High-Dimensional Slider-Based Preferential Bayesian Optimization With Mixed Local and Global Acquisition Strategies

Abstract

Preferential Bayesian optimization (PBO) is a framework for human-in-the-loop optimization to maximize black-box human preference functions such as seeking perceptually good visual designs. It is advantageous when consistently providing a degree of preference is challenging, but selecting the best option among multiple choices is relatively more straightforward. However, similar to conventional BO methods, PBO suffers from high dimensional problems, that is, finding a good solution becomes increasingly difficult as the dimensionality of the search space increases. In this paper, we focus on slider-based PBO (S-PBO), which is a variant of PBO generating a set of candidates aligned on a one-dimesional line segment, and propose novel acquisition strategies for high dimensional problems. Specifically, we propose two acquisition functions that have a different balance of local/global search abilities. In addition, we propose a simple yet effective randomized strategy that balances the local/global search abilities provided by the two proposed acquisition functions. Through empirical evaluation, we assess the effectiveness of our proposed strategy in improving high-dimensional problems.

Keywords

• Blackbox function optimization
• Gaussian process
• high dimensional optimization
• machine learning
• preferential Bayesian optimization
• preference learning