Managing constructive induction using optimization and test incorporation
Raj Seshu, Larry A. Rendell, David Tcheng · 2003
Constructive induction is the creation of new terms for describing and learning concepts more effectively. An approach to the problem that uses inductive test incorporation, for fast and effective assessment of terms, is presented. The authors investigate specific effects of system components using two domains: a well-known medical problem, and the relatively unsolved problem of protein folding. The new-term constructor borrows techniques from previous systems for constructive induction and multiple-objective optimization. It utilizes management and test incorporation techniques to control domain-dependent and domain-independent heuristics to extend the concept language. The system learns how to be effective by observing and apportioning credit to the various subcomponents used to generate terms. The system dynamically decides which terms to retain based on a flexible scheme for evaluating them at low cost; the scheme uses variable sampling and different measures of the effectiveness of new terms. Results are presented that show the extent to which each of the authors' techniques aids concept learning, in terms of accuracy and speed.>