High performance for expert systems: I. Escaping from the demonstrator class

An investigation was performed to elucidate the problems of converting a demonstrator expert system for leukaemia diagnosis into a large high-performance system. Analysis of the logical completeness of the demonstrator system revealed only minor errors and a limited representation of the domain. Most of its test cases were quite simple and correction of the minor errors raised the performance to 94%. However, the remaining cases were very complex and it is doubtful whether the system could handle them without fundamental changes to its representation of the domain. Demonstrator systems can function with an imperfect view of the domain since the small sacrifice of accuracy is of little consequence; but in a high-performance system the imperfections must be corrected before extra detail is added. The use of default mechanisms for making specific conclusions was harmful because they concealed errors, and a numerical method for handling uncertainty contributed nothing of value to the system.