Knowledge refinement in constraint satisfaction and case classification problems
Knowledge-Base Refinement (KBR) systems are an attempt to help with the difficulties of detecting and correcting errors in a knowledge-base. This thesis investigates Knowledge-Base Refinement within the two problem solving paradigms of Case-Based Reasoning and Constraint Based Reasoning. Case-Based Reasoners make use of cases which represent previous problem solving incidents. Constraint Satisfaction Problems are represented by a set of variables, the possible values these variables can take and a set of constraints further restricting their possible values. This thesis argues that if the problem-solving paradigms of Case-Based Reasoning and Constraint-Based Reasoning are to become truly viable, then research has to be directed at providing support for knowledge-base refinement, but aimed at the knowledge representation formalisms used by the two paradigms rather than more traditional rule-based representations. The CRIMSON system has been developed within the context of an industrial inventory management problem and uses constraint satisfaction techniques. The system makes use of design knowledge to form a constraint satisfaction problem (CSP) which is solved to determine which items from an inventory are suitable for a given problem. Additionally, the system is equipped with a KBR facility allowing the designer to criticise the results of the CSP, leading to knowledge being refined. The REFINER systems are knowledge-base refinement systems that detect and help remove inconsistencies in case-bases. The systems detect and report inconsistencies to domain expert together with a set of refinements which, if implemented would remove the appropriate inconsistency. REFINER+ attempts to overcome the problems associated with REFINER, mainly its inefficiency with large case-bases. The systems can make use of background knowledge to aid in the refinement process, although they can function without any. However, care must be taken to ensure that any background knowledge that is used is correct. If this is not the case, then the refinement process may be adversely affected. Countering this problem is the main aim of BROCKER, which further extends the ideas of REFINER+ to include a facility allowing incorrect background knowledge used to be refined in response to expert criticism of the system's performance. The systems were mainly developed making use of a medical dataset.