[2302.08298] Unleashing the Potential of Acquisition Capabilities in Excessive-Dimensional Bayesian Optimization

Obtain a PDF of the paper titled Unleashing the Potential of Acquisition Capabilities in Excessive-Dimensional Bayesian Optimization, by Jiayu Zhao and three different authors

Obtain PDF

Summary:Bayesian optimization (BO) is extensively used to optimize expensive-to-evaluate black-box this http URL first builds a surrogate mannequin to signify the target operate and assesses its uncertainty. It then decides the place to pattern by maximizing an acquisition operate (AF) based mostly on the surrogate mannequin. Nonetheless, when coping with high-dimensional issues, discovering the worldwide most of the AF turns into more and more difficult. In such circumstances, the initialization of the AF maximizer performs a pivotal position, as an insufficient setup can severely hinder the effectiveness of the AF.

This paper investigates a largely understudied drawback regarding the influence of AF maximizer initialization on exploiting AFs’ functionality. Our large-scale empirical research reveals that the extensively used random initialization technique typically fails to harness the potential of an AF. In gentle of this, we suggest a greater initialization strategy by using a number of heuristic optimizers to leverage the historic information of black-box optimization to generate preliminary factors for the AF maximize. We consider our strategy with a spread of closely studied artificial features and real-world functions. Experimental outcomes present that our strategies, whereas easy, can considerably improve the usual BO and outperform state-of-the-art strategies by a big margin in most check circumstances.