Loose coupling of failure explanation and repair: Using learning goals to sequence learning methods

Abstract

Because learning methods (i.e., knowledge repairs) can negatively interact, the arbitrary ordering of knowledge repairs can lead to worse system performance than no learning at all. Therefore, the problem of choosing appropriate learning methods given a performance failure is a significant problem for learning systems. Traditional case-based reasoners index learning or repair methods by specific failure characteristics so that once a failure is detected, a learning method can be brought to bear. Such tight coupling can be contrasted to a loose coupling in which the interaction between failure explanation and learning is mediated by the presence of learning goals generated by the learner. In an empirical study, the Meta-AQUA implementation performed significantly better under the guidance of learning goals (loose coupling) than under a condition in which learning goals were ablated (tight coupling). The conclusion is that unless repair interactions are known not to exist, a loose coupling is necessary for effective learning.