Modeling performance enhancement with constrained linear filters

Estimation of plant Jacobian is necessary for successful control of nonlinear systems using neural networks with the specialized learning scheme. Our previous study showed that neuro-emulators provide a better estimation of the plant Jacobian using a new cost function for learning during the course of dynamic modeling and control. This paper presents an approach for further enhancing the estimation of the plant Jacobian, where a constrained linear filter is proposed to improve the quality of Jacobian teacher signals for on-line modeling. Simulations, including both modeling and adaptive control of a unknown nonlinear system, were carried out to demonstrate the usefulness of the proposed strategy.