A multi-layer feed-forward neural network with dynamically adjustable structures

A general procedure for structure-level adaptation for multilayer feedforward networks is proposed. The general concept of structure-level adaptation for artificial neural networks is presented, and the algorithm for feedforward networks is introduced. The operation of the algorithm is demonstrated by computer simulation on a simple classification problem with time-varying statistics. The results confirm that the algorithm can find the correct structural representation for multilayer feedforward neural networks in time-varying environments. The addition of structure-level adaption to parameter adaptation provides an artificial neural network system with more complete adaptation power than a system that allows only parameter adjustments.<>