Abstract:

By analyzing particle swarm optimization (PSO), a feature selection algorithm is proposed which is on the ground of the simple PSO with adaptive mutationbased  t  distribution. As PSO relapses into local extremum easily, the adaptive mutationbased  t  distribution is used to break it away from local extremum effectively. Meanwhile, the mutual information is introduced for fear that initializing particles selected randomly may extend the searching time. The choosing probability of each feature is obtained by computing the dependence between the feature and class, by means of which an approximate optimum particle is generated. Therefore, the particle swarm can search along the reasonable direction and the evolutionary time is saved. Experiment results indicate the feasibility and validity of the algorithm with a support vector machine (SVM) as the classifier.