Selecting subsets of features for the MFS classifier via a random mutation hill climbing technique

The multiple feature subsets (MFS) classifier is a novel approach to one of the major problems in pattern recognition - feature selection. Instead of choosing one set of features, a number of sets of random features participate in voting for the final classification decision. We apply a stochastic strategy for improving accuracy of separate feature sets used in MFS. The experimental results suggest the attractiveness of the proposed idea.