Author: Christophe Theron
Date: 17:47:53 02/08/98
Go up one level in this thread
On February 08, 1998 at 14:32:26, Don Dailey wrote:
>Hi Stuart,
>
>This is still a mystery. Langs search seems different but surely he
>uses
>some of the same principles.
>
>I remember him once saying (years ago) that no one tries to be
>selectivity. He described his search as throwing out 10% 20% 30% etc.
>moves progressively at each level. He indicated that this was the
>principle, not the actual algorithm. I asked him what he did if a piece
>was lost after looking at only 10% of the moves and he said he simply
>looked at all the moves in this case! Duh!
>
>But this leaves out quite a bit. Like how can you find attacking moves?
> If I wrote a program that statically evaluated every single move,
>sorted them and then tapered them something like this, it would play
>horrible.
>
>But Bob once had a selective program that was not based on null move
>prunning.
>He will probably tell you it's no good but at least you might get an
>idea of what was tried several years ago.
>
>The Chess challenger 7 had a highly selective algorithm. It played
>pretty lousy but was not too bad for it's day and hardware. It had a
>non-existant endgame though. If you were a couple pieces down you still
>had a good chance of winning if you could get to an endgame! The
>documentation that came with it descibed how many moves were discarded
>at the various levels. But of course it said nothing about how these
>were chosen.
>
>I would be fun to write a truly selective program. I'll define
>selective as meaning no window considerations whatsoever. Some routine
>just looks at the moves and keeps a few threats and the positional moves
>that seem ok and throws out the rest! I wouldn't know how to do this
>well.
>
>- Don
This is exactly what I did in Chess Tiger 8. Sort all the moves by the
positionnal scores, mark the first N, look at the others and mark also
the agressive ones, and then search only the marked moves. I also
computed a "satisfaction" score, and if the score after searching the
marked moves was still below "satisfaction", I searched all the
remaining moves.
N was computed as a percentage of the number of available moves, and the
percentage depended on the distance from the leaves. Near the leaves, it
was about 5%, and increased to about 60% when it was 6 plies from the
leaves. I think I never went above 60%, so Tiger 8 could overlook some
combinations no matter how much time you let it think. This kind of
problem happened in 1 game out of 10 I can say.
Tiger 8 was a 16 bits DOS program, written with Borland C 3.10,
searching around 4000 nodes per second on a P100. In blitz, it was able
to search 6 or 7 plies, which was not so bad when you consider that it
was also doing a lot of extensions. I think that it was playing around
2150-2200 ELO on a P100. It had no hash tables nor thinking on
opponent's time.
Chess Tiger 8 made 3rd in the 1996 french computer championship, out of
10 competitors, sharing its ranking with Virtual Chess and Capture. The
championship version was released freeware on the companion disk of the
french revue "La Puce Echiquéenne".
I think Lang's search has nothing to do with such a technique.
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.