The effect of intermediate concepts in hierarchical relational learning

Abstract

The DARPA Bootstrapped Learning project uses relational learners to ladder concepts together to teach a final concept, essentially narrowing the search space at each step. However, there are many ways to structure background knowledge to teach a concept and it is uncertain how different ways of structuring knowledge affects the accuracy and performance of learning. In this paper, we examine the effect of having intermediate concepts when learning high level concepts. We used Quinlan's First Order Inductive Learner to learn target selection for a real-time strategy game and did cross-validation tests with varying levels of intermediate concept support. The results show that including intermediate concepts does not always improve performance and accuracy.