A probability estimation based criterion for model evaluation

Abstract

We develop a criterion based on the estimation of the joint probability density function (pdf) of the input and the error, and on the pdf of the input. It is made to decide when the couple inputmodel no longer fit together. The estimation of the pdf is made through a Probabilistic Radial Basis Function Network (PRBFN) which can also be used to estimate the given task. We compare the results when using a dedicated network, or when extracting the density value directly from the network which estimates the input-output mapping.

We develop a criterion based on the estimation of the joint probability density function (pdf) of the input and the error, and on the pdf of the input. It is made to decide when the couple inputmodel no longer fit together. The estimation of the pdf is made through a Probabilistic Radial Basis Function Network (PRBFN) which can also be used to estimate the given task. We compare the results when using a dedicated network, or when extracting the density value directly from the network which estimates the input-output mapping.