ABSTRACT
Over the past deeade, a number of Artifieial Intelligenee programs have been eonstrueted for solving problems in seienee, mathematies, and medieine. These programs, termed "Expert Systems" (Duda & Shortliffe, 1983; Feigenbaum, 1977) are designed to eapture what speeialists know, the kind of non-numerie, qualitative reasoning that is often passed on through apprentieeship rather than being written down in books. However, these programs are not generally intended to be models of expert problem-solving, neither in their organization of knowledge nor their reasoning proeess. Consequently, diffieulties have been eneountered in attempting to use the knowledge formulated in these pro grams outside of a eonsultation setting, where getting the right answer is mostly what matters. Their applieation to explanation and teaehing, in partieular, (Brown, 1977a; Claneey, 1983a; Swartout, 1981) has necessitated eloser adherenee to human problem-solving methods and more explieit representation of knowledge. That is, building expert systems whose problem-solving must be eomprehensible to people requires a elose study of the nature of expertise in people.
