Jia-Jie Zhu, Michael Hintermüller
01.01.2022 − 31.12.2023
Efficient and robust control of dynamical systems is a task of practical importance due to uncertainties and external disturbances arising is the real word applications. Classical approaches rarely provide performance guarantees and often rely on tuning the control system while more advances techniques can be challenging from both system-theoretic and algorithmic points of view. To effectively balance between computational complexity and mathematical guarantees one might be interested in establishing a family of methods of different properties arising from specific assumptions about the system.
A pressing challenge for data-driven control systems is the ubiquitous distribution shift. Building upon our previous theoretical works, this project will develop high-performance distributionally robust model predictive controllers, highlighting the theory of optimal transport and kernel methods. We expect our methodological innovations to enable new applications of robust machine learning and optimization to areas such as energy markets.
Central idea of the proposed methodology is to optimise the objective function as per stochastic model predictive control while taking special measures, i.e., distributional robustness, to handle system’s state uncertainty resulting from sensor noise, external disturbances or model mismatch.