Asynchronous batch Bayesian optimization with pipelining evaluations in experimental equipment-limited situations.

Area of Science:

• Computational Science
• Engineering
• Chemistry
• Biology

Background:

• Bayesian optimization (BO) is data-efficient and widely used in science and engineering.
• Batch BO accelerates optimization by parallelizing experiments but is limited by equipment constraints.
• Sequential methods become time-prohibitive when parallel experiments are scarce.

Purpose of the Study:

• To develop a novel method, Pipelining Bayesian Optimization (PipeBO), to reduce optimization processing time.
• To enable efficient optimization even with a limited number of parallel experiments.

Main Methods:

• PipeBO was inspired by CPU pipelining, dividing computational tasks into multiple processes.
• It achieves experiment parallelization by overlapping various experimental processes.
• PipeBO updates running experiments using results from completed ones, modeling experiments as asynchronous processes.

Main Results:

• PipeBO reduced average optimization processing time by approximately 56% for two-process experiments.
• Further time reductions were observed for experiments with more than two processes.
• The method demonstrated significant speedups for 20 out of 24 benchmark functions.

Conclusions:

• PipeBO effectively reduces optimization time in equipment-limited scenarios.
• It offers a viable solution for accelerating scientific and engineering optimization processes.
• PipeBO enables efficient Bayesian optimization even when full parallelization is not feasible.