Characterization of Default Knowledge in Ripple Down Rules Method

\Ripple Down Rules (RDR)" Method is one of the promising approaches to directly acquire and encode knowledge from human experts. It requires data to be supplied incrementally to the knowledge-base being constructed and new piece of knowledge is added as an exception to the existing knowledge. Because of this patching principle, the knowledge acquired strongly depends on what is given as the default knowledge. Further, data are often noisy and we want the RDR noise resistant. This paper reports experimental results about the e ect of the selection of default knowledge and the amount of noise in data on the performance of RDR using a simulated expert. The best default knowledge is characterized as the class knowledge that maximizes the minimum description length to encode rules and misclassi ed cases. The e ect of noise is sensitive at an earlier stage of knowledge acquisition where its performance strongly depends on the selection of default knowledge, but RDR eventually converges to its stable performance although the size of the knowledge base is strongly a ected by the chosen default knowledge.