UNIVERSITY OF SOUTHERN CALIFORNIA MARINA DEL REY INFORMATION SCIENCES INST
The use of knowledge in inductive learning is critical for improving the quality of the concept definitions generated, reducing the number of examples required in order to learn effective concept definitions, and reducing the computation needed to find good concept definitions. Relevant knowledge may come in many forms such as examples, descriptions, advice, and constraints and from many sources such as books, teachers, databases, and scientific instruments. How to extract the relevant knowledge from this plethora of possibilities, and then to integrate it together so as to appropriately affect the induction process is perhaps the key issue at this point in inductive teaming. Here we focus on the integration part of this problem that is, how induction algorithms can, and do, utilize a range of extracted knowledge. Preliminary work on a transformational framework for defining knowledge- intensive inductive algorithms out of relatively knowledge-free algorithms is described, as is a more tentative problem-space framework that attempts to cover all induction algorithms within a single general approach. The frameworks help to organize what is known about current knowledge-intensive induction algorithms, and to point towards new algorithms.
Prepared in cooperation with Univ. of Southern CA, Hill Center and AT and T Bell Laboratories.