A Portfolio Approach to Massively Parallel Bayesian Optimization.

Area of Science:

• Computational Science
• Optimization Theory
• Machine Learning

Background:

• Traditional optimization studies evaluate designs sequentially, which is time-consuming.
• Batch Bayesian optimization uses surrogate models for parallel evaluations but struggles with large-scale batching.

Purpose of the Study:

• To develop a scalable strategy for massive batch Bayesian optimization.
• To address the exploration/exploitation trade-off in large-scale parallel evaluations.
• To improve efficiency for noisy and multi-objective optimization problems.

Main Methods:

• Proposed a scalable strategy for massive batch Bayesian optimization.
• Focused on exploration/exploitation trade-off and portfolio allocation.
• Compared the approach with existing methods on noisy functions for mono- and multi-objective tasks.

Main Results:

• Achieved orders of magnitude speed improvements over existing methods.
• Maintained similar or better performance compared to current approaches.
• Demonstrated scalability for massive batching in noisy optimization.

Conclusions:

• The proposed scalable strategy significantly enhances the efficiency of parallel Bayesian optimization.
• This method is effective for large-scale, noisy, and multi-objective optimization problems.
• Offers a promising solution for accelerating complex simulation and design studies.