A crucial problem: validation of the knowledge base

Abstract

Knowledge bases of current expert systems are established by competent experts. They are not worked out in one go. During system development the expert is led to make numerous changes in the knowledge base, to modify certain rules, to change his mind about the way of expressing some data in the light of results obtained from tests. Most knowledge bases are perfected by testing and correcting the existing base against cases submitted to the system. This process is lengthy and development of the base is difficult even for the more experienced experts. In some cases the expert resorts to programs which help him to refine the knowledge rules; examples of this type of program are referred to by Davis and Teiresias (1980a), Politakis (1984) and Golmard (1984). The SEEK system developed by Politakis provides advice which leads to interactive refinement of the rules. It is founded on a basis of recorded cases whose purpose is to generalize or to specialize certain rules empirically, to obtain better performances. Generalization consists in weakening a rule by suppressing one or more of its conditions. This leads to more frequent release of the rule. Specialization strengthens a rule by adding one or more conditions which make its application less flexible. Golmard offers a development aid for the knowledge base of the SAM system (Gascuel, 1981) by a method of optimization of production rules weightings to cut down the development and test time which is a preliminary to time taken for validation of the rules base. These aids may be valuable but the knowledge base once constituted must be validated.