Preferential Bayesian Optimization with Inequality Constraints

Abstract

Preferential Bayesian Optimization (PBO) is a method for creating efficient human-in-the-loop optimization systems that treat human preferences as an objective function to be maximized. PBO has been successfully applied to simple design scenarios. However, design tasks often involve more complex challenges where finding the optimal design requires considering not only subjective preferences but also design constraints. This paper presents a new method to integrate additional criteria in the form of inequality constraints into PBO. We specifically propose a new acquisition function to enable this integration. Our evaluation using synthetic functions shows that our method identifies optimal solutions by effectively focusing on feasible solutions.