Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: Deep Blue Language

Author: Robert Hyatt

Date: 08:21:06 01/13/05

Go up one level in this thread


On January 13, 2005 at 01:56:23, Tony Werten wrote:

>On January 12, 2005 at 16:45:20, Daniel Márquez Lisboa wrote:
>
>>Hello Tony! Thanks. I would appreciate that you send me the paper they published
>>in "Artificial Intelligence".
>
>Done
>
>>
>>Yes, I have not read that book, and I don't know where get it.
>
>Proboably out of the bookshelves already. A too small audiance. The book is more
>about Hsu's experiences anyway than about technical details. I think the paper
>is more usefull for you.
>
>The paper contains some nice "hidden" details as well wich you might read over.
>
>One of them is an explanation about parameters that can be send to the hardware.
>
>One of them is "Depth of offset searches to detect singularity" wich (as opposed
>to what has been discussed here ) indicates that Deep Blue used the "cheapo" way
>of detecting singularity. ( else it would not need this parameter, because it
>could use "depth")

This is obviously talking about SD in _hardware_ not in software.  So far as I
had heard, at least through DB1, no SE in hardware was done.  But then I never
got into a technical discussion with Hsu or Campbell about that specific issue
so it might have been done and I didn't know.  The hardware search was
definitely different than the software search if you read his stuff however.
For example, only passing _one_ bound through the hardware rather than the
normal two.

Also, "depth of offset searches" is normal for _any_ SE implementation.  The way
to detect singularity is to search one move with normal window, remaining moves
with a reduced depth and lowered (offset downward) alpha/beta window to prove
all the remaining moves are worse (this is at a fail-high node, at a PV node you
don't do reduced depth offset searches in real SE).  So I really don't see where
the above suggests "cheapo SE" at all.  In fact, it suggests the more expensive
approach which does PV-singular tests as well as FH-singular tests where the
offset/reduced-depth searches are actually needed.  The cheapo approach does
_all_ offset searches to reduced depth and is less accurate.  And _far_ simpler
to implement.  I've done both, although I have not had much luck with either
approach in Crafty.  Hsu's SE worked just fine in Cray Blitz, but didn't work
well in Crafty.  Cheapo didn't work for me very well either although I still
have the code laying around and play with it from time to time.




>
>But that might be too technical for you :)
>
>Tony
>
>
>>
>>Thanks again.
>>
>>Daniel



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.