Author: Vincent Diepeveen
Date: 07:06:05 12/12/02
Go up one level in this thread
On December 11, 2002 at 19:36:19, Sune Fischer wrote: >On December 11, 2002 at 12:10:48, Vincent Diepeveen wrote: > >>On December 11, 2002 at 07:44:43, Sune Fischer wrote: >> >>What i meant to say Sune is that Uri doesn't look further >>ahead than a politician. He has 200 lines now, with more >>stubbornness than the entire CCC together he then concludes >>that 2000 is enough. then he has 2000 lines and he has to >>get shot down and lose game after game before he will conclude >>that he needs 3000. After that he will conclude 4000. > >You clearly underestimate Uri. >I think Uri is far ahead of us in some areas :) >On one has made me rethink the basics as much as Uri. >I feel very tempted to try a non-incremental move generator for example. I remember a conversation not so long ago with GCP at the ICQ where we regurarly talk. He spoke about how strong Sjeng tactically was. I tested deepsjeng at my dual at some positions and indeed it found stuff very quickly. Only then i realized that a year or 10 ago when programs used to win world championships based upon a combination of tactics and positional insight and that now that tactics do not matter anymore, that the 'amateurs' are doing fine tactically with their engines. I remember a big test of DIEP in start of 2002 done by Jan Louwman. About 1000 games 40 in 2, so even Hyatt should recognize that as considerable more significant than his HT tests, where he hasn't even posted output in node counts of crafty together with search times and scores (without node count of course the whole HT concept is not proven at all; because my opinion about HT and crafty has to do with a few things which go outside of the context here). About 500 games were done with DIEP version paderborn 2002. and about 500 games with DIEP with a selective search. the selective version searched on average 2 to 3 ply deeper than the diep paderborn 2002 version. So i was hoping for a big score improvement. My regret was big when i figured out that the selective search version was getting 20-25% less in score than the paderborn 2002 version. Only then i realized clearly that the search depth of the selective version, which at for example 14 ply could prune up to 7 ply positional, that this version basically saw tactics deeper and not so much positional deeper (of course to get decent mainlines i toyed some but that's not representative for 99.999% of the lines). It has taken over half a year for me to conclude then that just searching tactical deeper makes no sense at all. It's about positional search depth nowadays, not tactical. Therefore i completely disbelief any theory that there are better search algorithms than the current ones existing with regard to computerchess. Of course nullmove we could improve perhaps here and there by smarter reduction factors and such. Of course hashtable usage could be made more efficient in the future. Of course perhaps we could order moves better with some cool feature not yet invented which corrects for the nullmove alfa problem (in fail soft the nullmove causes for <= alfa nodes to return a score near alfa always so you never know what is the best move there). But that's not changing the big picture i have currently. And that's that many things which for sure do not work anymore are getting tried by some big amateurs now. incremental move generation is one of them. Ever realized why no commercial program uses that anymore? Branching factor rings a bell? >You lose a faction of speed, but in return you have vital knowledge that you can >use to decide when to nullmove, when to extend, when to prune etc.. >I can't say off hand that the overall return is going to be negative, can you? >>But he'll never conclude the right thing at once. >And you think that you do? If the evidence is that big sure! >Actually, I think it won't be long before Movei is whipping the floor with Diep. >Then you can take all your fansy patterns, all your SMP and simply get >out-pruned and out-extended by 200 lines of Uri brilliancy ;) I want to bet for a big amount of money and i give 2 to 1 that in a match in 2004 the program Movei will not win a long match against diep at say 6 games 120 0 level? Of course it is a bit unfair to give the movei only 2 years to implement 2000 lines of code to improve his evaluation, but well 2000 lines of evaluation for a chess program i could possibly write within a week. I've got some tens of thousands to bet for. Lemme know when we can appoint things! >>Even research from 1950 he will skip. It says that grandmasters >>know about 100000 patterns. Now we can discuss whether those >>100000 is a real 100000 or a human 100000 which means that >>for the current computers it is more like a million than 100k >>patterns. >> >>But it sure isn't 200 lines :) > >Classic mistake IMO, you want to throw a lot of patterns into the evaluation >thinking this is how humans do it and so it must be good for computers too. >It is easy to prove you wrong, as no computer has 100000 patterns, yet they >still kick GM's butt. No it is easy to proof me right because all commercial programs and also strong amateurs have put in a lot of extra chess knowledge the last years. I do not understand why you do not realize that. I am sure even Uri can proof this to you very clearly. They do not deny it even. Instead at computerchess events, where i by the way never see your face, they loudly discuss it. Brutus is an ultimum try from chessbase of course to create a program with a lot of knowledge. they put the knowledge in hardware which causes it to evaluate way quicker than i can do in software. So brutus is a very dangerous program. It gets millions of nodes a second at the loss of perhaps a factor 2 or 3 (because it has no hashtables in hardware; note that to limit the inefficiency in hardware it just searches 2 or 3 ply in hardware and not 4 or 5). I do not understand why people still are busy with 200 line evlauations here when the truth is so clear here. >I think the key to good chess programs is to be found in clever extensions and >solid pruning. IMO evaluation comes in third. > >>That Uri could conclude of course as he saw it himself, but even >>toying with Shredder doesn't get him to better conclusions. I wonder >>why with 2000 rating he misses that all games in the world top >>computerchess aren't dominated by tactics at all, but decided >>by anything *but* tactics. >> >>Of course the too weak last few participants not taken into account >>of the world champs 2002. I remember the game Diep - (some american >>program in the first round) which even made a 2 ply mistake. >> >>Of course you need to see those tactics, but even the thing from Uri >>should see those tactics. I wonder how he can conclude that he just >>loses tactical and that 2000 lines should fix everything for everyone. > >Everything is tactical, tactics govern chess. >If there is a danger, you can 1) try and evaluate this some ad hoc way, or 2) >extend here and figure out what is really going on. >I see no reason why 1) should beat 2). > >>Saying 'might' is not a good excuse. I'm not a lawyer. I see how people >>say statements. Just covering your ass from lawyer viewpoint is not >>what a discussion group is meant to be. >> >>You say a statement and having an escape route is not what was meant here. >> >>The statement as i read it was that 2000 lines is enough. > >Words don't mean much to you, Uri is different here. He is very precise in his >formulations, like a mathematician. He says what he means and not some variation >of it. When he says he believes it is possible to make a top program using only >2000 lines for eval, then that is what he means, not "you guys are so stupid >with all your evaluation, I am sure I can make a top level program using only >2000 lines for evaluation". That may be how you read it, but there is a >difference, quite a big one actutally.... > >>Fritz with 2000 lines of course is called fritz5.16 and >>doesn't make a chance in any world champs from 2001 to the next >>so many years. > >I don't believe that for a second, fritz 6 and 7 probably has many improvements >in faster code and better search. >No one, except perhaps you, would work entirely on the evaluation. > >>Just keep a match fritz5.16 versus shredder. remove the games that >>got decided on book. You'll see a clear picture. And it won't be >>tactics at all. Every idiot can see that, a 2000 player should. >> >>If not then i don't know how to call the person behind that 2000 >>player, but it won't be positive words. > >Most of the engine games I see are primarely tactical. >I seriously doubt you will be badly beaten if you are tactically superior. >I think Shedder is also outsearching fritz5.16. >Of course you can't tell this by looking at some meaningless ply number, because >it is _what_ they search that matters, not how deep they can search stupid >lines. > >-S. >>Best regards, >>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.