Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: Detecting repetition in a search......

Author: Roberto Waldteufel

Date: 10:33:03 10/01/98

Go up one level in this thread



On October 01, 1998 at 12:59:20, Bruce Moreland wrote:

>
>On October 01, 1998 at 07:58:57, Roberto Waldteufel wrote:
>
>>The only problem I have encountered with method 2 is that draws correctly
>>identified by the draw hash table will have an adverse affect on the scores of
>>ancestor nodes in the tree when they are entered into the transposition table
>>later on, and so a later hit on one of these ancestor nodes may give misleading
>>information. I don't really see a way to avoid this - it seems at least to be a
>>rare event.
>
>When you detect a repetition, just return instantly.  I don't know how your
>program is structured, but in mine, I can just return out of the search function
>at that point, and conveniently forget to store the result.


Yes, this is what I do now.


>Of course, if the draw score is a fail low, you'll get a fail high when you
>return, and you'll store that score.
>
>Which brings up a point I don't fully understand.  People are concerned about
>hashing draw scores because they don't want incorrect information in their hash
>table.  But you will always get incorrect information in your table, even if you
>have a constant value for a draw, and your hash store function has a hook in it
>that says never hash this score.
>
>When you detect a repetition, sometimes you'll have to select a different move
>in order to avoid it, and the score of this gets hashed.  Other times, when you
>find this position, and you have a choice between the move that was a repetition
>in the other position, and the move you chose instead, now you'll want to choose
>the repetition move, because it's no longer a repetition.  So what you get is an
>erroneous score in the hash table, but it's not the draw score in the table,
>it's the score of that second move.
>
>You get similar problems when you start doing extensions and pruning that are
>based upon alpha.  In one place you'll cut off and in another place you won't.
>So you're always going to have the general problem, I think, of positions that
>fail high when you search them, then fail low when you re-search them.

These are exactly the sort of problems I am worried about, but......

>
>I just ignore this problem and cross my fingers, and this seems to work fine.
>

Same here. I don't see any evidence of the search suffering, but it irks me to
know it could....:-)

Roberto



>bruce



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.