Incorporating domain specific knowledge into version space search

Abstract

While version space are useful in the conceptualization of an inductive concept learning problem, they are seldom used in practice. This is because the implementation of version spaces can use an amount of space that is exponential in terms of the amount of data presented, even for simple conjunctive learning problems. In this paper, an approach is developed that uses domain knowledge to infer an associated G set hypothesis for each active hypothesis in the S set. This allows the use of G set information while concurrently restricting the space required. A prototype, the Knowledge Based Candidate Elimination (KBCE) algorithm, solves the Boole problem using fewer examples than previous approaches, and is extended to a class of boolean functions that subsumes the multiplexer.