Defeasible Preferences and Goal Derivation

We present a logic for representing and reasoning with qualitative statements of preference and normality and describe how these may interact in decision making under uncertainty. Our aim is to develop a logical calculus in which goals or objectives can be derived in defeasible settings. This system employs the basic elements of classical decision theory, namely probabilities, utilities and actions, but exploits qualitative information about these elements directly. Preferences and judgements of normality are captured in a modal/conditional logic called QDT, for which we present a semantics and sound, complete proof theory. A simple model of action is incorporated into QDT for the purpose of deciding appropriate courses of action. Without quantitative information, decision criteria other than maximum expected utility are pursued. We describe how techniques for conditional default reasoning can be used to complete information about both preferences and normality judgements, and we show how maximin and maximax strategies can be expressed in our logic. We also describe a qualitative analog of the notion of value of information