Direct minimization of error rates in multivariate classification (Q1855617)

The authors propose and discusse two versions of a computer intensive method for linear dimension reduction that minimize the classification error directly. The method is based on a powerful modern optimization technique (simulated annealing optimization algorithm) and is a flexible tool for incorporating user preferences. Simulation and a real world example demonstrate the superiority of the proposed methods over the classical discriminant analysis. Special emphasis is given to the case when discriminant analysis collapses.