Model-based design of experiments based on local model networks for nonlinear processes with low noise levels

Most common methods for experiment design are classical, geometric designs and optimal designs. Both categories of methods don´t incorporate specific information about the process behavior into the design of experiments. In the case of optimal design often the underlying model structure is chosen as low order polynomial which is very restricted in its flexibility and causes problems, if used for higher-dimensional problems. Furthermore, the focus of these approaches lies on the minimization of the variance error. However, in many applications the process noise is negligible in comparison to the highly nonlinear behavior which usually causes a large bias error. Therefore, this paper presents the new algorithm HilomotDoE which is an active learning algorithm that aims to minimize the bias error of the model. This is achieved by an iterative refinement of a local model network and simultaneously the addition of a certain amount of measurement points. Demonstration examples and theoretical comparisons with the common D-optimal design show the usefulness of HilomotDoE for the mentioned problem class.