GENETIC ALGORITHMS IN FEATURES SELECTION

ILEANA, I.; ROTAR, C.; IOANA MARIA, I.; KADAR, M.; JOLDES, R.

Abstract

Pattern recognition applications often require a feature selection in order to reduce input space dimension, and therefore obtain a reasonable execution time. There are several “traditional” methods for operating this reduction (e.g. Karhunen-Loeve); however, in this paper we attempt to implement feature selection using genetic algorithms. We use neural networks for pattern recognition and the theoretical assumptions are verified when performing 2D image recognition. The genetic algorithms perform the reduction of the input vectors, retaining a smaller number of components, which are significant for and concentrate the essential characteristics of the classes to be recognized. Moreover, by using a genetic tool, the partial (small) vectors are recombined in order to obtain the right pattern (the appropriate pattern). The problem itself represents a complex, difficult problem: on one hand the developed algorithm should be able to recognize the important (characteristic) features and on the other hand, it should be able to reconstruct the patterns. We use for both tasks, evolutionary tools, which have been proved to be efficient instruments for solving these kinds of problems.

Coresponding author e-mail: iileana[at]uab[dot]ro

Session: Intelligent Control Systems