Simulating Active Travel Modes
Extending our massive agent-based transport models to achieve more detailed walking and cycling behaviour.
At Arup’s City Modelling Lab we build agent-based transport models of cities, regions and countries. The models aim to represent different people and their unique transport needs. One of the advantages of an agent-based approach, is that we can capture how changes impact individuals, in addition to the effects on the system as a whole. Some of the changes we consider are changes in transport infrastructure, public transit provision, and, with increasing appeal, interventions aspiring to encourage active mode travel.
Schemes encouraging active travel usually focus on improving cycling infrastructure, access to bicycles and promoting behavioural changes that encourage people to make shorter trips which they could more easily complete by walking or cycling (in 2021 the NTS reported that 17% of trips under a mile are completed by car or van — a distance of a 20 minute walk). Lately we have been working on being able to model the first two with MATSim.
MATSim has two extensions of interest in this space. The multimodal extension provides support for routing walking and cycling on the network with personalised travel times based on an agent’s age and gender, as well the slope of the link they are traversing. A newer extension, bicycle adds more functionality and uses additional data to more accurately model an agent’s cycling experience. As well as elevation, you can add different surface types or self-computed infrastructure factor to your network — all of which will affect the speed of the cyclist.
Improving active mode travel leads to impacts beyond a simple mode shift for trips of short distances. Cycling can increase the accessibility to public transport — with safe and convenient connections to bus and train it is possible to mode shift the far more difficult medium and long distance driving trips, with big emission savings.
Transport Network Representation
The first step to simulating an agent engaging in an active mode is providing them with the right environment. We build our networks using Open Street Map data and use tags reserved for walking and cycling where possible. The coverage of these tags is not perfect, so we make some assumptions like allowing people to walk and cycle on minor roads.
We currently use entirely separate graphs to represent the different modal networks, which means cyclists and cars cannot yet interact inside the simulation. We’d like to improve on this in the future so that we will be able to see increasingly realistic behaviour from our agents. One day, we would like to see cyclists avoiding roads with a lot of traffic (something the bicycle extension is close to achieving) and, having cars be slowed down by heavy bicycle traffic.
Aside from a well-maintained and segregated cycling lane, a significant contributor to a cyclist’s choice of route is slope. Unless you’re travelling with an e-bike, getting up that hill may be a deal-breaker. This is why we include elevation in our networks — if you want to add elevation data to your MATSim network, take a look at this script in GeNet (you will need a digital elevation file, which you can generate simply with QGIS and STRM data from here).
Capturing Agent Speeds
Active modes are called active for a reason. You need to use your body to propel yourself in the direction you want to go. In order to accurately model the real World, it’s important to represent people with varying degrees of ability to perform these actions. One approach to modelling this heterogeneity could be utility-based, by setting different perceived costs of an active mode based on agent attributes, those agents will evaluate their choices differently. The effects of demographics seem to differ based on location, with cycling in the Netherlands being least affected by those factors — our colleagues Theo and Paola are working hard on researching the topics surrounding transport equity and will be sharing their findings soon.
For now we rely on multimodal extension’s implementation of personalised speed factors, the authors of which rely on Parkin, J & Rotheram, J. (2010) for cycling and Weidmann, Ulrich (1992) for walking.
Through our endeavours with cycling and walking speeds, we contributed changes to MATSim’s core functionality. The fast Hermes MobSim (Mobility Simulator) module we rely on can now represent slower speeds, like those of active modes, much more accurately.
Bike availability
When we synthesise our agent population and their activity plans, we assign them attributes such as whether they own a car or bicycle. Agents will perceive cycling as more expensive if they don’t own a bicycle and instead need to spend some money on ‘renting’ one. For now we’re keeping it simple — we don’t model where they can rent from and whether there are any bicycles left at that location. A simplification for sure, but a model like this can highlight hotspots for demand for shared micro-mobility hubs without impeding agents from using the mode.
Conclusions
In this post we summarised our recent efforts in adding more detail to how active transport modes are simulated in our agent-based models. We have improved our network representation for walking and cycling infrastructure, which will enable us to model changes to it much more reliably. We have also put our agents on this network and made them use it in a way that is a little bit more personal to them. The agents have to make a choice whether they want to use one of the active modes for their travel, and now they can make a more informed decision.
We still have a way to go until we can say we are content with how active modes are represented in our models, but every detail added brings us closer to realising our ambitions. We have to stay focused and discerning however, as every addition could bring significant impacts on how long our models run and may not be contributing significantly enough to the quality of our outputs.
