A study of two neural network techniques for an automotive system

Neural network techniques, such as feedforward artificial neural networks, are naturally interpreted as performing multi-dimensional function fitting. They are commonly described as universal function approximations, but it is important to note that this claim must be qualified with the assumption that the standard algorithms used to train this type of network are actually able to converge onto a solution with zero error. Our investigation has made use of different neural network learning algorithms to obtain a model of a nonlinear system over a wide nonlinear operating region. In particular, the aim is to identify the best learning algorithm for an automotive application (although all the algorithms considered offer the capability of modelling nonlinear systems without a priori knowledge). In order to perform comparative studies of various learning algorithms, the problem of modelling the motion of a solenoid-operated exhaust gas recirculation valve is considered.