Data-driven control
Vlastní kurzOsnova témat
-
Updating a parameter (e.g. a reference for an internal feedback system or a controller coefficient) based on the response of the system to a harmonic perturbation added to the input. Model-free, online, continuous- and discrete-time.
-
For repeated finite time horizon tasks with identical references and disturbances (the latter unmodelled). Feedforward control based on the regulation error recorded over the whole previous finite time horizon (or also run/trial/pass/cycle/episode). Both partially model-based and model-free, online, discrete-time.
-
Design of a feedback controller within model reference framework reformulated into a system identification problem. Model-free, offline, discrete-time.
-
Taking advantage of the behavioral approach (introduced and promoted by Jan C. Willems). Past input-output data are directly used for prediction without resorting to system identification. The actual solution presented here has been authored by Florian Dörfler and his colleagues and named DeePC (Data-EnablEd Predictive Control).
-
Equivalent description of a finite-dimensional nonlinear (state-space) system using a linear but infinite-dimensional (Koopman) operator. Data-driven finite-dimensional approximation using (e)DMD. Enables linear predictive control. Offline (using experimental or simulation data), discrete-time (but the Koopman framework itself also continuous-time).
-
Where control meets machine learning. Based on (approximate) dynamic programming. Model-based and model-free, offline and online, discrete-time and continuous-time.