Richard Washington and Barbara Hayes-Roth
Optimal real-time planning at run time is impossible for complex domains, since planning is intractable. "Reactive" planning has opted for pre-colnpiled plans, but such plans cannot cope with changing goals, and the space of pre-compiled plans is too large to store for a complex domain. Heuristics can prune planning so that it is of more manageable complexity, but there is still no guarantee that a complete plan will be constructed in a given amount of time. We adopt the approach of building possibly incomplete plans at multiple levels of abstraction. Abstraction provides intermediate goals to guide lower-level planning and replanning, and the use of incomplete plans allows the planner to construct a plan adaptively, improving its plan as more time is allotted. In addition, we have incorporated a temporal representation into the planner. Traditional planning approaches have treated the world as static and discrete, which is an artificial simplification. Representing information temporally makes the planner more useful for interacting with the dynamic and continuous real world.