ANN modeling of Volterra systems

The authors describe ANN (artificial neural network) simulation experiments that were performed to try to gauge the abilities of ANNs to model systems possessing strong, higher-order nonlinearities. The target nonlinear system was always a Volterra system. The user could specify the degree of nonlinearity by selecting which Volterra terms to include in the series. The ANN architecture was either a feedforward or a recurrent architecture. The number of processing elements and connectivity were varied from experiment to experiment in an effort to establish a relation between Volterra series order and the ANN architecture which succeeded best in modeling the Volterra system. Two typical examples are discussed in detail. Current results indicate that a recurrent network architecture is comparatively more efficient in modeling specific Volterra systems