Reliability-Based Design Optimization Using Quantile Surrogates by Adaptive Gaussian Process

Abstract

It is of great significance to incorporate various uncertainties into the design optimization of structures and other engineering systems. Many reliability-based design optimization (RBDO) methods have been developed, but their practical applications can be limited if the reliability consideration entails a large number of evaluations of performance functions, especially for those requiring time-consuming simulations. To overcome the challenge, this paper proposes a new RBDO method that employs quantile surrogates of the performance functions to identify the admissible domain, termed the probability-feasible design domain. Gaussian process models of the quantile surrogates are updated adaptively through an exploration-exploitation trade-off based on inherent randomness and the model uncertainty of the surrogate. The method guides the computational simulations toward the domain in which the quantile estimation can make the greatest contribution to the optimization process. The validity and efficiency of the proposed RBDO method using quantile surrogates by adaptive Gaussian process (QS-AGP) are demonstrated using several numerical examples. The results confirm that QS-AGP facilitates convergence to a reliable optimum design with a significantly reduced number of function evaluations compared to existing RBDO approaches.