Abstract
This chapter describes a scheme for using metareasoning in intelligent agents for self-adaptation of domain knowledge. In particular, it considers retrospective adaptation of the content of intermediate abstractions in an abstraction network used for compositional classification when the classifier makes an incorrect classification. It shows that if the intermediate abstractions in the abstraction network are organized such that each abstraction corresponds to a prediction about a percept in the world, then metaknowledge comes in the form of verification procedures associated with the abstractions, and metareasoning invokes the appropriate verification procedures in order to first perform structural credit assignment and then adapt the abstractions.
