Formulation of tradeoffs in planning under uncertainty

The tradeoffformulation task is to identify the central issues in a decision problem by recognizing strategies that are inadmissible on qualitative grounds. SUDO-PLANNER formulates tradeoffs for an example medical decision problem by proving decisiontheoretically that certain plan classes are dominated based on a knowledge base of qualitative relations in the domain. The classical planning representation of goals as predicates is inadequate for choice among plans that may achieve objectives in part or with uncertainty. SUDO-PLANNER's dominance-proving architecture is a general framework for planning for partially satisfiable goals. Dominance-proving planners characterize the space of admissible plans by maintaining a specialization graph of plan classes annotated with dominance conditions derived from a domain model. Qualitative Probabilistic Networks (QPNs) are decision models expressing constraints on the joint probability distribution over a set of variables. Qualitative influences describe the direction of the relationship between two variables. Qualitative synergies describe interactions among influences. The probabilistic definitions of these constraints justify sound and efficient inference procedures based on graphical manipulations of the network. SUDO-PLANNER's dominance prover uses these procedures to establish dominance relations among plan classes. SUDO-PLANNER constructs decision models (QPNs) from a multilevel knowledge base describing the effects of actions and relations among events. The planning process alternates between model construction and dominance proving, producing a plan graph with dominance conditions ruling out the inadmissible therapy strategies for its medical decision example. Thesis Supervisor: Peter Szolovits Title: Associate Professor of Computer Science and Engineering