Develop adaptive control design methods based on model augmentation and reinforcing learning that are able to meet stringent operational safety standards during the learning phase and, as a result of learning, create a controller that ensures the guaranteed compliance with pre-set limits for the controlled system and the implementation of the performance requirements.
The increasing performance requirements and explosive hardware complexity in the automotive, mechatronic and chemical applications require the development of control algorithms that target the unique dynamics of a given device and can adapt to its changes, which of course can guarantee the guarantees and reliability of previous control technology. The research aims to develop a theoretical background for the technological change and to demonstrate its effectiveness in experimental devices and industrial applications.
In the case of complex systems or complex management tasks, where a physical model is not available and / or the management strategy cannot be fixed in advance due to the complexity of the task, reinforcement learning-based management planning methods are considered.
In these procedures, it is particularly important to comply with safety requirements during the learning phase, to design explorations that effectively map the operational space, and to ensure that the iterative learning algorithm converges to a management strategy that achieves the desired performance.
Expected results:
Development of self-learning, adaptive (predictive and RL) learning algorithms (toolchain) and related safe learning procedures, their implementation on real systems, demonstration of industrial applicability.