Dynamic bus operations optimization with REFLEX

High frequency bus operations in metropolitan areas should provide a reliable service to passengers by reducing their EWTs at bus stations. In several metropolis such as London and Singapore bus operators receive monetary incentives if they manage to reduce the EWTs of passengers or penalties if they fail to do so. However, optimizing the regularity of bus operations by preventing bus bunching is a computationally intractable problem and bus operators are not able to schedule the daily bus trips in an optimal way. Therefore, they rely on in-house expertise to manage their operations without fully exploiting the potential of applying operational control measures such as dispatching and bus holding at stations. For this reason, our work models the regularity-based bus operations and introduces REFLEX, an AI agent which enables the implementation of bus control actions. REFLEX uses a heuristic Sequential Exterior Point Greedy method for optimizing bus service operations and is tested in a trial with a major bus operator in Asia. REFLEX was able to optimize bus services with 200+ daily trips in just 1-2 minutes of computational time, while providing 17-35% theoretical service regularity improvement subject to a set of strict operational constraints such as adherence to layover times and departure frequencies ranges. Thanks to REFLEX rapid computation, bus operators can also simulate further service regularity improvements resulting from relaxing some operational constraints or adding more trips. Looking further, REFLEX can be combined with e-paper based timetable displays to provide a fully dynamic operational schedule and will be extended to support multi-modal interchange optimization.