The two main areas of application for decision-theoretic planners have been control (e.g. robot navigation) and generation of advice. To date, all work in the second area, with the exception of recent work by Hanks at the University of Washington, has assumed a kind of batch mode of interaction in which the planner is presented with a complete problem description and produces an optimal or near optimal plan. But many areas to which decision-theoretic problem solving techniques could be fruitfully applied require much tighter and more flexible user interaction. Examples include personal planning assistants for planning things like vacations, web-based information retrieval systems, and patient-centered medical advisory systems. Microsoft's office assistant, developed by their Decision Theory Group, is a nice example of the application of decision-theoretic principles to design of interactive systems.
In working with decision-theoretic planners, we have found (not surprisingly) that specification of each planning problem is typically extremely time consuming and error-prone. There are two components to problem specifications for decision-theoretic planners: the probability model and the utility model. The probability model specifies knowledge about the state of the world, the effects of actions, and domain relations. The utility model specifies the objectives to be achieved and the tradeoffs among them. While much work in AI has focused on providing representations and tools for elicitation of probabilities, relatively little work has addressed the elicitation of utility models. This imbalance is not warranted since the probability model is relatively stable across problem instances, while the utility model will typically be different for each instance. For example, in a system that helps users to plan vacations, the model specifying the likelihoods of weather conditions, the reliability of flight connections, and the geographic characteristics of the various destinations is not likely to change from day to day. Furthermore, this model would be built by the company providing the system and even though the effort would be large, the cost could be amortized over thousands of users. On the other hand, each user of the system would have a different set of preferences describing his or her ideal vacation, as well as cost and travel time constraints. While standard techniques are available for eliciting utility models, the elicitation task is typically time-consuming and tedious. Elicitation in decision analysis has focused on specification of a complete model, even though much of the model may be irrelevant to the problem actually being solved. Furthermore, decision-analytic elicitation requires the skill of an expert to identify what information is important and what simplifying assumptions are appropriate.
Once probability and utility models have been built and the appropriate analysis performed, problem solving results must be presented to the user in an easily intelligible form and one that facilitates communicating additional requirements to the system if the user is not satisfied with the results. Presentation of results and generation of explanations is particularly challenging in domains characterized by uncertainty.
My students and I are currently investigating techniques for incremental decision-theoretic problem solving. We have been developing techniques for representing and performing inference with partially specified probability models. We are examining methods for incrementally eliciting preference models, including case-based and constraint-based approaches. We are applying these techniques in building a Decision-Theoretic Interactive Video Advisor (DIVA). Other issues we are planning on investigating include incremental elicitation of probability models, inference with partial models, explanation generation, incorporation of user feedback, and representation of dynamic preference models.
contact info |