In this paper we address the important problem of optimizing regularization
parameters in neural network modeling. The suggested optimization scheme i
s an extended version of the recently presented algorithm [25]. The idea is
to minimize an empirical estimate - like the cross-validation estimate - o
f the generalization error with respect to regularization parameters. This
is done by employing a simple iterative gradient descent scheme using virtu
ally no additional programming overhead compared to standard training. Expe
riments with feed-forward neural network models for time series prediction
and classification tasks showed the viability and robustness of the algorit
hm. Moreover, we provided some simple theoretical examples in order to illu
strate the potential and limitations of the proposed regularization framewo
rk.