Active Preference Learning for Vehicle Suspensions Calibration

In control system design, the parameters of the feedback action are often computed as the solution of an optimization problem, in which the objective function is defined a priori by the end user. However, in many vehicle control applications, the driver preferences are subjective and hard to mathematically model, thus hours of driving are often needed for end-of-line (EoL) calibration. In this brief, we propose a purely model-free design strategy for vehicle suspensions based on active preference learning (APL). In this method, the controller is tuned only via few dedicated experiments, corresponding to different calibrations and the related user preferences. Since the lack of formal performance metrics might raise concerns about the optimality of the approach, a validation procedure is also proposed, in which an interpretable performance index weighting vertical acceleration and pitch rate is learned from data. This latent index is used to compare the proposed approach with state-of-the-art data-driven optimization.