Preferences, Planning and Control

Preference handling is a problem of much theoretical and practical interest. In planning, preferences arises naturally when one considers richer notions of goals, as well as oversubscribed planning problems. In knowledge representation, it is a core issue with much recent work on preference languages and algorithms. In system design, preferences can be used to control choices and provide a personalized experience or adapt to varying context. In this talk I will discuss some of my work, together with many colleagues, in these areas. I will consider some of the challenges we face when designing a preference specification formalism and describe a simple graphical input language, CP-nets which attempts to address some of these challenges. Surprisingly, CP-nets are closely related to an important analysis tool in planning the causal graph, and the problem of inference in these networks has important links to the question of the complexity of plan generation. Moreover, the problem of finding a preferred plan given a rich goal specification can be solved by using techniques developed for constrained optimization in CP-nets. But CP-network are inherently a propositional specification language, whereas many control applications require a relational language. Time permitting, I will explain why this problem arises naturally in intelligent control applications. I will show how some recent and richer relational languages can be used to address this problem, and how closely they are related to probabilistic relational models. Preferences play a central role in decision-making as they serve to guide our choice among alternative decisions. Planning problems are a special type of sequential decision problems. Thus, there are obvious relationships between the two areas. What is less obvious is that there are an important technical relationship between reasoning about preferences in some models and planning. This relationship has led to some new insights about the complexity of planning. In this talk, I will discuss some of these relationships: the more obvious relationships between preferences and the notion of a goal, and the less obvious relationship between the problem of planning and the problem of inference in preference reasoning, as well as between graphical structures for planning and preference. The work I will describe is the result of collaborations with quite a few people, including Craig Boutilier, Yury Chernyavsky, Holger Hoos, and Copyright c © 2008, Association for the Advancement of Artificial Intelligence (www.aaai.org). All rights reserved. David Poole, and most notably, Carmel Domshlak, whose thesis (Domshlak 2002) was the first to explore the relationship between causal graphs and CP-networks and made important contributions in both areas. Let us start with the obvious relationship between preferences and goals. Early work in AI focused on the notion of a goal—an explicit target that must be achieved—and this paradigm is still dominant in AI problem solving. But as application domains become more complex and realistic, it is apparent that the dichotomic notion of a goal, while adequate for certain puzzles, is too crude in general. The problem is that in many contemporary application domains, e.g. information retrieval from large databases or the web, or planning in complex domains, the user has little knowledge about the set of possible solutions or feasible items, and what she typically seeks is the best that’s out there. But since the user does not know what is the best achievable plan, she typically cannot characterize it or its properties specifically. As a result, she will end up either asking for an unachievable goal, getting no solution in response, or asking for too little, obtaining a solution that can be substantially improved. Of course, the user can gradually adjust her stated goals. This, however, is not a very appealing mode of interaction because the space of alternative solutions in such applications can be combinatorially huge, or even infinite. Moreover, such incremental goal refinement is simply infeasible when the goal must be supplied off-line, as in the case of autonomous agents (whether on the web or on Mars). Hence, what we really want is for the system to understand our preferences over alternative choices (that is, plans, documents, products, etc.), and home-in on the best achievable choices for us. But what are preferences? Semantically, the answer is simple. Preferences over some domain of possible choices order these choices so that a more desirable choice precedes a less desirable one. We shall use the term outcome to refer to the elements in this set of choices. Naturally, the set of outcomes changes from one domain to another. Examples of sets of possible outcomes could be possible flights, vacation packages, cameras, end-results of some robotic application (e.g, pictures sent from Mars, time taken to complete the DARPA Grand Challenge, etc.). Orderings can have slightly different properties, they can be total (i.e. making any pair of outcomes comparable) or partial, strict (i.e. no two outcomes are equally preferred) or weak but that is about it. A Proceedings, Eleventh International Conference on Principles of Knowledge Representation and Reasoning (2008)