Computer Chess Club Archives




Subject: Re: About False Fail Highs, professionals, and MTD searches

Author: Dieter Buerssner

Date: 13:17:22 04/12/02

Go up one level in this thread

On April 12, 2002 at 15:08:51, Gian-Carlo Pascutto wrote:

>Do they have a trick which they use to determine whether a
>fail high is real or not? I don't see how this could be

One explaination could be, that they just totally trust the hash table. If you
fail high at root, one ply deeper in the line, you will have an upperbound. If
you research, adjust beta with this upperbound (trust, that it is a real correct
upperbound), and just don't allow any better scores, so at the root, you will
never get back a score worse than before in the research.

BTW. I think, I have seen fail highs in Shredder 5, that seemed wrong (and
switched away again in the next depth). But, as you note, I also allways have
seen it keeping at the same ply.

I believe, this "total trusting in previous scores with the same search depth"
makes the branching factor smaller, but for me, it too often needed more depth
to find the same correct solution. This has IMO to do, with pruning decisions,
that are wrong (for example null move eats too many plies, to see the threat).
Researching with a different window, and null move may not fail high again in
the same position. And a search with a real move would show, that the lower
bound from the null move in the previous search was just wrong. But when you
trust the lower bound from the previous null move fail high and adjust alpha
accordingly ...


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.