Extrapolation and interpolation in neural network classifiers

Multilayer perceptrons have recently been shown by M.A. Kramer and J.A. Leonard (1990, 1991) to give anomalous behavior in a diagnosis application. It is shown that this unsatisfactory behavior indicates a mismatch of application and technique, rather than any deficiency in the multilayer perceptron. It is further argued that this mismatch is indicative of the differences between applications that require interpolation and those which require extrapolation in addition. Simulation results to indicate a satisfactory performance of multilayer perceptrons in suitable applications are presented.<>