About XRCE

7th December, 2015

WiML 2015
: talks by established researchers and graduate students, a poster session for graduate students, and a panel discussion to discuss careers in machine learning. The Xerox Research Centre Europe supports the workshop that is co-located with Neural Information Processing Systems (NIPS) 2015

Sofia Michel: Stochastic optimization of public transport schedules

Transfer waiting times are a big weakness of public transport systems. For example, on average in the UK, 23% of travel time is lost in connections for trips involving more than one mode. Studies have shown that out-of-vehicle times are more annoying for passengers than in-vehicle times, and transfer waiting times are even more irritating since they cannot be managed by passengers. Over the last two decades, numerous studies have focused on schedule synchronization, i.e. the design of schedules that minimize the delay caused by transfers. An important limitation of the standard operations research approaches is that the parameters of the problem are supposed to be known. However, the number of passengers making a given connection as well as the actual arrival and departure times of the vehicles at the stops are by nature non-deterministic. Only recently, some approaches that take into account the uncertainty of these parameters have been proposed.  

This work takes a data-driven approach to the schedule synchronization of a multi-modal transportation system. Using a real-world collection of transit instances in a public transport system, we propose a new model to alter existing schedules in order to minimize the waiting times of passengers during transfers. Our approach is a generalization of previous data-driven works: it is closer to the real use of the network because it is based on real passengers trips instead of queries to a journey planner; and it is more general and realistic
because we get rid of certain restrictive assumptions. We model the problem as a two-stage stochastic linear program with mixed-integer variables. We provide preliminary numerical results using transit data of the city of Nancy in France and show that our method reduces significantly the expected transfer waiting times.