On the Selection of Hidden Neurons with Heuristic Search Strategies for Approximation.pdf

ESANN2006 proceedings - European Symposium on Artificial Neural Networks Bruges (Belgium), 26-28 April 2006, d-side publi., ISBN 2-930307-06-4. On the Selection of Hidden Neurons with Heuristic Search Strategies for Approximation Ignacio Barrio Enrique Romero Llus Belanche Univ Politecnica de Catalunya Dept de Llenguatges i Sistemes Informatics Barcelona Spain Abstract Feature Selection techniques usually follow some search stra tegy to select a suitable subset from a set of features Most neural network growing algorithms p erform a search with Forward Selection with the ob jectiv e of nding a reasonably go o d subset of neurons Using this link b e tween b oth elds feature selection and neuron selection we prop ose and analyze dierent algorithms for the construction of neural networks based on heuristic search strategies coming from the feature selection eld The results of an exp erimental comparison to Forward Selection using b oth synthetic and real data show that a much b etter approximation can b e achieved though at the exp ense of a higher computational cost Intro duction The main ob jective of feature selection in inductive learning is to select the most suitable subset from a set of features Analyzing all p ossible subsets is usually unattainable and a heuristic search is often carried out in order to nd a go o d subset One of these search strategies is Forward Selection that starts with an empty subset and adds the most salient feature at a time as measured by some cost function keeping all the previously selected features in the current subset On t