Author: Vincent Diepeveen
Date: 11:59:15 05/11/98
Go up one level in this thread
On May 11, 1998 at 14:00:13, Robert Hyatt wrote: >On May 11, 1998 at 10:06:33, Vincent Diepeveen wrote: > > >>What i don't understand of Hsu is: if they can connect Deep Blue's >>processor >>to a main frame, why can't they add it to a PC? >> >>Processors are just a little bit of sand, nothing more. If you have >>already >>made more than one processor sucessfully, then it should be easy and >>cheap to mass produce it, and put it at a cheap PCI-card. >> >>I mean few million nodes a second in leafs, that is interesting to play >>against! >> > >You've already answered your question. "PC". the DB processor is one >thing, then the circuit board that holds these processors and interfaces >to the SP2 is another thing. In the case of DB, this "interface" is a >VME- >bus design, which has nothing to do with the incredible array of PC >standards, from ISA, EISA, PCI, etc... So *someone* would have to stop >and design a (probably) PCI circuit board for the chess processor, which >would take time. And the PC platform is "ever-changing" from 66mhz to >75mhz to 83 mhz to 100mhz... so which bus speed do they target? > >you get the idea... a real pain... > > > >> >>You're missing the whole point. I've written my fingers blue, so many >>bad and really horrible beginner moves are made by DB. Did you miss >>80% of the msg i wrote? Main part is below this line. >> >>And NO Bob, even with some errors in 8000 evaluation terms it >>cannot happen that it plays Bxg6. > > >sure it can. All it takes is one overweight king-safety term that says >that the missing h-pawn is really bad. And it might be that all the >modifications they did in the last 3 months before the match led to such >unexpected moves. But they can be tuned back out... Can't be in this position. h6 can be explained this way, but bxg6 not, because the black king is not at g8, but at e8, and if they search few plies then it's already at c8. >and I still continue to point out that they *beat* Kasparov with their >"bad" moves. It would be interesting to see how some of your "better Kasparov did average 5 horrible moves a game, which is way more than the 1-3 horrible moves average GM play, in fact i make usually less than 5 horrible moves a game. I'm not even IM. >suggestions" would pan out... particularly since any program can find >one or two better moves in some positions, but to find good moves for >the whole game is quite a bit different. What would you put DiepX's >chances at against Kasparov? I'd put mine at "0". So Deep Blue >obviously >does *something* right because they have won more than one game against >him which means that they are *not* playing "garbage"... Kasparov has learned his lesson now, but when i look at the underestimation of Kasparov which must have taken part in game 2 then i think my own program would have easily won that spanish game 2. In fact the Ra3 manoeuvre gets already found within 40 seconds at my P133 laptop i saw last Saturday, with Frans Morsch and Johan de Koning as eye-witness. So if i take my laptop to Kasparov, then this should be enough, assuming he'd play the same he did against Deep Blue. >IE who else has beaten (or drawn) Kasparov in a 40/2rh game? Against Deep Blue the human Kasparov made more errors than the entire summation of errors over the past 5 years he made against humans. >>ESPECIALLY with 8000 evaluation terms it cannot happen, because >>the other say 7800 evaluation terms will remove the fault after several >>tens of billion nodes. >> >>I'm having even more terms in Diep, and sometimes it makes weird moves >>because of evaluation faults, but the deeper it searches the more this >>gets corrected by the other evaluation terms, backupped by the search, >>and corrected because of search like >>some programmers might want to do believe us. Currently I'm debugging >>and expanding existing, and this really brings a lot. >> >>I'm doubting the knowledge that is in DB. It makes so many simple >>faults which are explained by lacking easy knowledge, that it makes >>me laugh a little, although i can forget sponsoring of >>chesscomputerevents forever because of >>that weak IM Kasparov. >> >> >>2 dubious moves are more than enough to lose a game. I've played >>myself games where my opponent or i myself (just little over 2220) lose >>because of 1 dubious move; a weak openingspositon like computers >>put at the board is already enough. >> >>And yes, kasparov played so horrible that GM v/d Wiel last Saturday, >>after explaining that Kasparov plays tactical himselve like a computer >>and >>didn't use his maingun: openingspreparation, >>called Kasparov's performance: "weak IM level". >> >>Fact is that DB didn't play 2 dubious moves a game, more like 5 dubious >>move AVERAGE a game. (See ICCA june 1997 the report of Seirawan) >> >>If FIVE horrible moves a game ain't enough to lose a game against a >>seriously playing and GM, then i eat my hat. >> > >I'll pass the salt for you. DB has won more games against Kasparov >that all the other programs put together. By a factor of *three* in >fact... We're analyzing the only thing DB has left besides some wild speculations by you. What DB has left can be analyzed. After analyzing it appears it made horrible mistakes which most PC programs don't make. See ICCA june 97. Did you ever read that article? Questionmark after questionmark. >>This is not true. PV is like Amir published based on 11 ply searches, >>and PV no more than 13 half moves at that depth. >> > > >please reread my explanation of two facts: (1) they construct the >PV from the hash table, which means it is often "short" because things >get overwritten before the PV is reconstructed when the search ends. >(2) their hardware does 4 plies of non-captures (at least) plus the >captures/etc... and this returns *no* PV at all... I construct the PV from hashtable too. And with probe > 4 it hardly gets overwritten, unless you have some bugs in your hashtable. >So at the very best, you can see their PV less the last 4 non-captures >+ captures, and if the hardware is following checks or whatever, that >"hidden" part of the PV can be quite deep. The SP2 is only doing about >7-8 plies of search before handing off these "near-leaf" positions to >the hardware. So at *best* you can see about 2/3 of the PV, At worst, >a small part... This is not true. I'm using probe=8, you are already for a long time telling the whole world how easy it is at these mainframes and supercomputers to use a huge probe. With a huge probe it never gets overwritten. The memory is huge at a mainframe compared to the number of tuples they need to store, so they chance they overwrite PV is near to zero. >>This is exactly what i experienced myself: if you search 11 ply then >>positionally >>you simply see not much deeper. Definitely not 30 ply. >Then you simply haven't done it right... Because selective extensions >do work. They work better and better as you have more hardware and can If they do work then why is DB the only program in the world where it works? I remember that hundreds of people have tried it, and no one got it to work, and suddenly DB got it to work? You told us about a year ago that DB team would publish some astonishing stuff. Still waiting for that publication. >afford non-important extensions, since it is difficult to tell the >difference >between those that are useful and those that are not, in a given >position. >They can search zillions of futile ones, to pick up that one very >important >one, and they *still* out-search us... zillions is not near the truth. they search around 36B nodes from which at least 18B gets wasted, then the processor is so fast that they cannot store this in shared memory, so we get few billions finally which they search. So if i allow my program to search a billion nodes with nullmove in a position where nullmove can be used (and that are all positions played that match), then my program sees way more than they did. That's very logical. >>Of course, mate in 15 ain't no problem, but let's discriminate between >>TACTICS, and positional depth. >singular extensions is not only about tactics. It can be about >positional >extensions just as easily. It depends on the singular margin you think >you >can live with. Most use big numbers to keep cost down. They are not >so constrained... This algorithm already goes wrong when a move is not singular. Singular moves are good to see some deep threats like mate. In opening for example after 1.d4,d5 2.e4?,dxe4 3.Nf3 this means that they cannot use this algorithm because you can cover e4 in more than 1 different way. So singularism is not the way to solve chess. Especially singularism is a tactical extension, just like chess. It is not positional. It has to do with threats where you only have 1 move to prevent that thread. If i threaten you, and you can only undo my threat in 1 way, then no matter the value i need to have only 1 way to undo that threat. Singular extensions is converting alphabeta back to minimax. A big problem of singular extensions is detecting them. To detect them i need to search all siblings which in fact i don't want to search because i want to give a cutoff here. searching it with a reduced depth is like a minimax algorithm where you search the biggest part of the tree with a reduction factor, but only a small reduction compared to the depth of the tree. >> >>Now don't make me laugh by saying that they use smart extensions: >>If you KNOW what to extend, then why extending, you can play the best >>move at once without searching. > > >You miss the whole point. They can *afford* "dumb" extensions. so that >if only one of every 100,000 extensions they try produces something >useful, >they can afford that. *we* can't... we'd be doing 4-5 ply searches. Yes they can afford to search minimax, but they won't search deeper than 11 ply then as it appears. >>>I disagree with the "clearly lacking knowledge".. It drew an ending I'd >>>bet every program around would lose. And it found ways to keep things >>>interesting... round 1 was a lucky win by Kasparov... one or two tempi >>>and things turn totally around... >> >>Can you base this on evidence, like what moves are so hard to find for >>our pc programs at analysis level (to compensate a little >>for their fast hardware and get more than 11 ply)? >> >>Fact is that DB did a bunch of bad moves, which for the major part are >>not done by commercial programs. >yes... it also did a bunch of *good* moves, most of which are *not* done >by commercial programs... Let me repeat the question to you WHAT moves are so well from DB? All ! moves (a4!,Bf5!) of DB in the 6th game for example are done very quickly by almost all PC programs. So any pc program would have won this game quickly.
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.