Abstract:

This paper presents a framework that justifies an agent's goal-directed behavior, even in the absence of a provably correct plan. Most prior planning systems rely on a complete causal model and circumvent the 'frame problem' by implicitly assuming that no unspecified relationships exist between actions and the world. In our approach, a domain modeler provides explicit statements about which actions have been incompletely specified. Thus, an agent can minimize its dependence on implicit assumptions when selecting an action sequence to achieve its goals.