Hybrid planning - from theory to practice

Abstract

This work lays fundamental groundwork for the development of so-called Companion Systems - cognitive technical systems that are capable to reason about themselves, their users and environment, and to plan a course of action to achieve their users' goals. They are intelligent devices that assist their users in operating them: instead of the user having to learn how to operate the respective system, the system is intelligent and flexible enough to provide its functionality in a truly user-friendly way.

To fully meet a user's demands, Companion Systems rely on a multi-facet of capabilities that stem from different disciplines, such as Artificial Intelligence (AI) planning, knowledge representation and reasoning, dialog management, and user interaction management, to name just a few. This thesis focuses on the relevant aspects of AI planning technology that are of importance for such systems. AI planning is the central technology for many Companion Systems as it allows to compute a course of action that, if followed by its user, achieves his or her goals and therefore serves as a basis of providing advanced user assistance. This thesis is concerned with hybrid planning - a hierarchical planning formalism that is especially suited for the basis of providing assistance to human users. Based on this formalism we will investigate the full endeavor of developing Companion Systems - from theory to practice.

The thesis presents a novel formalization for hierarchical planning problems, which has become a standard in the field. We present a categorization of different problem classes into which hybrid planning as well as other well-known problem classes fall. This formalization allowed to prove a series of novel complexity results that are of interest both for theoretical and practical considerations. For many of the identified classes we introduce novel heuristics that are used to speed up the solution generation process. Some of them are the very first for the respective problem class, and some are the first admissible ones, thereby allowing to find optimal solutions -- which is especially important when plans are generated for human users. We apply hybrid planning in a prototypical Companion System. It assists a user in the task of setting up a complex home entertainment system. Based on a declarative (planning) model of the available hardware and its functionality, the assistant computes a sequence of actions that the user simply needs to follow to complete the setup task. Several so-called user-centered planning capabilities are applied in this system, such as a technique for generating user-friendly linearizations of non-linear plans or the capability to answer questions about the necessity of actions - an essential property to ensure transparency of the system's behavior.

In conclusion: Most modern technical devices are still lacking true intelligence - since no research such as AI planning is sufficiently applied, so there is still huge potential in making such devices really smart by implementing them as cognitive systems that effectively assist their human users. Applying the research presented in this thesis is one step towards achieving this goal.