Decision-theoretic techniques were originally developed to help make careful choices in high-stakes situations such as making large business investments, planning military strategy, and choosing among medical treatment alternatives. Researchers in the uncertainty in AI community are asking whether these techniques can also be applied to help make high-quality decisions in more commonplace situations where the stakes are lower.
Decision theory is an attractive framework for building interactive systems for decision making or decision support, but a traditional decision-theoretic analysis requires both a probability model and a utility model, and it is typically time consuming and tedious to elicit either one. This overhead might not be justified by the importance of the problem being solved, especially if the elicitation cost cannot be amortized over many problem-solving episodes.
Often a single problem can be solved without eliciting a complete model in advance. For example, an automated travel agent would not need information about all of a user's travel preferences to build a single itinerary. Because it is usually impossible to ascertain ahead of time exactly what preference information will and will not be relevant to solving a particular problem, there is a need to interleave the elicitation of preference information with the problem-solving process itself.
Fortunately, the richness of the decision-theoretic framework provides valuable flexibility in problem representation. Unimportant portions of a problem space can be represented using coarse preference information or omitted altogether, but more important parts of the decision space can be represented more precisely. Symposium participants presented several techniques for representing and computing with partial or abstract models. Various techniques were …