Ultrasonic waveforms reflected by various kinds of defects were calculated by numerical analysis based on elastic wave theory. By using these calculated results as a knowledge base, development of nondestructive evaluation system which provides quantitative information, such as types, locations and sizes of the defects is expected. In previous studies, a neural network was applied to construct the knowledge base [1–3]. However, there are some problems in a neural network approach for the quantitative evaluation of defects. A neural network is able to characterize the waveform data effectively, but it does not necessarily interpolate the learning data accurately. In addition, there is an inevitable error caused by the different conditions in numerical analysis and experimental results.