Session: DAC-15-02-Multidisciplinary Design Optimization, Multiobjective Optimization, and Sensitivity Analysis

Paper Number: 91112

91112 - Bi-Objective Surrogate Feasibility Robust Design Optimization Utilizing Expected Non-Dominated Improvement With Relaxation 

Engineering design optimization problems often have multiple competing objectives as well as uncertainty. For these problems, quite often there is interest in obtaining multiobjective feasibly robust optimum solutions. Feasibly robust, here refers to solutions that are feasible under all uncertain conditions. In general, obtaining multiobjective feasibly robust solutions can be computationally expensive. Although surrogates have begun to be utilized to decrease the computational costs of multiobjective design problems, there is limited usage of Bayesian frameworks on problems of multiobjective optimization under interval uncertainty. This article seeks to formulate an approach for the solution of these problems via the expected improvement Bayesian acquisition function. The acquisition function is solved alternatingly with worst-case searches to find robust solutions. A method of iteratively relaxing the solutions to facilitate convergence to a set of non-dominated, robust optimal solutions. Additionally, a variation of the multiobjective expected improvement criterion is proposed to encourage variety and density of the robust Pareto optimal solutions. Several examples are performed and compared against another multiobjective robust optimization approach with surrogate utilization. It is shown that the proposed method performs well at finding robustly optimized feasible solutions with limited function evaluations. Finally, future directions for improving the proposed methodology are suggested.

Presenting Author: Randall Kania University of Maryland

Presenting Author Biography: Randall Kania attended Vanderbilt University for his undergraduate education, graduating in 2013. Afterward, he received his Master's Degree in Engineering from Carnegie Mellon University in 2015. Currently, he is a PhD Candidate at the University of Maryland, College Park.

Authors:

Randall Kania University of Maryland
Shapour Azarm University of Maryland