A generalization error estimate for nonlinear systems

Abstract

A new estimate (GEN) of the generalization error is presented. The estimator is valid for both incomplete and nonlinear models. An incomplete model is characterized in that it does not model the actual nonlinear relationship perfectly. The GEN estimator has been evaluated by simulating incomplete models of linear and simple neural network systems. Within the linear system GEN is compared to the final prediction error criterion and the leave-one-out cross-validation technique. It was found that the GEN estimate of the true generalization error is less biased on the average. It is concluded that GEN is an applicable alternative in estimating the generalization at the expense of an increased complexity.<>