Innovate UK: Building an Agent-Based Model as a Response to COVID-19

The City Modelling Lab was awarded an Innovate UK grant to build a model to inform COVID-19 response efforts

In the immediate wake of the UK’s lockdown in March 2020, we were awarded an Innovate UK grant as part of their Business-led innovation in response to global disruption competition. With that grant, we spent six weeks building an alpha Agent-Based Model (ABM) to test how quickly we could build the foundations of an ABM for a new city and what else might be required by a model to inform COVID-19 response efforts.

We wanted to build an alpha model that could act as a basis for future analysis, using publicly available data. We collaborated with Birmingham City Council and Transport for the West Midlands to deliver an alpha ABM of the West Midlands.

This post is the first of a series about the alpha model we built. The posts that follow will delve into specific lessons about data, the strengths and weaknesses of our model, and how an ABM approach could be used to support COVID-19 responses.

The multimodal network for Birmingham. (Yellow lines are smaller, urban roads. Blue lines are major roads and public transit routes.)

What is the pandemic doing to transportation planning?

COVID-19 presents unique challenges for our transport planners and operators: we are operating in an environment where there are few precedents and our baseline assumptions about how people will behave and how our networks will function are no longer valid.

As we move out of the pandemic, UK transport agencies will need to adapt and respond rapidly to potentially frequent changes in policy. They will need to plan services that minimise crowding while maintaining critical service schedules to serve those in need — maximising accessibility and minimising risk. The requirement to quickly test the impact of future scenarios, particulalry those unlike anything in the past, was a space where we thought that agent based models could fill a gap.

This has included the development of our open source Pandemic Activity Modifier (PAM), and we are honoured to contribute to the Royal Society’s Rapid Assistance in Modelling the Pandemic (RAMP) coalition to support efforts to model the pandemic and guide the UK’s response.

Why are ABMs relevant now?

We use transport models to understand transport networks and plan for changes in infrastructure and policy. However, traditional approaches are dependent on having reliable baseline indicators about how much traffic is on the network, where, at what time, and by what mode (car, public transit, cycle, etc.)

Existing models are suddenlystruggling due to rapid new changes in behaviour as a consequence of the pandemic. Collecting data and rebuilding traditional models would take months or even years as people’s behaviours continue to rapidly change. An ABM approach works well for understanding the pandemic’s impact on transportation for this very reason — the models are built on individual behaviour and can reflect changes in behaviour.

To plan for an equitable, safe release from lockdown and recovery from the pandemic, ABMs have advantages over traditional transport models in three key areas:

1. ABMs model individuals’ behaviour, which is what is changing. Traditional models look at aggregate groups of people and assume they have the same behaviours. The greater granularity allows us to model a wider range of impacts.

2. ABMs use integrated, multimodal representations of transportation. Traditional models look at each mode as a discrete system. The range and scale of the changes we are facing will impact across all modes.

3. ABMs are quick to build and are modular, meaning we can improve different components independently. Traditional models take more time to reconfigure and calibrate for new scenarios.

With the Innovate UK Grant, we could test these assumptions.

What have we been doing?

Activity locations derived from OSM. (Red points represent home locations. Blue points represent work locations)

Birmingham, as the UK’s second largest city by population and GDP was a prime location to test our approach. Over the last six weeks, we worked collaboratively with Birmingham City Council (BCC) and Transport for the West Midlands (TfWM), to build an agent based model for Birmingham and the wider West-Midlands region. Our overarching question was:

How quickly can we build the foundations of an ABM for a new city, and what would need to be done to help inform COVID-19 response efforts?

Within a short period of time, we:

• Built a multimodal network of the West Midlands region. We created the network using a mixture of open data from Open Street Map (OSM) and a proprietary set of timetable data for all the UK’s public transport services.
• We used census data to generate our population, and are grateful for the help, support, and data from our partners that have helped us test and understand how we could further improve the model to address their most pressing questions.
• Synthesised a population to reflect behaviour pre-pandemic, as a baseline behaviour model
• Modified this synthetic population to represent the impacts of the pandemic
• Ran pre-pandemic and pandemic simulations
• Visualised the outputs
• Developed a roadmap for how to iterate, and improve the sophistication and performance of the simulation

Reflections on the project

Alpha simulation output for a baseline simulation: this shows how busy each road in the network was at 9am as a proportion of total capacity. (Yellow lines represent roads with more traffic. Darker blue lines represent roads with less traffic.)

The simulation included approximately 200,000 individual agents (a 10% population) that can travel across a multimodal network of the West Midlands. We also applied the Pandemic Activity Modifier (PAM) which demonstrated the relative change in people’s behaviour, as the region entered lock-down. The simulation outputs benchmarked well compared with the conditions observed.

As well as the model and its outputs, we have learnt a lot from this experience, not least that we can build a working regional scale ABM in six weeks of technical work. While the scale and scope of a city or regional scale ABM may initially seem daunting, we have shown that we can approach model development in an incremental way, starting with an alpha model.

At Arup much of our work has changed significantly in response to the global pandemic. As a business we’ve been developing approaches to the pandemic, both internally and with our clients (Arup.com). Arup’s City Modelling Lab is no exception, and we have been looking for opportunities where we can help address the unprecedented challenges our cities face using our Agent Based Modelling (ABM) approaches.