Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: Couple of chess programming questions

Author: Robert Hyatt

Date: 10:40:22 09/11/02

Go up one level in this thread


On September 11, 2002 at 12:21:38, J. Wesley Cleveland wrote:

>On September 10, 2002 at 13:40:54, Robert Hyatt wrote:
>
>>On September 10, 2002 at 09:26:14, Eli Liang wrote:
>>
>[snip]
>>>(7) Dennis Breuker has a paper comparing 5 different transposition table
>>>replacement schemes.  Has any else been able to validate his results that
>>>replacement should be based on the number of nodes searched and not search
>>>depth?
>>
>>Hard to say.  The issue will become hash entry size.  You can store "depth" in
>>a byte.  You will need at least 4 (and really more) bytes for nodes searched.
>>Your table will therefore be smaller, and _that_ effect was not measured.  Also
>>I am not convinced that in endgames, nodes searched is the right idea because
>>of hash hits that cut trees off that have significant depth, but no nodes due
>>to the hash hit.
>>
>>I don't know that he addressed that case either, such as the one seen in fine
>>#70.
>
>If you saved nodes in the hash table, you could take the node value from the
>hash table and propagate it up the tree. This would also do wonders for your NPS
>;)

yes, although it would be wrong, just like saving the draw score in the hash
table.  IE in Cray Blitz our scoring window was opened up so that no _real_
scores could occur in the range 0.000 to 0.100 (we used millipawns as did most
programs back then).  Our draw score because DRAW+ply so that we preferred the
deepest draw (as opposed to the shallowest mate) to give our opponent a chance
to go wrong, or to give us a chance to search thru a two-fold repetition to
find a win before the three-fold repetition was reached.  We saw impossible
draw scores, such as draw in 99 plies, which would be impossible for CB as it
could not search beyond 64 plies deep, ever.

What doing the nodes trick would do would really produce a dynamically variable
NPS that would be almost a good random number.  :)

IE if you have two plies left, and get a hash hit on a position that has a
stored depth of 6 plies, the node counter would be incremented by a value _way_
too large...




This page took 0.02 seconds to execute

Last modified: Thu, 07 Jul 11 08:48:38 -0700

Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.