Abstract
AI research on metareasoning for agent self-adaptation has generally focused on modifying the agent’s reasoning processes (e.g., [2]). In this paper, we describe our ongoing work on the equally important problem of using metareasoning for modifying the agent’s domain knowledge. Since classification is a ubiquitous task in AI, we consider the problem of using meta-knowledge for repairing classification knowledge when the classifier supplies an incorrect class label. More specifically, we consider the subclass of classification problems that can be decomposed into a hierarchical set of smaller classification problems; alternatively, problems in which features describing the world are progressively aggregated and abstracted into higher-level abstractions until a class label is produced at the root node. This subclass of classification problems is recognized as capturing a common pattern of classification [1] [3]. We will call this classification task compositional classification, and the hierarchy of abstractions an abstraction network (AN).
