Learning Graph Configuration Spaces in Engineering Domains

Abstract

Background The configuration process for feature-oriented product lines is well-studied for Boolean and numerical feature configurations. However, many engineering domains face the challenge of optimising graph structures to represent product configurations effectively.

Research Objective Optimising complex graph structures to meet multiple objectives under various constraints demands a thorough understanding of the graph configuration space and the properties linked to each configured product. This thesis investigates two machine learning-based strategies to facilitate efficient exploration of graph configurations.

Method Using two engineering case studies (design of HVAC systems and traffic networks), this thesis first evaluates the applicability of established methods from feature-oriented software product line engineering to graph configuration problems. After initial promising results in learning graph configuration spaces, experiments were conducted with metaheuristics ---including local search, simulated annealing, and genetic algorithms--- to explore and optimise within these spaces.

Results The results demonstrate that graph embedding-based approaches lead to data-efficient learning. Additionally, learning-supported metaheuristics enable efficient exploration of graph configuration spaces and optimisation within those spaces, significantly reducing the need for costly evaluations of every graph configuration.

Conclusion This work demonstrates that, with appropriate adaptations, strategies from software product line engineering can effectively support metaheuristic graph configuration space exploration. Future work is necessary to (i) fully leverage the potential of learning graph configuration spaces by improving sampling, learning, and evaluation strategies for enhanced training efficiency, prediction accuracy, and interpretability, and (ii) extend these methods to tasks such as dynamic configuration, constraint mining, and evolution.