QUALITATIVE PREFERENCES AND DECISION MAKING

For some four decades, artificial intelligence and economics paid little heed to each other's approaches to decision making. Microeconomics and decision theory viewed rational action as acting to maximize expected utility, a quantitative measure combining Bayesian probability with utility judgments. With few exceptions, AI viewed acting as mediated by goals, sentential conditions taken as targets for problem-solving reasoning and search. Each approach offered advantages and disadvantages. Economics offered a precise and deep mathematical theory, but one suited only to labor-intensive analysis of specific decisions. AI offered a practicable approach to formulating and making decisions across a broad range of problems, but one lacking much in the way of formal justification.

Recent work has sought to obtain the respective advantages while avoiding all the disadvantages by developing elements of a ``qualitative'' decision theory that melds both the existing qualitative mathematical foundations of decision theory with the generic comparisons and inferences of reasoned problem solving and deliberation. This talk first surveys the two approaches and their advantages and disadvantages, then explores in more detail a decision-theoretic interpretation of goals as preferences ceteris paribus, that is, as conditions preferred to their opposites, other things being equal. This interpretation sheds new light on traditional decision-analytic methods, and provides the first formal proofs of several traditional methods for reasoning about goals.

The Qualitative Decision Theory web site, http://www.kr.org/qdt, provides further information about this area.