An efficient algorithm for output error model reduction

A new technique is presented for reduced-order modeling of linear discrete-time systems, using an output error criterion. A closed form expression is developed for the gradient of the cost function with respect to the model parameters. A computationally efficient algorithm for computing this gradient is derived. A Fletcher-Powell optimization procedure utilizes the gradient vector to compute the reduced-order model parameters. A special initialization procedure is proposed, and the stability of the reduced-order system is monitored. The performance of the algorithm is illustrated by some numerical examples. Results seem to indicate that the proposed technique performs better than many existing model-reduction methods.