Case-base reduction using learned local feature weights

Case-base reasoning (CBR) systems making use of previous cases to solve new, unseen and different problems have drawn great attention in recent years. It is true that the number of cases stored in the case library of a CBR system is directly related to the retrieval efficiency. Although more cases in the library can improve the coverage of the problem space, the system performance will be downgraded if the size of the library grows to an unacceptable level. This paper addresses the problem of case base maintenance by developing a new method to reduce the size of large case libraries so as to improve the efficiency while maintaining the accuracy of the CBR System. To achieve this, we adopt the local feature weights approach. This approach consists of three phases. The first phase involves partitioning the case-base into different clusters. The second phase involves learning the optimal local feature weights for each case and the final phase involves reducing the case-base based on the optimal local weights. This paper focuses on the last two phases. To justify the usefulness of the method, we perform an experiment which uses efficiency, competence, and ability to solve new problems as the benchmark to verify our design.