Author: Robert Hyatt
Date: 06:44:38 05/22/00
Go up one level in this thread
On May 21, 2000 at 23:40:46, blass uri wrote: >On May 21, 2000 at 21:19:23, Robert Hyatt wrote: > >>On May 21, 2000 at 19:56:01, blass uri wrote: >> >>>On May 21, 2000 at 19:02:26, Robert Hyatt wrote: >>> >>>>On May 21, 2000 at 18:49:00, blass uri wrote: >>>> >>>>>On May 21, 2000 at 17:02:02, Marcos Christensen wrote: >>>>> >>>>>>On May 20, 2000 at 07:11:31, Terje Vagle wrote: >>>>>> >>>>>>>1r6/5kp1/RqQb1p1p/1p1PpP2/1Pp1B3/2P4P/6P1/5K2 b - - >>>>>>> >>>>>>>After 45. Ra6, Fritz suggests Qe3 for black and evaluates the position as 0,94. >>>>>>>It does not seem to find the famous draw-line for Kasparov. >>>>>>>10 hours analysis on PIII-600, and 28006383 KN evaluated >>>>>>> >>>>>>>Does any other program find the draw-line? >>>>>>> >>>>>>> >>>>>>>Regards >>>>>>> >>>>>>> >>>>>>>Terje >>>>>> I thought that even deeper blue didnt saw the draw! Maybe I'm wrong >>>>> >>>>>Deeper blue did not see the draw but it does not change the fact that people >>>>>expect other programs to be better than deeper blue. >>>>> >>>>>I believe that the estimate that program will need 10 years to get to the level >>>>>of deeper blue is wrong because deeper blue did serious search errors like not >>>>>seeing the draw and the fact that other programs do the same search error is >>>>>going to change in the future. >>>>> >>>> >>>>This will _never_ change. Programs will _always_ have "search errors". I don't >>>>think a program will find that draw in the next 10 years. The reason is pretty >>>>obvious... to find the quiet queen move deep in the tree means the tree is going >>>>to be _huge_ since many other "quiet" (but very useless) moves will be searched. >>> >>>I know that part of the program can find the draw if you let them to play >>>against themselves. >>> >>>You need only to let programs to play against theirselves them and learn from >>>the games in order to find the draw. >> >>I don't believe this will work. The tree is huge. And _every_ pathway has to >>be followed to prove the thing is a draw. Just because two programs start at >>the position and end up in a draw does not mean (a) the position really is a >>draw; (b) that they will find the draw OTB as who knows if white really puts >>up the best resistance based on a bunch of shallow searches tacked on to each >>other end-to-end. > >In the case of the last game after you go forward,backward enough times the >evaluation will be draw because of learning. True... but what if it takes only 25,000,000 games to do this? > >White has only a small number of lines when programs believe that it >has advantage(otherwise people could not prove by chess programs that it is a >draw). No... white has a large number of lines that have to be proven bad one at a time... by the time you get way out beyond 30 plies, when you can really only search to half of that, you have a _lot_ of tree left to work on. > >After going forward and backward enough time programs can learn that it is a >draw. Again I agree. But I believe our definitions of "enough" are way different. IE I know a person that tried to have crafty solve "wild 7". This is a king and three pawns vs king and three pawns ending with a well-known way to force a win by white. White has pawns at a2/b2/c2, black has pawns at f7/g7/h7. White's king is at d1. Black's king is at e8. White to play and win. He played thousands of games. learning by "position" as crafty does it, is hurt by something known as "a local maxima" which stifles learning beyond that point. This position is way easier to solve than the DB position, yet it seems impervious to learning approaches. Ande search. Yet it is so easy for a human, once he understands the idea. (Hint: it is all about zugzwang). > >You need to decide about algoritem for the number of moves in the games against >yourself(it is not logical to do long games when the moves are not forced) but I >am sure that it is doable because I proved it in a similiar way. > > > > >> >> >> >>> >>>programs do not use even 10% of their time for doing it and I believe that it is >>>a mistake at long time control and that it is going to be a mistake at >>>tournament time control because the hardwrae is getting better. >>> >>>I also do not agree that the move Qc1 is a quiet move. >>>It is not a check but it is easy to see that this move threat to take the >>>bishop. >> >> >> >>Not to a computer it isn't easy. Just seeing checks are hard enough... a >>non-check that deep is not going to make it past any reasonable move selection >>algorithm... The search is going to have to be able to reach that position >>where Qc1 is required, and still have a full-width ply left so that the move >>is tried... And _then_ it has to see deep enough _beyond_ that move to see that >>the draw is forced. >> >>That is a truly daunting task for a computer. Maybe not for a human. But for >>a computer... it is difficult enough to almost be considered impossible. >> >> >> >>> >>>A real quiet move do not have a threat to take material in the next ply. >> >>In any position on the board, almost, the side on move can make a move that >>threatens a piece of the opponent. How to weed those out (to prevent the >>tree from exploding) while keeping the Qc1 move _in_??? >> >> >> >> >> >>>> >>>>Hang on to the position for a while. Let's see when it is really solved (IE >>>>the program won't play the Rook move because it sees the forced draw.) >>>> >>>> >>>>>We will never know when programs are at the same level of deeper blue because >>>>>deeper blue had the advantage of being unknown when it played against kasparov >>>>>when other programs will not have the same advantage(kasparov will not agree to >>>>>play them in the same conditions. >>>>> >>>> >>>>We will have a fair idea.. when a program can play reasonably level with >>>>Kasparov. >>>> >>>> >>>> >>>>>The first program+hardware to win kasparov must be clearly better than deeper >>>>>blue. >>>>> >>>> >>>>How/why?? >>> >>>The reason is that I believe that kasparov is not going to repeat the mistake of >>>playing against something unknown. >>> >>>I believe that Deeper blue could win kasparov in a match because of the fact >>>that kasparov could not see games of the same machine and could not buy >>>something similiar to the same machine to learn about the weaknesses of deeper >>>blue. >>> >>>Uri >> >> >>Wouldn't he have the same problem with Crafty? I change it all the time. >>Wouldn't he have the same problem playing Anand? He would go into a tank for >>several months to plan some new opening strategies. And didn't he find a >>possible weakness in one game, only to find the weakness 'repaired' in a later >>game? Nothing says a program must be "static" any more than there is a rule >>that says a human can't "learn" or "prepare"... > >He will not have the same problem with crafty because the source code of >crafty is known and even if you do some changes that are not known he can know >almost everything. > >In the case of the game against deeper blue he simply could see nothing about >the machine. > >No games with something close to be the machine. > >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.