Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: The Null Move Killer Killed (and an announcement)

Author: Stuart Cracraft

Date: 13:09:07 10/01/04

Go up one level in this thread


On October 01, 2004 at 14:41:48, Robert Hyatt wrote:

>On October 01, 2004 at 12:21:51, Stuart Cracraft wrote:
>
>>On September 30, 2004 at 20:00:57, Robert Hyatt wrote:
>>
>>>On September 30, 2004 at 18:28:51, Stuart Cracraft wrote:
>>>
>>>>On September 30, 2004 at 18:04:05, Robert Hyatt wrote:
>>>>
>>>>>On September 30, 2004 at 14:25:34, Stuart Cracraft wrote:
>>>>>
>>>>>>On September 30, 2004 at 09:35:09, Robert Hyatt wrote:
>>>>>>
>>>>>>>On September 30, 2004 at 02:53:16, Stuart Cracraft wrote:
>>>>>>>
>>>>>>>>The null move killed, win-at-chess 141, has itself
>>>>>>>>finally been killed, vanquished with the help of
>>>>>>>>two board contributors whose combined suggestion
>>>>>>>>led to a 17-fold reduction in time-to-solve.
>>>>>>>>
>>>>>>>>This posting announces those winners. First the
>>>>>>>>stats!
>>>>>>>>
>>>>>>>>Now solved in 5.49 seconds on a P3 @ 1ghz it would be
>>>>>>>>solved in under 2 seconds on more modern equipment.
>>>>>>>>Formerly it took 95 seconds to solve.
>>>>>>>>
>>>>>>>>That's good enough for me. And it's good enough to win
>>>>>>>>the $50 contest posed recently since it broke the
>>>>>>>>10-second-and-under-barrieras posed in the contest
>>>>>>>>posting.
>>>>>>>>
>>>>>>>>The search:
>>>>>>>>
>>>>>>>>Alpha=-1332 Beta=-531 Maxdepth=9999999 MaxTime=99999
>>>>>>>> 1/ 9  g2f1  0.00 -953      511 g2f1 f4d5
>>>>>>>>                                g2f1 f4d5
>>>>>>>> 2/ 9  g2f1  0.01 -953      884
>>>>>>>>                                g2f1 f4d5 c1g5
>>>>>>>> 3/12  g2f1  0.06 -953    11929
>>>>>>>>                                g2f1 f4d5 c1g5 d5f6
>>>>>>>> 4/16  g2f1  0.39 -953    72781
>>>>>>>>                                g2f1 f4d5 b3d5 c6d5 f1g2 d6e7
>>>>>>>> 5/24> g2f1  3.83 -552   978925
>>>>>>>>                                g2f1 b5b4 b3a4 f4d5 f6g5 d5e7
>>>>>>>> 5/25  c1f4  5.49 2260  1420038 c1f4 d6f4 h4h5 g6h5 h1h5 f4h6 h5h6 c7g3 g2g3 d7d
>>>>>>>>6
>>>>>>>>                                c1f4 d6f4 h4h5 g6h5 h1h5 f4h6 h5h6 c7g3 g2g3 d7d
>>>>>>>>6
>>>>>>>> 6/25  c1f4  6.06 2260  1519145
>>>>>>>>                                c1f4 d6f4 h4h5 g6h5 h1h5 f4h6 h5h6 c7g3 g2g3 d7d
>>>>>>>>6
>>>>>>>>
>>>>>>>>And with it the announcement -- because of the contribution
>>>>>>>>of Will Singleton in indicating that null move should be
>>>>>>>>avoided before leaves in the main search (and the sense
>>>>>>>>of a comparison in an old commented out section of the
>>>>>>>>code associated with disabled null move verification having been
>>>>>>>>intended to do what Will suggested but having been miscoded
>>>>>>>>by me and then #ifdefed out months ago) and Uri Blass'
>>>>>>>>comments about my recaptures being too free and easy,
>>>>>>>>the program went from a total of 95 seconds
>>>>>>>>for wac 141 to 5.49 after these two suggestions were
>>>>>>>>implemented.
>>>>>>>
>>>>>>>I doubt null-move is the problem.  I do null-move _everywhere_ and Crafty has no
>>>>>>>problem solving wac 141 doing so...
>>>>>>>
>>>>>>>
>>>>>>>>
>>>>>>>>So Will and Uri are the winners, if they wish to accept,
>>>>>>>>of the divided $50 prize. Because Will's contribution was
>>>>>>>>more significant but less work for him and Uri's contribution
>>>>>>>>was less significant but with more work for him, but in either
>>>>>>>>case without the change from the other's suggestion the result
>>>>>>>>would not have been as dramatic getting down to <= 10 seconds
>>>>>>>>as stated in the earlier contest challenge a day or two ago,
>>>>>>>>the award has been divided in half for the 2 winners.
>>>>>>>>
>>>>>>>>Will and Uri are welcome to send me, and only if they wish
>>>>>>>>to collect, their postal mail addresses, to cracraft@cox.net
>>>>>>>>and a check for $25 will be sent out to each.
>>>>>>>>
>>>>>>>>In the future, more contests will be held like this whenever
>>>>>>>>I run into a huge roadblock but I see none looming presently,
>>>>>>>>including a rather unusual one that I am not ready to announce.
>>>>>>>>
>>>>>>>>Thanks everybody for the help on 141 -- and thanks to Will
>>>>>>>>Singleton and Uri Blass.
>>>>>>>>
>>>>>>>>Stuart
>>>>>>
>>>>>>What is your quiescence like? Do you investigate moves-that-check
>>>>>>at the first ply of quiescence?
>>>>>
>>>>>
>>>>>My q-search has _no_ checks or check-evasions whatsoever.  Just captures, and
>>>>>the captures have to appear to be at least equal using SEE or they get discarded
>>>>>as well...
>>>>
>>>>What if a capture is a check or check evasion? Acceptable?
>>>>
>>>>Stuart
>>>
>>>
>>>Yes, but Crafty doesn't notice this nor handle it differently as they are "just
>>>captures" in the q-search...  q-search doesn't detect mate stalemate or draw at
>>>all either...
>>
>>How about check? I assume you don't hand off an incheck position to
>>quiesce, saving the function call and keeping it investigated by
>>the main search for one ply extension only at all places in the tree.
>
>That is why I extend on _giving_ check rather than on _escaping_ check.
>Therefore I will _never_ drop into the q-search with the side on move in check,
>unless I have extended so far that I don't extend the check a full ply...  There
>I just accept the errors since most likely a capture of the checking piece will
>be the only valid move anyway...
>
>
>>
>>But in quiescence what if you end up in check? What then?
>
>I ignore it.  If I can capture the checking piece, or if the king can capture
>any piece to move out of check, it'll do so.  Otherwise it will have to stand
>pat...
>
>
>>
>>Stuart

Based on this post I disabled my check-evasion extension in my
quiescence and my main search and instead instituted in main search
only an extension on checking move. Seemed to produce a lot fewer
nodes, but in about the same time and only very slightly worse
results on WAC (<1% worse). M

My routine to determine whether incheck is expensive.

I will play with this extension for a few days and see if any other twiddle
can get it doing better given the overhead.

Stuart




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.