Computationally complex, nonlinear systems modeling using neural networks

Neural networks, a powerful machine learning paradigm, have been successfully applied to a wide spectrum of practical problems. This dissertation discusses the modeling of three nonlinear systems using neural networks. Although they are different problems in different fields, they share a common factor---they are all complex nonlinear systems and they all utilize neural networks to model the system and to solve the problem. The first article describes research employing a neural network inverse modeling approach as an improved method of balancing the low rotor of aircraft jet engines. The objectives are to establish propulsion system inverse models that can analyze sensor data and diagnose the unbalance state of the engine, when faced with nonlinearities, which make current engine balancing techniques break down.; The second article discusses the use of an artificial neural network to model the dynamics of complicated gene networks and to learn their parameters. The method is tested by examining small synthetic data sets, generated from a known, randomly generated network. The effectiveness of the method is also demonstrated by recreating the SOS DNA Repair network of Escherichia coli bacterium, previously discovered through experimental data.; The third article describes the use of a multi-stream Extended Kalman Filter (EKF) for the CATS benchmark (Competition on Artificial Time Series) in the time series competition of the International Joint Conference on Neural Networks 2004, to train Recurrent Multilayer Perceptrons (RMLP). A weighted bidirectional approach is adopted, to combine forward and backward predictions, and to generate the final predictions of the missing points.