Evaluation of neural networks for identification of parameters in mathematical models

Mathematical models of systems and instrumentation are often used in conjunction with numerical algorithms to identify parameters that obtain optimal fits of models to data. These methods are very useful, and they will continue to provide pertinent information. Occasionally, however, they fail to produce physically realistic information when the data contain information that is not included in the model of interest. Neural networks may be more reliable in parameter identification problems than conventional methods. Neural networks are recognized to be robust, and they can approximate any continuous function to any specified accuracy. A group of 26 data sets from measurements on subcritical assemblies was analyzed with several neural networks and with a nonlinear minimization algorithm. A comparison of these results for two neural networks is shown. Networks with a variety of linearly independent transfer functions should be more useful for parameter identification than those with only sigmoidal functions. This claim will be evaluated after the required software development is completed