Hierarchical agent supervision

Agent supervision is a form of control/customization where a supervisor restricts the behavior of an agent to enforce certain requirements, while leaving the agent as much autonomy as possible. To facilitate supervision, it is often of interest to consider hierarchical models where a high level abstracts over low-level behavior details. We study hierarchical agent supervision in the context of the situation calculus and the ConGolog agent programming language, where we have a rich first-order representation of the agent state. We define the constraints that ensure that the controllability of in-dividual actions at the high level in fact captures the controllability of their implementation at the low level. On the basis of this, we show that we can obtain the maximally permissive supervisor by first considering only the high-level model and obtaining a high- level supervisor and then refining its actions locally, thus greatly simplifying the supervisor synthesis task.