COVID-19 has brought modelling and simulation of complex systems to the centre of public debate and the policy-making process. As countries across Europe return to lockdown, it’s clearer than ever that policies need to be informed by the best possible science in order to maximise impact and minimise disruption.
The efforts from across academia and industry to meet this demand for science and get the right information to the right people as quickly as possible has, in the UK and elsewhere, been extraordinary. It’s been amazing to see so many institutions and individual specialists from a vast range of disciplines step up and volunteer their knowledge and resources.
I’m incredibly proud of the small but significant contribution that Improbable has been able to make.
Better science for better planning and policy design
The Rapid Assistance in Modelling the Pandemic (RAMP) initiative was convened by the Royal Society. It brings together interdisciplinary modelling expertise from across academia and industry to support the work of the Government’s scientific advisory groups such as the Scientific Pandemic Influenza Group on Modelling (SPI-M) and the Scientific Advisory Group on Emergencies (SAGE).
The Urban Analytics group leads one of the main projects within RAMP. They’re responsible for combining models of epidemic spread with detailed data on behavioural patterns, including transport systems, shopping and leisure activities, into a single simulation. This is designed to allow policy-makers to analyse how COVID spreads, explore a wide range of nuanced interventions, and understand the likely consequences of each one.
The Urban Analytics group is led by Professor Mark Birkin from the University of Leeds and The Alan Turing Institute, with the help of Professor Nick Malleson. It comprises experts from Leeds, Exeter and Cambridge universities as well as University College London, who specialise in spatial microsimulation, urban flows, social dynamics, transport modelling and disease transmission.
Improbable sent two research scientists, Rory Greig and Charlie Shenton, to join the team. They both have years of experience working on mathematical and computational aspects of large scale simulations, and combine expertise in computer science, game engines and scientific modelling of complex systems. Recently their research focus has been on high-performance programming and visualisation techniques for exactly the kind of models that are commonly used by epidemiologists to study how viruses spread.
Modelling the pandemic: the technical challenge
The Urban Analytics group focused initially on building a model for Devon, with the goal of scaling to the whole country at a later stage. Their model is agent-based, and simulates behavioural patterns of a detailed synthetic representation of Devon and a population of 700,000 individuals. It incorporates a vast amount of data and research into how real-world populations spend their time, move around and interact.
This granular level of detail makes it possible to model the likely effects of a wide range of interventions such as local lockdowns, placing partial restrictions on specific sectors, and restrictions relating to travel, retail, hospitality, education and industry.
At the same time, due to the model’s complexity, the team was faced with computational challenges typical of large-scale microsimulations. A simulation at national scale that might project disease spread a month into the future would take days to complete. Exploring the full distribution of possible outcomes under multiple potential plans, which would involve thousands of simulation runs, would have required an enormous amount of computational resources.
Why microsimulation matters
Improbable’s defence and national security business has spent the last few years developing methods and technologies to solve precisely this kind of problem. Our scientists and model engineers specialise in taking complex models like RAMP’s and integrating them to form massive microsimulations and multi-domain synthetic environments. These help enhance situational awareness, planning and decision making for organisations like the British Army and UK Strategic Command.
To understand the scale and nature of the challenge facing the Urban Analytics group, it’s worth spending a few words on microsimulations in general.
Microsimulations are computational models that represent a piece of the real world at high resolution, breaking down the system in terms of its individual components, and aiming to reconstruct its overall behaviour from the bottom up. This approach allows researchers to combine large amounts of fine-grained data and ask questions about how the system might respond to a variety of stimuli. For example, social scientists might model whole populations down to individual people, and use these models to explore how policy options might impact society as a whole.
A powerful aspect of high-resolution models is that, independent of the specific context they were developed in, their designs tend to converge on comparable representations of the world. This, at least in principle, creates opportunities to couple together models from different domains to form increasingly rich simulations, bringing together knowledge and data from disciplines that are often siloed.
This is useful to teams like RAMP because different organisations and research institutions tend to have models and data that are very specific to their aims. Some models are designed to help people analyse physical systems like the weather, supply chains, transport systems and critical national infrastructure like power grids, telecoms networks and water supplies. Other models are engineered to give insights into abstract systems like economics, information spread and opinion dynamics. Coupling many of these models together into a unified, cohesive simulation presents a number of methodological and computational challenges.
That’s where the core of our expertise and technology at Improbable lies.
Faster models for better science
In just five weeks, Charlie and Rory worked alongside the Urban Analytics researchers to integrate disparate models written in different programming languages and re-write the entire simulation in highly optimised form. This improved performance by a factor of 10,000. A month-long simulation that would previously have taken an hour to run now takes less than half a second.
The team also also built a user interface that allows researchers and policy designers to visualise and interact with the detailed information generated by the model in an intuitive way. This UI, combined with the performance boost, drastically shortens the development cycle and makes entirely new analysis possible, including sensitivity analysis, uncertainty quantification and automatic calibration.
The model code has now been fully open sourced and integrated into the project’s codebase.
In summing up the outcomes of this collaboration, Professor Mark Birkin, Programme Director for Urban Analytics at The Alan Turing Institute, said:
The pandemic has demonstrated unequivocally the need to mobilise the best in mathematical modelling and data science to inform policy in the public interest. Massive enhancements in the power of our models through the collaboration with Improbable open up new horizons in the evaluation of multiple lockdown scenarios, assimilation of complex data relating to social and behavioural patterns and health outcomes, as well as opening the door to new academic insights in addition to enhanced decision-support.
Want to join the team?
Improbable’s Defence and National Security business operates globally and combines our parent company’s software engineering experience with expertise in computational modelling, AI and data analytics. Our work focuses on adapting and extending Improbable’s multiplayer gaming technology to enable the most sophisticated military simulations and synthetic environments ever experienced.
You can find out more about what we do at https://improbable.io/defence, and at https://defense.improbable.io/.
If you’re interested in joining us, that’s great news. Because we’re hiring.