Investigating a wrapper approach for selecting features using constructive neural networks

This paper investigates the problem of feature subset selection using a wrapper approach implemented using genetic algorithm and a constructive neural network. The main goal of the experiments conducted is to investigate whether the subset of features identified by the wrapper approach, implemented using the DistAl constructive neural algorithm, can also improve the accuracy of other constructive neural algorithms, namely, Tower, Tiling and Upstart algorithms. The results show that, in spite of the wrapper being directed by DistAl, the feature subsets selected can improve the accuracy of the other constructive neural algorithms as well.