The paper introduces the concept of dynamic output elements (DOE), a novel approach to generate continuous control signals within a self-learning neural control architecture. Using DOEs the control signal is not determined by the decision of the control unit directly, but rather dependent on the temporal sequence of these decisions. Thus the basic mechanisms of decision making and learning are preserved. The benefits of the DOEE-architecture are shown on a challenging benchmark of learning to control a highly nonlinear chemical reactor.
展开▼
