Beyond Classical Planning: Procedural Control Knowledge and Preferences in State-of-the-Art Planners

Abstract

Real-world planning problems can require search over thou-sands of actions and may yield a multitude of plans of dif-fering quality. To solve such real-world planning problems, we need to exploit domain control knowledge that will prune the search space to a manageable size. And to ensure that the plans we generate are of high quality, we need to guide search towards generating plans in accordance with user pref-erences. Unfortunately, most state-of-the-art planners cannot exploit control knowledge, and most of those that can exploit user preferences require those preferences to only talk about the final state. Here, we report on a body of work that extends classical planning to incorporate procedural control knowl-edge and rich, temporally extended user preferences into the specification of the planning problem. Then to address the en-suing nonclassical planning problem, we propose a broadly-applicable compilation technique that enables a diversity of state-of-the-art planners to generate such plans without ad-ditional machinery. While our work is firmly rooted in AI planning it has broad applicability to a variety of computer science problems relating to dynamical systems.