Constrained backpropagation for feature selection and extraction of logical
rules

W\lodzis\law Duch, Rafa\l Adamczak and Krzysztof Grabczewski
Department of Computer Methods, Nicholas Copernicus University,
Grudziadzka 5, 87-100 Torun, Poland.
E-mail: duch,raad,kgrabcze@phys.uni.torun.pl 


A new architecture and method for feature selection and extraction of logical
rules from neural networks trained with backpropagation algorithm is presented.
The network consists of nodes that discover linguistic features and nodes that
discover logical rules. Most weights are constrained to $\pm 1$ or zero values.
The relevant input features are automatically generated and selected by the
network. Rules are generated consecutively, from the most general, covering
many training examples, to the most specific, covering exceptions only.
Automatic weight pruning ensures that a minimal number of logical rules is
found. Results for the Iris classification problem illustrate the efficiency of
this method.


Proc. of the First Polish Conference on Theory and Applications of Artificial
Intelligence, d, December 1996}} 
