Name of the participant:: Johannes Westermann
Description of the IT research project: The increasing complexity of mechatronic systems is also reflected in the resource requirements of development. Control concepts must be developed, implemented and optimized for each system, for which expert knowledge and experience are usually indispensable. In addition, the modeling of such complex systems often does not provide sufficiently accurate system models, which form the basis of conventional controller design.
Methods from the field of machine learning can help. Self-learning concepts such as Reinforcement Learning (RL) have received increasing attention in control engineering in recent years. These methods learn by interaction with the system to control it optimally with respect to a given quality requirement. Nevertheless, only a few control concepts are used outside simulation and laboratory environments. The reason for this is the necessity of learning RL methods by trial and error – similar to babies who initially perform awkward, random movements when controlling their extremities.
In this project, RL methods are to be developed which use as much previous knowledge as possible in order to make the initial training phase safe and to be able to use it on real systems. Thus, the resource requirements of the controller design are concentrated on the one-time development of self-learning concepts, which can then automatically learn and apply control laws for complex systems.
Software Campus partner: KIT, IAV
Implementation period: 01.01.2020 – 30.04.2021
