Metric Systems for Performance Evaluation of Active Learning Kriging Configurations for Reliability Analysis

Abstract

In active learning Kriging (AK) reliability-based analysis, a surrogate model is trained in a stepwise manner and used to evaluate the reliability of the desired system by reducing the computational cost of analysis. While extensive studies were conducted on advancing the AK reliability methods by developing new learning functions, limited work studied the effect of AK configuration on the accuracy, efficiency, and consistency of the AK reliability analysis. AK configuration is defined herein as a unique set of Kriging correlation, Kriging regression, learning function, and AK reliability method for the AK procedure. This paper presents six metric systems to evaluate the performance of AK reliability analysis based on AK configurations including the comprehensive metric system (CMS), the weighted metric system (WMS) with local optimized weights or average optimized weights (LOW or AOW), and modified desirability function, and two original desirability functions used for multiple response optimization. The ranking optimizes four scaled indexes as measures of accuracy, efficiency, and consistency of the reliability analysis. The metrics are developed and applied to four diverse examples, where a total of 14,400 AK reliability analyses were considered. The results show the validity of the metric systems to rank AK configurations based on their performance.