constructive algorithms

[john moody]

In addition to the Cascade Correlation and SONN networks, there is a
large body of work by Barron and Barron and by Russian authors on
synthesizing multilayer polynomial networks. The synthesis algorithms
come under the general title GMDH (Group Method of Data Handling),
and have been developed since the early 1960's. They have been the subject
of large amounts of both theoretical analysis and empirical testing.

Two useful references are:

S.J. Farlow, Self Organizing Methods in Modeling: GMDH Type Algorithms,
New York: Marcel Dekker 1984.

A.R. Barron and R.L. Barron, "Statistical learning networks: a unifying view,"
Computing Science and Statistics: Proc. 20th Interface Symposium, Ed Wegman,
ed. Am. Statist. Assoc. Washington DC, pp 192-203, 1988.

Furthermore, an impressive new algorithm for generating tree-structured
polynomial networks will be presented as an invited talk at NIPS*90 in
three weeks.  The algorithm called MARS (Multivariate Adaptive Regression
Splines) was developed by Jerome Friedman, Chairman of the Stanford
Statistics Department. I believe that a paper on the algorithm is about
to appear in one of the statistics journals. Professor Friedman's email
address is jhf at playfair.stanford.edu.

Hope this helps,

John Moody

-------