Structured and Unstructured Induction with EDAGs

One objective of knowledge acquisition using inductive modeling has been the development of correct and effectively computable descriptions that can also be assimilated and used by a human being. On substantial problems neither trees nor rules satisfy this objective, both being generally difficult to understand as overt knowledge. Exception directed acyclic graphs (EDAGs) are knowledge structures that subsume trees and rules but can be substantially more compact. The comprehensibility of manually constructed and induced EDAGs is investigated by reconstructing Shapiro’s “structured induction” of a solution to the pawn versus rook end game. It is shown that the induced EDAG is very similar to that produced through consultation with experts, and that both are small, comprehensible solutions to a problem that is difficult for people.

• Type: Conference
• Year: 1995
• Where: KDD
• Authors: Brian R. Gaines
• Keywords: Data Mining, KDD 1995, Knowledge Acquisition, Overt Knowledge, Pawn Versus Rook