Manual tuning of performance-critical controller parameters can be tedious and sub-optimal. Bayesian Optimization (BO) is an increasingly popular practical alternative to automatically optimize controller parameters from few experiments. Standard BO practice is to evaluate the closed-loop performance of parameters proposed during optimization on an episode with a fixed length. However, fixed-length episodes can be wasteful. For example, continuing an episode where already the start shows undesirable behavior such as strong oscillations seems pointless. Therefore, we propose a BO method that stops an episode early if suboptimality becomes apparent before an episode is completed. Such early stopping results in partial observations of the controller’s performance, which cannot directly be included in standard BO. We propose three heuristics to facilitate partially observed episodes in BO. Through five numerical and one hardware experiment, we demonstrate that early stopping BO can substantially reduce the time needed for optimization.

Climate-controlled cabins have for decades been standard in vehicles. Model Predictive Controllers (MPCs) have shown promising results in achieving temperature tracking in vehicle cabins and may improve upon model-free control performance. However, for the multi-zone climate control case, proper controller tuning is challenging, as externally, e.g., passenger-triggered changes in compressor setting and thus mass flow lead to degraded control performance. This paper presents a tuning method to automatically determine robust MPC parameters, as a function of the blower mass flow. Constrained contextual Bayesian optimization (BO) is used to derive policies minimizing a high-level cost function subject to constraints in a defined scenario. The proposed method leverages random disturbances and model-plant mismatch within the training episodes to generate controller parameters achieving robust disturbance rejection. The method contains a postprocessing step to achieve smooth policies that can be utilized in real-world applications. First, simulation results show that the mass flow-dependent policy outperforms a constant parametrization, while achieving the desired closed-loop behavior. Second, the robust tuning method greatly reduces worst-case overshoot and produces consistent closed-loop behavior under varying operating conditions.