New dynamic events-based public transport router for agent-based simulations

Public transport route choice is an open problem studied for many years and it is well regarded that many factors influence the choice of stages making up a public transport trip. In addition to factors such travel time, fare, service frequency, number of transfers and transfer waiting time, the focus of research shifted more and more to dynamic factors such as crowdedness and travel time variability. This paper describes a method to simulate public transport route choice for a whole population in a city or even a country taking into account all factors mentioned above using the MATSim framework, a multi-agent and activity-based transport simulator. A new public transport router is proposed using a time dependent shortest path algorithm in a graph with links representing in-vehicle travel, waiting and walking (from activity locations to public transport facilities and between public transport facilities). The cost of the path can be specified to depend on the fares, travel times, occupancy levels, waiting times and number of transfers. In addition, the agent-based nature of the simulation model allows to define agent-specific parameters for each attribute to account for preference heterogeneity within the agent population. Simulated dynamic effects such as bus bunching and overcrowded services lead to additional waiting and travel times, compared with the public transport schedule. Within the MATSim evolutionary algorithm, this information is saved after each simulation of the same day, and used to re-route agents. The router is implemented and tested for Singapore, whose modeled public transport system features bus and subway. As in most major cities, the model shows that the stability and schedule reliability of bus services can at times only hardly be maintained, while crowdedness at the peak hours is an issue for both bus and mass rapid transit and light rapid transit. We illustrate how the new approach allows agents to adapt to these dynamic complexities. The new approach, along with the original scheduled-based router of MATSim are compared for computation time, as well as for agent learning rate, i.e. the time needed to reach a particular solution state. With a better capacity to model observed public transport route choice, more than 10% of improvement in computation time per simulation iteration, and needing much smaller number of iterations to reach user equilibrium, the proposed alternative results a Pareto-dominant improvement