Case recognition and strategy classification

In a conventional expert system shell, the process of automated reasoning is generally determined by four components: the current fact situation, a goal, the content of the knowledge base, and control knowledge. This paper focuses on control knowledge. The approach taken is to obtain control knowledge from precedents which resemble the current fact situation. Control knowledge, which is here understood as both referring to search and strategy, is important because control knowledge: 1.-facilitates efficient problem solving; 1.-enables context sensitive and adequate questioning; and 3.-may actually determine the outcome of the reasoning process. With regard to the latter, we believe that one of the manifestations of the discretion a judge is provided with, is that judges often choose a particular control strategy in order to obtain a certain result. As a consequence the strategy followed is often very case-specific. If the aim of a LKBS is to simulate or predict the outcome in a particular problem situation, the system has to simulate or predict the control strategy in this specific case. The model presented uses precedents similar to the case at hand to obtain the applicable control strategy. Neural networks are used to find similar cases. One of the characteristics of the system presented, is, however, that the inferences made by neural networks are not a part of the final reasoning chain. The model is flexible in various ways. Firstly, the precedents may be applicable to a wider range of cases then the precise precedent represented in the case base. Secondly, since the precedent is a model for the control strategy rather than the solution of the case, the amount of problem solving that has to be one after case recognition may vary. This means that the precedent is a model for the type of problem solving and that the specifics of the current case may still determine the precise outcome. Thirdly, since the control strategy determines which questions the system will ask, the recognition method is able to suggest a similar case with the applicable control strategy in a very early stage of problem solving, i.e. before all information is available.