Home > Research Groups > Junior Research Group DART > Research > Bayesian Optimization for the commissioning of controls

Bayesian Optimization for the commissioning of controls

In one of our main research areas, we are investigating the potential of Bayesian Optimization (BO) in control engineering, especially in the interactive implementation on the real system, which is very relevant in practice. A control system designed by a physical model requires fine tuning with respect to the real system. This is e.g. due to the choice of the modeling depth, the identification of system parameters and also because it is generally not possible to capture all real occurring phenomena in the model (simulation-to-reality gap). In most cases, fine-tuning is based on easily implemented approaches, such as manual adjustment by an domain-expert or grid-based heuristics in the sense of a design-of-experiment approach. When implementing complex control structures with several degrees of freedom, however, these approaches quickly reach their limits due to their poor scalability. Furthermore, the objectivity regarding the evaluation of the control is not guaranteed when searching manually. We will solve this problem by using Bayesian optimization during the implementation of open-loop and closed-loop controllers.