Reusing optimization histories to accelerate engineering design optimization

Cimpoesu, Petru-Cristian, Toal, David, Wang, Leran, Keane, Andy, Gregory, Jonathan and Nunez, Marco (2025) Reusing optimization histories to accelerate engineering design optimization. In AIAA SCITECH 2025 Forum. AIAA International.. (doi:10.2514/6.2025-0775).

Abstract

A well-known limitation of surrogate-based optimization concerns the sparse sample available to train the initial model. Given an unlucky training set, the convergence of the process can slow down or be prevented altogether. This limitation is commonly addressed by adding an element of exploration into the infill strategy. However, the full information gathered by balanced infill strategies is rarely reused in future optimization tasks, as most samples are discarded due to infeasibility or sub-optimal performance. In the context of limited computational budgets, such samples can be invaluable. Their effective use may help cover unexplored regions of the design space and thus significantly reduce the sparsity of the initial sample at no additional cost. The current paper presents a method by which historical data can be integrated into new design tasks to improve performance in the early stages of design optimization. The method covers sequential transfer problems with constant parameterization, objectives and constraints. It exploits correlations between data source outputs and, where possible, attempts to improve such correlations through input calibration. Results indicate the method is highly effective if the data sources are well correlated or their relationship can be represented by linear mapping.

This record has no associated files available for download.

Contributors

Author: David Toal

Author: Andy Keane

Download statistics