Enhancing generative graph-based optimisation: evaluating the impact of simulated annealing on genetic algorithm efficiency

Al-Harasis, Dania ORCID: https://orcid.org/0000-0001-9102-8882 and Jabi, Wassim ORCID: https://orcid.org/0000-0002-2594-9568 Enhancing generative graph-based optimisation: evaluating the impact of simulated annealing on genetic algorithm efficiency. Presented at: eCAADe Conference 2026: Informed creativity and fabrication in architecture and engineering, Lübeck, Germany, 7-11 September 2026. eCAADe 2026. Item availability restricted.

	PDF - Draft Version Restricted to Repository staff only Download (10MB)
	PDF (Provisional file) - Accepted Post-Print Version Download (17kB)

Abstract

This paper evaluates a hybrid optimisation pipeline that integrates Simulated Annealing (SA) into a Genetic Algorithm (GA) for the generative design of graph-based spatial layouts. Graphs encode spaces as nodes and adjacencies as edges, with semantic attributes attached to nodes. Standard GAs are effective global searchers but prone to premature convergence in high-dimensional, constrained design problems. To address this, a stagnation-aware hybrid GA-SA framework is developed in which SA is triggered when GA progress stalls and is applied as a final refinement stage. Both pipelines share the same graph encoding, composite similarity fitness function, and a stratified export mechanism that produces datasets for graph machine learning. Performance is evaluated against a baseline GA over 30 independent runs on a benchmark autism-friendly educational layout comprising 55 spaces and 64 adjacencies. The hybrid GA-SA significantly improves optimisation quality without additional runtime cost, achieving a mean final fitness of 0.751 (best 0.826) compared to 0.472 (best 0.490) for the baseline, with markedly higher constraint satisfaction. The study contributes: (1) a graph-aware hybrid GA-SA pipeline for spatial layout optimisation; (2) empirical evidence of improved convergence and constraint compliance; and (3) a stratified dataset-generation strategy that supports downstream machine learning in architectural design.

Item Type:	Conference or Workshop Item - published (Paper)
Status:	In Press
Schools:	Schools > Architecture
Uncontrolled Keywords:	Genetic algorithms, simulated annealing, generative tool, optimisation, graph-based generation
Date of First Compliant Deposit:	19 May 2026
Date of Acceptance:	15 May 2026
Last Modified:	19 May 2026 09:39
URI:	https://orca.cardiff.ac.uk/id/eprint/187075

Actions (repository staff only)

Edit Item