Complex transport systems such as the London Underground, used by millions of people per day, must be prepared to respond fast and effectively to unpredictable unplanned disruptions that cause stations or parts of a line to close down. This project concerns the development of models for passenger movement and reaction to closures so that we can learn how demand and usage fluctuate over different spatial regions of a transport system and different times of the day. This will allow for a better understanding of demand that has not been met when passengers decide to stop their journey earlier, and how delays can be propagated by crowds changing their route. This has direct implications on the safety of passengers by providing tools to improve crowd management within a system like the Underground and by better compensating predicted excess demand with external vehicles such as buses.
The methodological basis is probabilistic machine learning, where we model the behaviour of a fast rail transport system in terms of a network tomography problem: we observe part of the passengers’ behaviour and infer unobserved quantities of interest. In particular, latent passenger trajectories and train availability are quantities which can be inferred indirectly and partially validated with data graciously provided by Transport for London. These are hard computational problems, and the methodology will rely in particular on stochastic variational inference. One of the most crucial aspects of the research will advance the state of the art on how to leverage historical data from past disruptions to understand what happens under unseen situations. This will combine background knowledge on causal effects of disruptions on passengers extracted from the tomography of the system, along with modern predictive modelling.
Part of The Alan Turing Institute-Lloyd’s Register Foundation Programme for Data-Centric Engineering.