Optimal recurrent backpropagation networks for real-time identification

The author first defines process identification (ID) and then discusses real-time ID by neural networks, which is done in a manner involving fully-connected recurrent networks but with a novel use of internal feedback of the activation levels of all neurons at the end of the previous sampling interval. A simple new method, optimal recurrent backpropagation I, for backpropagation training of recurrent nets is presented, and then adapted to identification application. Finally, the subject of optimally efficient and (subject to hardware limitations) arbitrarily rapid training of recurrent networks is treated