Directional Decomposition of Multiattribute Utility Functions

Abstract

Several schemes have been proposed for compactly representing multiattribute utility functions, yet none seems to achieve the level of success achieved by Bayesian and Markov models for probability distributions. In an attempt to bridge the gap, we propose a new representation for utility functions which follows its probabilistic analog to a greater extent. Starting from a simple definition of marginal utility by utilizing reference values, we define a notion of conditional utility which satisfies additive analogues of the chain rule and Bayes rule. We farther develop the analogy to probabilities by describing a directed graphical representation that relies on our concept of conditional independence. One advantage of this model is that it leads to a natural structured elicitation process, very similar to that of Bayesian networks.