Creativity and learning in a case-based explainer

Explanation-based learning (EBL) is a very powerful method for category formation. Since EBL algorithms depend on having good explanations, it is crucial to have effective ways to build explanations, especially in complex real-world situations where complete causal information is not available. When people encounter new situations, they often explain them by remembering old explanations, and adapting them to fit. We believe that this case-based approach to explanation holds promise for use in AI systems, both for routine explanation and to creatively explain situations quite unlike what the system has encountered before. Building new explanations from old ones relies on having explanations available in memory. We describe explanation patterns (XPs), knowledge structures that package the reasoning underlying explanations. Using the SWALE system as a base, we discuss the retrieval and modification process, and the criteria used when deciding which explanation to accept. We also discuss issues in learning XPs: what generalization strategies are appropriate for real-world explanations, and which indexing strategies are appropriate for XPs. SWALE's explanations allow it to understand nonstandard stories, and the XPs it learns increase its efficiency in dealing with similar anomalies in the future.