Author: Bernd Nürnberger
Date: 08:18:56 05/04/04
Go up one level in this thread
On May 04, 2004 at 10:30:33, rasjid chan wrote:
>On May 03, 2004 at 09:02:52, Bernd Nürnberger wrote:
>
>>Hi Volker,
>>
>>On May 03, 2004 at 08:18:48, Volker Böhm wrote:
>>
>>>On April 30, 2004 at 08:26:59, Bernd Nürnberger wrote
>>[...]
>>>Hi,
>>>
>>>you can nearly never stop searching "earlier" except if and only if there is a
>>>recapture that proves to be the only move not loosing material.
>>>Even if alphabeta don´t varry much your engine could just find a tactical
>>>problem in the next millisecond.
>>>
>>In advance, thank you for your answer!
>>
>>Ok, this sounds reasonable to me. But how does some engines manage it
>>to play a move really quickly then? What I mean is, some engines does
>>some moves *really* fast, when considering the time control. And the
>>corresponding moves are recaptures only in fewer cases.
>>
>>>There are two point where you should search (much) longer.
>>>
>>>1. Your best move calculated from the previous interation fails low. At this
>>>point you should spend enough time to be sure that all other moves are searched
>>>too.
>>>2. If after your iteration the result drops below your initial alpha value. Then
>>>you should spend enough time to search again with a larger window. When the
>>>result is much smaller than before, you should consider to search the next
>>>iteration too.
>>
>>Are these points also important, if I do a search with a -INFINITY, INFINTY
>>window from the root (the current position)? So I should never get a fail high
>>on the first move, shouldn't I?
>>
>>Greetings, Bernd
>
>I too have some difficulty in understabding Volkers reply about failing low
>in an iteration.
>
>When you start with -inf,+inf you never fail low at a root iteration.
>
Does that also apply, if I do PVS? Clearly I research if I get a
score between alpha and beta (it is only meant to get rid of all moves
that are just too bad).
>With aspiration search we always start with with a guess x and use
>x-width, x+width with width about 1/2 pawn and research if FL/FH
>and gets our 1st alpha. Maybe he meant aspiration search FL requiring a
>research.
Probably that's what he meant. I will add aspiration search to my engine
the next time, because I think, it will do better than PVS. Maybe I
can do aspiration search before the first alpha exceed and then
do PVS (failing returns to aspiration).
>
>About maybe 2 weeks ago there was a long thread relating to something
>about ... when is move good enough ...and we can return before using
>the full allocated time(I failed to retrace the thread
>for you earlier). Just happened no mention was made about
>same best move repeating for say N iteration. I'm sure many not too
>sophisticated like me returns after N == some number(8). It may not have
>validity as it MAY change after 2N iteration.
Thank you very much for this information. This thread sounds really
very interesting for my concerns.
Do you have some expirience by yourself on this topic (stop search earlier)?
On the WAC suite, my engine performs worse when N (see above) is too
small. I did not do tests with larger N's so far.
Greetings,
Bernd Nürnberger
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.