Ensemble Techniques for Determining Globally Optimal Designs for Problems with Broadly Stated Design Objectives

Abstract

Abstract : In this project we have investigated the use of global optimization techniques for device design problems in photonics and electronics. Using several concrete device design problems as guides, we have defined a sub-class of global optimization problems with high-dimensional design space as the main focus of our effort. We have evaluated optimization approaches including Genetic Algorithm and Simulated Annealing for this class of problems. This evaluation confirmed our previous experience in using these techniques in global optimal design problem. In fact these methods do not effectively use all available local information during the search for the global optimal solution and can easily be trapped near locally optimal solutions. We have subsequently developed an Ensemble Global Optimization (EnGO) technique that explicitly defines region of exclusions around already examined points in the design space to allow maximal coverage of the design space through randomized search. At the same time local gradient information is used to accelerate convergence to local and global optimal solution. The method is being tested successfully on prototype problems in relatively high dimensional design space.