Efficient Optimization of Engineering Problems using Multi-Fidelity Models

Resumo

Optimization methods can be employed to find the optimum design of engineering structures. Due
to their ease of implementation and robustness, bio-inspired optimization algorithms have been widely applied
to solve complex optimization problems. However, these methods require a large number of expensive function
evaluations. For a more efficient process, surrogate models can be used to provide a cheaper estimate of the
structural responses. These models are built from a small set of true responses, and their approximated surface
assists in the selection of promising trial designs. Efficient Optimization can be performed by iterately improving
the model by the addition of new points in regions of interest, thus improving the accuracy of the model near
the optimum location. In this work, we study the use of Multi-Fidelity models for the Efficient Optimization of
engineering problems. Kriging and Hierarchical Kriging models are employed, and the selection of new points is
performed using variations of the Expected Improvement and Probability of Improvement criteria. The obtained
results are compared in terms of accuracy, the number of evaluations, and computational efficiency. Results show
that Multi-Fidelity approaches are able to find optimal results using fewer high-fidelity evaluations.