Document Type

Technical Report

Publication Date






Technical Report Number



This paper investigates a different foundation for decision theory in which successive model refinement is central. The idea is to modify utility so that it can sometimes be calculated for an outcome without considering all the relevant properties that can be proved of the outcome, and without considering the utilities of its children. We build partially ordered heuristic utility functions. We treat the analysis of personal decision trees like heuristic search of game trees (taking expectations instead of doing minimax). Analysis of decision then becomes a process of constructing and evaluating defeasible arguments for decision. This leads to an iteratively improving computation of decision, or what Dean and Boddy have dubbed an "anytime algorithm" for decision. An axiomatization of this idea is simple in an existing system of defeasible reasoning. As a special case of defeasible reasoning, computing defeat among decision trees is also simple. The axioms for preference that lead to metric utility can be retained if we take the defeasibility to be a result of the epistemic problem of individuating objects of value. We say nothing yet about the specification of actual search strategies for particular forms of heuristic utility functions, though it is clearly a matter for further research.


Permanent URL: