TR2019-047

Data-Driven Control Policies for Partially Known Systems via Kernelized Lipschitz Learning


    •  Chakrabarty, A., Jha, D., Wang, Y., "Data-Driven Control Policies for Partially Known Systems via Kernelized Lipschitz Learning", American Control Conference (ACC), July 2019.
      BibTeX Download PDF
      • @inproceedings{Chakrabarty2019jul,
      • author = {Chakrabarty, Ankush and Jha, Devesh and Wang, Yebin},
      • title = {Data-Driven Control Policies for Partially Known Systems via Kernelized Lipschitz Learning},
      • booktitle = {American Control Conference (ACC)},
      • year = 2019,
      • month = jul,
      • url = {https://www.merl.com/publications/TR2019-047}
      • }
  • MERL Contacts:
  • Research Areas:

    Control, Machine Learning


Generating initial stabilizing control policies that satisfy operational constraints in the absence of full model information remains an open but critical challenge. In this paper, we propose a systematic framework for constructing constraint enforcing initializing control policies for a class of nonlinear systems based on archival data. Specifically, we study systems for which we have linear components that are modeled and nonlinear components that are unmodeled, but satisfy a local Lipschitz condition. We employ kernel density estimation (KDE) to learn a local Lipschitz constant from data (with high probability), and compute a constraint enforcing control policy via matrix multipliers that utilizes the learned Lipschitz constant. We demonstrate the potential of our proposed methodology on a nonlinear system with an unmodeled local Lipschitz nonlinearity.