Preference Factoring for Stochastic Trees

Stochastic trees are extensions of decision trees that facilitate the modeling of temporal uncertainties. Their primary application has been to medical treatment decisions. It is often convenient to present stochastic trees in factored form, allowing loosely coupled pieces of the model to be formulated and presented separately. In this paper, we show how the notion of factoring can be extended as well to preference components of the stochastic model. We examine updateable-state utility, a flexible class of expected utility models that permit stochastic trees to be rolled back much in the manner of decision trees. We show that preference summaries for updateable-state utility can be factored out of the stochastic tree. In addition, we examine utility decompositions which can arise when factors in a stochastic tree are treated as attributes in a multiattribute utility function.

• 저널Management Science · 46(3) · 389–403
• 토픽Bayesian Modeling and Causal Inference · Artificial Intelligence
• DOI10.1287/mnsc.46.3.389.12067
• 저자Gordon B. Hazen