THE EQUAL-VALUE SEARCH: ACCELERATING SEARCH IN FUNCTION INDUCTION

Abstract

Function induction is a major component of many learning systems. Its purpose is to extract information in the form of a functional relationship from a given set of examples. Since learning systems strive for generality, search emerges as the best candidate for tackling the task of function induction. However, it is inevitably less efficient than specialized problem-solving methods. Previous researchers have sacrificed generality by developing strategies that utilize domain-specific knowledge to improve efficiency. This thesis presents a different approach, the "equal-value" method, which directly improves search performance while maintaining its generality. The result is a new search strategy that is both general and efficient. Experiments suggest that, in the case of numeric functions, performance can increase by several orders of magnitude compared to generic exhaustive search. While the strategy was developed specifically to address the function induction problem, it is possible that a similar approach applies to other induction problems. In any case, the equal-value search provides a powerful new technique for general function induction.