Penalization method to convert Bayesian optimization methods into batch multi-objective Bayesian optimization methods

Area of Science:

• Engineering
• Computer Science
• Optimization

Background:

• Bayesian optimization is a powerful tool for engineering design.
• Sequential methods are common, but batch methods are more efficient for parallelizable tasks.
• Existing batch multi-objective Bayesian optimization (B-MOBO) methods are often built from scratch, limiting reuse of sequential acquisition functions.

Purpose of the Study:

• To develop a generalizable methodology for converting sequential Bayesian optimization methods into B-MOBO methods.
• To create a novel composite acquisition function for efficient batch selection in multi-objective optimization.
• To improve the real-time performance of Bayesian optimization in parallel engineering design.

Main Methods:

• A transformation methodology is proposed to convert sequential Bayesian optimization into B-MOBO.
• A Euclidean distance-based composite acquisition function with multi-objective penalization averaging is introduced.
• The methodology is applied to Expected Improvement (single-objective) and Quality-Based (multi-objective) acquisition functions to create new B-MOBO variants.

Main Results:

• The developed B-MOBO methods demonstrate comparable solution quality to existing approaches.
• The new B-MOBO methods show significant improvements in real-time performance.
• The transformation methodology effectively leverages sequential acquisition functions for parallel settings.

Conclusions:

• The proposed methodology provides a flexible way to create effective B-MOBO methods.
• This approach enhances the efficiency of Bayesian optimization for parallel engineering design.
• The new B-MOBO variants offer a practical solution for reducing computation time in real-world applications.