A Modified Efficient Global Optimization Algorithm for Maximal Reliability in a Probabilistic Constrained Space

Abstract

Design optimization problems under random uncertainties are commonly formulated with constraints in probabilistic forms. This formulation, also referred to as reliability-based design optimization (RBDO), has gained extensive attention in recent years. Most researchers assume that reliability levels are given based on past experiences or other design considerations without exploring the constrained space. Therefore, inappropriate target reliability levels might be assigned, which either result in a null probabilistic feasible space or performance underestimations. In this research, we investigate the maximal reliability within a probabilistic constrained space using modified efficient global optimization (EGO) algorithm. By constructing and improving Kriging models iteratively, EGO can obtain a global optimum of a possibly disconnected feasible space at high reliability levels. An infill sampling criterion (ISC) is proposed to enforce added samples on the constraint boundaries to improve the accuracy of probabilistic constraint evaluations via Monte Carlo simulations. This limit state ISC is combined with the existing ISC to form a heuristic approach that efficiently improves the Kriging models. For optimization problems with expensive functions and disconnected feasible space, such as the maximal reliability problems in RBDO, the efficiency of the proposed approach in finding the optimum is higher than those of existing gradient-based and direct search methods. Several examples are used to demonstrate the proposed methodology.