Learning best concept approximations from examples

This paper addresses the problem of learning the best approximation of a concept from examples, when the concept cannot be expressed in the learner's representation language. It presents a method that determines the version space of the best approximations and demonstrates that for any given approximation of the target concept there is a better approximation in this version space. The method does not depend on the order of examples and has an almost monotonic convergence. This method was developed for the Disciple learning agent that can be taught by a subject matter expert how to perform complex problem solving tasks.