Computer Chess Club Archives




Subject: Evaluating piece-values by TD Learning

Author: guy haworth

Date: 08:30:57 01/01/01

Go up one level in this thread

Don Beal did a paper in ICCA_J some time ago to evaluate the values of the
pieces by TD Learning.

Various 'runs' came out with different values ... the surprise for me was the
strength of the B over the N.  With the P at '1', I think the Q came out ~9.5
and the R about 4.5.

Other experiments have shown that the learning process is so inefficient as to
be arguable ineffective.  Michael Buro favours Bayesian statistical analysis of
features and I agree ... if you can find the features to analyse

An experiment that I proposed in ICCA_J in 1998 (v21.4) is to take an
established linear evaluation function as the start-point for a 2-level
artificial neural network and see how the ANN's parameters evolve.

In other games, investigators have replaced a single Linear Weighted Sum by
multiple LWSs.  There is an argument that line pieces grow in value, e.g., as
more space becomes available.

This page took 0.04 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.