Feature selection with particle swarms

Feature selection is widely used to reduce dimension and remove irrelevant features. In this paper, particle swarm optimization is employed to select feature subset for classification task and train RBF neural network simultaneously. One advantage is that both the number of features and neural network configuration are encoded into particles, and in each iteration of PSO there is no iterative neural network training sub-algorithm. Another is that the fitness function considers three factors: mean squared error between neural network outputs and desired outputs, the complexity of network and the number of features, which guarantees strong generalization ability of RBF network. Furthermore, our approach could select as small-sized feature subset as possible to satisfy high accuracy requirement with rational training time. Experimental results on four datasets show that this method is attractive.