Neural network topologies and training algorithms in nonlinear system identification

Nonlinear system behavior is not always well characterized by linear or linearized system models, especially if the system is rapidly time varying and/or is chaotic. Model paradigms that are themselves nonlinear, such as neural networks, potentially offer more accurate and more robust models for these nonlinear systems. This research studies the use of a neural network structure to model a linear system and two nonlinear systems, a quadratic system and a chaotic system. Several training algorithms are used, including traditional back propagation, an evolutionary programming approach, and a hybrid approach. Net architectures studied here consist of a traditional feed forward topology and a radial basis topology. Modified back propagation training using a feed forward network proved adequate for modeling the linear and quadratic systems, but these were hopelessly inadequate in modeling the chaotic system. The radial basis net fared better, but was still a poor performer for projecting the chaotic system beyond the observed data.