4. Summary and discussion
Various methods of initialization of adaptive parameters in MLP networks
have been briefly discussed. Initialization of MLPs is still
done more often by randomizing weights [8], although initialization
by prototypes based on initial clusterization presented in [5] and
network construction based on discriminant techniques should give
much better results enabling solutions to complex, real life problems.
Introduction of this methods of network construction should
allow for creation of robust neural systems requiring little optimization
in further training stages. However, for problems requiring
sharp decision borders MLPs trained with the gradient-based techniques
may not be the best models.