Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: hash collisions

Author: Russell Reagan

Date: 20:19:12 07/09/02

Go up one level in this thread


On July 09, 2002 at 22:55:02, Robert Hyatt wrote:

>I will report more as I test.  The obvious question is "how many false matches
>can be tolerated?"  My gut feeling is "zero" because I can print out a small
>tree, and change _one_ score of my choosing, and change the entire PV and
>score.  But in reality, the search appears to be much more robust than I would
>have guessed...

1 in 1000 is only 1/10th of a percent. I don't have much experience with this
stuff, but 0.1% sounds like it might not damage the search enough to change it's
search results.

In a 4-ply search with a branching factor of about 5-6, you would search about
1000 nodes. If there are only a handful of good moves at each node, we're
talking about only 20-40 nodes that actually matter, maybe up to 80 even. The
odds are pretty good that if there is a false match, it will fall in the other
nodes that are bad to begin with.

There is a low chance it will fall on a node that will matter, and then even if
it does, there's still a chance that it might not change anything.

That's my common sense analysis for you from my zero experience :)

Russell



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.