A preliminary study of the effect of feature selection in evolutionary RBFN design

M.D. Pérez-Godoy, J.J. Aguilera, F.J. Berlanga, V. Rivas, A.J. Rivera.

In this paper, the effect of the inclusion of a feature selection stage previous to the RBFNs design is analyzed. Two different RBFNs design algorithms have been used: a cooperative-competitive scheme, where each individual is a single neuron, and a Pittsburgh evolutionary scheme, where each individual is a complete network. On the other hand, six different feature selection algorithms (three filter and three wrapper) have been considered. The experimentation shows the generalization ability of the obtained RBFNs (with and without applying feature selection). Furthermore the inclusion of an FS stage leads to less complicated network structure and thus increases the simplicity of the system and the efficiency in processing data.

PDF full paper