Author: Gian-Carlo Pascutto
Date: 13:40:43 05/18/01
Go up one level in this thread
On May 18, 2001 at 14:23:01, Miguel A. Ballicora wrote:
>Of course, I just threw some ideas that might be considered to determine
>if the algorithm it is worth to be triggered. In the case of mate threats
>or checks to extend more than one ply (maybe),
I went back and forth on this and have it as a settable option, to
allow more than 1 ply of extensions per nodes.
Standard Crafy does not allow more than 1 ply per node of extensions.
A system like maximum 2 ply per 2 nodes, as described in the DB papers,
would work much better. It would not blow up the search so much like my
methode now does.
>I agree, it has to be statically determined.
You will have to use an advanced form of SEE in that case.
>>of nearly 2 ply. In the LCT2 which I used that lost many other solutions.
>
>2 ply? ouch. My intuition was not bad. How does it play overall?
I have not tested actual gameplay.
>The latter is a nice trick. Similar to R.H. suggestions some time ago to
>use as negative extensions (reduce depth after several moves have been tried).
>Kind of risky though.
We will only extend too much, so there is no risk involved at all.
>
>What I did so far it is very simple. It is the most limited case I could
>imagine: I triggered SE when in check (after it was extended already)
>This is the pseudocode:
>
>if (incheck) {
> depth +=1;
> if (1==moves_available) {
> depth += 1; /* single response extension */
> } else {
> enable_SE = TRUE; /* it will test a singular check evasion */
> }
>}
>
>generate_moves();
>bestmove = loop_moves(&best);
>
>if (enable_SE && best > original_alpha) {
> SE_alpha = lower_of (original_alpha, best-PAWN_VALUE);
> secondbest = research_moves_but_bestmove (depth-2, SE_alpha, SE_alpha+1)
> if (secondbest > original_alpha) {
> best = research (bestmove, depth+1);
> }
>}
>
>return best;
Don't you mean depth-3, assuming you already do depth-1 on normal moves?
You dont want to research if SE_alpha < original_alpha, because you will
not do anything with the secondbest value anyway.
Another nitpick is that you should research _all_ moves and not only
the best move. If it fails low because of the extension you have to
pick up a new best one, and if needed, start the process all over again.
This is how SE detects horizon type moves.
--
GCP
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.