Design Preference Elicitation Using Efficient Global Optimization

We seek to elicit individual design preferences through usercomputer interaction. During an iteration of the interactive session, the computer presents a set of designs to the user who then picks any preferred designs from the set. The computer learns from this feedback and creates the next set of designs using its accumulated knowledge to minimize a merit function. Under the hypothesis that user responses are deterministic, we show that an effective query scheme is akin to the Efficient Global Optimization (EGO) algorithm. Using simulated interactions, we discuss how the merit function form and user preference sensitivity can affect search efficiency and hence the time to complete an interactive session. We demonstrate the proposed algorithm in the design of vehicle exteriors. NOMENCLATURE D the design space with p dimensions. D+ the most preferred region in D . x a design in D . y the response for x. ŷ the prediction of y(x) for a certain model. l the number of designs in each design set during an interaction. t the maximum number of iterations during an interaction. u(x) the utility function defined on D . {βh}h=1 coefficients in the ordinary kriging model. ∗Address all correspondence to this author. β̂ the maximum likelihood estimation in the simple kriging model. R the correlation matrix in the kriging model. λ kriging the Gaussian spread used in R. 1 a column vector where all elements are 1’s. w, b parameters for a linear decision function. C the soft-margin SVM weight. ξi the classification error for sample i. K the l× l kernel matrix. λ SVM the Gaussian spread used in K. α solutions to the soft-margin SVM. s2 the mean square error in a kriging model. w1,w2 weights in the weighted sum merit function. λ utility the Gaussian spread used in u(x).