Robust direct data-driven controller tuning with an application to vehicle stability control

摘要

In direct data-driven controller tuning, a mathematical model of the plant is not needed, as the control law is directly derived from experimental data. Because the most widely used data-driven techniques are based on the assumption that the underlying dynamics - albeit unknow - is linear, the performance of the resulting controller may not be acceptable with systems whose operating region vary along the time. In this paper, we discuss how to robustify linear data-driven design by exploiting the features of scenario optimization. More specifically, we carry out a modified version of the well known virtual reference feedback tuning approach where probabilistic performance guarantees are given also when the current operating condition is different from the one observed in the controller identification experiment. We validate the proposed approach on a vehicle stability control problem, via a thorough simulation campaign on a multibody simulator. The experimental results show the effectiveness of the proposed approach in a complex real-world setting.