Reinforcement Learning-Basierte Worst-Case Analyse zum Testen von Flugregelungssystemen

This dissertation investigates the use of Reinforcement Learning techniques for optimization-based flight control law clearance. The idea is to train an agent to drive a simulation model of the system under test as close as possible to - or even beyond - the admissible limit performance. This allows for the efficient identification of design weak points or, more generally, regions of presumed worst-case performance. The proposed worst-case analysis approach allows the simultaneous consideration of time-varying system inputs and static parameters. Additionally, the thesis provides COLIBRY, a MATLAB software toolbox for worst-case analysis.