Optimization Strategy for a High Dimensional and Heavily Constrained Expensive Black-box Problem

This presentation was made at the 2019 NAFEMS World Congress in Quebec Canada

Resource Abstract

Design optimization involving expansive engineering simulations is time-consuming. Furthermore, there are heavily constrained black-box problems with over 100 design variables which demand many optimization iterations and a large turnaround time. With the technology advances of High Performance Computing (HPC), the total turnaround time of optimization could be greatly reduced if a large number of design evaluations could run in parallel. When runs are in parallel, the overall turnaround time instead of the total number of design evaluations becomes a better benchmark criterion for comparing optimization algorithms. This paper introduces a trust region based surrogate-assisted Genetic Algorithm (GA). At each iteration, the trust region is first determined based on previous iterations, and then the GA exploits the trust region with a relatively large number of designs in parallel. The trust region adaptively expands, shrinks, or remains depending on the solution quality of previous iterations. The optimization strategy is tested on a well-known benchmark problem with 124 design variables, 68 constraints, and one objective. The result shows that steadily improving solutions can be achieved every few iterations, and the entire optimization can be completed within a practical and reasonable time frame.