Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: MTD(f) and killer heuristics

Author: José Carlos

Date: 04:26:33 07/12/01

Go up one level in this thread


On July 12, 2001 at 07:19:02, Marcus Heidkamp wrote:

>Has anyone experience with the killer heuristics in an MTD(f) driven search? I
>am wondering if it would make any sense, because good cutoffs are stored into
>the trasposition tables anyway, and each call to the driver either fails high or
>low. So my (rather vague) guess would be that the cutoff ratio for killers will
>be small or even zero, because transposition cutoffs are done first.
>
>With history heuristics it's different, because history directly relates to move
>ordering, where killers seem to me to affect mostly the cutoffs.
>
>I already did a quick and dirty implementation of killers in my program, and it
>seemed not to outweight the overhead of move validation. But maybe I had a bug
>in there... :-(.
>
>Comments are very welcome.
>
>Marcus

  I haven't tried MTD(f), but I don't see why killers should be useless there.
Hash cutoffs occur before searching, when the current position has a lower bound
in the hash table that is > Beta. If you don't get a cutoff from hash table you
have to search. Then, you try the hash move first, good captures then...
whatever you chose as move order. Somewhere in that order, you put the killer
moves (for example, after good captures). So the killers improve your move
ordering, and have (in principle) nothing to do with a hash cutoff in a given
node.
  This said, I might miss something because I don't have experience with MTD(f).

  José C.



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.