Toward a winning GP strategy for continuous nonlinear dynamical system identification

System identification is the scientific art of building models from data. Good models are of essential importance in many areas of science and industry. Models are used to analyze, simulate, and predict systems and their states. Model structure selection and estimation of the model parameters with respect to a chosen criterion of fit are essential parts of the identification process. In this article, we investigate the suitability of genetic programming for creating continuous nonlinear state-space models from noisy time series data. We introduce methodologies from the field of chaotic time series estimation and present concepts for integrating them into a genetic programming system. We show that even small changes of the fitness evaluation approach may lead to a significantly improved performance. In combination with multiobjective optimization, a multiple shooting approach is able to create powerful models from noisy data.