Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: search extension

Author: Ulrich Tuerke

Date: 08:12:23 07/08/98

Go up one level in this thread


On July 08, 1998 at 09:55:58, Robert Hyatt wrote:

>On July 08, 1998 at 06:20:18, Inmann Werner wrote:

> ...
>
>I think the idea is one that has "been tried and found wanting".  It has
>problems for a couple of reasons:
>
>1.  many programs use PVS or Nega-scout algorithms, which means that 99.999%
>of all nodes searched have alpha=beta-1, making it difficult to return a
>value "very much lower than alpha" to trigger this.
>
>2.  once you try it for a while, you find it has simply cost you a lot of
>time.  I don't know of anyone using this method any longer, including
>Chrilly...

IMHO, another draw-back of the originally suggested method is that the
extensions depend on the current alpha-beta window. So, it could happen that a
verification search resulting from such kind of extension does not reproduce the
condition for the extension (as result of the shifted alpha-beta window).

A unstable search could result (fail-low in fail-high verification search, ...).

The corresponding modification for mate values only is free from this draw-back.

>
>But there is one useful thing you can do...  at ply N+1, you *can* return
>an accurate score, namely "mate".  And this can be useful.  Bruce was the
>first I know of that used this method, and it works well.  If you do nothing,
>and you get mated, then there is obviously a "threat" in the position and it
>is worth searching deeper.  I do this too, but *only* for mate, because I
>can't detect when the score is "well below alpha" because it can't happen in
>PVS.



This page took 0.04 seconds to execute

Last modified: Thu, 07 Jul 11 08:48:38 -0700

Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.