🚀 go-pugleaf

A new version (October '95) of the Atree 3.0 Educational Kit is now
available.

It can be found on ftp.cs.ualberta.ca in pub/atree/atree3/atree3ek.exe
(binary mode, 829,794 bytes)

The software runs on PCs under Windows.

It now has a "pause" command in the shell language, and there
are a few other improvements, including better multitasking during
decision tree generation.

EK is not intended for large projects, since it is limited to three
variables -- two inputs and one output.  However one can plot the
resulting functions with the built-in plotting facilities and get
insight into the properties of ALNs fitting data.

**************************************

If anyone, student or prof, is using the Educational Kit in courses, I
would very much appreciate hearing about it, either by email or better
yet on the alnl mailing list or in comp.ai.neural-nets.  Suggestions
for changes or improvement are welcome.

**************************************

An interesting ALN result

One recent application of ALNs is in the area of cardiology.  Mark
Polak has used an ALN on data from resting electrocardiograms to
determine whether or not a person has ischemia (A shortage of blood
getting to the heart).  Taking a resting electrocardiogram is more
convenient and less costly than the standard test whereby the patient
wears a portable monitor.  Mark used a very simple ALN of two linear
pieces as a discriminant.  What is unusual about his result is that it
was better than the method of Parzen windowing, which is (in the limit
of an infinite sample) optimal.  This result makes me thing the project
suggested below will be successful.

****************************

SMALL PATTERN CLASSIFICATION PROJECT WITH EK

An interesting EK project for a student would be to take randomly
generated, smooth, 2-dimensional probability density functions, sample
them, use Parzen windows to get estimated density functions, then
learn these densities with an ALN.  If the ALN smooths the bumps in
the estimator, it may give better generalization than the Parzen
estimator by being a better approximant of the original density.

Actually it is better, I think, to take the log of the probability.
That way, a Gaussian would turn into a quadratic whose function graph
is convex-up and easily approximated by an ALN with a single AND node.

Getting an idea of how well sampled distributions are approximated by
this "Parzen plus ALN "method would be interesting.

Another advantage of the using ALN decision trees to compute the class
conditional densities is, of course, the speed of classification.



--
* Prof. William W. Armstrong,		arms@cs.ualberta.ca
* Computing Science Department,		tel(403)492 2374 fax (403)492 1071
* University of Alberta,		home (403) 438 1103
* Edmonton, Alberta, Canada T6G 2H1	http://web.cs.ualberta.ca/~arms