Getting the Most from Flawed Theories

摘要

This paper introduces a new classification technique called degree-of-provedness classification, or DOP-classification. This technique exploits information implicit in the structure of a possibly incomplete or incorrect domain theory in order to improve classification accuracy. It is also shown how DOP-classification can be used to identify theories for which theory revision is unnecessary (because the unrevised theory can be used directly by DOP-classification to achieve near-perfect classification accuracy) or insufficient (because the initial theory is so flawed that it would be preferable to induce a new theory directly from examples).