A private-sector bus operator, responsible for a significant portion of the London bus network.
The London bus network does not run to a predetermined schedule. Instead, the operator is incentivised to provide a regular service, which it aims to do by regulating departure times from each end of the route and holding buses at bus stops along the route.
The current arrangement involves a team of some 160 route controllers who manage services by using GPS technology and radio communication with drivers 24 hours a day. The operator charged Faculty with establishing whether it would be possible to have a computer model use real-time data to advise controllers on the best action to take.
Analysis of historical data yielded many patterns in bus routes that were influenced by factors like time of day, day of the week and public holidays. We built and tested increasingly complex models to analyse these patterns, using a number of different methods, including decision trees and Support Vector Regression (SVR). To predict future journey information, the models incorporated the patterns discovered from historical data, journey information from several preceding buses, and the journey information from the route so far.
Improvements of 38 percentage points above the benchmark were seen for the most complex models.
A graphic showing buses (white and grey lines) for one of the studied routes travelling along their route, from left to right. The highlighted routes (white) are exhibiting ‘bus bunching’.
All the models we developed exceeded the benchmark for performance (i.e. how well-spaced the buses are) used by the human route controllers. The best model, using SVR, delivered a 38% improvement against the benchmark. The operator is now preparing to roll it out on hundreds of bus routes across London, which will result not only in significant savings for them, but also in a more regular and reliable bus service for Londoners.