Design Preference Elicitation, Derivative-Free Optimization and Support Vector Machine Search

In design preference elicitation, we seek to find individuals’ design preferences, usually through an interactive process that would need only a very small number of interactions. Such a process is akin to an optimization algorithm that operates with point values of an unknown function and converges in a small number of iterations. In this paper, we assume the existence of individual preference functions and show that the elicitation task can be translated into a derivative-free optimization (DFO) problem. Different from commonly-studied DFO formulations, we restrict the outputs to binary classes discriminating sample points with higher function values from those with lower values, to capture people’s natural way of expressing preferences through comparisons. To this end, we propose a heuristic search algorithm using support vector machines (SVM) that can locate near-optimal solutions with a limited number of iterations and a small sampling size. Early experiments with test functions show reliable performance when the function is not noisy. Further, SVM search appears promising in design preference elicitation when the dimensionality of the design variable domain is relatively high.