In Dublin, like many cities globally, public transportation is promoted as the preferred mode of travel to address growing traffic and congestion problems. The bus network is the work horse of the transportation system, facilitating mass movement of people across the city.
To facilitate this, many bus service providers operate routes with a high frequency to ensure high mobility within the city area. One problem that comes along with highly frequent bus routes is called bus bunching, also known as platooning, which refers to the phenomenon when two buses of the same route arrive at the same time at a bus stop. See http://setosa.io/bus/ for visual explanation of the phenomenon.
In 2013, I was part of an IBM Research team that partnered with the city of Miami to tackle this problem using predictive analytics. The system we piloted focused on four key routes along South Beach and aimed at providing real-time predictive alerts on when bunching was likely to occur, giving controllers ways to manually intervene to slow down or speed up buses along the corridor. Over the course of the pilot, we generated several thousand alerts – far too many for controllers to deal with manually. We wanted to take this solution further to automatically propose corrective actions to drivers, i.e. eliminate the need for human intervention.
Last year, Matthias Andres joined us as an intern from the Department of Mathematics at the University of Kaiserslautern, Germany. With his background in applied mathematics and control theory, this seemed like a perfect opportunity to blend specialized intern-expertise and a real-world problem. Over the course of his visit, he developed an automatic control mechanism to address bunching with predictive analytics. The work formed the basis of his thesis.
The main algorithmic contributions are also reported in a recent paper that Matthias and I, as his mentor, published entitled “A predictive-control framework to address bunching.” In this paper, we focus on two main areas of research based on the data-driven prediction of bunching events and on corrective control strategies, creating a predictive-control framework based on real data of a busy bus route in Dublin.
First, the challenging problem of short term prediction is addressed by considering headways, which measure how far apart two buses are, either in time or distance. The method works by interpreting headways measured continuously in time as time series and building a model to predict these directly. Typical approaches rely on predicting arrival times at stops and then looking to analyze the different predictions to estimate future headway. We avoid this intermediate step and forecast headways directly.
In the second step, Matthias investigated how one type of control strategy, called holding, can use this future information. Holding refers to the alteration of buses’ dwell times at stops after the boarding process is finished. The goal of the holding strategy is not to adhere to a schedule, but to regularize the headway distribution and with that to improve the service provider’s performance and the customers’ experience. The paper shows that such an approach is stable and has significant improvements in both operator and customer metrics.
The data we used in this paper is freely available at Dublin’s open data portal Dublinked. We used the two datasets, one on the real-time position information of the bus fleet, and the second set that provides the schedule and infrastructure data of the transit system, such as stops and route configurations. This data can be also be used for better passenger information system, route performance management, and optimizing schedules.
While we demonstrated the control mechanisms in simulated environments, in a real system the buses will need appropriate onboard devices to guide drivers on how to best manage their vehicle, taking into account the context of how the entire system is performing. Instead of regulating headways, as we have done, one could consider controlled segment travel times between stops. In this case the driver could, based on their knowledge and experience of the route, decide if the current traffic situation allows the bus to reach the suggested (optimal) segment travel time by waiting at a stop or if they should flow with the traffic, what might also increase the segment travel time in case of congestion.
After his internship with IBM in Dublin, Matthias returned to his university TU Kaiserslautern and finished his Master’s program. he was one of seven young researchers honored in June 2017 for his outstanding thesis and received an award from the Kreissparkassen-Stiftung für die Technische Universität Kaiserslautern, a foundation which promotes and supports innovations in science and research which serve the domestic industry. Matthias is now pursuing his Ph.D. in the area of mathematical modeling and optimization of medical applications.
The internship program is one way IBM Research engages with the next wave of researchers. Through the program, I enjoyed the opportunity to mentor Matthias who brought tremendous energy and ideas to the problem. In true collaborative fashion, it was a great learning experience all around.