Metareasoning-Based Learning for Classification Hierarchies

This paper takes a metareasoning-based approach to classification learning, framing the learning problem as one of self-diagnosis and self-adaptation. Artificial Intelligence (AI) research on metareasoning for agent self-adaptation has generally focused on modifying the agent´s reasoning processes. In this paper, we describe the use of metareasoning for retrospective adaptation of the agent´s domain knowledge. In particular, we consider the use of meta-knowledge for structural credit assignment in a classification hierarchy when the classifier makes an incorrect prediction. We present a scheme in which the semantics of the intermediate abstractions in the classification hierarchy are grounded in percepts in the world, and show that this scheme enables self-diagnosis and self-repair of knowledge contents at intermediate nodes in the hierarchy. We also provide an empirical evaluation of the technique.