Author: Ulrich Tuerke
Date: 09:15:36 04/02/01
Go up one level in this thread
On April 02, 2001 at 11:46:41, Dieter Buerssner wrote:
>On April 02, 2001 at 04:47:10, Ulrich Tuerke wrote:
>
>>quite impressive performance by Yace again !
>
>Uli, thanks for the nice words.
>
>>It's probably a question of choosing smart search extensions ?
>
>Yes, I think this is correct. Perhaps not smart extensions, but I think I have
>extensions that fit this position. I believe the main problem is to find the
>refutation for 1. Ba3 Qxa3. Uri has allready sent a line for this. I get:
>
> 814247 5.553 4.59 8t 2.Nh5+ gxh5 3.Qg5+ Kh8 4.Qxf6+ Kg8 5.Qf7+ Kh8
> 6.g3H Nxd4H 7.e7H Qc1+H 8.Kg2H Qc2+H 9.Kh3H
> Qf5+H 10.Qxf5H Nxf5H 11.e8=Q+H Kg7H {HT} {421}
>
>Unfortunately, there are many moves from the hash table (all the moves, that
>have "H" appended); these are not allways reliable. However, here they look
>right. The line starts with 4 checking moves by white in a row. Also, later,
>there are 3 checking moves in a row for black. For such situations, I do some
>rather aggressive extensions. Of course, you will be well aware of the fact,
>that more aggressive extionsions will in general give a bigger search tree, and
>so in many positions can hurt.
I think that it's really a challenge here to find the quiet move g3 .
The subsequent series of "revenge checks" make this position even more
difficult, shifting things beyond the horizon.
I do also care for series of repeated checking moves by one side, but mainly if
it's not always the same piece which is checking. I thought when several pieces
are involved, then it's perhaps an indicator for a dangerous attack and worth
the extensions.
But in this position it's always the queens keeping on the checks. So. "bad
cards" for comet.
>
>Another thing might be quiescence search. I think, the moves starting with 10.
>Qxf5 are from quiescence search.
>
>Yace does not have the knowledge mentioned by Uri about the knight moves, that
>are needed to stop a passed pawn. I have thought of something like this for
>endgames, where the opponent has only a knight and pawns. But I found it too
>difficult even there, because the knight may gain tempos by checking moves. With
>Qs on board, this looks even more difficult.
>
>BTW. Out of curiosity, I made a small experiment with null-moves. Yace uses
>something similar to the double null-move, that was mentioned in this forum
>quite a few times. I disabled this. The solution time for this position did not
>vary significantly. But when I am using more aggressive null-move (R=3 when
>still high search depth is required in the tree, R=2 otherwise; by default Yace
>uses R=2 allways), the solution Ba3 is found at depth 12 instead of depth 11.
Some time ago, I have also experimented a bit with R=3. I came to the conclusion
that the win in search depth is not really worth the risk of missing some
tactics. I think that the real win in null move was from R=1 to R=2.
Perhaps the recursive application of the null move search lowers the win from
R=3. With R=2 you can in some cases apply the null move more often in one line
than with R=3 because of the limited line length. Applying 3 times R=2 in one
line may give a similar win as applying R=3 twice ?
Thanks for your comments.
Regards, Uli
>
>-- Dieter
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.