Two connectionist schemes for selecting groups of features (sensors)

Suppose for a given classification or function approximation (FA) problem data are collected using l sensors. From the output of the i^th sensor n_i features are extracted, thereby generating p = Σ_i=1^l n_i features. So for the task at hand we have X ⊂ R^p as input data along with their corresponding outputs or class labels. Here we propose two novel connectionist schemes that can select the sensors that yield redundant or bad features online and also do the required task, say, FA or classification. One of the schemes is based on the Radial Basis Function network and the other uses the Multilayered Perceptron network. Simulations show that the methods can detect the bad groups of features online and can also eliminate the effect of these bad features while doing the task.