RDS-DeePC: Robust Data Selection for Data-Enabled Predictive Control via Sensitivity Score
Jiachen Li, Shihao Li
TL;DR
RDS-DeePC tackles the dual problems of computational burden and data-quality sensitivity in Data-Enabled Predictive Control by introducing a sensitivity score derived from influence functions. By selecting the K lowest-sensitivity trajectory segments, the method achieves substantial real-time speedups and automatic outlier filtering without requiring data labels. The approach extends from LTI systems (offline sensitivity analysis plus online reduced optimization) to nonlinear dynamics via a two-stage locality-then-robust selection, accelerated by LiSSA. Experimental results on a DC motor and an inverted pendulum validate significant robustness improvements under data corruption and dramatic computational savings, highlighting practical impact for data-driven predictive control.
Abstract
Data-Enabled Predictive Control (DeePC) offers a powerful model-free approach to predictive control, but faces two fundamental challenges: computational complexity scaling cubically with dataset size, and severe performance degradation from corrupted data. This paper introduces Robust Data Selection DeePC (RDS-DeePC), which addresses both challenges through influence function analysis. We derive a sensitivity score quantifying each trajectory segment's leverage on the optimization solution, proving that high-sensitivity segments correspond to outliers while low-sensitivity segments represent consistent data. By selecting low-sensitivity segments, RDS-DeePC achieves computational efficiency and automatic outlier filtering without requiring data quality labels. For nonlinear systems, we extend the framework through a two-stage online selection approach accelerated by the LiSSA algorithm.