%%%%%%%%%%%%%%%%%%%%%%%% %%%%%%% %%%%%%% Still to do: %%%%%%% --More participant names, esp from outside AI. %%%%%%% --Recent conferences/workshops outside of AI %%%%%%% --Complete ``Pollock'' name %%%%%%% --Possible fourth organizer %%%%%%% %%%%%%%%%%%%%%%%%%%%%%%% ===================== 1994 Spring Symposium Proposal ===================== Title: Decision-Theoretic Planning ===================== Organizing Committee: Steve Hanks---University of Washington (chair) Stuart Russell---University of California, Berkeley Michael Wellman---University of Michigan ===================== Description: Both AI planning and decision theory are devoted to the problem of how an agent chooses a good course of action, based on information about the world, about the agent's capabilities, and about the agent's preferences. The AI community has concentrated on the task of synthesis: given a description of a situation, schematic descriptions of actions and methods for composing them, and some objectives, generate a course of action that furthers those objectives. Classical planning algorithms have for the most part assumed that the agent has perfect information about and control over the world, and that the objectives are described by a symbolic goal state that either is or is not satisfied. These simplifying assumptions conspire to make it difficult to reason about tradeoffs in the planning process, in particular tradeoffs involving the relative likelihood and desirability of possible plan outcomes. Under the classical assumptions, both plan success and plan quality are all-or-nothing propositions. Decision theory provides a language for expressing richer notions of success and quality, along with a normative criterion for making tradeoffs between the likelihood of achieving an objective and the cost of doing so. Decision theory lacks a computational model of how to generate those plans in the first place, however, thus does not address the synthesis task. This symposium aims to unify current lines of research in the AI planning community with research in related disciplines---decision analysis, economics, control theory---by exploring how the richer constructs offered by the decision-theoretic methodology for describing preference and uncertainty can be applied to the problem of plan synthesis. Although AI and decision theory have sometimes been viewed as competing approaches, a growing number of researchers have begun to appreciate their complementary character, and are starting to address the challenge of integrating these ideas. The call for participation will solicit contributions pertaining to all aspects of decision-theoretic planning, particularly those making substantive connections between the two fields. Topics will include, but are not limited to: -- Representing and reasoning about preferences How can the notion of planning goals be extended to richer utility models? How do we represent concepts such as partial goal satisfaction, the cost of consuming resources, multiple objectives, and so on? -- Uncertain effects of actions How can causal or action models be extended to incorporate richer models of change, taking into account uncertainty about the state of the world, the effects of actions, and exogenous events? How do we cope with the complexity of conditional planning? -- Model construction How can representations and techniques from the decision sciences and other disciplines, e.g., graphical decision models such as influence diagrams and solution techniques such as policy iteration, be applied to the problem of plan generation? In particular, how can we exploit these technologies without unduly restricting the expressiveness of our representation for actions and their effects? -- Specialized problems and representations What is the relationship between decision-theoretic techniques for path- and motion-planning problems that generally describe their state spaces and operators numerically and the problems involving symbolic state spaces and operators more typically addressed by AI planning research? -- Decision-theoretic meta-level reasoning How can decision-theoretic criteria be applied to the problem of controlling the reasoning of an agent so that it behaves rationally without (necessarily) using decision-theoretic calculations to make its decisions. Possible techniques include the use of decision theory to control the plan-generation process itself, and using the methodology to pre-compile rational behaviors. Various groups in AI, in the decision sciences, in control theory, and in economics have been pursuing research efforts using the concepts and techniques underlying decision-theoretic planning. The efforts have varied in the issues they consider important and in the simplifying assumptions they make. We want to provide a forum to explore these differences and arrive at a better understanding of the common enterprise. =============================== Evidence of Interest: We have observed an increased interest in the AI planning community over the past several years in issues of uncertainty and partial goal satisfaction, as well as a renewed openness to consider alternative paradigms such as decision theory. At the same time, researchers in uncertain reasoning and the decision sciences have begun to focus far more directly on computational and representational issues salient to plan generation. -- Several recent workshops and conferences have had panels on this topic: * UAI-92 (Conference on Uncertainty in AI) * AAAI-91 workshop on knowledge-based construction of probabilistic and decision models * 1990 Spring Symposium on Planning in Uncertain... Domains * 1989 Spring Symposium on AI and Limited Rationality -- Recent book by Dean and Wellman directly addresses this area. -- Eight relevant papers in AIPS 92, as well as some in recent NCAIs and other AI conferences and workshops address the topic. -- (D)ARPA/Rome Lab Planning Initiative has an explicit component in decision-theoretic planning. -- A significant number of papers in =============================== Possible participants: Fahiem Bacchus, Waterloo Jack Breese, Microsoft Tom Dean, Brown Ed Durfee, Michigan Keiji Kanazawa, UBC Mark Boddy, Honeywell Peter Haddawy, UWM Jon Doyle, MIT Eric Horvitz, Microsoft Ross Shachter, Stanford Max Henrion, CMU Mark Peot, Rockwell Marie desJardins, SRI Paul Lehner, GMU Bruce D'Ambrosio, Oregon State Ron Loui, WashU Greg Provan, UPenn Dave Smith, Rockwell Josh Tenenberg, Indiana James Allen, Rochester Nat Martin, Rochester Reid Simmons, CMU Pollock ????????????/ Piotr Gmytrasiewicz, Hebrew University Jeff Rosenschin, Hebrew University Lonnie Chrisman, CMU Peter Ramadge, Princeton Pravin Varaiya, UCB Ron Howard, Stanford Anil Nerode, Cornell Wolf Kohn, Cornell