MPC-based high-performance virtual driver tuning using Non-Stationary Bayesian Optimization

The effectiveness of numerical optimal controllers, such as Model Predictive Control (MPC), is highly dependent on the proper selection of hyperparameters. In real-world or realistic systems, this selection is particularly challenging, as even minor variations could drastically alter the closed-loop behavior. This is often due to induced numerical instabilities or an amplification of model-plant mismatches, which may ultimately lead to failures in the closed-loop experiment. Among the available automatic tuning strategies, Bayesian Optimization (BO) has demonstrated strong potential due to its ability to achieve high performance with a limited number of experimental evaluations. Traditionally, classical BO employs stationary Gaussian Processes (GPs) to model the performance function. However, the stationarity assumption limits the model’s capacity to capture discontinuous behaviors that often arise near the boundaries of high-performance regions. To address this limitation, we propose the use of non-stationary GP Bayesian Optimization for calibrating a performance-driven, MPC-based virtual driver. The non-stationary GP enables more accurate modeling of performance landscapes, leading to improved estimates and, ultimately, more effective MPC calibration.