Recompiling Utility Functions in a Changing World

Our previous work considers a method for building a numeric utility function out of qualitative ceteris paribus preferences, or preferences other things held equal. Dynamic domains embody changes in preferences. This can come in many forms. The preferences themselves may change, the variables over which preferences are expressed may change, or the forms of utility independence that hold in the domain may change. We consider the consequences for each type of change to our system, and conclude that inherent ambiguities in our representation allow for simpler handling of change than might otherwise be the case. Decision Support and Preference While classical decision theory can appear too limited or too rigorous to support decision-making in dynamic domains and under changing conditions, many decision-theoretic formalisms have great conceptual benefit. Researchers in artificial intelligence have made several formulations of qualitative decision theory (Wellman & Doyle 1991; Tan & Pearl 1994; Boutilier et al. 1999), applicable in domains where accurate probabilities and time consuming preference elicitation techniques are either unavailable or undesirable. Such qualitative preference formulations allow decision makers to make natural statements, representing user’s intuitions about the decision space and easing the preference elicitation task. Including qualitative preferences in decision support systems allows personalization. Decision support systems that aim to help more than one decision maker need a way to represent the differences between different users. The most direct method is to represent their preferences. By preferences we mean the desires, tastes, and priorities that a person has, and that furthermore differ from person to person, causing two different people faced with the same decision to make different choices. Ceteris Paribus Preferences in Dynamic