Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: For the Record .. Deeper search & Eval Mate in 14

Author: Tony Werten

Date: 23:08:15 12/04/05

Go up one level in this thread


On December 04, 2005 at 06:17:18, Uri Blass wrote:

>On December 04, 2005 at 04:49:55, Tony Werten wrote:
>
>>On December 03, 2005 at 12:27:56, Robert Hyatt wrote:
>>
>>>On December 03, 2005 at 11:12:23, chandler yergin wrote:
>>>
>>>>New game
>>>>[D]3Q4/bp3pp1/6k1/8/7R/6p1/B2p1PP1/q1rn2K1 w - - 0 1
>>>>
>>>>Analysis by Shredder 8:
>>>>
>>>>1.Qd3+ f5 2.Qxg3+
>>>>  +-  (#14)   Depth: 10/10   00:00:00
>>>>1.Qd3+ f5 2.Qxg3+
>>>>  +-  (#14)   Depth: 12/12   00:00:00
>>>>1.Qd3+ f5 2.Qxg3+
>>>>  +-  (#14)   Depth: 12/12   00:00:00
>>>>1.Qd3+ f5 2.Qxg3+
>>>>  +-  (#14)   Depth: 12/12   00:00:00
>>>>1.Qd3+ f5 2.Qxg3+
>>>>  +-  (#14)   Depth: 12/12   00:00:00
>>>>1.Qd3+ f5 2.Qxg3+
>>>>  +-  (#14)   Depth: 12/12   00:00:00
>>>>1.Qd3+ f5 2.Qxg3+
>>>>  +-  (#14)   Depth: 12/12   00:00:00
>>>>1.Qd3+ f5 2.Qxg3+
>>>>  +-  (#14)   Depth: 12/12   00:00:00
>>>>1.Qd3+ f5 2.Qxg3+
>>>>  +-  (#14)   Depth: 12/12   00:00:00
>>>>1.Qd3+ f5 2.Qxg3+
>>>>  +-  (#14)   Depth: 12/12   00:00:00
>>>>1.Qd3+ f5 2.Qxg3+
>>>>  +-  (#14)   Depth: 12/12   00:00:00
>>>>1.Qd3+ f5 2.Qxg3+
>>>>  +-  (#14)   Depth: 12/12   00:00:00
>>>>1.Qd3+ f5 2.Qxg3+
>>>>  +-  (#14)   Depth: 12/12   00:00:00
>>>>1.Qd3+ f5 2.Qxg3+
>>>>  +-  (#14)   Depth: 12/12   00:00:00
>>>>1.Qd3+ f5 2.Qxg3+
>>>>  +-  (#14)   Depth: 12/12   00:00:00
>>>>1.Qd3+ f5 2.Qxg3+
>>>>  +-  (#14)   Depth: 12/12   00:00:00
>>>>1.Qd3+ f5 2.Qxg3+
>>>>  +-  (#14)   Depth: 12/12   00:00:00
>>>>1.Qd3+ f5 2.Qxg3+
>>>>  +-  (#14)   Depth: 12/12   00:00:00
>>>>1.Qd3+ f5 2.Qxg3+
>>>>  +-  (#14)   Depth: 12/12   00:00:00
>>>>1.Qd3+ f5 2.Qxg3+
>>>>  +-  (#14)   Depth: 12/12   00:00:00  26kN
>>>>1.Qd3+ f5 2.Qxg3+
>>>>  +-  (#14)   Depth: 12/12   00:00:00  26kN
>>>>1.Qd3+ f5 2.Qxg3+
>>>>  +-  (#14)   Depth: 12/12   00:00:00  26kN
>>>>1.Qd3+ f5 2.Qxg3+
>>>>  +-  (#14)   Depth: 12/12   00:00:00  26kN
>>>>1.Qd3+ f5 2.Qxg3+
>>>>  +-  (#14)   Depth: 12/12   00:00:00  26kN
>>>>1.Qd3+ f5 2.Qxg3+
>>>>  +-  (#14)   Depth: 12/12   00:00:00  26kN
>>>>1.Qd3+ f5 2.Qxg3+
>>>>  +-  (#14)   Depth: 13/13   00:00:00  41kN
>>>>1.Qd3+ f5 2.Qxg3+
>>>>  +-  (#14)   Depth: 14/14   00:00:00  75kN
>>>>1.Qd3+ f5 2.Qxg3+
>>>>  +-  (#14)   Depth: 15/15   00:00:00  153kN
>>>>1.Qd3+ f5 2.Qxg3+
>>>>  +-  (#14)   Depth: 16/16   00:00:00  247kN
>>>>1.Qd3+ f5 2.Qxg3+ Kf6 3.Qd6+ Kg5 4.Rh5+ Kxh5 5.Qxd2
>>>>  +-  (#11)   Depth: 17/18   00:00:01  596kN
>>>>1.Qd3+ f5 2.Qxg3+ Kf6 3.Qd6+ Kg5 4.Rh5+ Kxh5 5.Bf7+ Kg5 6.Qg3+ Kf6 7.Qg6+ Ke5
>>>>8.Qe6+ Kf4 9.g3+ Kg4 10.Qg6+
>>>>  +-  (#14)   Depth: 18/54   00:07:23  166686kN
>>>>1.Qd3+ f5 2.Qxg3+ Kf6 3.Qd6+ Kg5 4.Rh5+ Kxh5 5.Bf7+ Kg5 6.Qg3+ Kf6 7.Qg6+ Ke5
>>>>8.Qe6+ Kf4 9.g3+ Kg4 10.Qg6+ Kf3 11.Qxf5+
>>>>  +-  (#14)   Depth: 19/55   00:07:28  169067kN
>>>>1.Qd3+ f5 2.Qxg3+ Kf6 3.Qd6+ Kg5 4.Rh5+ Kxh5 5.Bf7+ Kg5 6.Qg3+ Kf6 7.Qg6+ Ke5
>>>>8.Qe6+ Kf4 9.g3+ Kg4 10.Qg6+
>>>>  +-  (#14)   Depth: 20/53   00:07:35  171711kN
>>>>1.Qd3+ f5 2.Qxg3+ Kf6 3.Qd6+ Kg5 4.Rh5+ Kxh5 5.Bf7+ Kg5 6.Qg3+ Kf6 7.Qg6+ Ke5
>>>>8.Qe6+ Kf4 9.g3+ Kg4 10.Qg6+ Kf3 11.Qxf5+
>>>>  +-  (#14)   Depth: 21/56   00:07:42  174826kN
>>>>1.Qd3+ f5 2.Qxg3+ Kf6 3.Qd6+ Kg5 4.Rh5+ Kxh5 5.Bf7+ Kg5 6.Qg3+ Kf6 7.Qg6+ Ke5
>>>>8.Qe6+ Kf4 9.g3+ Kg4 10.Qg6+ Kf3 11.Qxf5+
>>>>  +-  (#14)   Depth: 22/36   00:07:49  178326kN
>>>>1.Qd3+ f5 2.Qxg3+ Kf6 3.Qd6+ Kg5 4.Rh5+ Kxh5 5.Bf7+ Kg5 6.Qg3+ Kf6 7.Qg6+ Ke5
>>>>8.Qe6+ Kf4 9.g3+ Kg4 10.Qg6+ Kf3 11.Qxf5+
>>>>  +-  (#14)   Depth: 23/38   00:08:03  184315kN
>>>>1.Qd3+ f5 2.Qxg3+ Kf6 3.Qd6+ Kg5 4.Rh5+ Kxh5 5.Bf7+ Kg5 6.Qg3+ Kf6 7.Qg6+ Ke5
>>>>8.Qe6+ Kf4 9.g3+ Kg4 10.Qg6+ Kf3 11.Qxf5+
>>>>  +-  (#14)   Depth: 24/40   00:08:26  194168kN
>>>>1.Qd3+ f5 2.Qxg3+ Kf6 3.Qd6+ Kg5 4.Rh5+ Kxh5 5.Bf7+ Kg5 6.Qg3+ Kf6 7.Qg6+ Ke5
>>>>8.Qe6+ Kf4 9.g3+ Kg4 10.Qg6+ Kf3 11.Qxf5+
>>>>  +-  (#14)   Depth: 25/43   00:08:56  207920kN
>>>>1.Qd3+ f5 2.Qxg3+ Kf6 3.Qd6+ Kg5 4.Rh5+ Kxh5 5.Bf7+ Kg5 6.Qg3+ Kf6 7.Qg6+ Ke5
>>>>8.Qe6+ Kf4 9.g3+ Kg4 10.Qg6+ Kf3 11.Qxf5+
>>>>  +-  (#14)   Depth: 26/43   00:09:43  229005kN
>>>>1.Qd3+ f5 2.Qxg3+ Kf6 3.Qd6+ Kg5 4.Rh5+ Kxh5 5.Bf7+ Kg5 6.Qg3+ Kf6 7.Qg6+ Ke5
>>>>8.Qe6+ Kf4 9.g3+ Kg4 10.Qg6+ Kf3 11.Qxf5+
>>>>  +-  (#14)   Depth: 27/41   00:11:01  263041kN
>>>>
>>>>(, MyTown 03.12.2005)
>>>> Which is what I said; deeper search confirms.
>>>>I see no Bug, the Program found a Mate in 11 in .01 seconds
>>>>I cut the analysis short. My error, not the Program.
>>>
>>>
>>>You may not see the bug, or you may not understand the bug.  But it is
>>>absolutely a bug.
>>>
>>>Pick up _any_ AI textbook.  Alpha/Beta is _guaranteed_ to return the same score
>>>as a pure minimax search, only far faster.  If a minimax search says "mate in
>>>11", and there is no mate in 11 present, then that is a bug, nothing more,
>>>nothing less.  Many programs will find a mate that is sub-optimal (longer) but
>>>if one finds one that is shorter than can be forced, it is simply a bug.
>>
>>Why would checkmate scores have to be exact ?
>>
>>You can choose to return (heuristic)checkmate scores from eval, with some bound
>>on iteration depth.
>>
>>
>>Tony
>
>I agree that in theory checkmate score do not have to be exact but I think that
>practically  when programs say shorter mate than the real mate it is because of
>bugs.
>
>Note that usually evaluation is only an estimate for the real score and if a
>program says 0.00 it does not mean that it is a draw.
>
>A program may have a function to see wins in pawn endgame(I do not have it) and
>it may evaluate the following position as mate in 20 for white during the
>search.
>
>[D]8/p3k3/8/8/8/8/PPPPK3/8 w - - 0 1
>
>The reason that I believe that no program does it is because I see no practical
>advantage from doing it because even if you evaluate it as winning and use it
>for pruning you can give it a bonus of 50 pawns and have a rule to prune lines
>when the evaluation is more than 50 pawns advantage unless you already found
>advantage of more than 50 pawns for one side and you search for mate.

I know at least 2 programs using it.

Giving 50 pawns bonus is based on the same idea. The problem is what to do with
different wins. Add a rook for white, and score would go up but how much ?

Returning a checkmate score, where the distance to mate is based on the
difficulty of the win is a lot easier.

When searching, iteration depth will go up, and real checkmates pop up, because
they are the only ones having a dtm below or equal to the iteration depth.

One big point of attention though: You should first check wether the move leads
to a draw by repetition before you return the heuristic checkmate score. Else
you might allow a 3 fold repetition in a heuristicly won position. Very
embarrasing when that happens at, oh say, a world championship.

Tony

>
>
>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.