A Parallel Learning Strategy for Adaptive Kriging-based Reliability Analysis Methods

Adaptive Kriging-based reliability analysis methods have shown great advantages over conventional methods for their computational efficiency and accuracy. However, the widely accepted learning strategies such as Expected Feasibility function and U function can select one training point for each iteration, and therefore are not suitable for parallel processing. To address this limitation, the uncertainty of the failure probability is estimated through Adaptive Kriging with probabilistic classification-based Monte Carlo simulation based on the fact that the total number of failure points follows a Poisson Binomial distribution. By maximally reducing the uncertainty of the estimated failure probability, the theoretically optimal learning strategy is derived in this paper. Due to the computational difficulty in implementing the optimal learning strategy, a pseudo optimal parallel learning strategy is proposed to closely reach the optimal solution. The efficiency of the proposed parallel learning strategy is investigated here by implementing two benchmark reliability problems. Results indicate that the total number of evaluations to the performance function through the proposed parallel learning strategy can be even close to the approach based on single training point enriching.This research has been partly funded by the U.S. National Science Foundation (NSF) through awards CMMI-1462183, 1563372, and 1635569. Any opinions, findings, and conclusions or recommendations expressed in this paper are those of the authors and do not necessarily reflect the views of the National Science Foundation