Ensemble methods for nonlinear system identification with uncertainty quantification on example of locally linear-affine multi-models

• Applicant: Professor Dr.-Ing. Andreas Kroll
• Subject Area: Automation, Mechatronics, Control Systems, Intelligent Technical Systems, Robotics
• Term: since 2024
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 541311230

In general, dynamic models obtained by nonlinear system identification methods provide a single output value ("single-value prediction") for the target time. However, a predictor should provide confidence information in addition to the predicted value in order to be able to take into account the trustworthiness or the uncertainty of the predicted value when making decisions. Uncertainty in modeling can arise for model structure, model parameter values, measurement/ observations and initial values, with the treatment of structural uncertainties being particularly difficult.The aim of the project is to research ensemble methods for uncertainty quantification of the prediction in nonlinear system identification, on example of locally affine time-discrete multi-models of the Takagi-Sugeno type. For this purpose, improved estimation and initialization strategies are examined for single-value predicting Takagi-Sugeno models as ensemble members / base learners. Such base learners are subsequently extended by the prediction uncertainty. The next step is to design methods for arranging and for interaction between the ensemble members. The extension to other modeling approaches is also subject to investigation. Since formal proofs cannot be provided for general nonlinear systems and realistic random processes, the methods’ properties are evaluated statistically using a portfolio of case studies with about 10 different dynamical systems incl. test stands of the applicant, so that particular statements are avoided and the basis for transferability to other systems is laid.

• DFG Programme: Research Grants