Floating search methods in feature selection

Sequential search methods characterized by a dynamically changing number of features included or eliminated at each step, henceforth "floating" methods, are presented. They are shown to give very good results and to be computationally more effective than the branch and bound method.