Dynamic version spaces in machine learning

Abstract

A hybrid learning algorithm for discovering concepts with multiple disjuncts is presented. The algorithm, HYBAL, in incorporating both version spaces and genetic algorithms, extends the work of R.G. Reynolds (1990) to learning of Boolean concepts from an exponentially growing hypothesis space. Learning is accomplished via factoring the underlying version space into tractable subspaces, and then dynamically deriving concepts for the corresponding S set and G sets. In delaying the specification of a concept language until run time, it is demonstrated that HYBAL is capable of solving a larger class of Boolean functions than with traditional version spaces, where concepts are specified at compile time.<>