Reducing the computational demands of continually online-trained artificial neural networks for system identification and control of fast processes

This paper describes many of the generic factors which influence the computational demands of continually online-trained backpropagation artificial neural networks (ANNs) used to identify and control fast processes. The limitations of even parallel hardware in meeting these demands is discussed. An adaptive online-trained ANN induction motor stator current controller is considered as a typical fast process. Various modifications are made to the ANN structure to lower computational demands and increase ANN parallelism. The effects of these modifications on the overall controller stability and performance are illustrated by means of simulation results.