Learning Graph Configuration Spaces with Graph Embedding in Engineering Domains

In various domains, engineers face the challenge of optimising system configurations while considering numerous constraints. A common goal is not to identify the best configuration as fast as possible, but rather to find a useful set of very good configurations in a given time for further elaboration by human engineers. Existing techniques for exploring large configuration spaces work well on Euclidean configuration spaces (e.g., with Boolean and numerical configuration decisions). However, it is unclear to what extent they are applicable to configuration problems where solutions are represented as graphs – a common representation in many engineering disciplines. To investigate this problem, we propose an adaptation of existing techniques for Euclidean configurations, to graph configuration spaces by applying graph embedding. We demonstrate the feasibility of this adapted pipeline and conduct a controlled experiment to estimate its efficiency. We apply our approach to a sample case of HVAC (Heating, Ventilation, and Air-Conditioning) systems in 40,000 simulated houses. By first learning the configuration space from a small number of simulations, we can identify 75% of the best configurations within 7,508 simulations compared to 29,725 simulations without our approach. That is a speed-up of 4.0 \(\times \) and saves more than 15 days if one simulation takes about one minute, as in our experimental set-up.