| The optimization of classification systems is most of the time performed by a human expert. Classifier complexity may be reduced through feature subset selection (FSS), and a recent trend is to combine several classifiers into ensemble of classifiers (EoC). This thesis proposes a feature extraction based approach to optimize classification systems using a multi-objective genetic algorithm. The approach optimizes feature sets using the Intelligent Feature Extractor methodology. After this stage, the selected classifier can have its complexity reduced through FSS, or the entire IFE result set is used to optimize an EoC. This thesis also details a validation strategy to control over-fit, inspired by classifier training. Finally, a stopping criterion based on the approximation set improvement is proposed and tested. An experiment set is performed on isolated handwritten symbols demonstrate that the approach to optimize classification systems outperforms the traditional approach, also confirming the global validation strategy and the stop criterion. |
