Automatic tuning of control engineering algorithms with Bayesian optimization

  • Automatisiertes Tuning Regelungstechnischer Algorithmen mit Bayes’scher Optimierung

Stenger, David; Abel, Dirk (Thesis advisor); Trimpe, Johann Sebastian (Thesis advisor)

Aachen : RWTH Aachen University (2023)
Dissertation / PhD Thesis

Dissertation, Rheinisch-Westfälische Technische Hochschule Aachen, 2023


Control engineering algorithms typically rely on various tuning parameters. Correctly setting these is critical to the performance of the closed control loop. However, manual tuning can be tedious, heuristics sub-optimal, and analytical tuning rules are typically based on idealized assumptions that often do not hold in practice. Instead, this thesis formalizes the tuning task by formulating an episodic black-box optimization problem that is approximately solved using Bayesian optimization (BO). This way, the closed loop is optimized w.r.t. arbitrary high-level performance criteria in a sample-efficient manner, i.e., requiring only a few possibly expensive or time-consuming experiments. Three methodological contributions are proposed in the context of BO for control: First, BO-PSA combines BO with the parameter space approach leveraging model knowledge to achieve safe exploration. Second, BO-VDP introduces virtual data points to address failed experiments (crash constraints) in combination with other challenges, such as constraints and multiple objectives. Third, PC-CMES combines multiple approaches to address constrained optimization for various operating conditions (contextual optimization), crash constraints, and dynamic domain extension. A comprehensive benchmark of the sample efficiency of BO for control has not been performed yet. Therefore, this thesis compares eleven variants of BO (including BO-VDP) with seven other black-box optimizers on ten simulative, deterministic, and unconstrained control engineering tasks. Pattern search (PS) is observed to be the most sample efficient for a problem dimensionality d of d = 2. In contrast, for 3 ≤ d ≤ 5, Bayesian adaptive direct search (BADS), a combination of BO and PS, is the most sample efficient. Using PS and BADS outperforms random sampling by, on average, 6.6% and up to 16%. The practical impact of BO on control is demonstrated in seven case studies from different areas of applied control engineering research. Parameters are optimized automatically for various algorithms, e.g., model predictive control and Kalman filter, while addressing different optimization problem structures. For example, PC-CMES is demonstrated experimentally on a physical three-tank system. Furthermore, fault diagnosis and navigation filters are optimized offline using experimental data. Additionally, it is shown in simulations for an underwater vehicle that simultaneously optimizing state estimation, path planning, and control can resolve the dependencies between the different algorithms. In summary, BO effectively addresses various challenges in the practical tuning of controlengineering algorithms while often outperforming competing optimizers.


  • Chair and Institute of Automatic Control [416610]