Nonlinear system identification: From multiple-model networks to Gaussian processes

Neural networks have been widely used to model nonlinear systems for control. The curse of dimensionality and lack of transparency of such neural network models has forced a shift towards local model networks and recently towards the nonparametric Gaussian processes approach. Assuming common validity functions, all of these models have a similar structure. This paper examines the evolution from the radial basis function network to the local model network and finally to the Gaussian process model. A simulated example is used to explain the advantages and disadvantages of each structure.