A new method for feature selection using radial basis function neural networks based on fuzzy set theoretic measure is proposed. The network's input values are all the membership of f feature values in a certain sample appertaining to C lcass (f: the amount of features, C: the amount of classes). Here, the fuzzy set theoretic π measure based on the normal distribution is used for computing the membership. Hence, there are f x C π measurements that are used as the inputs of the neural network. A radial basis function neutral network with increasing hidden nodes is trained for classification, which is believed to be able to perfectly simulate the nonlinear relevance among the inputs. And then, we set zero to the C input nodes concerning one feature (we call this input vector the revised input vector), which means as far as this feature is concerned, it belongs to none of the classes, which is considered to be the real delinking. The deviation between the output corresponding to the revised input vector and the expected output corresponding to the unrevised one is thought to denote the impact and the importance of this feature. Through this way, we may rank the features and select a suitable feature subset. Effectiveness of this algorithm is demonstrated on several sets of data, and compared with the effect of the feature-bsed node-prune MLP neural network method.