Learning to plan in continuous domains

摘要

In this paper, we propose an approach to planning in domains with continuous world features. We argue that current models of world change (including traditional planners, reactive systems, and many connectionist systems) implicitly adopt a discrete action assumption which precludes efficient reasoning about continuous world change. A formalism for continuous world change is outlined, and an ideal continuous domain planner is defined. An implemented computationally tractable approximation to the ideal planner is discussed and its behavior is described. Empirically, the implementation is shown to exhibit some of the important design features of the new planning approach. Learning plays a central role in this approach. With experience, accuracy is increased and planning time is reduced even though the system's background knowledge of the world is only approximate or “plausible”. The acquired planning concepts are most accurate in situations similar to the ones in which they are most exercised. Thus, the approach possesses a natural adaptation to systematic properties implicit in the observed distribution of problems.