Evolutionary feature selection for artificial neural network pattern classifiers

This paper presents FeaSANNT, an evolutionary procedure for feature selection and weight training for neural network classifiers. FeaSANNT exploits the global nature of evolutionary search to avoid sub-optimal peaks of performance. FeaSANNT was used to train a multi-layer perceptron classifier on seven benchmark problems. FeaSANNT attained accurate and consistent learning results, and significantly reduced the number of data attributes compared to four state-of-the-art standard filter and wrapper feature selection methods. Thanks to the robustness of evolutionary search, FeaSANNT did not require time-consuming re-tuning of the learning parameters for each test problem.