Simple feed-forward type neural networks have been investigated to obtain those learning abilities and to obtain effects of initial values of network weights by back-propagation (BP) algorithm. Only restricted types of solutions are obtained by the BP algorithm out of many of them.

A method for constituting new multiple-simulation-type embedding networks (EM networks) is proposed. These EM networks have the multilayer structure, and each layer can be regarded as an ordinary EM network. By using this method, to constitute the complex EM networks including many different conversion parts becomes much easier, and the number of the required immittance converters decreases remarkably.