NONLINEAR SYSTEM IDENTIFICATION BASED ON EVOLUTIONARY DYNAMIC NEURAL NETWORKS WITH HYBRID STRUCTURE

Abstract

The paper presents a novel dynamic neural architecture that allows a flexible and compact representation of the nonlinear processes. The suggested neural topology considers local internal recurrence and a heterogeneous structure of the hidden layer. It allows the cooperation between different types of hidden units, such as perceptrons, sigmoidal neurons with functional links, radial basis function structures and/or gaussian neurons with complex weights. An evolutionary multiobjective procedure assists the automatic design of appropriate neural networks. It searches for accurate neural models, characterised by good generalisation capabilities. The experiments reveal that the presented approach is suitable for system identification.