Soft computing techniques for case knowledge extraction in CBR system development

Abstract

The performance of a case-based reasoning (CBR) system depends on its problem-solving quality, efficiency and competence. In a case base, a case can be defined as a piece of contextual and specific knowledge. The more the cases, the better the competence (coverage) of the problem domain, and therefore larger CBR systems tend to provide better solutions than the smaller ones. However, this is not always true because not all the cases collected in the system are useful for problem solving. For example, cases may be in conflict with each other; many cases may be redundant because of their close similarity; some cases may be noises in the system because they are not offering any help in the problem solving, and sometimes may even cause confusion. Another important aspect of CBR system is its efficiency (or speed) in providing helps. The purpose of this research is to examine closely these two aspects, and develop feasible computational techniques that will facilitate the development of CBR systems. This research question leads us to think deeply what constitute the problem solving ability of a CBR system; and also how to strike a balance between efficiency and problem-solving quality. Furthermore, in many real world situations, data and information collected are always incomplete, uncertain and vague, thus, the use of soft computing principles to achieve tractability, robustness and low solution cost is inevitable. Having the above understanding in mind, we then built up a set of soft computing based techniques for the extraction of case knowledge from data. They aim at (i) removing the redundancy and noises; (ii) reducing the size of the case base; and (iii) preserving the problem solving ability (or competence in CBR terminology). The developed algorithms deal with the processes of feature selection and reduction; similarity learning among features; case selection and case generation; and competence model development. Specific concepts and techniques, like approximate reducts; GA-based case-matching; redefined case coverage and reachability measurement; boundary cases with NN guiding principle; fast rough set-based feature reduction; rough LVQ based case generation; fuzzy integral-based case base competence model, are developed, tested and compared with traditional methods such as KPCA and SVM. The experimental results are very promising, and support our objective of trying to develop a compact and competent CBR system through case knowledge extraction.