On-line optimizing neural networks for Kalman filtered real-time identification

The questions addressed in this paper concern the applicability of system parameters recursive identification using neural networks as well as a method how the exposed problems can be reduced to a standard formulation using dual heuristic dynamic programming and back-propagation training. Some background materials are presented on the recursive least-square methods and task-determination of neural network structure, formulated in terms of generalized linear quadratic control. The results arise with such representation and their training capability could be useful in dynamic systems for adaptive identification and control.