Rough set based feature selection using swarm intelligence with distributed sampled initialisation

In this paper two evolutionary computation techniques viz. Particle Swarm Optimization (PSO) and Intelligent Dynamic Swarm (IDS) have been implemented for feature selection. In this paper, a population initialization method for PSO and IDS has been proposed, which uniformly samples the search space. Proposed initialization method shows improved performance compared to random initialization techniques. Use of proposed distributed sampled(DS) initialization of seed solutions in PSO and IDS yields significant improvement in selected subset size, execution time and classification accuracy, as compared to randomly initialized PSO and IDS. A pre-processing is also done on all the datasets before applying proposed feature selection method. The fitness function for selection of subset of features is rough dependency measure of any feature or set of features to class labels. Results of the experiments show that with the help of proposed initialization, PSO and IDS are able to select the best set of features with less execution time.