Author: Christophe Theron
Date: 23:40:01 12/25/97
Go up one level in this thread
On December 25, 1997 at 20:38:54, Don Dailey wrote:
>Let's see if we can figure out who is doing attack extensions
>and see what they are doing with them:
>
> Attack extensions are really dubious and you should
> not be using them in your program. If you do, you
> must surely be a bad person.
Nasty Don! :)
Well, I used to do some kind of attack extensions in Tiger. In fact it
was a "threat evasion" extension. And it was not an extension. When I
detected a threat in the first plies of the quiescence search, I set the
score to alpha instead of Eval(Pos), and tried every possible move (with
an optimization possible here if Eval(Pos) very < alpha)... Ok, it's
like not being in QSearch, yes.
With some tuning it's possible to avoid to overload the tree. For
example, restrict the extensions to positions resulting from MORE
threatening moves, don't do it when the threats did not increase with
the last move.
There are some problems to deal with when you do it:
* It doesn't work well with futility pruning
* You have to use at least a SEE to evaluate the threats, and a special
"mate threats" detector
I gave up using these "threat evasion extensions", because I don't use a
SEE any more, and it was not clear that it strengthened my program.
Surprisingly, the improvement, if any, was a positional one. Not
tactical.
Jean Christophe Weill describes some threats extensions (entropic
extensions) for the old Joker program in his PHD thesis (in french...).
He also gave up using them.
Christophe
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.