A design preference elicitation query as an optimization process

We seek to elicit individual design preferences through human-computer interaction. During an iteration of the interactive session, the computer queries the subject by presenting a set of designs from which the subject must make a choice. The computer uses this choice feedback and creates the next set of designs using knowledge accumulated from previous choices. Under the hypothesis that human responses are deterministic, we discuss how query schemes in the elicitation task can be viewed mathematically as learning or optimization algorithms. Two query schemes are defined. Query type 1 considers the subject's binary choices as definite preferences, i.e., only preferred designs are chosen, while others are skipped; query type 2 treats choices as comparisons among a set, i.e., preferred designs are chosen relative to those in the current set but may be dropped in future iterations. We show that query type 1 can be considered as an active learning problem, while type 2 as a black-box optimization problem. This paper concentrates on query type 2. Two algorithms based on support vector machine and efficient global optimization search are presented and discussed. Early user tests for vehicle exterior styling preference elicitation are also presented.