A Comparative Study of Learning Functions in Adaptive Kriging for Reliability Analysis

As engineering structures have been becoming more and more complicated, many uncertain factors should be considered in engineering design process in order to meet the design requirements. Reliability analysis methods which can quantify the effects of the uncertainties has been widely used. However, they still need high computational cost in complex problems. For the purpose of reducing computational cost, the adaptive Kriging (AK) model has gained a lot of attentions in the reliability analysis field. However, the choice of learning function has a great influence on its performance. In this paper, we compare three classical learning functions, i.e., the U-, EFF- and H- learning functions, through five examples. The results indicate that the performance of the U-learning function is greatly influenced by the initial samples, while the initial samples have nearly no effects on that of the EFF-learning function. In addition, for the low dimensional problems, the U-learning function reduces the computational cost significantly, whereas the EFF-learning function has a better performance for high dimensional problems in terms of robustness and efficiency.