A dynamic approach to support outbreak management using reinforcement learning and semi-connected SEIQR models

Area of Science:

• Epidemiology
• Computational Science
• Public Health Policy

Background:

• COVID-19 containment measures severely impacted global economies.
• Balancing public health interventions with economic stability is a critical challenge.

Purpose of the Study:

• To develop and evaluate a reinforcement learning (RL) algorithm for optimizing COVID-19 control strategies.
• To assess the algorithm's ability to minimize disease spread while mitigating economic disruption.

Main Methods:

• A novel RL environment was created, simulating four interconnected Japanese regions using a Susceptible-Exposed-Infected-Quarantined-Removed (SEIQR) model.
• The RL agent learned optimal policies for individual movement and screening by interacting with the simulated epidemic.
• The trained agent was tested against historical COVID-19 epidemic data from Tokyo, Osaka, Okinawa, and Hokkaido.

Main Results:

• The RL agent significantly reduced peak infectious cases (e.g., from 165 to 35) and epidemic duration (e.g., from 148 to 131 days for the 5th wave).
• Policy analysis revealed adaptive strategies: movement restrictions increased with rising cases, while screening eased during declines.
• Okinawa showed unique adaptive responses, with tightened screening during rapid case increases.

Conclusions:

• Reinforcement learning demonstrates significant potential for informing public health policy decisions.
• The developed SEIQR model effectively simulates cross-regional human flow dynamics for epidemic modeling.