Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: Compute Power Affects Search Heuristics

Author: Robert Hyatt

Date: 07:39:10 12/24/98

Go up one level in this thread


On December 23, 1998 at 00:56:41, Dave Gomboc wrote:

>On December 22, 1998 at 15:15:29, José de Jesús García Ruvalcaba wrote:
>
>>	Well, I tried to say exactly the opposite of what you think I think.
>>Perhaps we have a language problem, I am not a native English speaker. Let me
>>restate my point: while algorithms have improved, they have also adapted to
>>faster hardware.
>[snip]
>
>This is a very important point, if I may go off on a tangent.
>
>I've heard some comments from people to the effect that "if other people had the
>hardware that the Deep Blue team did, they wouldn't be searching everything and
>extending like they do now, instead they'd be using null move with r=2 just like
>the rest of us".  But you know, it's not exactly difficult to implement
>null-move, it's been done many times before, and though the Deep Blue team
>consists of a bunch of pretty bright people, it was their choice to do things
>differently.  I think it'd be fairly easy for them to test out a null-move
>implementation, so would they really just ignore it?  My opinion is that they
>probably did give it a try and, at the nodes per second they were searching,
>found the extra speed not worth the accuracy loss.
>
>Comments?
>
>Dave Gomboc


I definitely think that technology drives the search *and* the evaluation.  IE
if you can only search 1-2 plies deep, you had *better* teach your evaluation
something about 'forks' because it won't be able to find them with the shallow
searches.  But once you reach 4-5-6 plies, you can forget about forks because
the search finds them, and evaluating them is not needed.  This idea carries on
to other parts of the tree.  IE this is why I added singular extensions to CB
a few years ago.  I felt that we were going deep enough for normal positional
type ideas, but that we needed to follow some forcing lines deeper than we
were.  I tried singular extensions somewhere around 1978 or so, and while they
found some nice tactical plans, they cost me a ply.  In 1978 we are talking
about going from 5 plies back to 4.  That hurt and I gave up on them and removed
the code just before the 1978 ACM event.  And I stayed away from them even after
Hsu reported good results, because he had a *lot* more horsepower than I did and
he could afford to lose the ply to gain the depth along selected lines.  Once
we hit the 1/2M node per sec speed, I reconsidered, added the code again, and
was much happier dropping from 11 to 10 plies than I was when I dropped from 5
to 4 plies...

micros are now getting to the speed point where this is perfectly doable...



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.