Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: developing Junior (and other pro programs)

Author: Robert Hyatt

Date: 10:18:41 09/02/02

Go up one level in this thread


On September 02, 2002 at 05:16:57, Gian-Carlo Pascutto wrote:

>On September 02, 2002 at 04:29:46, Ricardo Gibert wrote:
>
>>In comparison with alpha-beta, with no research, you examine 50% as many nodes.
>>With one research you about break even. With 2 researches, 150% as many, etc.
>
>Uhh?
>
>As you said above, most of the searches are single zero window tests (also in
>alphabeta/PVS), in which case there is no performance difference of MTD over
>alphabeta. If you need researches (you always need at least 2 with MTD), the MTD
>method without memory loses out.
>
>Basically, the MTD method _always_ loses.
>
>--
>GCP


This is not quite right.

If you only do two searches, hash is _irrelevant_.  The first fails low or
high, the second fails high or low.  No hash assistance whatsoever there, so
the M part has no use...

In the case of DB, if you have read much of Hsu's stuff, they could also do
the mtd(f) type searches in parallel to accurately "bracket" the score and
quit early once they have an upper and lower bound, which is often good enough
(near the tips) to discard things efficiently.  A normal mtd with hashing would
offer nothing over a DB-type mtd in such conditions...



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.