A description length-based decision criterion for default knowledge in the ripple down rules method (Q1606556)

The `Ripple Down Rules (RDR)' method is a promising approach to directly acquiring and encoding knowledge from human experts. It requires data to be supplied incrementally to the knowledge base being constructed, each new piece of knowledge being added as an exception to the existing knowledge base. Because of this patching principle, the knowledge acquired depends strongly on what is given as the default knowledge, used as an implicit outcome when inference fails. Therefore, it is important to choose good default knowledge for constructing an accurate and compact knowledge base. Further, real-world data are often noisy and we want the RDR to be noise resistant. This paper reports experimental results about the effect of the selection of default knowledge and the amount of noise in data on the performance of RDR, using a simulated expert in place of a human expert. The best default knowledge is characterized as the class knowledge that maximizes the description length of encoding rules and misclassified cases. We confirmed by extensive experimentation that this criterion is indeed valid and useful in constructing an accurate and compact knowledge base. We also ascertained that the same criterion holds when the data are noisy.