Machine Learning for Modelling and Control

Focus on data-driven modelling (identification) and control of complex physical/chemical systems, in particular in the high-tech and process technology domains.

The research activities aim at efficiently addressing modelling and control of nonlinear/time-varying behavior of systems in these domains by developing a fusion of system identification, control and machine learning methods. The resulting methods automatically construct dynamical models capturing user specified aspects of the system behavior. In terms of control, policies/algorithms are automatically synthesized that realize a desired behavior of a system by manipulating its actuators. A strong emphasis is put on data-driven structural exploration of the underlying system dynamics, like identification of structured nonlinear systems, and data-driven synthesis of control polices. In this exploration, learning the associated model accuracy/control performance versus complexity trade-off plays an important role. Another focus of the research activities is the development of automated methods that use of surrogate models with linear, but varying dynamical representation concepts, such as linear parameter-varying models, to facilitate technological evolution of currently wide-spread methodologies based on the linear time-invariant framework in engineering.