Online Coreset Selection for Learning Dynamic Systems

With the increasing availability of streaming data in dynamic systems, a critical challenge in data-driven modeling for control is how to efficiently select informative data to characterize system dynamics. In this work, we design an online coreset selection method under the framework of set-membership identification for systems subject to process disturbances, with the objective of improving data efficiency while ensuring convergence guarantees. Specifically, we first propose a stacked polyhedral representation that over-approximates the feasible set of system parameters. Leveraging a generalized Grünbaum's inequality, we design a geometric selection criterion for constructing the coreset. To reduce computational complexity, an online double-description-based constraint reduction method is introduced to simplify the polyhedral representation. Finally, we analyze the convergence of the feasible set with respect to the coreset and derive upper bounds on the selection probability and the expected number of data in the coreset. The effectiveness of the proposed method is demonstrated through comprehensive simulation studies.