Our Research Field
AI-Based Dynamics Modeling
What Motivates Us
While AI-based modeling has been demonstrated and applied to stationary physical systems with great success, accurate modeling of dynamic systems is still challenging. Scalable models, high accuracy demands, uncertainty estimates, modular and reusable models as well as limited computational resources make this a challenging application domain. Typical applications of these models lie in understanding system behavior as well as optimizing system design, control structures and controller parameters.
To tackle the challenges associated with physics-based modeling in a dynamic environment, we focus on developing accurate dynamics models, such as Gaussian process-based models or recurrent neural networks, that can capture dynamic effects on different time scales. Applications also require the modeling of the system’s unobservable inner states and propagation of uncertain states over time. Furthermore, efficiently learning AI-based models requires optimal data acquisition without destroying the system during measurement, requiring research in the context of safe exploration for active learning. In this context, we also examine how to integrate expert knowledge into or transfer learned knowledge to new models, in particular when data acquisition is cumbersome. Finally, we investigate machine learning approaches in order to optimally control the system while guaranteeing controller stability.
Recent exhaust gas legislation and demands to environmental protection require clean emissions of combustion engines in any given driving scenario. AI-based methods enable efficient development of solutions by providing insights into the system behavior. Many further Bosch applications exist that require optimization of system parameters via dynamics models, such as battery management or manufacturing processes.
Doerr et al.“Probabilistic Recurrent State-Space Models”
- Authors: Andreas Doerr, Christian Daniel, Martin Schiegg, Duy Nguyen-Tuong, Stefan Schaal, Marc Toussaint, Sebastian Trimpe
- Published at ICML in 2018
Schillinger et al."Safe Active Learning and Bayesian Optimization for Tuning a PI-Controller"
- Authors: Mark Schillinger, Benjamin Hartmann, Patric Skalecki, Mona Meister, Duy Nguyen-Tuong, and Oliver Nelles
- Published in IFAC in 2017
Schreiter et al."Efficient Sparsification for Gaussian Process Regression"
- Authors: Jens Schreiter, Duy Nguyen-Tuong, and Marc Toussaint
- Published in Neurocomputing in 2016
Hartmann et al."Online-Methods for Engine Test Bed Measurements Considering Engine Limits"
- Authors: Benjamin Hartmann, Ernst Kloppenburg, Philip Heuser, René Diener
- Published in ISSAM in 2016
Vinogradska et al."Stability of Controllers for Gaussian Process Forward Models"
- Authors: Julia Vinogradska, Bastian Bischoff, Duy Nguyen-Tuong, Henner Schmidt, Anne Romer, and Jan Peters
- Published in ICML in 2016
Schreiter et al."Safe Exploration for Active Learning with Gaussian Processes"
- Authors: Jens Schreiter, Duy Nguyen-Tuong, Mona Eberts, Bastian Bischoff, Heiner Markert, and Marc Toussaint
- Published in ECML in 2015
Schillinger et al."Safe Active Learning of High Pressure Fuel Supply Systems"
- Authors: Mark Schillinger, Benedikt Ortelt, Benjamin Hartmann, Jens Schreiter, Mona Meister, Oliver Nelles
- Published in EUROSIM in 2016
Schreiter et al."Sparse Gaussian Process Regression for Compliant, Real-time Robot Control"
- Authors: Jens Schreiter, Peter Englert, Duy Nguyen-Tuong, and Marc Toussaint
- Published in ICRA in 2015
Bischoff et al."Learning Throttle Valve Control Using Policy Search"
- Authors: Bastian Bischoff, Duy Nguyen-Tuong, Torsten Koller, Heiner Markert, and Alois Knoll
- Published in ECML in 2014
Tietze et al."Model-based Calibration of Engine Controller Using Automated Transient Design of Experiment"
- Authors: Nils Tietze, Ulrich Konigorski, Christian Fleck, Duy Nguyen-Tuong
- Published in ISSAM in 2014