Evolving Neural Network Classifiers and Feature Subset Using Artificial Fish Swarm

As a novel simulated evolutionary computation technique, artificial fish swarm algorithm (AFSA) shows many promising characters. This paper presents the use of AFSA as a new tool which sets up a neural network (NN), adjusts its parameters, and performs feature reduction, all simultaneously. In the optimization process, all features and hidden units are encoded into a real-valued artificial fish (AF), and give out the method of designing fitness function. The experimental results on several public domain data sets from UCI show that our algorithm can obtain an optimal NN with fewer input features and hidden units, and perform almost as good as even better than an original complex NN with entire input features. And also indicate that optimizing a network classifier for a specific task has the potential to produce a simple classifier with low classification error and good generalization ability