Combining Case Based Reasoning with Neural Networks

This paper presents a neural network based technique for mapping problem situations to problem solutions for Case-Based Reasoning (CBR) applications. Both neural networks and CBR are instance-based learning techniques, although neural nets work with numerical data and CBR systems work with symbolic data. This paper discusses how the application scope of both paradigms could be enhanced by the use of hybrid concepts. To make the use of neural networks possible, the problem’s situation and solution features are transformed into continuous features, using techniques similar to CBR’s definition of similarity metrics. Radial Basis Function (RBF) neural nets are used to create a multivariable, continuous input-output mapping. As the mapping is continuous, this technique also provides generalisation between cases, replacing the domain specific solution adaptation techniques required by conventional CBR. This continuous representation also allows, as in fuzzy logic, an associated membership measure to be output with each symbolic feature, aiding the prioritisation of various possible solutions. A further advantage is that, as the RBF neurons are only active in a limited area of the input space, the solution can be accompanied by local estimates of accuracy, based on the sufficiency of the cases present in that area as well as the results measured during testing. We describe how the application of this technique could be of benefit to the real world problem of sales advisory systems, among others.