Author: Robert Hyatt
Date: 21:15:44 02/25/02
Go up one level in this thread
On February 25, 2002 at 21:05:03, Christophe Theron wrote: >On February 25, 2002 at 00:02:50, Robert Hyatt wrote: > >>On February 22, 2002 at 06:36:45, Uri Blass wrote: >> >>>On February 22, 2002 at 05:43:49, Gian-Carlo Pascutto wrote: >>> >>>>On February 22, 2002 at 03:27:21, Yar wrote: >>>> >>>>>That's very famous position from Deep Blue-Kasparov match. The last played move >>>>>by Deep blue ra6?? was not correct. More interesting that in this position >>>>>Kasparov resigned !! He missed a draw by perpetual check. This position was >>>>>deeply analyzed by Karpov in his book (in 2000) >>>> >>>>I know, I know :) >>>> >>>>The point of Fritz's analysis is that a modern program CAN find the >>>>draw, which was believed to be impossible as it is very deep. >>>> >>>>-- >>>>GCP >>> >>>Bob Hyatt believe that it is impossible to find it. >> >>I don't "believe" it is impossible to find it from the Kf1 position, I >>_know_ it is impossible for a program to find this perpetual. I have >>given the reasons. I think Ed (at one point) had the analysis on his >>web site which gave a couple of the critical variations which included >>the "quiet moves" on the end. Those quiet moves are the problem... And >>they are a _big_ problem. >> >> >> >> >>>I did not. >>> >>>Hyatt also talked about the position some plies earlier but I think that in this >>>case every top program is going to find Kh1 after a long search and not Kf1 >> >> >>I can flip a coin to choose between the two moves. But if I can't see that >>Kf1 leads to a perpetual while Kh1 does not, then choosing either move for >>any other reason is wrong. And is based on luck, not chess skill... >> >> >> >> >>> >>>The evaluation after the alternative Kf1 may not be a draw but the main reason >>>for it is the fact that programs do not choose later deeper blue moves and they >>>believe that Ra6 is a mistake. >>> >>>Every top program today can expect the draw line of deeper blue inspite of the >>>fact that the evaluation is not going to be a draw(in the worst case after long >>>search and in the best case after a short search of at most few minutes) >>> >>>Only deeper blue could not see it and expected trading queens >>> >>>If you combine it with the fact that there is no impressive move of deeper blue >>>in the games then you can understand the impression that deeper blue's search >>>was inferior relative to the search of the top programs of today. >>> >>>I also do not think that you need to search something close to 60 plies even >>>without extensions in order to see the draw. >>> >>>You need to see more than 40 plies forward but not 60. >> >>You _definitely_ need to follow one key line to 60 plies or 30 full moves, >>because that is how deep the final perpetual is. _all_ lines don't have >>to be searched that deeply, but the _critical_ one does. That is the line >>that breaks _everybody_ because they reach a position with no checks or >>captures, and say "ok, nothing going on here, do the evaluation" and they >>find a score of some sort, without noticing that if the right "quiet move" >>is played at the critical point, the opponent can't avoid the perpetual. >> >>But even if it were only 30 plies, trying to search _through_ a couple of >>quiet moves breaks even the best chess program. Just look at how they follow >>deep series of checks and conclude "draw" when it isn't... or they find a >>mate in 30 but can't find the shorter mate in 15 because the first 6 moves >>are non-checks in the shorter mate. > > > >I'm not writing this just to contradict you Bob. I believe you are probably >right that current chess programs do not see the draw for the good reasons, and >that the couple of quiet moves breaks current programs' searches. > >However I would like to point out that I have made numerous experiments with >searches that did not simply use the concept of QSearch in a simplistic way >(like most programs including Crafty and Fritz do) and in these experiments the >couple quiet moves mentionned above would *not* break the search at all. > >One of these versions was actually so close to the efficiency of a simple >QSearch and so spectacular in the very deep lines it was able to spot almost >*instantly*, that I was about to make it the standard Chess Tiger version. > >Finally I refrained, but I can tell you for sure that some search algorithms >would not be blocked at all by the quiet moves in the drawing sequence. > >If you want to see something similar, install Genius 5 and set it to the >"instant" level. Play a game and try to figure out how the program is from time >to time instantly able to come up with 12 plies best lines. Have a closer look >at the best line, and you will see that they are not composed only of non-quiet >moves. > > > > Christophe I studied Genius very carefully a long while back because it was quite amazing on slow machines. However, I _never_ saw variations with 30+ moves followed by a couple of quiet moves, followed by 10 more moves that might be checks or captures. I can't personally imagine any search rules that would pull that off without totally blowing up the tree on all the _other_ positions where a couple of quiet moves followed by a lot of checks was simply wasted space. I have seen lots of programs have quiet moves in the middle. But those are usually inside the first 20 moves or so at best (20 plies). Singular extensions can drive this out a bit, but at a cost... I think that for Fritz, it is simply a matter of not knowing which side is winning and making a bogus assumption about that...
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.