Author: Robert Hyatt
Date: 14:01:04 11/23/98
Go up one level in this thread
On November 23, 1998 at 12:59:36, Bruce Moreland wrote: > >On November 23, 1998 at 11:50:01, Amir Ban wrote: > >>On November 23, 1998 at 09:37:25, Robert Hyatt wrote: > >>>what you are overlooking is the point that junior (and all the other programs) >>>look at a fat, shallow tree. >> >>I am quite sure that the opposite is true. All PC programs have a much smaller >>effective branching factor than DT/DB. This is because they all do forward >>pruning, many of them aggressively, while DT/DB did none, and they do >>extensions, most at least as much as DT/DB, and at least in Junior, much more >>aggressively than DT/DB. > >I think that they are doing some sort of tree size limiting, but I don't know >for sure or how. > >>> Because *they* are searching 10 times deeper than I am on most >>>moves, thru their "singular extensions" (and other extensions). >> >>This doesn't agree at all with what Deep Thought and others found about this, or >>even with common sense. Even if singular extensions are always a great thing to >>do (doubtful even according to the inventors), it covers only one side of the >>game. Most tactical situations are non-singular, so SE has nothing to do there. >>To argue, as you do, that it effectively adds several plies to the search depth >>is quite a hyperbole. > >Dunno. I do know, and the math from Amir is "broken". IE how deep does junior search on axb5, after 7 minutes. How deep did DB search after 7 minutes? How would he explain that huge disparity (Junior probably searches "deeper" yet it is 1/10,000th as fast overall, yet it goes "deeper" than deep blue if you look at nominal search depth?) Something's wrong in his analysis. Branching factor is "noise". The point is how do they search 250M nodes per second and get to ply X, he searches 250K nodes per second and gets to ply Y, where Y is > X? Unless they do something "different" and search very deeply along some of those lines? > >>To say that today's top program are not only unable to discover the c5 line, but >>even to find that any move within this line is singular is beyond their >>capabilities, is one of the greatest exaggerations yet seen on this newsgroup. > >Maybe Bob isn't remembering this right, but he seems pretty sure so I'm >listening. He says they had +2 there. +2 is a lot to have there, and I think >it's pretty unlikely that Bob is mis-remembering. So I think it's at least >likely that DT found something there. I can't find anything there, can you? So >maybe this is a case where their search worked. I'm not remembering this wrong. Remember that it happened to *me*. as I watched... > >>> We've already >>>seen that in the Deep Blue vs Kasparov game two, Dark Thought and Ferret have >>>searched axb5/Qb6 to depth 20 or 21 without seeing anything to cause it to fail >>>low, yet we know deep blue did. At 1/2 that depth. So it might take a program >>>like junior *fifty* plies to find what is going on there for all I know at >>>present. And if I could somehow give you a PV to get you down to the point >>>where Junior sees this, it would be so deep, probably, that it would be easy >>>to say "but this isn't the best move, white or black should try this instead. >>>And we end right back up at square zero. >>> >> >>Quite an exaggeration, don't you think ? So we can calculate for many centuries >>and not find it ? Why didn't you say so before we already wasted about a week on >>this ? >> >>But besides, this is a switch you are pulling here: We put all the computers on >>this position to vindicate DB's axb5. Now it's become a tautology that it's best >>? No, tis not I pulling the word games... tis you... I suggested that maybe someone might find that qb6 is not so good, and might be able to find that axb5 is better. And that would certainly suggest that DB was on the up-and-up. I've since found out (private source) that both of these moves lead to the same final position. But you don't prove a "negative"... so finding that a program would play axb5 would have eliminated all controversy because there would be no reason to doubt DB could also find that. But just because we *don't* find that Qb6 is not winning a pawn certainly doesn't mean that *they* didn't. So it wasn't a "waste of time." at the start... it was a faint hope that one of us might find what they found. we didn't (at least so far). So what? Do we go back to the stupid Kasparov claim once again? He mentioned it on ICC Sunday night. So I guess he still has his head stuck up somewhere where it doesn't belong, and where it is *very* difficult to see anything... > >It may not be any better than Qb6. In order to DB to find it, it had to think >that it was better, for just one ply. > >If someone finds axb5, that is a pretty convincing argument that it is possible >to find it. > >If someone gets a big score drop on Qb6, well, this at least shows that a >program can understand Qb6 to some degree. > >Seirawan says in the ICCAJ that "It is intriguing to understand how DEEP BLUE >could reject a line that wins two Pawns by force." I think that we're all >capable of seeing that it only wins one pawn, but I'm hoping that we can find >that it wins zero pawns. > >I'm working on ply 20, it's been going for a few days, when it's done I think >I'm going to stop. > I let this run for a while here on a single P6... I just finished 19 and my score for Qb6 is down to +.97, on the other machine, axb5 is still holding on at +.6, but that is one ply behind the machine with Qb6 and has stopped because of a power failure kicking that machine in the head. >>It should also be remembered that we don't even have a hypothetical line that >>Deep Blue saw to make it dislike Qb6. >> >> >>>There are just some things they can see at 250M+ nodes per second that we won't >>>ever see... >>> >> >>Why ? The computers are now thinking for tens of billions of nodes. They should >>be able to see as far as DB now. Probably much farther because of their deeper >>nominal depth. > >[snip] > >This is always an interesting question about DT/DB. If your hardware was 1000x >faster, would you change your search? I doubt I would, but I'd have to have the >hardware in order to know. I can only answer for myself. And the answer is "yes". If I were suddenly 1,000 times faster, I'd likely change quite a few things about how I extend, including going back to the full-blown singular extensions I used on Cray Blitz. Because I'd much rather search forcing moves to 30 plies than arcane moves to 15... > >An argument about DT is that things are different when you go deeper, and your >search techniques need to change substantially. I haven't found this to be true >even though hardware has sped up by a factor of at least 20 since I've been >doing this. We're getting to be fairly close to the older DT now, aren't we? > I'd guess not. Remember the c5 move was played by the *old* deep thought in 1988. >[snip] > >>We didn't think it was, quite the opposite: The conclusion from taking you at >>face value is that back in the late 80's and early 90's computers were playing >>at at least 1000 ELO stronger than today, to the extent that today's top >>programs are not able to reconstruct their moves, or even to follow their >>reasoning. >> >>All I wanted to see, by the way, was a demonstration of singular extension in >>action by DT/DB. Surely something like this exists. > >Might be hard to generate this, since they didn't run without it. They may have >test suite position with and without it. They've written articles about testing >procedures, but I can't recall having seen them publish a position that they >solve faster with singular extension than without it. > >>Amir I have the 1994 *socrates game bruce posted... and a few comments about his recollections from Hsu. Let me analyze this and perhaps we can study what they did there to see how programs would handle that... again that being an old deep thought version of course.. and not deep blue.
This page took 0.01 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.