Selecting useful groups of features in a connectionist framework

Suppose for a given classification or function approximation (FA) problem data are collected using 1 sensors. From the output of the ith sensor, n(i) features are extracted, thereby generating P = Sigma(l)(i=1) n(i) features, so for the task we have X subset of R-p as input data along with their corresponding outputs or class labels Y subset of R-c. Here, we propose two connectionist schemes that can simultaneously select the useful sensors and learn the relation between X and Y. One scheme is based on the radial basis function (RBF) network and the other uses the multilayered perceptron (MLP) network. Both schemes are shown to possess the universal approximation property. Simulations show that the methods can detect the bad/derogatory groups of features online and can eliminate the effect of these bad features while doing the FA or classification task.