Quantitative study on effect of center selection to RBFNN classification performance

In pattern classification problems using a RBFNN classifier, the selection of the number of clusters and their corresponding centers influences the network´s ability to generalize unseen data. In this paper, we evaluate different RBFNN architectures by a quantitative measure - RBFNN sensitivity measure, which is defined as the absolute expectation plus standard deviation of network output perturbations with respect to input perturbations. Numerical comparisons of a number of different RBFNN architectures are given using two of UCI datasets. The experiments show that the sensitivity measure would be correlated to the testing error for the unseen samples and simpler classification problem may have smaller sensitivity measure.