On training radial basis function networks as series-parallel and parallel models for identification of nonlinear dynamic systems

Identification of linear and nonlinear dynamic systems can be performed with a series-parallel or parallel model. In this paper both approaches are compared. While many applications require the model to run in parallel to the process, usually the identification procedure is carried out with a series-parallel model. This paper shows that optimization of a series-parallel model does not necessarily lead to a good parallel model. Furthermore a decrease of the error in one configuration may result in an increase in the other.