Case-Based Re nement of Knowledge-Based Neural Networks

Knowledge-Based Neural Networks (KBNN) are concerned with the use of domain knowledge to determine the initial structure of Neural Networks(NN). KBNN are shown to classify better unseen examples than randomly initialized NN. In this paper we study the potential of Case-Based Reasoning (CBR) for further improvement of a trained KBNN. The idea is to apply CBR only for correction of KBNN solutions that seem to be wrong. Potential corrections are searched by matching the current situation against the stored cases formed from the KBNN training set examples. The approach is tested on ve well-known machine learning (ML) benchmarkdata sets described with numerical attributes glass , diabetes, liver disorders, breast cancer and iris. Experiments indicate an increased classiication accuracy in the case of three data sets and practically the same performance in the rest 2 data sets.