Fine-Tuning Activity Plans in Agent-Based Transport Models
Making behaviourally consistent transport demand models
In the City Modelling Lab, we use agent-based simulations to inform transport planning policy.
Using a behaviourally reasonable starting point for our simulations is very important. It can reduce simulation run times, generate more valuable insights from our post-processing analysis, and widen the possibilities for scenario testing and policy recommendations. For this reason, we work hard to ensure the transport plans of the agents in our simulations are as realistic as possible.
This post will explore how we assign optimal locations to activities in agent travel plans.
Traditional transport modelling vs an activity-based approach
For decades, the most popular approach in transport modelling has been focussed on modelling individual car trips by assigning them to a transport network and deriving aggregate traffic flows along each road. Although proven and tested for many years, this approach omits a critical element from the equation: the individual users of the transport system itself. Their individual needs, perceived constraints and idiosyncrasies are not taken into account.
As transport-related problems and questions become ever more complex, attention has started shifting from aggregate models to more detailed representations of individual mobility behaviour.
Activity-based modelling (AcBM) has emerged in recent decades as an approach that aims to understand individual mobility-related sensitivities and constraints. The main principle of AcBM is that the demand for travel stems from the need for individuals to participate in activities in order to fulfil their everyday needs.
For example, an individual living in the suburbs might commute to work in the city centre in the morning using public transport, then later walk to a social activity close to work, and then shop for groceries on the way back home again using public transport. Identifying the most likely locations for those out-of-home activities and the means for travelling there, given individual preferences and constraints, becomes an important step towards developing an AcBM.
Location choice problem
We divided the location choice problem into two pieces: long-term choice of significant activities, e.g. where people work, and short-term choices for things that have more flexibility on a day-to-day basis, e.g. shopping, leisure, etc. We initially focussed on tackling the longer-term choices.
The choice of transport mode is also essential and interlinked with location choice. For example, an individual with a car is more likely to drive to a workplace in an area with few to no car access restrictions. On the other hand, individuals with no car in their household might choose a workplace in a city centre with better public transport, or a place that is more easily accessible by walking/cycling.
Household characteristics can also play an important role in influencing the perception of travel time and the allocation of resources, e.g. cars, to household members.
Work location choice model
Using travel surveys, we analysed individuals’ choices to identify how sensitive those choices are to travel time and cost, and how those sensitivities might differ across different demographics. We also want to capture the impact of employment opportunities as a measure of attraction for commuting.
To this end, we estimated a range of behavioural models (Multinomial Logit models) aiming to understand work location and mode choices in a joint fashion, acknowledging the interdependences of those two choice dimensions.
As a measure of validation, the resultant estimated trade-offs of time and cost (Values of Travel Time) were comparable with the official values used for appraisal purposes. We performed further validation tests, focussing on the model’s generalisability and prediction performance across different years (temporal transferability) and regions (spatial transferability).
That gave us the confidence to apply the estimates from that model in different application contexts, both spatially and temporally. The model was able to provide valuable insights from a policy perspective, such as:
- The higher likelihood, relative to urban dwellers, of individuals in rural areas using their car
- The increased preference of male individuals, relative to females, for cycling
- The likelihood of part-time workers performing shorter commuting trips than full-time workers
- The lower sensitivity to travel costs of higher-income individuals
The model described can allow us to perform scenarios involving changes in land use (e.g. adding additional jobs to an area) and analyse how individuals’ work location choices and overall mobility patterns are affected.
Furthermore, the model can accommodate the analysis of a wide range of scenarios focused on different factors, for example, changes in population demographics (e.g. age and/or income), or the impact of fuel cost increases as a stand-alone policy scenario or in combination with public transport fare decreases (demand elasticity/cross-elasticity analysis).
Discretionary locations
Estimated work locations, combined with home locations generated during population synthesis, provide anchor points within which employed individuals can choose the locations for their remaining secondary activities. These choices are a function of time-space constraints and destination attraction.
All else being equal, the further the work location from home, the less likely it is — due to time-space constraints — that the individual deviates from the straight path between home and work to perform “discretionary” activities.
However, individuals are more likely to deviate from that straight home-work path to reach a more attractive destination, such as one offering more shopping options.
In the following figure, the shopping location with 50 stores presents a more attractive destination for the decision-maker, but it would require a larger deviation from the home-work path. In contrast, the two shopping locations with 30 and 10 stores are equally reachable, but the former presents a more attractive alternative and is thus more likely to be chosen.
Conclusions
The goal of our simulations is to better represent the complexity of people’s choices and decision-making. Our approach to location selection in agent activity plans means that we can create more realistic inputs to our simulations, enhancing our ability to understand and improve the future of our cities.
The combination of approaches optimised for both long-term and short-term locations allows us to include more meaningful activity locations in our synthetic population, significantly reducing running times in the subsequent agent-based simulations.
The work location model and the time-space approach are part of a wider activity synthesis pipeline that also consists of an activity generation/scheduling component, which will be described in a future post.
The described activity-based pipeline leads to a synthesised travel demand consistent with observed mobility behaviour, while also being responsive to changes in network conditions, land use and demographic attributes. The outputs from that process are then used as input to our agent-based simulations to address policy questions around equity, reducing emissions and car dependency, among others.
