Kösters, Tarek

Preferred name
Kösters, Tarek

Settings

Now showing 1 - 3 of 3
  • Some of the metrics are blocked by your 
    consent settings
    PublicationAccess status: Metadata only ,
    Density Estimation in the Regressor Space for Nonlinear System Identification
    (2023)
    Peter, Timm J.
    ;
    Kösters, Tarek
    ;
    Nelles, Oliver
    In this paper, the functioning and performance of a recently proposed weighting approach for dynamic datasets is demonstrated for nonlinear system identification. The overall goal of the new method is to improve data-driven model performance. This is achieved by increasing the weight of training data points in areas of the regressor space that are underrepresented during training, while decreasing the weight for overrepresented data points. The method is based on estimating the parameters of a model using weighted error norms, where the weighting is carried out with the inverse values of the kernel density estimated in the regressor space. The performance of the new method is investigated using nonlinear ARX models for the identification of (i) an artificial Hammerstein process and (ii) a nonlinear real-world process. Overall, the new weighting method shows significantly improved performance on the simulation error. It is demonstrated that performance is improved most notably in areas that are underrepresented in the training data.
    Source Type:Article
    DOI:10.1016/j.ifacol.2023.10.078
      1
  • Some of the metrics are blocked by your 
    consent settings
    PublicationAccess status: Metadata only ,
    Handling of time delays in system identification with regularized FIR models
    (2023)
    Kösters, Tarek
    ;
    Illg, Christopher
    ;
    Nelles, Oliver
    System identification with finite impulse response (FIR) models has been recently revised by introducing specific forms of regularization in the estimation. By utilizing prior knowledge of the dynamic process under investigation different kernels for regularization can be derived. However, the dynamical processes considered are mainly restricted to time-delay free systems. Therefore, in this paper we propose two novel methods to handle the time-delay case. Both methods incorporate the time-delay estimation in the hyperparameter optimization of the regularized FIR models. The first method is built on a single kernel (SK) and the second utilizes a multiple kernel (MK) approach. Simulations on different dynamical systems show the superior behavior of both methods in comparison to standard regularization methods, whereupon the MK outperforms the SK approach.
    Source Type:Article
    DOI:10.1515/auto-2023-0007
      1
  • Some of the metrics are blocked by your 
    consent settings
    PublicationAccess status: Metadata only ,
      3