An empirical improvement of the accuracy of RBF networks

Neural networks have been developed for machine learning and data mining tasks, and because data mining problems contain a large amount of data, sampling is a necessity for the success of the task. Radial basis function networks are one of representative neural network algorithms, and known to have good prediction accuracy in many applications, but it is not known to decide a proper sample size like other data mining algorithms, so the task of deciding proper sample sizes for the neural networks tends to be arbitrary. As the size of samples grows, the improvement in error rates becomes better slowly. But we cannot use larger and larger samples for the networks, because there is some fluctuation in accuracy depending on the samples. This paper suggests a progressive resampling technique to cope with the situation. The suggestion is proved by experiments with very promising results.