The Generality of Overgenerality

Several recent papers Cohen, 1990c; Flann and Dietterich, 1989] have presented learning systems that use, as background knowledge, an initial domain theory that is overgeneral: that is, a theory that deenes a concept that is a superset of the concept to be learned. Learning then becomes a problem of specializing the initial theory. Such an overgeneral theory is, of course, only one of many possible ways in which background knowledge can be used in learning. This paper argues that many diierent types background knowledge can expressed as overgeneral theories, albeit overgeneral theories of a somewhat artiicial nature. A consequence of this is that theory specialization techniques are suprisingly general: in particular, we will argue that many diierent types of learning tasks involving background knowledge can be solved using existing theory specialization techniques.