Author: Uri Blass
Date: 02:25:19 04/03/02
Go up one level in this thread
On April 03, 2002 at 05:02:27, Sune Fischer wrote: >On April 03, 2002 at 02:45:54, Uri Blass wrote: >>>> >>>>No >>>>I do not think that low rated humans have better >>>>positional understanding than computers. >>>> >>>>There are positions that they understand better than computers but there are >>>>positions when the computer evaluation function is superior and in most of the >>>>positions that practically happen the computer's >>>>positional evaluation is better. >>>> >>>> >>>>Humans may overevaluate or underevaluate some advantage when >>>>computers often know better to give the right weights >>>>for the different factors in the evaluation >>> >>>Absolutely not. >>>If we talk only of a static eval, then it is much harder for a program to >>>"understand" when a knight is better than a bishop. >>>The problem with programs is they add so many small things together and do not >>>understand the critical point of a position, e.g. they don't understand the >>>fight over a dominating square, humans know when a passed pawn is strong and >>>when it is weak, >> >>It is not truth and there are cases they think they know only to find later that >>they were wrong. >> >>There were cases when I had a wrong evaluation of the position during a game and >>only after I came home to analyze the position with my chess program it could >>understand the position better by it's static evaluation and I am clearly better >>than 1600. > >I suspect you are not really talking about static eval the way I do. >Did you run you program on a position and let it search only 1 node?? I do not rememebr thew relevant position now but I remember that in at least one of my games I believed that had an advantage when I had no advantage and the computer knew immediatly that I have no advantage(I could not tell it to search 1 node but I could see that even at small depths it does not see an advantage for me. > >If the computers had as good a static eval as humans, they would surely be >completely unbeatable because they only eval() at the leafs where some noise can >be accepted. >Even the programs that have hundreds of specific cases programmed are very crude >in estimating the specific position, when I play blitz with Fritz7 it wants to >play c4 in this position: >[D]rnbq1rk1/pp3ppp/3bpn2/2pp4/3P1P2/2PBPN2/PP1B2PP/RN1QK2R b KQ - 0 7 > >however if i play O-O instead of Bd2, fritz wants to play b6 or Qb6. >I think it is clear that fritz does not understand that playing c4 will close >the queen side, and be much harder for black to use his terrain when there is a >blocking pawn on c4, it is a positional error IMO. White obviously never >intended to play on the queenside, so the black pawn will protect white. >Fritz *understands* the position better when I castle, probably because the eval >parameters are calibrated by assuming I would castle. >I can even get fritz to advance the a and b pawns, a complete waste of time, but >to fritz the pawns are worth more the closer they get to queening. >[D]3r1rk1/4nppp/3bp2q/pp1p1b2/2pPnP2/2P1PN2/PP2BNPP/R1B1QRK1 b - - 0 20 >Any human here would probably play b4! it is clear that 20...a4 21.a3 will just >close the position completely - fritz played a4. >That is not a 2700 level player, even a 1400 level player would either not move >the pawns until more pieces had been activated on the queenside, or just gone >for b4 straight away. I doubt if the moves of Fritz are significant positional errors. >Fritz killed me later on the king side, I can't hold the tactics, the best I can >do is to get fritz to play 20 nonsense moves, but it always finds a tactical >shot. If it finds a tactical shot then it means that maybe the moves are not nonsense. There is a simple rule in chess that when your position is really bad you have no opportunity for a tactical shot. > >You cannot put every single individual case in an eval, often it is good to >block a bishop, but there always will be exceptions. Think of the knowledge a GM >has, if we could put all that into the eval, and still run at a good kns, the >program would be very strong indeed. > > >> this is all very hard to put in a static eval. Much of this is >>>just intuition, if we can't even put that stuff into words, then how are we >>>supposed to program it? >>>I imagine there must also be great problems in programming what kind of endgames >>>that win or lose. This is something you need to know way in advance before going >>>into the endgame. >> >>I agree that programs are relatively weak in understanding endgames that are not >>tablebases postions but the main problem in evaluation is evaluating the middle >>game and I think that the top programs are often better than humans in >>evaluating these positions(I am talking about players of strength 2000 and not >>about GM's but you also talked about low rated players in your post). >> >> Complex pawnstructures can only be approximated in a static >>>eval, the GM will have a much more intuative feel about what is drawn and what >>>is won. >>> >>>IMO programs get by with the use a tactics, you never see them sac. >> >>Not truth. >> >>There are programs that sacrifice material for positional gain. >> >>Junior,Gambittiger are 2 names of programs that do it and even more >>materialistic programs do it. >> >>In the israeli league Rebel refused to play a combination that wins a pawn >>against a GM because it prefered the positional advantage. >> >>I do not think that the decision was right but the point is that programs >>sacrifice > >Of cause you can make it so, it is easy go give a higher score for positional >gain, but programs do not know when it is correct to sac, you said yourself you >disagreed with the program :) Humans also play sometimes wrong positional sacrifices. I gave the wrong move only as an example to prove that programs can sacrifice. Both humans and machines may be wrong or right in sacrificing but if I give cases of a right sacrifice you may claim that maybe the program saw everything in advance so it is not a sacrifice(something that is not always correct). Uri
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.