Minimum Square-Error Modeling of the Probability Density Function

Training of normalized radial basis function neural networks can be considered as a probability density function estimation of the experimental data. A new unsuper-vised method of probability density function estimation is proposed. The method is applied to a multivariate Gaussian mixture model. Batch-mode learning equations are derived and some simple examples are given. Training method is called a minimum square-error modeling of the probability density function. It is similar to the maximum-likelihood method but is numerically less demanding.