Bayesian Optimization for Cascade-Type Multistage Processes.

Area of Science:

• Multistage decision-making processes
• Optimization in science and engineering

Background:

• Complex scientific and engineering processes often involve multistage decision-making.
• Cascade processes, where output of one stage feeds into the next, are common.
• Exhaustive parameter searching is often infeasible due to high stage costs.

Purpose of the Study:

• To develop an efficient optimization framework for cascade processes.
• To address the challenge of optimizing expensive multistage decision-making.
• To introduce novel acquisition functions for Bayesian optimization in this context.

Main Methods:

• Formulation of cascade process optimization as an extension of Bayesian optimization.
• Proposal of two acquisition functions: one based on credible intervals, another on expected improvement.
• Investigation of theoretical properties of the proposed acquisition functions.
• Inclusion of a 'suspension setting' for practical applicability.

Main Results:

• Demonstrated effectiveness of the proposed acquisition functions through numerical experiments.
• Validation of the extended Bayesian optimization framework for cascade processes.
• Successful application of the method to a solar cell simulator test problem.

Conclusions:

• The proposed Bayesian optimization extension offers an efficient approach for optimizing costly cascade processes.
• The novel acquisition functions and suspension setting enhance practical utility.
• This method provides a valuable tool for complex multistage decision-making in science and engineering.