From Urban Mobility to Epidemic Dynamics: A Mixture-of-Experts Framework with Preference Alignment for Policy Scenario Simulation

This document introduces UrbanShare-MoE-PA, a new data-driven, agent-level framework designed to simulate policy scenarios, specifically focusing on the impact of non-pharmaceutical interventions (NPIs) on epidemic dynamics. The framework models how NPIs influence individual behaviors, such as daily time allocation, activity, and mobility trajectories, which then propagate through a calibrated behavior-driven SEIR simulator to predict outcomes. This approach moves beyond simple aggregate mobility reductions, emphasizing the behavioral reallocations driven by policy changes.

Why it matters

This framework offers a more sophisticated tool for predicting the effects of policy interventions, particularly in public health and urban planning contexts. By integrating granular behavioral modeling with epidemic simulation, it enables more accurate scenario planning and impact assessment for policy makers, moving beyond simplistic assumptions about population-level responses.

Key insights

• Non-pharmaceutical interventions (NPIs) impact epidemic risk primarily through behavioral reallocations rather than just aggregate mobility reductions.
• Effective NPI analysis requires a behavioral layer that translates policy calendars into realistic activity and mobility trajectories.
• UrbanShare-MoE-PA is a data-driven, agent-level framework for mapping NPI calendars to daily time-allocation trajectories.
• The framework propagates these behavioral changes through a calibrated behavior-driven SEIR simulator to model downstream epidemic outcomes.
• Its behavioral engine decomposes agent-day actions into travel share, Point-of-Interest (POI) category allocation, and travel-mode allocation.
• The approach combines a structural model with preference alignment to better simulate real-world responses to policies.

Source

arXiv — Computers and Society — https://arxiv.org/abs/2608.20512