Empirical Bias for Version Space

Abstract

The ability to generalize remains one of the central issues of concept learning. A general generalization algorithm -the Candidate Elimination Algorithmexists but practical applications of this algorithm are still limited, due to its low convergence. The issue has shifted to the design of a useful bias limiting the size of the Version Space. This paper proposes a new kind of bias, called empirical bias, and a new general algorithm, ICE, for generalization in presence of bias. This proposition is founded on the concept of focus set, which provides a very flexible way to express expectations or constraints on the space of generalizations.