Two-layer random forests model for case reuse in case-based reasoning

Abstract

Case reuse is important for case-based reasoning (CBR) because without it, a CBR system degrades to a case retrieval system, and a retrieved case generally cannot solve a problem directly. To improve the accuracy of case reuse, a two-layer random forests model is proposed and the framework of the corresponding CBR system is presented. First, clustering analysis is used to organize the cases in the case base, and gravitational self-organizing mapping algorithm is adopted to automatically detect the cluster number on the basis of the structure of the data. Then, a two-layer model scheme is proposed to model the mapping in every cluster. In this scheme, the first layer model obtains the pre-estimate of the output feature value of the query case, and the second layer model models the error of the pre-estimate, which is used to decrease the error generated by the first layer model. Random forests algorithm, which is a popular ensemble learning model, is adopted as a model in the two layers to improve accuracy and stability. Several benchmark datasets are used to validate the proposed two-layer model scheme, and the results demonstrate that it can improve the case reuse accuracy and stability. The proposed two-layer random forests model is applied to hydraulic generator design, and the results confirm that the proposed model is effective for case reuse.