Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: Diep fullwidth vs Deep Blue fullwidth

Author: Robert Hyatt

Date: 19:58:32 12/21/99

Go up one level in this thread


On December 21, 1999 at 18:31:58, Vincent Diepeveen wrote:

>On December 21, 1999 at 17:44:06, Robert Hyatt wrote:
>
>>On December 21, 1999 at 17:02:53, Greg Lindahl wrote:
>>
>>>On December 21, 1999 at 16:18:27, Albert Silver wrote:
>>>
>>>>That's correct, he says it was done through software: "During the 1997 match,
>>>>the software search extended the search to about 40 plies along the forcing
>>>>lines, even though the nonextended search reached only about 12 plies." He also
>>>>mentions that "The software portion of the search can be arbitrarily selective
>>>>without slowing down the system."
>>>
>>>If you read the beginning of that paragraph, Hsu explicitly says that the 8
>>>plies of software search included forcing. Hsu doesn't say if the final 4 plies
>>>of hardware search included forcing by droping back to software or not. Given
>>>that the chess chips seem to operate in an embarrassingly parallel fashion, I
>>>would suspect that there was no forcing for those plies. Someone could always
>>>ask Hsu...
>>>
>>
>>
>>
>>Old news.  The first 4 plies (+ whatever extensions were used) were done on a
>>single SP processor.  The next 4 plies + whatever extensions were triggered were
>>done in parallel on the SP, which (if stated simply) says that the first 8
>>plies, plus all the extensions, are done on the general-purpose SP hardware.
>>The _final_ 4 plies, plus the capture search were done on the chess processors.
>>The chess processors _did_ do extensions, but not singular extensions.  IE Ken
>>Thompson did the usual in-check and recapture extensions in Belle, and the
>>first deep thought (chiptest) chip was nothing more than "belle on a chip".
>>
>>Also, chess is _far_ from "embarassingly parallel".  It is one of the more
>>difficult-to-program parallel algorithms, because alpha/beta is a strictly
>>defined sequential algorithm.  Doing it in parallel invites a lot of extra
>>work that can't be avoided.
>
>I couldn't say it better Bob!
>
>>>How important is forcing in shallow plies verses deeper plies? That's easy to
>>>examine using a program.
>>>
>>>-- greg
>>
>>
>>Hsu would _like_ to have been able to do singular extensions in hardware.  But
>>there was simply not enough space on the chip as things get _very_ complex
>>compared to a simple alpha/beta hardware design...
>>
>>But you have to ignore some of Vincent's ramblings about DB's search depth.  I
>>once posted a position where they found a forced win of material OTB vs Cray
>>Blitz, as but one example of their extreme tactical strength.  _NOBODY_ found
>>that win OTB, or overnight.  Many liked the move, but _nobody_ saw the tactical
>>consequences that were forced.  They have done this _many_ times over the
>>years.  So I'd say their "11-ply search" is _far_ better than our 14-15 ply
>>searches, no questions asked...
>
>Bob, i explained that the c5 move was because of a horizon effect in deep blue.
>Diep had the same horizon effect too when i added the extension.

No, you explained _your_ programs failing there.  I _saw_ their PV and eval.
It was +3.  Not +1.1 or something like that.  It saw the bishop being taken
in the PV output.  Remember that I was sitting there, talking with Murray and
Hsu during the game, when they failed high.  Murray commented "DT seems to want
to pick on your bishop, but I don't see why that would fail high unless there is
a trick I don't see."  A bit later (less than a minute) Murray said "Hey, it
is winning your bishop outright..."  Our score was still good.  It remained
good for at least another 5 moves before it started to drop.  And by the time
10 moves had passed we were at -3.

But they saw it from the _beginning_.






>
>Secondly an optimistic evaluation function or a program doing the Rxa4
>in the quiescencesearch is gonna find things a lot faster than a smart
>qsearch that is not doing Rxa4 in the qsearch.
>
>Before this gets a blindfolded discussion, first here the position:
>
>black timeleft=27:46.40.00
> r = - = - b k =   Qa7-e3   20    Rf8-a8
> = - = - = r o o   c2-c3    21    Qb4-b7
> - B o o b o - =   Rf1-f2   22    Qb7-a7
> = - = - n - = -   Qe3xa7   23    Ra8xa7
> N = - = - = - =   Bf4-e3   24    Ra7-a5
> = O O - = - = -   Be3-b6   25    Ra5-a8
> O = B = - R O O   Bd3-c2   26    Be7-f8
> = - = - R - K -   Rd1-e1   27    ...
>white timeleft=27:46.40.00
>black to move
>
>Note that it's smarter to get the last few moves for repetitions,
>that speeds up the search proces *considerable*, as last few moves
>were not exactly the most exciting moves.
>
>r4bk1/5rpp/1Bppbp2/4n3/N7/1PP5/P1B2RPP/4R1K1 b - -
>
>the main idea in this position is that after c5 the bishop on b6 is more
>or less hung. Now this doesn't mean that black sees the forced win,
>because seeing a forced win here means +5.0.
>

I count a bishop as 3.  Their eval was over +3.  They _saw_ it.




>What is the problem here? Well the big problem for nowadays smart programs
>is that after a couple of moves there goes a rook to b7 and
>white for some reasons must move a rook to b-file. all bishops are
>gone then from their places which allows Rxa4 bxa4 Rxb1 in quiescencesearch.
>However not all programs play Rxa4 in qsearch.
>
>Just get away the c2 bishop and the b6 bishop and put the white rook to
>b2 and the black rook to b7. Not many programs will play Rxa4 here in
>quiescencesearch, as that's covered by a pawn, meaning an exchange gets
>lost.
>
>The score of Deep Blue here is however based upon a horizon effect which
>i posted a couple of time ago. This doesn't take away that *all* programs
>will play c5 here with a good score for black. Nowadays DIEP is very
>optimistic in such endgames too. Let's see whether i can get the same
>high score with the normal version of DIEP here searching fullwidth...

I certainly can't get it with Crafty.  And I didn't get it with Cray Blitz.
No one else that tried this last year got it either...  although several got
the right move, but way wrong eval/PV.



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.