Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: History based pruning question

Author: Robert Hyatt

Date: 21:44:42 08/26/05

Go up one level in this thread


On August 27, 2005 at 00:28:59, rasjid chan wrote:

>On August 26, 2005 at 15:38:06, Robert Hyatt wrote:
>
>>On August 26, 2005 at 15:27:51, Alvaro Jose Povoa Cardoso wrote:
>>
>>>>My first thought is that the number of "fail lows" is irrelevant.  What you
>>>>really want to avoid is a reduction on a move that might fail high.  Any move
>>>>will fail low in some situations, but you want to handle the "typical" case
>>>>correctly and not reduce if there is a reasonable chance the reduction will hide
>>>>something.
>>>
>>>So, you are saying history based pruning is not a safe thing to do?
>>>Or are you saying I should use tipical history values instead of counters?
>>>
>>>Alvaro
>>
>>I don't know how safe it is.  Bruce Moreland and I played with this idea back
>>around 1996 or so, but we were not paying any attention to the actual moves
>>themselves (for example, never reducing a checking move perhaps) so that what we
>>fooled around with never worked very well in actual games.
>>I'm not sure what I would use, should I try to play with this again.  In fact, I
>>have it on my "to do list" since such ideas (futility) have been used in Crafty
>>with good results already, and this is just another variation on that sort of
>>theme...
>
>A related problem is when it is good to extend or reduce.
>
>In a past thread "why is fruit so strong?" many(including Fabien) only made
>general(good) comments and only Thomas of Toga was definite - "That's easy to
>answer..." and he mentioned :-
>  Fruit only extends pv nodes and don't reduce pv nodes(I'm not sure if it is
>  without exception).
>  Reductions only done in non-pv nodes.
>I come to figure that there may be a point for this.
>
>All child nodes of pv-nodes usually have positions with one piece position
>differing. Static evaluations of these child nodes have higher average
>probability of being close to the pv score. Nodes far away from the pv have
>a high average probabilty of having scores away from the pv score. In any
>non-critical stage of a game, the pv scores from one iteration to the next do
>not change much and this means the set of  pv node moves all have greater
>probability to be in the next pv and hence Thoma's observation.
>
>This may affect any pruning scheme so that a workable scheme may seem not to
>work.
>
>Rasjid


That will take some thought.  I don't personally like the idea of doing one
thing to PV nodes and something else to non-PV nodes.  That means that all moves
but the first are searches more recklessly, making it harder to find a sudden
killer-type move after searching the first move...




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.