Feature selection and classification in noisy epistatic problems using a hybrid evolutionary approach

Abstract

A hybrid evolutionary approach is proposed for the combined problem of feature selection (using a genetic algorithm with Intersection/Union recombination and a fitness function based on a counter-propagation artificial neural network) and subsequent classifier construction (using strongly-typed genetic programming), for use in nonlinear association studies with relatively large potential feature sets and noisy class data. The method was tested using synthetic data with various degrees of injected noise, based on a proposed mental health database.allResults show the algorithm has good potential for feature selection, classification and function characterization.