The “ Multi-Modal ” Contribution

MATSim’s standard mobsim, QSim, has recently been enabled to model multimodal scenarios as shown in Section 4.6. In this chapter,1 an earlier approach to handle multimodal scenarios, the multimodal link contribution, is presented. As shown below, it is a very e cient approach, that considers persons’ biking and walking speeds to improve the teleportation estimates for these modes, whereas mode interactions are not taken into account.


Introduction
MATSim's standard mobsim, QSim, has recently been enabled to model multimodal scenarios as shown in Section 4.6.
In this chapter, 1 an earlier approach to handle multimodal scenarios, the multimodal link contribution, is presented. As shown below, it is a very e cient approach, that considers persons' biking and walking speeds to improve the teleportation estimates for these modes, whereas mode interactions are not taken into account.

Multi-modal Link Contribution
Figure 21.1 shows the implementation's basic concept-a multimodal contribution is added to each link object in the mobsim.
While tra c ow dynamics are simulated by MATSim's mobsim using a queue model, these ows are not taken into account in the multimodal contribution. Examining typical pedestrian and cyclist tra c ows shows that congestion is very rare compared to vehicular tra c, justifying application of this simplistic approach over a scenario. For regions with higher tra c ows, this simple model loses accuracy, but still outperforms the teleportation approach, which MATSim uses by default.
Each multimodal link contribution uses a priority queue to manage all agents traveling on that link using a non-motorized mode. The queue orders the agents based on their scheduled link leave time (see Figure 21.2). This time is calculated when an agent enters a link and is based on parameters like the agent's age and gender, as well as the links' steepness. In each time step, it is checked whether the queue contains agents who have reached their link leave time and thus must be moved to their route's next link. An agent's position on a link is not determined by the model. However, under the assumption that agents move with constant speed, their position can be interpolated. This approach is computationally very e cient, because computation e ort is created only when an agent enters or leaves a link but not when it is traveling along a link. Additionally, agents can travel at di erent speeds, so can overtake each other.   Figure 21.3(a)). Finally, to calculate the person's travel time on a speci c link, in uence of the link's steepness on the person's speed is taken into account (see Figure 21.3(b)). The combination of person-speci c attributes and link steepness is shown in Figure 21.3(c). As a result, a person's speed on plain terrain is calculated as:

Travel Times
A link's steepness is incorporated as: v person walks on link = v person, walk · f steepness (21. 3) The speed of cyclists is determined using results from Parkin and Rotheram (2010). Starting point is, again, an individual's speed based on a normal distributed (N (6.01, 1.17)) reference speed. Once more, a person's speed is calculated by accounting for age and gender (see Figure 21.4(a)).
When calculating the steepness factor, one must de ne whether a link goes uphill or downhill. When going uphill, the person's speed is reduced by a factor based on the grade and a reference factor of 0.4002 meters per second, which is scaled by the same factor as the person's reference speed. i.e., the speed drop of slow people is lower than the drop of fast people. When bike speed drops below walk speed, which happens at a grade of approximately 12 %, it is assumed that the person switches to walking (see Equation (    Another parameter a ecting pedestrian and cyclist speed is the crowd density of the link where they are physically present. Data to take this e ect into account is, again, presented by Weidmann (1992). However, to calculate crowd density of a link, its geometry has to be taken into account, as discussed by Lämmel (2011).

Conclusions and Future Work
The multimodal contribution allows the tracking an agent's movement in detail, essential for studies related to topics like evacuations, e-bikes, car sharing or public transport. Experiments testing the implementation and demonstrating its capabilities are described by Dobler (2013). An application's required level of detail strongly in uences the modeling approach selection. A simple model including agents' age and gender, but not incorporating agent-agent interactions, might be detailed enough for some studies (e.g., e-bikes or public transport). However, for other studies, a more detailed model, also simulating agent interactions, might be necessary.
A rst implementation of a pedestrian simulation module for MATSim, which also supports agent-agent interactions, was presented by Lämmel and Plaue (2014) introducing a force-base model. The agents' high-level planning (i.e., route and destination choice) was performed on a graph representing the transport system (e.g., a MATSim network), while the low level behavior (i.e., physical interaction between the participants) was simulated with a force-based model. Due to the intense computational e ort of the underlying physical model, the scenario size was limited to a few thousand agents. An attempt to bypass this limitation was presented by Dobler and Lämmel (2012). They combined the force-based pedestrian simulation module with the multimodal link contribution, creating the opportunity to simulate large-scale scenarios, by staying highly resolved where needed and being more aggregated where possible.