Active Control of Complex Systems via Dynamic (Recurrent) Neural Networks

In this work the synthesis of artificial neural networks is examined from the perspective of statistical estimation of functions, and development of synthesis algorithms is centered on new tools for building dynamic (recurrent) neural networks that incorporate internal feedbacks and time delays. The DynNet algorithm is described; it learns the feedforward and feedback structure of a nonlinear dynamic neural network and optimizes the coefficients therein. Applications of the algorithm are presented for the following areas: time-series predictions related to an advanced turbopropulsion combustion process rapid predictions of the responses of a synchronous generator to changes in its input and load conditions predictions of the behavior of a deterministic chaotic process on-line, real-time, optimal two-point boundary-value guidance of an air-to-air missile. The report outlines the advantages of dynamic neural networks and probes the issues related to their synthesis and use.