Computer Chess Club Archives


Search

Terms

Messages

Subject: Article on automatic tuning of evaluation weights

Author: David Eppstein

Date: 12:49:00 07/13/99


Readers of this group may be interested in my paper "Setting parameters by
example", which can be found at http://xxx.lanl.gov/abs/cs.DS/9907001

The paper is mostly about some combinatorial optimization problems like shortest
paths and minimum spanning trees, but there's a section towards the end
describing possible applications of the same methods to tune chess evaluation
function weights. First, a big warning: this has not been tried experimentally,
it's all just speculation. Basically, the idea is to apply the ellipsoid method
from linear programming to find a set of weights such that a shallow search
using those weights plays correctly on each problem from a test suite. Each time
the ellipsoid method wants to test a set of weights, you use those weights to
run the test suite. If one of the test suite positions is played wrong, you can
search that one position more carefully to find two positions A and B such that
A should be evaluated better than B but actually is evaluated worse and this
misevaluation is causing the position to be misplayed.  Then you add a
constraint to your linear program saying that the evaluation of A should be
greater than the evaluation for B, and go on to the next ellipsoid iteration.
Read the paper for (a few more) details.

Yes, I know test suites are not very representative of actual play (especially
the tactical ones that are the most commonly available).  So one of the ways
this could go wrong would be that it finds evaluation weights that work for the
test suite but are bad for actual play.  Another way it could go wrong would be
that it finds that there is no assignment of weights that correctly plays all
test suite positions; but this might be useful too, either in finding bad
positions in the test suite or in helping understand what important terms might
be missing in your evaluation.

Apparently, for some other games, there is much more sophisticated use of
automatic methods for incorporating game knowledge (e.g. the big libraries of
edge patterns used in good Othello programs) and I am interested in to what
extent this is possible for chess.  The advantage of test suites is that they
are annotated with correct moves, and (should be) filtered for cooks, but it
would be even better if there were some way to learn more than just opening
lines from the big collections of thousands of high-quality pgn games.


By the way, I'd recommend http://xxx.lanl.gov/ for anyone seeking a good way of
making electronic copies of computer science papers.  It's very reliable, it's
easier to find papers there than on thousands of individual or departmental web
sites, it sends out announcements of new papers to a wide mailing list, and it
provides convenient format-conversion services (e.g. pdf as well as postscript,
which is great for me since the postscript viewers on the mac all suck rocks).

Bob Hyatt, you mentioned (on the crafty list) making a recent paper on book
learning available on your ftp site -- maybe you could put it on xxx instead,
that way people not on the crafty list might notice it too.



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.