Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: The new version of Junior is significantly better

Author: Komputer Korner

Date: 08:11:16 08/05/98

Go up one level in this thread


On August 05, 1998 at 06:37:51, Robert Hyatt wrote:

>On August 04, 1998 at 20:32:37, blass uri wrote:
>
>>
>>On August 04, 1998 at 19:55:55, Fernando Villegas wrote:
>>
>>>Hi both:
>>>Sorry if a miserable non programmer intervenes, but usefulnes of full width
>>>search does not depend on how much nodes your machine can calculate? I presume
>>>that with an ideal, infinitely fast machine, the best you can do is to calculate
>>>absolutely all, no null move, no extensions, no prunning. In less than ideally
>>>fast machines equation alters very much, but Deep Blue, although not infinetly
>>>fast, is enough fast for being near the area of validity of that reasonning.
>>>Maybe if you surpass certain threshold in speed, the old technique recover some
>>>sense. In fact, DB games against K are a prooof of it.
>>>Excuse me :-)
>>>Fernando
>>The games do not prove it because maybe deeper blue could play better
>>with null moves.
>>
>>The idea of no extensions is not realistic because sometimes you must see
>>30 plies and even 60 plies forward to go to the right decision
>>and even a machine 1000 times faster than deeper blue cannot see
>>30 plies in the brute force depth.
>>
>>Uri
>
>
>I did the math a while back, and concluded that if you take DB at 200M nodes per
>second, and Crafty at 200K, searching to 12 plies in a normal middlegame, that
>it should reach 12+log2.5(1000) plies on DB hardware.
>
>So yes, I'd love to see what a real null-mover could do on such hardware.
>
>ANd one last point...  don't forget that DB is not a "no extension" program.
>They do more than any of the rest of us...

Something is wrong here Bob.
The log 2.5(1000) is .39794 x 1000 = 397.94  So you are saying that Crafty on
Deep Blue hardware would search 410 plies deep. Obviously the ply depth cost per
move equation is not scaling linearly like you would like it to here. We need a
more accurate scaling factor for very deep depths of search with null move and
extensions. Does Crafty do any forward pruning besides null move?

Below is a previous answer by Bob on the subject:

"I can't answer.  For Crafty, each successive iteration takes about 2x-3x
the time of the last search in the middlegame.  Rebel seems to be about
the same, as does Genius.  I don't know about DB, but suspect they are
up at 5x still, because they don's use null-move (at least according to
last reports) and they don't do forward-pruning.  Ed has more Rebel
data, but I assume his numbers are similar to what I've seen when running
Rebel 8 here.

In the endgame it is generally significantly lower then 2.5x-3x...  I've
seen many positions where it is < 2x.

however, when you analyze, log(base 5) is not a lot different than
log(base 2.5) when the target number is huge..."


--
Komputer Korner



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.