Reference: Pfleger, K. & Hayes-Roth, B. Using Abstract Plans to Guide Behavior. Knowledge Systems Laboratory, January, 1998.
Abstract: Plan representation and the method of plan interpretation by execution modules of an agent are critical architectural issues. In contrast to the traditional model of plans as executable programs dictating behavior precisely, we advocate the notion that autonomous agents inhabiting dynamic, unpredictable environments should use plans that only abstractly describe their intended behavior. Similar ideas have been discussed elsewhere, notably by Agre and Chapman, but have seldom been operationalized in complex agents. We present a detailed agent architecture utilizing this idea with a specific useful plan representation and working mechanisms for plan following. Experiments in a number of different domains demonstrate the generality of our approach. In addition to robust improvisation of goal-directed behavior in response to dynamic situations, our approach carries other beneficial run-time advantages and has design and configuration implications as well. The core ideas of the approach are natural extensions of fundamental ideas in software engineering.
Full paper available as ps.