Author: Robert Hyatt
Date: 21:20:08 01/19/00
Go up one level in this thread
On January 19, 2000 at 21:56:48, Bruce Moreland wrote: >On January 19, 2000 at 18:00:09, Robert Hyatt wrote: > >>I don't remember any problems. I found Be4 after a 21 ply search that took >>almost 2 days. Your program found h5 that Kasparov said no program would play. >>I didn't (myself) spend a lot of time on axb5 but didn't see anything remarkable >>about that.... >> >>So what strange move can't we explain or produce? > >I don't remember this. I remember that Be4 had approximately the same score as >the other move, but I don't think that anyone "found" it with their stock >program. > >This whole thing is some serious bad deja vu. If Amir hasn't said it yet, he's >going to say that the move he's interested in is axb5, not Be4. We went through >this all before, at least twice. > >Mine couldn't find either of the DB moves but I don't think that it matters. >The initial suggestion was something to the effect of: "How could a computer do >that! It's giving up a chance to win three pawns!" But I do know that the >score is not anywhere near +3 if you search long enough, and I don't think it is >unbelievable that a computer could do what DB did. The value of the position >closes on zero if you search long enough, and I'm willing to believe that DB >burped somewhere on the way to zero, and out came something a little strange. > >Certainly this is more believable than the alternative, which is some deep >conspiracy to cheat. That's just irresponsible, I think. I can't believe that >some human analyzed some other line, got insanely worried that the computer >would fall into some awful tactical line, cheated, and then covered it up. If >you've been working on something like this for a long time you aren't going to >lose confidence in its ability to spot tactics, and you certainly aren't going >to overrule it on a tactical issue based upon your own analysis, especially when >you can't verify your analysis using DB, since it is playing. > >bruce I never got anywhere with axb5... but I did get Be4 at depth=21, after a day or two. What I did was run on two machines, one searching Be4 _only_ and the other searching Qb6. And at depth 21 Be4's score was a tad higher than Qb6 when it got to depth=21. Qb6 was not _draw_. Both were positive. I only concluded that Be4 was 'possible'. At least with that version of crafty, at depth=21, with that particular evaluation code. Which might have had plenty of errors and bugs for all I know. axb5 was another issue, but I had thought that the general consensus was "OK, king safety could explain this." From my perspective, it doesn't matter whether _any_ move can be reproduced by a PC, since no PC is the same class of chess engine as DB. So what if they can not find what it found. I reported Cray Blitz's WAC results before. All solved in under 3 secs... and only 2 took anything other than 0 seconds (where 0 is obviously anything < 1 no matter how close to 1). I haven't seen _any_ program that could do that. I tried wac230 on that 16 cpu alpha last summer, and it took over 20 seconds. So it is 'close' maybe with crafty on hardware that might be fairly close to the hardware we used for CB 10 years ago (the C90). But I would not expect anyone else to solve those that quickly. And I wouldn't expect anyone to say that a human was helping it. Larry Kaufman can tell you about how fast CB ate up a tactical test he gave it... Micros were taking minutes... he was disturbed about the 0 second times we were getting, and wanted more accuracy, as 0 times were screwing his rating prediction formula. So if we can't produce every move of DB, or if we can't produce _any_ move of DB, it still doesn't follow that the only explanation must therefore be cheating. At least it doesn't follow for most of us. "some" are another story altogether, of course. It's about as sensible as saying that the Dodge Viper really isn't the fastest American-made production car. That Chrysler is really cheating to break 190mph. They are cheating ok... they are using a huge V10 with a bunch of horsepower... just like IBM did with DB. There are some places where 'cheating' explanations just don't fly. This is one of them. IMHO of course.
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.