Using a-priori information in networks

For control and supervision of dynamic systems, accurate prediction of the outputs and states is important. In the paper a new method is presented for prediction in the case where no, or only an inaccurate, model of the system is available. A feedforward neural network is useful in accurately predicting the output of systems for which a low-order input-output mapping exists. Different feedback configurations have also been suggested. If a part of the dynamic system is known a priori, this information can be exploited to require fewer nodes, fewer weights, and less training to achieve a given accuracy in predicting the system output. The a-priori information also assists convergence toward the global minimum and yields networks that extrapolate better, as demonstrated in this work. A-priori information of most real systems is usually available as a continuous-time nonlinear state-space model. Feedforward networks can utilize a-priori information only when it is available in input-output form. For feedback networks, to the authors' knowledge, no use of a-priori state-space model has been reported so far. In the paper, a new strategy is proposed for fully exploiting state-space information that involves unmeasured system states.<
>