Author: Dieter Buerssner
Date: 10:53:42 03/19/01
Go up one level in this thread
On March 19, 2001 at 01:04:32, Ferdinand S. Mosca wrote:
>Here is the log on P3 667
[..]
>I think it is similar.
Thanks. Yes, looks almost identical.
>On my new setting, Yace does not even consider Qe7, after Kg7, here it is:
[...]
> 66395034 3:50.9 -2.86 14t 1...Kg7 2. Qd7+ Kg6 3. h4 gxh4 4. Qe8+ Kg7 5.
> Qe7+ Kh8 6. d7 Qc6+ 7. Kh2 Qc7+ 8. Kh3 Qc3+ 9.
> Qe3 Qf6 10. Qe8+ Kh7 11. d8=R Qxf2 12. Qg8#
> {-260}
>The changes are:
>icheck 0.6
>irepcheck 1.0
To readers not aware of Yace or its documantation. These are the extensions Yace
does in case of check and "repeated check". Default values are 1.0 and 0.6.
I am sorry, changing icheck (the fracional incncrement for search extensions
when in check) won't work well right now. A value smaller 1.0 may cause an
artifact seen above (and values bigger 1.0 get more or less truncated to 1.0 in
most search situations). In the search routine, Yace assumes, that quiescence
search will not be entered, when in check. But I think exactly this happens
above, and you see a score of -2.8 for a line that ends with mate.
So, I fear the use of icheck.
Yes, this is a bug, and I will fix it.
>My idea at first is to discourage Yace from making some checks, Yace is very
>sensitive to checks, especially at endings, and I raise the value of irepcheck
>on the idea that there is a serious bussiness when there is repetition of check.
I agree with you, that Yace often does too many checking moves in endgames.
However, decreasing check extensions (which Yace allready does in the background
in certain endgame situations) my not yield in what you intend.
Look at the following situation:
1. almost any white move black advances passed pawn (PP)
The program as white won't like this. So it gets the idea
1. stupid check move out of check, by placing the black K even slightly better
3. any white move black advances PP.
At the end, black has not only advanced the PP but also put his king to a better
place.
So, how much search depth is needed, to see, that it is no good idea to check
the black K. Without any extensions, it is 4 ply. (After 3 ply, the black PP is
not advanced yet, and the position with slightly better placed black K may look
better for the program, than the more advanced PP). With a full extension for
checks, the same will be seen with a search depth of 3. Because this is the only
extension in the line, even a fractional extension would be enough here.
Actually, when I have thought about search extensions first, i did not think
about this at all (checking for seemingly delaying a threat, btw. instead of
check also a trade of a piece may be inserted). And perhaps it is totally wrong,
that Yace sometimes does not extend checks in endgames (depends on the position
in the search tree, I can give you details in email).
>If you can I need some explanation on extension parameters.
I can do here. I fear, it might not be of too much public interest and we can
switch to mail.
>BTW a position from a game played by Yace:
>
>8/5k2/2p5/2Pq4/5PpP/4Q1K1/8/8 w - - 0 1
>solution fastest to see f5
>
>On a new setting Yace sees f5 quickly than the default.
>12.860 1.44 12t+ (new setting)
>2:35.8 1.87 13++ (default)
Interesting. I can add, that here Yace needs even 14 plies depth ...
But this may also be due to a different set of TBs. Only changeing irepcheck,
and the solution will be found here at depth 12, additionally disabling check
extensions all together, and the solution is found really fast. Hmmmm - seems to
totally contradict, to what I have said above.
Perhaps other readers have also thought about this, and are willing to share
their ideas. It may be, that queen endgames are special in this regard, or that
my ideas are totally bogus.
> This is also the setting where Yace tried not to like Rc3 from my other post.
I will also look into this again.
I will send you a version, that does not have the above mentioned bug.
Regards,
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.