Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: Nolot Bxh7 and deep thought/deep blue

Author: Vincent Diepeveen

Date: 11:33:42 01/12/00

Go up one level in this thread


On January 12, 2000 at 12:55:34, Robert Hyatt wrote:

>On January 12, 2000 at 10:02:08, Vincent Diepeveen wrote:
>
>>Let's have another thread about the Nolot positions.
>>>
>>>Here are my results as of now:
>>>
>>>   DT     Ferret
>>>          4x450
>>>   ------ ------
>>>1  6hr
>>>2  2min   31sec
>>>3
>>>4  2.5hr
>>>5  2hr
>>>6
>>>7  6hr
>>>8
>>>9  9min
>>>10 2min   11sec
>>>11 5min   48sec
>>
>>Let's zoom into position #8 not found by Deep Thought
>>
>>r3rbk1/ppq2ppp/2b1pB2/8/6Q1/1P1B3P/P1P2PP1/R2R2K1 w - - bm Bxh7; 8
>>
>>This position is of course in nolot as PC programs don't extend such
>>lines too deeply. Just like the other nolots it only has to do with
>>king safety and a lot of forced moves around the king.
>>
>>[c] ==> check extension
>>[s] ==> singular best move extension
>>[f] ==> forced move (a forced move is singular too) extension
>>
>>the TACTICAL line needed to find Bxh7:
>>
>>Bxh7+     Kxh7[c]
>>Qh5+      Kg8[c]
>>Rd4[s]    gf6[f]
>>Rg4+[s]   Bg7[c]
>>Qh6[s]    Kf8[f]
>>Rxg7      Rec8
>>Qxf6      Be4[f]     (otherwise rh7-h8 mate)
>>Rae1[s]   Bf5[f]
>>g4[s]     Bxc2[f]
>>Rxe6[f]   Bd3
>>Rg6[c]    Qc1+
>>Kh2[c]    Qc7+
>>f4[c]     Qc2+
>>Kg3[c]    <any legal move>
>>Qh8 mates in quiescencesearch.
>>
>>So totally 10 plies don't carry an extension and therefore need
>>to be seen by brute force search depth.
>>
>>Note that diep is already positive for white after Rxe6, but
>>score is not high enough for white to change to Bxh7 in the root,
>>as diep is *very* happy anyway with this position.
>>
>>Diep evaluates after Rg6 the position as +1.598 for white.
>>
>>Now you can tell me a lot of crap about singular extensions, but
>>if you search 7 ply and use singular extensions as described by Hsu,
>>then you need just 7 ply to find Bxh7.
>
>Vincent, this is showing that you do _not_ understand Singular Extensions.
>
>Your definition is whenever one move is clearly better than the others.  DB's
>definition is when one move is better than the others, but the others are
>searched to a much shallower depth than normal.
>
>That changes things.  I have a version of Crafty that uses SE.  Mike Byrne
>ran it on ICC a lot.  It found some things faster, but more were actually
>slower, and I haven't 'accepted' that result yet.  PV singular moves are
>very accurate extensions...  which means that if your best move is going to
>win something, you will likely see it when it is searched at the start of an
>iteration.  Fail-High singular moves are much harder to handle, because you
>have to search every branch at a fail-high position when you would normally
>stop after the fail-high move.  This turns the tree into pure minimax, which
>can't be tolerated.  So, to prevent this, the moves that are normally pruned
>by the fail-high are, instead, searched to reduced depth.  And this will miss
>some singular cases...

I know *exactly* the difference. I wrote down some time ago a long
article how big the overhead is of non-PV singular extensions versus
others. the moves i marked as singular or forced will be even with a
near to single ply search singular in this position.

>
>
>>
>>If Deep Thoughts eval is a bit worse than DIEP's, like mainly counting
>>pawns, then it needs an additional few ply. It needs another 3 ply
>>to see the checking sequence, so it totally needs 11 ply then to
>>see the Bxh7 trick with a score of over +10.0.
>

Bob,
you said yourself that deep thought was very materialistic compared to
deep blue.

So i apply your description of the evaluation of the
machine. Deep blue as played against
kasparov obviously is using big eval terms. Where here
deep thought with singular extensions is the subject.

It says 'dt' in the table, so i assume deep thought.

>Your comment is funny, because in one thread you claim they have way large
>king safety terms, then here you assume they  have very small positional
>terms.  It can't be both ways, of course.
>
>
>
>>
>>If it doesn't do checks in quiescencesearch
>>then it needs 11 ply totally to find this combination.
>>
>>Vincent
>
>
>We are talking about deep thought here, as that was the results Bruce posted.
>We know it didn't do checks in the q-search.  That was an error Hsu thought
>needed fixing in later versions when DB was actually designed.  But all of those
>results were from the old, original, Deep Thought hardware, _not_ Deep Blue.
>DB was far different...



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.