Active Learning in Correcting Domain Theories: Help or Hindrance?

In active learning, the learner uses its current knowledge to guide the choice of future training instances. This dependency between the agent's knowledge and its subsequent training set provides the active learner with a possibility for improved learning and at the same time a potential pitfall that can hinder learning. The opportunity for improved learning comes from choosing closely related instances, in such a way that they help the learner identify why one instance is positive and another is negative. The potential problem is that incorrect early learning can inhibit the learner from seeing instances that will lead it to correctly learn the target concept. These two problems are explored within the task of learning domain knowledge. I propose a method for avoiding the problem and taking advantage of the opportunity, with experimental evidence to verify that this improves the performance of a particular active learner, IMPROV.