Comparison of parallel infill sampling criteria based on Kriging surrogate model

Area of Science:

• Computational Science
• Engineering Optimization
• Aerodynamics

Background:

• The efficiency of the efficient global optimization (EGO) algorithm is significantly influenced by the selection of infill sampling criteria.
• Existing criteria can suffer from issues like local optima and point concentration, limiting global search capabilities.

Purpose of the Study:

• To compare common efficient parallel infill sampling criteria for EGO algorithms.
• To introduce and evaluate a pseudo-expected improvement (EI) criterion and an adaptive distance function to enhance optimization performance.
• To assess the effectiveness of these improved criteria on actual engineering problems, specifically the inverse design of RAE2822 airfoil.

Main Methods:

• Comparative analysis of established efficient parallel infill sampling criteria.
• Introduction of a pseudo-expected improvement (EI) criterion applied to minimizing the predicted (MP) and probability of improvement (PI) criteria.
• Development of an adaptive distance function to prevent update point concentration and bolster global search.
• Validation using seven test problems and application to the RAE2822 airfoil inverse design.

Main Results:

• The pseudo method demonstrates applicability to PI and MP criteria.
• The proposed pseudo-expected improvement (PEI) and dynamic minimizing the predicted (DMP) criteria emerged as the most efficient and robust.
• Infill criteria incorporating the MP criterion showed higher optimization efficiency in the RAE2822 airfoil inverse design.

Conclusions:

• The enhanced infill sampling criteria, particularly PEI and DMP, offer significant improvements in efficiency and robustness for EGO algorithms.
• The adaptive distance function effectively addresses point concentration, enhancing global search capabilities.
• The applied methods prove effective for real-world engineering optimization challenges, as evidenced by the airfoil design results.