Author: Robert Hyatt
Date: 16:57:15 10/22/02
Go up one level in this thread
On October 22, 2002 at 19:38:58, Nagendra Singh Tomar wrote:
>Hi All,
> Following is the snippet of the classical PVS code
>
>/* zero windo search */
>value = -PrincipalVariation(pos, -alpha-1, -alpha, depth);
>
>/*if the zero win search fails-high do a research */
>if(value > alpha && value < beta) <=====
>{
> /* do a wide window search here */
>}
>
>My question is regarding line marked <=====
>
>Should'nt the check be (value >= alpha && value < beta)
No. If you are doing a null-window search _already_ because you were called
from
an earlier node that is searching alpha,alpha+1, then you want the above test to
fail. IE
if you get a score > alpha and < beta, you _know_ that (a) you are not doing a
null-window
search as otherwise that condition could not be true and (b) the value is not
too high so that a
re-search is pointless as it would still fail high...
> ^
>
>This is because the BETA value in the zero window call is -alpha and inside that
>call it will fail-high if score >= beta i.e score >= -alpha
>
>So a return value of -alpha from the zero window call also signifies a fail
>high. So we should do a re-search for this also.
>
>Is that OK.
>
>Another question is that if the wide window search fails-low due to transref
>tables or some window dependent extensions, what do we normally do. Do we do a
>re-search with a still wider window or do we ignore the fail-low. I remember
>from some Bob's post that we ignore it.
It depends on where you mean. If you get a fail high on the null-window search,
and
then when you re-search with the wider non-null-window and fail low, then I just
ignore
the original fail high (and the accompanying fail-low) completely...
>What if this condition (fail-high followed by a fail-low) occurs at the root (in
>aspiration search). How can we ignore it there ?
You can't ignore it there. You have to lower the alpha value and re-search to
find
the best move...
You can only ignore fail lows when they happen right after a fail high...
>
>
>regds
>tomar
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.