Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: Announcing the first public release of C.A.P. data!

Author: Jay Scott

Date: 12:20:02 01/05/99

Go up one level in this thread



On January 04, 1999 at 18:32:24, Dann Corbit wrote:

>Since this analysis is clearly fairly superficial, the most interesting part
>will be in the large ce evaluations.  For instance, gambit openings are quite
>likely to show a score of +/- 100 centipawns for obvious reasons.  For openings
>like this, I would not be nearly as concerned about the centipawn score as the
>win/loss percentage and how some of the greats have played it.  But for
>positions which have a ce greater than 2 the results are quite interesting.
>After enough data is gathered, I will be able to perform simulated annealing on
>the database.  At that time, I expect the value of analysis to improve
>exponentially.

I assume the "simulated annealing" is some algorithm which attempts to
back up scores from leaves up through intermediate nodes without totally
obliterating the scores calculated at higher levels. How does that work?

To me, the obvious way to analyze an opening tree is from the leaves up.
When all positions below a given node have been analyzed, you stash their
scores into a persistent hash table. Then when you analyze this node, the
search is essentially looking to see whether there's a superior alternative
to the book moves that are already analyzed. The search can take advantage
of transpositions into the already-analyzed portion of the book to return a
more knowledgable evaluation of those alternatives.

The last search is of the root position, and it discovers, "hmm, 1. f3 is
not as strong as 1. d4." :-) When that's finished you have a fully-scored
tree, probably with many alternate suggestions that you can wade through
for interesting TN's. If you later go back and analyze some leaf node more
deeply, you may have reanalyze higher nodes as well to back the score up the
tree.

Obviously this method is a lot harder to coordinate if you have lots of
volunteers doing the searches!

  Jay



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.