Discovery of nominally conditioned polynomials: RF6.2 algorithm

摘要

Given data containing both nominal and numeric values, this paper considers discovering a law in the form of a rule set of nominally conditioned polynomials. Recently, a connectionist method called RF6 was proposed to solve problems of this type; however. for real complex problems, RF6 can suffer from a combinatorial explosion in the process of restoring rules from a trained neural network. This paper eliminates the above drawback of RF6 by inventing an efficient restoring procedure, where the number of distinct polynomials is reduced by vector quantization with a model selection criterion, and a set of nominal conditions is extracted by decision tree generation. Our experiments using four data sets showed that the new version of RF6, called RF6.2, works well in discovering very succinct interesting laws even from data containing irrelevant variables and a small amount of noise.