Learning Graph Configuration Spaces to Support Road Network Design Optimisation
Michael Mittermaier, Takfarinas Saber, Goetz Botterweck · Proceedings of the Genetic and Evolutionary Computation Conference · 2025
Genetic algorithms (GA) allow us to optimise graphs according to multiple objectives while considering many different constraints. These population-based algorithms assess the fitness of a high number of genomes. In the case of optimising road networks, a high number of fitness assessments leads to high computational costs of traffic simulations. In this work, we explore the application of learning graph configuration spaces to make efficient use of these traffic simulations by using learning model predictions for the majority of fitness assessments. In a controlled experiment, we compare the quality of GA optimisations with and without learning model predictions on the same simulation budget. Our results indicate that although we lose accuracy in the fitness assessments with predictions, the GA reliably finds road networks with better traffic flow and lower overall road length while using the same number of traffic simulations. We show that learning models can support GAs to make efficient use of the simulation budget and thus improve the optimisation. Future work is necessary to confirm these results for larger road networks.