"Least Squares Fitting" Using Artificial Neural Networks YARON DANON and MARK J. EMBRECHTS

the neural net for pattern p and output neuron k. This process has a similar minimization objective to that used in the well known least squares fitting method (LSF), where we normally would have k = 1 for the least squares curve fit. The difference between fitting data points with a neural net and fitting with the LSF technique is that with a neural net the fitted function is represented by the network and does not have to be explicitly defined as in the LSF method. The learning process changes the internal parameters (weights) of the network such that the neural net can represent the given set of points in the best way by minimizing the error function. In this paper the results from a backpropagation fit to various continuous functions will be presented, showing properties of neural network fitted functions. This technique is then used to fit the neutron time dependent background for neutron time of flight cross section measurements, where the theoretical shape of the curve is unknown. The distinct advantages of using a neural net over the LSF method will be discussed for this application. error :II: