Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: right now, at this very moment wac 141?

Author: Robert Hyatt

Date: 21:01:02 09/30/04

Go up one level in this thread


On September 30, 2004 at 23:02:42, Uri Blass wrote:

>On September 30, 2004 at 18:19:01, Robert Hyatt wrote:
>
>>On September 29, 2004 at 19:54:50, Stuart Cracraft wrote:
>>
>>>On September 29, 2004 at 18:34:40, Uri Blass wrote:
>>>
>>>>On September 29, 2004 at 18:20:45, martin fierz wrote:
>>>>
>>>>>On September 29, 2004 at 16:21:43, Uri Blass wrote:
>>>>>
>>>>>
>>>>>>I think that it is clear that extending every nate threat by a full ply will
>>>>>>blow your search.
>>>>>>
>>>>>>If you do not use partial extensions then I suggest that you do not extend mate
>>>>>>threats unless  you have some conditions to extend them only near the root and
>>>>>>not every where.
>>>>>>
>>>>>>Uri
>>>>>
>>>>>i extend mate threat by a full ply always. and my search isn't blowing up, at
>>>>>least not that i notice. then again, WAC 141 shows some strange behavior with my
>>>>>latest version:
>>>>>
>>>>>MUSE 0.89.10 UCI 30 MB:
>>>>>
>>>>> 1   00:00 -1.65   Kg2g1
>>>>> 2+  00:00 -2.36   Kg2f1
>>>>> 2   00:00 -2.36   Kg2f1 Nf4d3
>>>>> 3   00:00 -1.22   Kg2f1 Re8b8 Kf1e1
>>>>> 4   00:00 -2.13   Kg2f1 Kg8h7 Bb3d1 Nf4d3
>>>>> 5   00:00 -2.16   Kg2f1 Re8e2 Kf1g1 Nf4d5 Kg1f1
>>>>> 6+  00:00 -1.66   Qc1xf4
>>>>> 6   01:24  M6  Qc1xf4 Bd6xf4 Rh4xh5 g6xh5 Rh1xh5 Bf4h6 Rh5xh6
>>>>>
>>>>>it sees Qxf4 after 0 seconds (this is on a slowly 1.4GHz P4), but needs a very
>>>>>long time to resolve it. but at least it sees Qxf4 in 0 seconds :-)
>>>>>
>>>>>cheers
>>>>>  martin
>>>>
>>>>If you extend also checks by  a full ply then there are cases when there is a
>>>>long sequence
>>>>check ,escape threat mate,check,escape threat mate and it means that you may
>>>>have big problems to search deep in position when one side threats mate but the
>>>>opponenthas a lot of checks because he tries to do perpetual checks.
>>>>
>>>>You can easily search some lines to more than 100 plies at small depth and  I do
>>>>not believe that it is a good idea.
>>>>
>>>>Uri
>>>>
>>>>
>>>>Uri
>>>
>>>This is very useful. Okay, in main search I can just start off with
>>>revaluing check extension to 0.75 instead of 1 but I'd expect that
>>>since I have few extensions, very few checks would be searched.
>>>How would you handle reducing check extension value in such an
>>>arrangement.
>>>
>>>Also, my quiescence search always investigates all check evasions
>>>to any depth. Should I be doing that?
>>
>>Absolutely not.  IE you capture at first q-search ply where you can also stand
>>pat.  Your opponent captures at the next ply and you try all moves to get out of
>>check and eventually find a deep mate.  It isn't forced as you can just stand
>>pat at the first capture and all that searching was wasted...
>
>The searching was not wasted unless standing pat is better because you learned
>that the capture is not good.
>Let take 2 examples
>
>[D]r5k1/6pp/4b3/7q/8/4R3/5PPP/N2B2K1 w - - 0 1
>
>If you search check evasion you can find that BxQ is bad and return the stand
>pat score that is bad for white.
>
>[D]r5k1/6pp/4b3/7p/8/4R3/5PPP/N2B2K1 w - - 0 1
>
>In this case of course you do not search after Bxh5 RxN+ because there is no
>check evasion to do the score above alpha and alpha in the beginning was the
>standpat score or something higher.
>
>Uri


I've said this hundreds of times...  I don't want my q-search to deal with that
at all.  Whether it does only captures or captures+checks+check evasions is
_still_ way too selective.  I can construct just as many positions where the
best move is not a capture or a check, and the q-search returns the wrong value.
 So I prefer to reduce the q-search to a minumum and hopefully let my normal
search go a ply or two deeper and handle that stuff correctly by looking at
everything, not just a very selective subset of moves...

My own personal preference of course...  but WAC141 gives me absolutely no
probelms either...



This page took 0.02 seconds to execute

Last modified: Thu, 07 Jul 11 08:48:38 -0700

Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.