Transport AI 2024

Authors: Ferran Torrent, Nuria Toribio, Monica Dominguez, Gavin Bailey, Athina Tympakianaki 

Carles will present Aimsun’s Network Mobility Forecaster (NMF) submodule, explaining the capabilities of supply forecasting when coupled with online learning techniques, and the potential benefits.  

Supply forecasting is usually addressed in the literature as an offline regression task. However, most architectures neglect long term contextual information, spatial dependencies and the impacts of data drift when transposing historical data statically into current and future data scenarios. The application of online learning attempts to address these issues by dynamically adapting to new data. 

The online learning algorithm automatically and periodically supervises whether new mobility patterns have emerged triggering, if needed, the update of the NMF which follows three main steps: (i) addition and deletion of long-term dependencies, (ii) update of the time and spatial layer, and (iii) the update of the output layer. Within the context of the Horizon 2020 TANGENT project, we tested our NMF with 4-5 years of flow data from two cities. Experiments with online versus offline learning in city A improved MASE in 0.12 for all time horizons and in city B the improvement ranged from 0.21 to 0.28. 

The online model was designed to be re-trained with a few weeks of data, hence performance is kept high even in the context of frequent data drift. The application of online learning techniques can deliver significant improvements compared to traditional supply forecasting which in-turn can extend the use of network mobility forecasting to a broader range of transport network types and planning requirements.