Multi-Fidelity Optimization Strategies using Genetic Algorithms and Sequential Kriging Surrogates

摘要

Many engineering problems use high-fidelity simulations for analysis; these simulations may have long runtimes and may have inherent discontinuities. Engineering design problems may also involve discrete variables. Population-based methods, such as the genetic algorithm, can perform design optimization for these kinds of problems. In spite of its advantages in handling discontinuities in the functions and/or variables, a notable disadvantage of using such optimization methods is the large number of function evaluations (cost) associated with them. This study approaches the issue by formulating and implementing two multi-fidelity optimization strategies that use a genetic algorithm framework in which sequentially-updated Kriging surrogates provide low-fidelity function evaluations to reduce the computation time for the optimal search. Space-filling sampling methods help to obtain good design space coverage for the initial Kriging models. The two multi-fidelity strategies are compared to the binary-coded genetic algorithm on a number of analytic test functions and engineering problems, including an aircraft-sizing problem. Results showcase how the two strategies perform as well as the binary-coded genetic algorithm in finding solutions at a fraction of the computational cost.