Planning for the Real World: Soft Constraints and Incomplete Models

  

Subbarao Kambhampati
Arizona State University

Providing an automated agent the ability to "plan" -- i.e., convert its high-level goals into an executable course of action -- has been a long-standing quest in Artificial Intelligence. For much of the history of automated planning, the dominant research theme has been efficient synthesis of plans under increasingly expressive system dynamics (classical, temporal, stochastic etc.). While this focus has lead to impressive gains in recent years, it needs to be broadened. Real world planning often involves more than efficient synthesis of feasible plans under complete and correct models.

Recent work in my group has focused on two orthogonal themes--soft constraints and (domain) model incompleteness. The former focuses on planning scenarios that admit soft constraints and allow partial satisfaction of goals, thereby foregrounding the plan quality considerations. Work in this direction has lead us to investigate quality-sensitive plan synthesis under a variety of cost and utility interactions between planning goals.

The second theme--model incompleteness--is motivated by the fact that in many real domains the main impediment to the adoption of planning technology is the difficulty of getting complete and correct models of domain physics and user objectives in the first place. Pursuit of model incompleteness has lead us to investigate approaches for generating diverse/multi-option plans, and combining learning and planning techniques to support planning under incomplete and evolving domain models.

I will discuss our progress in these directions, share some challenges and lessons learned, and will try to evangelize broader participation for these research directions.


Subbarao Kambhampati is a professor of computer science and engineering at Arizona State University, where he directs the Yochan
research group. His research and teaching interests are broadly split between automated planning and intelligent information integration,
and he has published over a hundred papers in these areas. Kambhampati is the recipient of an NSF Research Initiation Award
(1992), an NSF Young Investigator Award (1994), a College of Engineering Teaching Excellence Award (2002) and an IBM Faculty Award (2004). In 2004, he was elected a Fellow of American Association of Artificial Intelligence. He was an associate editor of Journal of AI Research and co-chaired the 2000 Intl. conference on Automated Planning and Scheduling, and was the program co-chair of the National Conference on AI (AAAI-05). Many of his former students are in influential R&D positions (including USC/ISI, CMU, Xerox PARC, IBM, Microsoft Research and SRI).