Multi-fidelity optimization strategies using genetic algorithms and sequential kriging surrogates

Nithin J Kolencherry, Purdue University

Abstract

Engineers have used numerical methods for optimizing simulations representing real world problems. Many such engineering problems, especially those in the aerospace field, use high fidelity non-linear simulations for analysis that are associated with long run times. Since many real world simulations have inherent discontinuities or deal with discrete variables, they cannot be optimized using gradient-based methods. Such problems are well handled by population-based optimization methods such as genetic algorithms. However, these methods are associated with a very large number of function evaluations. This thesis approaches this issue by presenting two multi-fidelity optimization strategies using genetic algorithms that considers the simulation as a high-fidelity function evaluation and uses a sequentially updated Kriging surrogate for a low-fidelity function evaluation. In order for these strategies to build a global surrogate model and find the global minimum, good design space coverage is required, which is obtained by means of space-filling sampling strategies. The described strategies are tested on two analytical test functions and on two benchmark engineering test problems and its performance is compared to that of a binary coded genetic algorithm. The strategies are then used to optimize an aircraft design for minimum fuel consumption on a certain medium range mission using FLOPS. The experimental results presented showcase the ability of the two strategies at efficiently locating the global minimum for a number of optimization problems. The limitations associated with the two methods and potential future improvements are also discussed.

Degree

M.S.A.A.

Advisors

Crossley, Purdue University.

Subject Area

Aerospace engineering|Operations research

COinS