Comparison of Candidate Generation Algorithms in Human-in-the-loop Bayesian Optimization

Abstract

We aim to develop a method that can systematically handle factors that are difficult to incorporate into requirements in the design of products and services, such as people’s aesthetic senses and preferences. We focus on the framework of Comparative Bayesian Optimization, in which the system repeatedly asks a human (oracle) to select a favorite among multiple objects, and the system models the human’s preference based on the results. Typically, a pairwise comparison is used to ask users to compare two objects, but the number of questions tends to get too large for accurate preference modeling. Hence, other methods of asking questions have been considered. How to design an acquisition function is also important in comparative Bayesian optimization. In this paper, we focus on “how to ask questions” and “design of the acquisition function”. In particular, we apply a framework called “batch acquisition function” to the design of acquisition functions, and propose one that can be used in a unified manner for various types of questioning. In order to find a practically useful option among the various system designs, we vary the questioning methods, and confirm the results through simulation experiments.