Research On Case-Based Reasoning With Artificial Neural Networks

Case-based reasoning(CBR),compared with its rule-based counterpart, is a new pattern of reasoning having been developed in the recent two decades. It is also an important method of problem solving and learning, which emphasizes the recurrence of past experiences and predecessor's wisdom. Artificial neural networks(ANN) can be used to simulate the visual thinking process of experts, and many of its advantages make itself have some natural connection and complementation with CBR. Thus, it could get good results employing models and techniques of ANN to realize the method of CBR.In this thesis, the history of CBR is first reviewed, followed by the introduction of current research state of CBR as well as its characteristics and application fields. The basic knowledge of CBR and its key techniques are also discussed in detail. After the introduction of the fundamental theory and models of ANN, the possible combination of CBR and ANN is discussed by the former related work in three respects of CBR process, which is case retrieval, case base maintenance and case adaptation.As one of the strategies of case base maintenance based on pattern induction, clustering analysis is used to maintain the case base in CBR system. After the introduction of related basic concepts and the discussion of enough algorithms for clustering analysis, the fuzzy clustering algorithm of cell pruning has been put forward. The experimental results has proved the feasibility and better performance of the algorithm. Followed the clustering analysis for case base, the construction of a new base is discussed.The non-linear relationship between the speed of similar case retrieval and the scale of case base is one of the speciality of the similar case retrieval method based on ANN. Considering the features of general radial basis function (RBF) network and the actual requirements of similar case retrieval and case adaptation, the thesis improves the structure of general RBF network to get the non-linear transition function in output layer. On the basis of clustering analysis of the case base, the model for similar case retrieval based on RBF network is proposed and the corresponding process of retrieval is given. The experimental results of the proposed model are convincing. With respect to several types of problem, the strategies of case adaptation are discussed. Finally, the prototype model of CBR system based on ANN is proposed on the basis of acquired results in the former parts of the thesis.