Author: Vincent Diepeveen
Date: 11:26:20 05/11/98
Go up one level in this thread
On May 11, 1998 at 12:01:00, Robert Hyatt wrote: >On May 11, 1998 at 10:15:15, Vincent Diepeveen wrote: > >> >>On May 11, 1998 at 09:46:09, Amir Ban wrote: >> >>>On May 11, 1998 at 08:41:11, Vincent Diepeveen wrote: >>> >>>> >>>>On May 11, 1998 at 07:04:23, Amir Ban wrote: >>>> >>>>>On May 10, 1998 at 18:51:35, Vincent Diepeveen wrote: >>>>> >>> >>>>> >>>>>This position was analyzed much more deeply than this. >>>>> >>>>>After 36.Qb6, Rd8 is indeed best, but DB did not consider it but >>>>>36...Qe7 (I posted the complete analysis recently). After 36... Rd8 >>>>>37.axb5, a micro will quickly see that white is in trouble, but after >>>>>36...Qe7 37.axb5, black is a full pawn worse because of the need to >>>>>protect the bishop on d6. Justifying 36.axb5 if you do not consider Rd8 >>>>>is more than a few tenth of a pawn to justify, actually it's about a >>>>>full pawn. >>>>> >>>>>Why didn't DB consider Rd8 ? Probably it saw 36. Qb6 Rd8 37.Be4 ! and it >>>>>seems black is screwed. But black has a fantastic resource: 37... a5! >>>>>38.axb5 axb4!! sacing a piece, to get the queen to the first rank and >>>>>force a draw on perpetual threats (echo of the final position, but more >>>>>complicated). >>>> >>>>It is clear that DB didn't see the piece sac, because this line is way >>>>longer >>>>than the 23 ply (Diep needs) to see that Kf1? leads to a draw and Kh1! >>>>wins the game. >>>> >>> >>>Right. A human may play it on a hunch that things will work out, >>>specially if he's desperate. >>> >>> >>>>In the line you posted i see that DB score doesn't get to zero, but >>>>just goes down few tens of a pawn. So that'll be some king safety, >>>>no doubt. >>>> >>> >>>King safety no doubt, but how much ? This was already discussed here >>>last year. >> >>If you look to the final position that Db prints out, then you see that >>the king ain't covered by pawns. In the past i gave this terrible >>penalties. >> >>Also something i still do is: giving freepawns that are covered by a >>pawn >>huge bonuses. It is clear in this game that DB doesn't do this (the fact >>that it played Nf5! is my proof, when i remove some bonuses for >>pawncovered freepawns, then Diep plays Nf5 too). >> >>The both of 2 gives you this result. >> >>>>I was surprised seeing in the lines you posted here that DB just got 11 >>>>ply. >> >>>>Although i admit that when i would search fullwidth with SE and all kind >>>>of check/threat extensions i doubt whether i would get more with >>>>around 70 Billion nodes. >>>> >>> >>>Yes. Actually the log says iteration not ply, but my guess is that this >>>is the brute-force ply depth. We are spoiled by null-move depths, but DB >>>don't do that, and they don't believe in pruning either. I think a micro >>>with full width search + regular extensions would just barely reach 8 >>>ply in this position. 11 ply is slightly lower than I expect, but >>>remember this is a tough position for DB with the eval going up and down >>>like crazy. Some people suggest you should add 4 ply for the leaf >>>processors, but arithmetic tells me that 15-ply brute force is >>>stretching credulity. >> >>Well if i allow diep to search a night with nullmove turned on, then it >>gets more like 18-20 ply here. >> >>The better your evaluation, the better your branching factor, is the >>experience with Diep, although objectively a branching factor should >>be independant from the evaluation. Guess it has to do with >>that you have a more accurate score for leaves, where programs not >>getting this, have q-search problems: another ply changes score from >><= alfa to >= beta. >> >>So if they say that after this 11 ply they get another 4 ply from the >>processors, then this still says nothing. I get 18 ply easily after a >>night, >>and no extension makes up for that extra positional insight. >> >>Also in all these positions you can correctly nullmove, because there >>are >>no threats that cannot be seen, assuming you do checks in your q-search. >> >>>Amir > >This is baloney. I suppose you are ready to sit down and play someone >like Karpov a match, and thoroughly trounce him? Because "his selective >search extensions can in no way compensate for the extra plies of >full-width >you get over him?" >That's silly. To characterize Crafty, I generally hit 12-13 plies in Your comparision is not fair. We cannot compare human intelligence with computer intelligence, you know that very well. If Anand needs 200 nodes to see a deep combination, then we can bet that it might take DB tens of billions to see it, or more. So we cannot compare a selective search of a computer with that of a human, so 11 ply of DB will NEVER in my life in these positions make up for 18 ply of DIEP, positionally spoken. >the >middlegame of a 40/2 game, using my 4 processor machine. I search to an >average depth during those 12 ply iterations of maybe 18-20 plies. This >counts the many paths that are extended, plus the captures tacked on to >the end. If you extend several stupid lines, which makes the average depth deeper than what you search right now. >DB goes about 10 plies deeper on average. Still 11-12-13 plies >depending >on the position, but their extensions are *much* more significant >fraction >of their total search than mine are. Since you haven't played them >ever, >your comments are forgivable. But once you have played against them, >you >realize *just how much more they are seeing than you do*... by watching >when *they* fail high, then seeing how many full moves before *you* fail >high. Remember the bishop trap against Cray Blitz in 1989, where they >saw >it 10 full *moves* before CB did. And in 1989 CB was *not* slow by any >standard of measure. This is not true. If they search 11 ply, and if i search 11 ply then positionally our results are veyr well comparable. If they are so stupid to extend some threats, then they see tactically little more, but i get 2 additional plies for that at least. With 200M nodes x 180 seconds = 36Billion nodes or something i search easily 20 ply. If they search fullwidth, then YES they search less, around 11 ply it appears. Now don't tell us that DB is the only program in the world that selectively sees the same which only Humans see. That is of course not true. If they search 11, then they search 11 ply. That's it. If i get that 11 ply with some tricky pruning, then i might miss something, but if i get 18 ply, then you can forget that i miss something. ALL variations that make sense of DB are of course contained in that. Besides, all proof is also indicating in that direction. >In Cape May, *Socrates was out-searching them by 1-2 plies, yet when it >got "interesting" they saw something maybe 10 moves before *socrates* >did, >because they were searching deeply along the *right* paths. They search >deeply along lots of wrong paths to be sure, but they have the >horsepower >to get away with that where we can't... I'm asking for PROOF. Positions with moves, and their output and score. We can get masses of info out of Principal variations. We all know the weak evaluation function of some parallel searchers. When Cilkchess played here at my home at 1 processor a testmatch against Diep (rapid chess), in a vague position suddenly diep claimed +5.xx, and it took Cilkchess many many moves before they even realized that they were in danger. Diep searched about 6 ply, and it appeared to be king safety which caused this score. When i play with Diep at the internet i very much see that my evaluation is totally different from the opponent. Sometimes diep is correctly a pawn up in evaluation compared to the opponent. No vague arguments Hyatt, i presented proof why DB is a simplistic program, which mainly relies on speed, now YOU need to present some proof, not some vague arguments, like "Somewhere at Mars there is a program playing stronger than we all can imagine, because they're faster and way smarter". The way you present the DB program is the opposite of what they're doing. You present them as the smartest PC program which coincidentely is also way faster than ours, where the facts say that they're fast and stupid, and positionally not searching deeper than we do, and probably they're tactically not that strong like we think too. In fact when i remember it well, then you presented them first as having 1000 adjustable parameters. When i replied that this is hardly more than the piece square tables in rgcc, they suddenly had 8000 parameters. Weird. For some reason i still think they have 1000 parameters and no single parameter more, because moves like Bxg6?, h6? Qa5? Bd6-c7? are all beginners faults. To kick an open door, suppose they are stupid, but tactical strong then why if they solve for example all Nolot test set positions, then why didn't they publish it? Is good commercial for IBM too. Somehow i missed these logfiles. It's easy to attach one processor to a computer (they've done it before), give it a different name, and let it search some positions. >But they are *most impressive* when you sit across from them and watch >*their* output and *your* output. And you realize just how "small" your >search tree is compared to theirs... Yes they search more stupidiness, that's very clear now. DB searches 99.9999% stupid lines which are not relevant, Diep searches 99.99% stupid lines which are not relevant. This is how the facts are. They played bunches of bad moves. Every game they play average FIVE bad moves. The average GM who loses a game plays 1-3 bad moves. They played few strong moves, which as appears are found easily by most programs, or easy to see why they find it and we didn't. Their endgame is relatively good, which is just a matter of hard work what has been done. Also their freepawn evaluation seems very well debugged. Greetings, Vincent
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.