A general-purpose approach to multi-agent Bayesian optimization across decomposition methods.

Area of Science:

• Optimization
• Machine Learning
• Distributed Systems

Background:

• Bayesian optimization (BO) is effective for optimizing expensive black-box functions.
• Distributed optimization often involves agents with unknown local costs and shared constraints.
• Existing multi-agent BO methods may require extensive local data sharing.

Purpose of the Study:

• To propose a general-purpose multi-agent Bayesian optimization (MABO) framework.
• To enable coordination among agents without sharing local cost data.
• To provide a versatile framework adaptable to various decomposition methods.

Main Methods:

• Augmenting traditional BO acquisition functions with coordinating terms.
• Developing a general framework applicable to diverse decomposition techniques.
• Conducting regret analysis for the proposed MABO framework.

Main Results:

• The proposed MABO framework facilitates coordination among subsystems via shared variables or constraints.
• Regret analysis shows cumulative regret is the sum of individual regrets, independent of coordinating terms.
• Numerical experiments confirm the framework's effectiveness across different decomposition methods.

Conclusions:

• The general-purpose MABO framework offers a versatile solution for distributed optimization problems.
• The approach effectively handles agents with unknown local costs and shared dependencies.
• The method's adaptability ensures broad applicability in complex optimization scenarios.