New method of discriminant analysis by utilizing genetic algorithm.

Abstract

The methods and techniques of multivariate analysis are often used in “analysis of impressions”. However, almost all methods and techniques of multivariate analysis respectively premise some kinds of hypotheses or assumptions, such as the variants distributing with the normal distributions, the variants being independent mutually, the objective functions of the analyses being differentiable. Are the actual data to be analyzed conformable with these hypotheses or assumptions? And if the data to be actually analyzed are not conformable with these hypotheses or assumptions, how do the results of the analyses change? This paper shows that the conventional “analytical”method and algorithm does not always guarantee the finding of the most discriminative solution for “discriminant analysis”, which is often used for the “model for estimating impressions”. And it also proposes the new “computational”method and algorithm of discriminant analysis, that can truely find the most discriminative solutions, corresponding to the properties of the data to be actually analyzed and the purposes in its applications of the models.