A GENETIC APPROACH FOR SELECTION OF (NEAR) OPTIMAL SUBSETS OF PRINCIPAL COMPONENTS FOR DISCRIMINATION

Principal Component Analysis (PCA) is being used both in the preproces sor to a feed-forward neural network and in the Subspace Pattern Recog nition Method (SPRM). Most of the classifiers based on PCA use the fir st few Principal Components (PCs) associated with dominant eigenvalues of the pattern covariance matrix. Recent investigations reveal that c onsidering PCs corresponding to the non-dominant eigenvalues will resu lt in the design of better classifiers in certain application domains: the PCs that are most useful for discrimination may fall in the entir e spectrum of PCs. Finding an optimal subset of PCs which maximizes th e classification rate of a selected classifier is computationally expe nsive as the search space increases exponentially with the increase in either the dimensionality or the number of classes. In this paper, an approach based on genetic algorithms is proposed to search for an opt imal subset of PCs. SPRM is used as the classifier because of its comp utational simplicity. The proposed approach is tested on two real data sets, and the results obtained are better than those obtained with Cl afic, an algorithm used to choose the basis vectors in the SPRM. Altho ugh the PCs selected were from the entire spectrum, a considerable red uction in dimensionality was still present.