Computer Chess Club Archives




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

Author: Gian-Carlo Pascutto

Date: 15:27:37 04/12/02

Go up one level in this thread

On April 12, 2002 at 18:10:30, Josť Carlos wrote:

>>That can't be it. If a move would be, say, only 0.03 better,
>>you'd fail low and have to do another research to figure out
>>if the move is a real fail high or not. Moreover, the moves
>>that fail high the first time always end up being the PV.
>>Essentially, Fritz has no false fail highs. I'm clueless as
>>to what's happening here.
>  I'm not sure what you mean in this last paragraph,

What it boils down to is that no matter what you have to
open up the window to get a variation back and to be sure
the fail-high is 'true'.

>but the behaviour you
>mention seems logical to me. I'll tell you what I do: when I fail high in a null
>window search, I don't print anything yet. I open the window a bit (in my case,
>half the aspiration size), and research with Alpha,Beta+AS/2.

I do the same, though I just use alpha,beta.

> I have two search
>routines, one with high prunning level and one "careful search". I use the
>latter here. If the search returns something above alpha, I know the fail high
>is good, and print a fail high (with a 1 move pv).

This, I don't understand. If you do the research with alpha,beta+x, and
get something back that's below beta+x, the move is already confirmed good and
you should have a PV to go with it, so why only the 1 move pv?

I.e. what you explain would perfectly be Fritz behaviour, but only printing
1 move when you have a full PV that's trustworthy makes just no sense.

If first search gets a fail high, i.e. indicating at least beta, I
treat it as a normal (aspiration) fail high, trust the move immediately and
research with an open window.

> Then, I open the window to
>150 points above the returned val (val,val+150) and search carefully again.

Do you also do this if the first research comes back with a value below
beta+x ?


This page took 0.06 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.