Efficient Global Optimization Using Hybrid Genetic Algorithms

Abstract

The optimization of many realistic large-scale engineering systems can be computationally expensive. The evaluation of a single design configuration can take minutes or hours, and although computing power is steadily increasing, the complexity of the analysis codes continues to keep pace. In this paper a novel hybrid optimization method is introduced to efficiently find the global optimal of complex, highly multimodal systems. The motivation lies in the fact that to optimize many realistic engineering systems often requires numerous computationally expensive analyses to be performed. Heuristic optimization algorithms such as Simulated Annealing or Genetic Algorithms often can locate near optimal solutions but can require many function evaluations. Local search algorithms, including both gradient and non-gradient based methods, are quite efficient at finding the optimal within convex areas of the design space but often fail to find the global optimal in multimodal design spaces. The hybrid optimization approach presented in this work switches between global and local search methods based on the local topography of the design space. The global and local optimizers work in concert to efficiently locate quality design points better that either could alone. To demonstrate the usefulness of the approach presented in this paper, two case studies of differing complexity are considered.