Bayesian System Identification of Nonlinear Systems: Informative Training Data through Experimental Design

This paper addresses the situation where one is performing Bayesian system identification on a nonlinear dynamical system using a set of experimentally - obtained training data. To be more specifi c, an investigation is performed to find the optimum form of excitation that should be used during generation of the training data. To that end, the Shannon entr opy is used as an information measure such that, through analysing the information content of t he posterior parameter distribution, the `informativeness' of different sets of training data can be assessed. In the current work the form of excitation is parameterised thus allowing the choosing of an appropriate excitation to be phrased as an optimisat ion problem (where one is aiming to maximise the information content of the training data)