Agent Based Models through the eyes of a traditional transport modeller
Seb Bukowski is a transport modeller and data scientist at Arup, specialising in strategic transport modelling, automation and transport data analytics
I’ve built strategic transport models for multiple cities. All of those models have been traditional models, dominated by the 4-step method. That changed in spring 2020 when I built an Alpha Agent Based Model (ABM) for Birmingham, with the aim of supporting cities with their COVID-19 responses. In this post, I share my experience developing and working with an ABM from a perspective of a traditional, strategic modelling practitioner. I explore key differences between traditional models and ABMs for key aspects of a transport model lifecycle.
This work was supported by an Innovate UK grant for Business-led innovation in response to global disruption (read more here).
Agent based models compared with traditional strategic models
ABMs simulate the behaviour of individuals, including their different characteristics, needs, and resources (e.g. how many people have a driver’s license or own a car.) What’s very exciting about ABMs is that they are very flexible because of this granularity. We can understand where people are travelling more precisely; compared with traditional models that represent flows of travel between different zones, ABMs represent travel behaviour between facilities (e.g. a residential block, a hospital, a school, an office, etc.) We can understand when people are travelling, second by second. Traditional models aggregate travel into about 5 time periods across a day. And we can more easily adapt ABMs to extend their use cases. It is possible to repurpose ABMs in a way that we cannot with traditional models, which are largely restricted by assumptions made during the development stage.
Model Purpose
Traditional strategic transport models are typically built with a set of use cases defined prior to the development of a model. These models become a reliable tool to assess the impacts of different scenarios that fit within these defined use cases. However, in most cases these models struggle to assess scenarios that are outside of those defined use cases. Adapting traditional models can take a significant amount of time.
The world changed because of COVID-19. Traditional transport models will struggle to cope with the new changes due to the assumptions, characteristics and granularity represented in the models. Traditional transport models are developed on the basis of historical trends and represent typical average conditions for a given time period. One of the impacts of COVID-19 is that we are uncertain of what the new ‘normal’ will be. We can no longer depend on historic travel patterns to forecast how people will behave tomorrow, let alone further into the future. This is where ABMs step in. ABMs have the potential to fill a gap in our modelling suite because we can simulate changes in individual behaviours.
Model Build Time
Digital transformation is well underway within transport modelling. Whilst traditional modelling is adapting to take advantage of automation and improved quantity and quality of data, the model development phase can still take 1–2 years before the models can be put into use.
ABMs may be able to provide us with insights faster. Take the Birmingham model as an example: we built an Alpha ABM in six weeks. This Alpha ABM consists of a core network and a synthetic population.
This Alpha ABM has not yet gone through a calibration exercise. However, in fewer than six weeks we represented changes in travel behaviour that compared favourably with our benchmark data. We could not have built a traditional model with comparable outputs in the same timeframe.
Model Development Guidance and Best Practice
Agent Based Modelling is a very exciting field to be part of as it has high potential to address limitations within traditional strategic modelling and deliver better decisions for governments at a city, regional, or national level. On the other hand, traditional modelling is a well-defined process that has been mostly unchanged for many years. Models are kept consistent and are trusted because of robust modelling guidance. Because the limitations of 4-step models are well understood, they have been effectively used in decision making for years, giving decision makers confidence.
Widespread adoption of ABMs will be constrained until further research and definition of best practice are developed, and accepted by the wider transport modelling community. This will need to address aspects of the models such as model stability and model convergence. Currently, it’s still an open field and there is no single authority or standards on how to formally assess ABMs. This is currently one of the core development areas within Arup’s City Modelling Lab.
Model Outputs
The range of possible outputs that can be generated from a model is definitely an advantage of ABMs. All typical outputs such as network performance or standard plots (e.g. vehicle flows, volume/capacity) can be produced using an ABM model.
ABMs stand out from traditional models through their ability to model and visualise each individual decision made on the network at a much more granular level of detail, both temporally and spatially. These outputs make a significant difference to how model outputs can be communicated and enable a much more informed and detailed conversation around modelled scenario results.
Summary
Working on an ABM for the first time was a very insightful experience for me, allowing me to gain general knowledge about this new modelling capability and explore the differences, benefits and limitations of this approach in comparison to the traditional modelling. There is still a lot of work ahead to develop better awareness and understanding of this capability within the transport modelling community, but ABMs definitely have a role to play in the transport modelling field over the coming years. I’m excited to be part of this journey.
