J. Jarmulak, S. Craw, and R. Rowe, “Genetic Algorithms to Optimise CBR Retrieval”, in: E. Blanzieri and L. Portinale, editors, Advances in Case-Based Reasoning: Proceedings of EWCBR-2K, Trento, Italy, 2000, pp. 136-147. (pdf)
Abstract: Knowledge in a case-based reasoning (CBR) system is often more extensive than simply the cases, therefore knowledge engineering may still be very demanding. This paper oers a rst step towards an
automated knowledge acquisition and renement tool for non-case CBR knowledge. A data-driven approach is presented where a Genetic Algorithm learns eective feature selection for inducing case-base index, and
feature weights for similarity measure for case retrieval. The optimisation can be viewed as knowledge acquisition or maintenance depending on whether knowledge is being created or rened. Optimising CBR retrieval is achieved using cases from the case-base and only minimal expert input, and so can be easily applied to an evolving case-base or a changing environment. Experiments with a real tablet formulation problem show the gains of simultaneously optimising the index and similarity measure. Provided that the available data represents the problem domain well, the optimisation has good generalisation properties and the domain knowledge extracted is comparable to expert knowledge.