Combining an Asynchronous Multi-Point and a Multi-Fidelity Infill Strategy for Surrogate-Based Optimisation
An Empirical Evaluation on Unconstrained Problems
I.L. Oostmeijer (TU Delft - Aerospace Engineering)
M.F.M. Hoogreef – Mentor (TU Delft - Aerospace Engineering)
Matthijs Langelaar – Mentor (TU Delft - Mechanical Engineering)
M. Slebioda – Mentor (TU Delft - Mechanical Engineering)
G. la Rocca – Graduation committee member (TU Delft - Aerospace Engineering)
O. Nejadseyfi – Graduation committee member (TU Delft - Mechanical Engineering)
More Info
expand_more
Other than for strictly personal use, it is not permitted to download, forward or distribute the text or part of it, without the consent of the author(s) and/or copyright holder(s), unless the work is under an open content license such as Creative Commons.
Abstract
Engineering design frequently involves optimisation problems incorporating computationally expensive analysis tools. Surrogate-based optimisation has been shown to reduce the runtime of these optimisations. Multi-point (MP) and multi-fidelity (MF) infill strategies can further accelerate convergence, yet their combination has not been studied extensively.
This thesis introduces and evaluates an asynchronous MP MF infill strategy, benchmarked against eleven unconstrained MF numerical problems, using expected runtime (ERT).
The asynchronous MP single-fidelity strategy (16 ranks) reduced the geometric mean ERT by 72.7% compared with the baseline Efficient Global Optimisation with Expected Improvement. The single-point MF strategy, on the other hand, increased the ERT by 21.2%, degrading performance. The combined asynchronous MP MF strategy achieved a 68.2% reduction relative to the baseline.
These results show that the asynchronous MP strategy substantially improves performance. In contrast, the selected MF strategy proves detrimental, indicating that a revised MF strategy is required to yield further gains.