A Utility Uncertainty Approach to Multi-Attribute Bayesian Optimization

Abstract

We consider multiattribute Bayesian optimization, where each feasible design is associated with a vector of attributes that can be evaluated via a time-consuming computer code, and the optimizer has not been provided with a utility function over attributes. Past work on multiattribute and multiobjective optimization has focused on estimating the Pareto frontier, measuring performance without a clear link to the value derived from the estimated frontier. We present a new decision-theoretic way to value the information derived from sampling in such multiattribute optimization problems. We assume that a decision-maker has a private utility function over attributes, according to which she may select attribute vectors, but which she cannot easily articulate to the optimizer. We model this utility function as drawn from a Bayesian prior distribution. The decision-maker will use this private utility function to select her most preferred design from a set identified by the sampling algorithm. The algorithm’s goal is to identify a set of designs that maximizes the expected utility of this most preferred design. We develop a novel algorithm using this approach, and show that it is better able to focus sampling effort on designs with attribute vectors that are more likely to be preferred.

Date
Jan 7, 2019 9:00 AM — 9:30 AM
Location
Room 200C, Knoxville Convention Center
701 Henley St, Knoxville, TN, 37902, United States