Graph-Based System Architecture Optimisation
Study of Graph-Based Surrogate Modelling Approaches for Black-Box Optimization of Aerospace System Architectures
F.J. Latała (TU Delft - Electrical Engineering, Mathematics and Computer Science)
E. Isufi – Mentor (TU Delft - Electrical Engineering, Mathematics and Computer Science)
Jasper Bussemaker – Mentor
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
System Architecture Optimization (SAO) problems are typically mixed-discrete,
hierarchical, and expensive to evaluate, motivating the application of Bayesian
optimization, which is known for its sample efficiency. Current approaches for
SAO use Gaussian processes (GPs) operating on encoded design vectors. These
encodings can obscure similarities in component connectivity, making it harder
for the GP to capture performance correlations between evaluated and uneval-
uated architectures and potentially reducing optimization sample efficiency. To
address this, we propose a GP surrogate that incorporates similarities between
architecture graphs through the Weisfeiler–Lehman optimal assignment kernel
to better capture shared structural patterns. We also introduce extensions that
facilitate adapting the neighborhood scale at which architecture graphs are com-
pared, as well as the specificity with which structural patterns are distinguished.
This allows learning those properties during GP fitting rather than tuning them
beforehand, which is infeasible for SAO, as new problems typically lack prior
simulation data. Across four benchmarks against encoding-based GPs, an alter-
native graph-kernel, graph neural networks, and model-free baselines, our method
achieves the best optimization performance on connectivity-driven problems. It
reduces regret by up to 40% and final normalized hypervolume error by up to
50% relative to the strongest encoding-based baseline. This comes at a higher
computational cost, but the additional overhead is offset once architecture evalu-
ations take approximately 4–25 seconds. On the remaining problems, our method
remains competitive with state-of-the-art encoding-based approaches. These find-
ings suggest that graph-based approaches could help engineers find better system
architectures faster when simulations are expensive.