Reliability-Based Design Optimization Using Buffered Failure Probability

摘要

Reliability-based design optimization (RBDO) seeks the best design for a structural system under uncertainty. Typically, uncertainty arises from random loads such as wind pressure and random material properties such as yield stress. A reliable design must account for uncertainties to ensure safety. Various methods have been proposed to solve the nonlinear optimization models that RBDO uses. However, these methods are theoretically and computationally troublesome as they involve constraints on failure probability, and failure probability is difficult to handle in optimization algorithms. This thesis considers an alternative approach to RBDO that uses the 'buffered failure probability,' and develops four new solution algorithms based on sample average approximations. Buffered failure probability is more conservative than failure probability and it is much easier to handle in optimization algorithms. We test the algorithms on six engineering-design examples from the literature. The examples range from simple systems with two design variables to complicated ones with ten. Results show that the new algorithms may reduce solution time by an average factor of 560 compared to an existing algorithm. Furthermore, they can handle problem instances with two orders of magnitude larger sample sizes, which may be important for reasons of accuracy.