Original title: Multi-Objective Bayesian Optimization with Active Preference Learning
Authors: Ryota Ozaki, Kazuki Ishikawa, Youhei Kanzaki, Shinya Suzuki, Shion Takeno, Ichiro Takeuchi, Masayuki Karasuyama
The article addresses multi-criteria optimization problems, seeking specific solutions without exploring the entire Pareto front due to excessive search costs. Their solution involves a Bayesian optimization approach tailored for identifying the most preferred solution amidst costly objective functions. They introduce a Bayesian preference model, learning the decision maker’s preferences interactively through pairwise preferences and improvement requests. Their approach integrates uncertainties in objective functions and decision maker preferences within an acquisition function, aiding in exploring the preferred solution. To reduce interaction costs, they propose an active learning strategy for preference estimation. Empirical tests, including benchmark function optimization and machine learning model hyper-parameter optimization, validate the method’s effectiveness. This novel approach streamlines finding preferred solutions in multi-criteria optimization, providing promising pathways for tackling real-world optimization challenges.
Original article: https://arxiv.org/abs/2311.13460