University of Washington
Computer Science & Engineering
Research in Decision-Theoretic Reasoning
We are studying a variety of ways to apply representational
and algorithmic techniques from AI, especially AI planning
techniques, to problems where there is uncertainty in the
domain model and the decision maker has preferences more
complex than simple goals.
Current projects include:
We can study probabilistic sequential decision making by assuming a goal-directed agent acting in a world where the current state is not known exactly, sensors and effectors are sometimes incorrect, and exogenous events change the world
in ways that cannot be predicted exactly.
MDP and Planning Algorithms
Probabilistic planners and Markov Decision Process (MDP) algorithms solve very similar problems, yet the assumptions, representations, and algorithms have evolved quite differently. We are comparing the two bodies of work both theoretically and empirically.
We are exploring more complex ways to model a system's behavior over time, again based on traditional AI state and action representations. Our current project involved medical diagnostic reasoning.
Another major limitation of classical planning work is that it assumes the decision maker is motivated by its goals alone, and is satisfied only when every goal is fully satisfied (without regard to how long it takes or how much it costs). We are working on formal frameworks and practical algorithms that allow the planning problem to be defined using a value or utility model, and plans are generated that are optimal with respect to that model.
As domain and preference models become more complex, it
is unreasonable to expect that these models can be elicited
fully ahead of time: doing so forces the user to supply a
tremendous amount of information, most of which will turn out to be irrelevant to solving the particular problem at hand. We are working on systems that interleave model elicitation with problem solving, asking for preference information from the user only as it becomes relevant to solving the current problem. Our current application is automating the process of building airline itineraries.