Orthogonal projection pursuit using genetic optimization

Projection pursuit is concerned with finding interesting low-dimensional subspace of multivariate data. In this paper we proposed a genetic optimization approach to find the globally optimal orthogonal subspace given training data and user defined criterion on what subspaces are interesting. We then applied this approach to human face recognition. Suppose face recognition is done by simple correlation, a subspace is obtained using our approach that achieve the lowest error rate of face recognition given FERET data set as training set. As Yambor [W.S. Yambor, et al., 2000] showed in experiments that PCA subspace is a pretty good subspace for correlation-based face recognition, we compared the performance of the sub-space we obtained with that of PCA subspace. Experiment result showed this subspace outperformed PCA subspace.