Feature selection vs theory reformulation: A study of genetic refinement of knowledge-based neural networks

Expert classification systems have proven themselves effective decision mak ers for many types of problems. However, the accuracy of such systems is of ten highly dependent upon the accuracy of a human expert's domain theory. W hen human experts learn or create a set of rules, they are subject to a num ber of hindrances. Most significantly experts are, to a greater or lesser e xtent, restricted by the tradition of scholarship which has preceded them a nd by an inability to examine large amounts of data in a rigorous fashion w ithout the effects of boredom or frustration. As a result, human theories a re often erroneous or incomplete. To escape this dependency, machine learni ng systems have been developed to automatically refine and correct an exper t's domain theory. When theory revision systems are applied to expert theor ies, they often concentrate on the reformulation of the knowledge provided rather than on the reformulation or selection of input features. The genera l assumption seems to be that the expert has already selected the set of fe atures that will be most useful for the given task. That set may, however, be suboptimal. This paper studies theory refinement and the relative benefi ts of applying feature selection versus more extensive theory reformulation .