Author: Uri Blass
Date: 23:11:49 09/12/01
Go up one level in this thread
On September 12, 2001 at 22:58:27, Robert Hyatt wrote: >On September 12, 2001 at 18:43:39, Dann Corbit wrote: > >>On September 12, 2001 at 17:57:42, Uri Blass wrote: >> >>>On September 12, 2001 at 17:04:13, Dann Corbit wrote: >>> >>>>On September 12, 2001 at 16:07:45, Uri Blass wrote: >>>>[snip] >>>>>I think that your program understands the position better than >>>>>Deeper blue. >>>>> >>>>>It is more logical to assume that Deeper blue is wrong and not that >>>>>all the top programs are wrong. >>>> >>>>Try all the top programs on this one: >>>>2b5/1r6/2kBp1p1/p2pP1P1/2pP4/1pP3K1/1R3P2/8 b - - bm Rb4; id "WAC.230"; >>>> >>>>6 billion votes "AYE!" >>>>1 vote "NAY." >>>> >>>>The one vote may be right, and the 6 billion wrong. >>>>[snippity-snip] >>> >>>I remember that there was a discussion about this position and it was not clear >>>if Rb4 wins. >> >>It is unclear. It is the only move with winning chances, but it is possible it >>is only a draw. There are other test problems as well that no chess engine can >>answer correctly. >> >>>I do not say that the majority is always right but I tend to believe the >>>majority if I have no evidence that it is wrong and in the case of the draw >>>against kasparov in game 2 I tend to believe that programs see 0.00 some plies >>>after the root for the right reason. >> >>It is certainly possible that modern chess programs might do better in some >>positions. Certainly not on all of them, especially those plagued by null move. > > >I don't like _any_ of this part of the discussion. DB played in 1997. That >is only 4 years ago. I haven't seen any remarkable advances come along in the >last 4 years that would make a new program "modern" and a 4-year old program >"ancient".... > > > >> >>>program of today also can find Qe3 when programs like Genius3 could not do it >>>so I tend to believe that deeper blue's evaluation is similiar to some old >>>programs that cannot find Qe3 and not to the new programs. >> >>That is an interesting theory. >> >>>I also tend to believe that the new programs are better than the old programs >>>and the fact that they can find Qe3 (even without 0.00 evaluation) when old >>>programs like Genius3 or Deeper blue cannot do it is not luck. >> >>Genius 3 and Deep Blue are not comparable, in my opinion. We know that there >>were 4000 tunable parameters in the eval. Does *any* modern program have this? >> > > >I believe they claimed 8,000 at several presentations they made. I don't >have 8,000 eval terms, yet Uri thinks that I understand the position better >than DB. I can't follow that logic, yet... > > > > >>>The new programs are simply better. >>>They have better search rules and better evaluation function. >> >>I think that there are surely some improvements in searching and also in >>pruning. A 500-1000x improvement in speed cannot simply be discounted. >> >>Surely, the trees searched are very different in shape. In fact, they probably >>could have made a speculative version of Deep Blue that would search much >>deeper. However, given the horsepower at their disposal, saftey margins allowed >>them the luxury of much more forceful searching. This sort of decision could >>truly cause them to miss some solutions that could be found by deep searching. >>I have not seen any actual evidence of this, but it is certainly possible. > > >I haven't seen any either... > >For a point of reference, to test "old programs" I played thru several of the >1986 WCCC games as played by Cray Blitz. I used Crafty, on hardware that was >a bit faster than Cray Blitz of that time. Crafty found _zero_ tactical errors >by that "ancient program". And it found very few places where it disagreed with >positional moves where there was an important positional point at stake. > >So much for the "old/dumb" argument, IMHO. 1)If cray blitz won most of these games then it is possible that the opponent played so weak that Cray blitz did not have the opportunities to do blunders. 2)The point is that in a match of 6 games I expect to see a move that can be found but is hard to find. We have a match of 6 games between Heubner and Deep Fritz. Heubner also played quite lines against Deep Fritz. I believe that if you use old programs to analyze this match you are going to find moves that are hard to find. I have not time to analyze these match by other programs and I hope that someone will do it to prove that there are moves that all programs can find Deep Fritz's move but need time to do it and it is not obvious decision. 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.