Neuro-fuzzy approach for maintaining case bases

In practical use of case-based systems, there are always changes in the reasoning environment. Overtime, the case library may need to be updated in order to maintain or improve the performance in response to these changes. The larger the case library, the more the problem space covered. However, it would also downgrade the system performance if the number of cases grows to an unacceptably high level. In order to maintain the size of a case-based system as well as preserving its competence, we propose an approach of selecting representative cases using soft computing techniques. The approach consists of three phases. In the first phase, we determine the degree of membership of each case record to different classes using a neural network. This phase will generate a fuzzy set defined on the cluster space for each record. The second phase is to refine these fuzzy sets by a transformation, where the transformed coefficients are determined by gradient-descent technique, such that the degrees of membership can be as crisp as possible. The last phase uses the fuzzy class membership value of each record to formulate the deletion policy. The case density of each record is computed to determine the percentage of record to be deleted. Using this approach, we could maintain a reasonable size of the case-base without loosing significant amount of information.