Constraint-Directed Improvisation for Complex Domains

Current approaches to reactive planning are limited in their ability to perform well in domains characterized by complexity and significant variability, their ability to perform in domains with which they are less than completely familiar, and by their reliance on local information for decision making. In this paper, we present a novel architecture known as Waffler, based on constraint-directed reasoning. This architecture allows an agent to perform in complex, variable domains in real time by improvising on a routine method of accomplishing an activity using the background knowledge from which the agent's routine is derived. Agents employing this approach can follow a routine in the face of uncertainty and variability, and can apply a routine in a situation with novel aspects, satisficing to the degree that time is available. This paper describes the Waffler architecture's basis in constraint-directed reasoning, it's knowledge structures and processing mechanisms, and an implementation in a simulated environment.