Reinforcement learning for policymaking in epidemic control: A scoping review

Area of Science:

• Epidemiology
• Computer Science
• Public Health Policy

Background:

• Managing epidemics requires rapid policy adaptation, which conventional rule-based interventions struggle to provide.
• Reinforcement learning (RL) offers a potential solution for designing adaptive non-pharmaceutical interventions (NPIs).
• Existing research on RL for NPI design is fragmented across diverse models and reporting styles.

Purpose of the Study:

• To systematically map the application of RL in epidemic NPI design.
• To describe common modeling choices, algorithm architectures, and evaluation practices in RL for NPIs.
• To identify trends and research gaps in this field.

Main Methods:

• A systematic literature search was conducted across major databases (IEEE Xplore, ACM Digital Library, ScienceDirect, Scopus) for peer-reviewed studies from 2014-2025.
• Studies applying deep RL to select NPIs were included, supplemented by reference list scanning.
• Data on epidemic and RL models, experiments, validation, and outcomes were charted and descriptively summarized.

Main Results:

• Thirteen studies met the inclusion criteria, utilizing various RL methods (value-based, policy-gradient, hybrid) and simulation models (compartmental, agent-based).
• Reward functions typically balanced pandemic severity against socio-economic costs, with discrete restriction levels as common action spaces.
• RL policies demonstrated superior performance over baselines in reducing infections, deaths, or lockdown duration while minimizing economic loss.

Conclusions:

• Reinforcement learning holds significant promise for adaptive epidemic control strategies.
• Standardization of benchmark environments, evaluation metrics, and reporting is crucial for robust comparison and validation.
• Future research should focus on incorporating realistic economic/behavioral models, uncertainty-aware RL, and prospective validation for real-world deployment.