From Urban Mobility to Epidemic Dynamics: A Mixture-of-Experts Framework with Preference Alignment for Policy Scenario Simulation
This paper introduces UrbanShare-MoE-PA, a data-driven agent-level framework that uses a mixture-of-experts architecture with preference alignment to translate non-pharmaceutical intervention policies into realistic mobility and activity trajectories, thereby enabling more accurate simulations of epidemic dynamics and activity trade-offs.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
When a city faces a sudden health crisis, the most immediate question for leaders is not just how to stop the virus, but how to keep the city functioning. The answer lies in understanding human behavior. For decades, scientists have known that people do not move randomly; our daily lives are structured routines of work, shopping, and travel, constrained by time and habit. When governments impose restrictions like lockdowns, they do not simply reduce the total amount of movement. Instead, they force a complex reshuffling of where people spend their time and how they get there. A person might stop visiting a crowded office but continue to drive to a grocery store, or switch from a busy subway to a private car. These specific choices determine how much exposure to the virus remains, and how much economic activity survives. The challenge for policymakers has been to predict these shifts before they happen, rather than just reacting to them after the fact.
A team of researchers at University College London has developed a new way to simulate these shifts, moving beyond simple counts of how many people are moving to a detailed map of how they are spending their time. Their work, centered on data from Singapore during the first wave of the pandemic, creates a digital mirror of daily life that can be tested against different policy calendars. Instead of guessing how a lockdown might change behavior, the researchers built a system that learns from real people's actual movements to predict what would happen if the rules changed. They found that the timing of a restriction matters far more than its length, and that the specific mix of activities people keep doing is just as important as the total number of trips they take.
The researchers started with a massive dataset of real-world movements from 911 individuals in Singapore, tracked over 184 days from March to August 2020. This period covered the entire arc of the country's first wave, including the strict lockdown, the gradual reopening, and the shifting phases of restriction. Rather than treating these people as a single, blurry mass, the team looked at each person's day as a series of choices: how much time was spent traveling, where they stopped when they weren't traveling, and what mode of transport they used. They noticed that when policies changed, people did not just move less; they moved differently. Some stopped going to restaurants but kept visiting parks; others stopped taking public transit but continued to drive. The researchers realized that to predict the future, they needed a model that could understand these specific, individual trade-offs.
To build this model, the team created a digital engine that breaks down every day into three parts. First, it calculates how much of the day is spent in transit. Second, it decides how the remaining time is divided among different types of places, such as homes, offices, shops, and schools. Third, it determines how the travel time is split between walking, driving, taking the bus, or using the train. The key innovation was teaching this engine to recognize that different people react to rules in different ways. Some people might be able to work from home easily, while others cannot. Some might switch to driving immediately, while others wait. The model uses a "mixture of experts" approach, which is like having a team of specialists inside the computer, where different specialists take the lead depending on the situation. One specialist might handle the behavior of people who stay home during a lockdown, while another handles those who continue to go to work. This allows the system to capture the messy reality that not everyone follows the same script.
The researchers then added a layer of "preference alignment" to ensure the model behaved logically when the rules changed. They taught the system to prefer actions that matched the current policy phase over actions that belonged to a different time. For example, if the policy shifted to a lockdown, the model learned to prioritize staying home, even if the person had a history of going out. This was crucial for simulating "what-if" scenarios. The team wanted to know what would have happened if the lockdown had started a week earlier, ended a week later, or lasted for a different duration. By running these simulations, they could generate a complete, day-by-day forecast of how the population would behave under these alternative rules, long before the virus spread to those new conditions.
The results of these simulations revealed a clear and powerful pattern: the timing of a restriction is more critical than its duration. When the researchers simulated an earlier lockdown, starting a week before the actual event, the peak of infections dropped significantly. The virus was cut off before it could build up a large wave of cases. Conversely, a delayed lockdown allowed the infection to spread further before restrictions kicked in, leading to a higher peak and more total cases. Interestingly, simply making a lockdown longer did not always help. If a lockdown was extended after the peak of the epidemic had already passed, it did little to reduce the total number of infections but did reduce the amount of time people spent in public spaces. This suggests that once the worst of the outbreak is over, extending restrictions yields diminishing returns for health while continuing to cost the economy.
The study also highlighted the delicate balance between controlling the virus and keeping the city alive. The researchers created an index to measure how much economic activity remained under different scenarios. They found that a shorter lockdown preserved the most economic activity, but it came with a small increase in the number of infections. An early lockdown was the most effective at stopping the virus, but it came with a larger drop in economic activity. A late lockdown kept activity levels high but failed to control the disease effectively. There was no perfect solution; every choice involved a trade-off. The model showed that the best outcome depends entirely on what a society values most at a specific moment: minimizing the immediate health crisis or preserving the maximum amount of daily life.
What makes this work significant is that it does not rely on broad assumptions or aggregate numbers. It builds its predictions from the ground up, using the actual habits of real people. By understanding that a lockdown changes the structure of a day rather than just the volume of movement, the model provides a much clearer picture of the consequences of policy. It shows that a policy calendar is not just a list of dates, but a set of instructions that reshapes the daily rhythm of millions of people. The researchers demonstrated that if you want to predict the outcome of a policy, you must first understand how that policy changes the way people spend their time.
The study concludes that for future crises, the ability to simulate these behavioral shifts will be essential. Policymakers will need to know not just how strict a rule is, but how it fits into the flow of daily life. The model suggests that acting early is often more powerful than acting for a long time, and that the specific mix of activities people continue to do matters as much as the total number of people moving. By turning the complex, chaotic reality of human behavior into a structured, predictable simulation, the researchers have provided a tool that can help leaders navigate the difficult choices between health and economy. The work does not offer a magic bullet, but it does offer a clearer map, showing that in a crisis, the timing of a decision can be just as important as the decision itself.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.