Design of Self-Optimizing Agent-Based Controllers

The dynamic behaviour of a mechatronic system can be influenced and optimized in many different ways. One way is to use reactive systems that process measurements of the real system and compute control vectors using a preprogrammed stimulus-response behaviour. The optimization is performed indirectly or by learning procedures which will assess a possible success or failure of the real system. A second way is to employ "classical" controllers that are based on differential equations and a model-based optimization that relies on physical models of the plant, including excitation and weighting models. Decoupling of the optimization process from the actual control and the plant allows unimpaired operation of the technical system, which is of vital importance for systems with high safety requirements. This paper presents a combination of both approaches, an agent-based self-optimization of controller systems that allows a foresighted planning of the system behaviour for a certain period of time as well as short-term adaptations of the dynamic behaviour to an altered environment.