On-line estimation of the final prediction error via recursive least-squares method

This paper starting from the very first principle presents a derivation of an equation estimating of the final prediction error for a neural network under the recursive least square framework. The equation is in the form:〈〈PE〉F〉T=〈TE〉TN+d1N-d2,where d1d1 and d2d2 are some values determined by the gradient of the nonlinear mapping at the true system parameter. A cheap way of estimating such prediction error based on the information obtained via the recursive least square training method is suggested.