Fast learning in symbolic/neural models using external constraints and automatic re-structuring

The author has developed an effective method for modelling very complex systems, called n-port symbolic neural (NPSN) modeling. Both production-rule sets and back-propagation multilayer neural nets are used as building blocks. This method is effective for the development of `trained´ models of real physical systems both for simulation of the system as built (with possible faults), and for the modeling of human control behavior. Methods have been devised to speed up the notoriously slow learning process in back-propagation networks. It is shown that if large networks are broken down into architectures of smaller networks to make the input-output data of each subnetwork strictly monotonic, this speeds up the training process. An NPSN model is being created to model the pilot in controlling an airplane in instrument-flying situations. The nature of NPSN models is described, and an example NPSN design cycle is considered