Short description and objective of the project: The estimation of traffic flows resulting for the traffic assignment of OD matrices can be performed through Graph Neural Networks, a type of machine-learning model, but it requires a suitable training pipeline and a dedicated training dataset. Still, GNNs, as most machine-learning models, usually learn from historical data and may fail if the tested scenarios follow different distributions than the ones seen during training (e.g. when future traffic patterns strongly differ in their distribution from the historical ones). This is a problem for practitioners that may want to use these tools for traffic forecasting in scenarios that include disrupted versions of the network, for example under heavy flooding conditions that modify the structure of the underlying graph. These emergency scenarios are though the ones where the fast inference of machine-learning tools is valuable as long as the traffic predictions are reliable. The main issue is that, being these extreme scenarios, little training data may be available on an everyday basis. Besides, the number of possible disruptions on the graph constitutes a combinatorial problem that scales heavily with the number of links and nodes in the network. For these reasons, it is challenging for practitioners to obtain training data that includes all the traffic flow distributions resulting from demand patterns and network disruptions. This, in turn, makes it difficult to build GNNs that can be used in emergency or extreme scenarios.
The aim of the project is to build a GNN architecture that will prove robust to disrupted network configurations (e.g. multiple closures due to heavy flooding). If needed, synthetic data obtained through fast macroscopic simulations may be used to build a training dataset including multiple recordings under disruption. Besides, other methodological approaches may be tested to increase the robustness of the surrogate (e.g. meta-learning).
