the q- search. All using 500K-1M node per second searches on a C90. They are 250 times faster, yet they search no deeper *nominally*. Obviously they extend more than we did in CB. I'd bet that CB extends at least as much as you are, but we could compare notes to quantify this. >Second, because it does not make sense for Deep Blue to take an avant-garde >position on this. Their greatest asset was raw computational ability, they had >NPS to spare, and they had every reason to take a conservative position on the >search and not risk their advantage. This has been said by several people before >to explain some things they did, like for example SE, which other developers >didn't find worthwhile, but when you start with a 100 to 1 computing advantage >over your opponent may be worth just for the slight extra insurance. > Did you read their ICCA paper on search extensions? Did that sound "conservative"? >According to all appearances, they did in fact take this conservative approach, >and perhaps even carried it too far. Hsu's prejudice against forward pruning is >just that, a prejudice, which cost them expensively in computing power, but they >didn't feel the heat and could indulge themselves. Against equal hardware, they >would quickly forget this prejudice like all of us. > that I wouldn't doubt. Hsu didn't like my speculative pruning in the q-search of Cray Blitz, then later told me that they did something like that in the Deep Blue hardware. They aren't dumb... by a long shot. They even ran lots of experiments with null-move (again, in the ICCA paper), but didn't use it in real life for reasons of their own... >Another way of saying this is that the search algorithms of Junior arose not so >much because I wanted to do them in any theoretical sense, but because of the >need to be competitive against opponents who seemed to outperform me. I can't >imagine anyone who doesn't feel this kind of pressure doing the same, and from >their point of view, doing this would probably appear irresponsible. > >Amir possibly... and that's an important point. and where they seem to be is in a position where they make *no* mistakes during the first 10-11 plies since they have no forward pruning and no null-move to introduce or hide errors... yet they can do this and *still* have enough horsepower to push the average search depth to 30 plies or so for the interesting lines... but only because of their "horsepower"...
This page took 0 seconds to execute
Last modified: Thu, 15 Apr 21 08:11:13 -0700
Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.