Abstract
Researchers at the University of Southern California have developed CampusPandemicPlanR, a reinforcement learning (RL) simulation tool designed to assist university administrators in navigating campus operations during a pandemic. By employing a Q-learning agent, the tool identifies optimal classroom attendance strategies that balance the conflicting objectives of maximizing in-person learning and minimizing COVID-19 transmission. This paper demonstrates the feasibility of using RL to model complex campus dynamics and provide data-driven decision support.
Introduction
The COVID-19 pandemic presented educational institutions with a difficult choice: remain fully open or transition to entirely remote learning. In practice, campus management is a nuanced optimization problem. Administrators must frequently adjust student density in classrooms based on fluctuating community risk levels and internal infection rates. The core challenge lies in the trade-off between educational benefit—which is significantly enhanced by in-person interaction—and public health, which necessitates social distancing. This work addresses the need for a decision-support tool that models these dynamics to suggest "smart" operational strategies.
Methodology
The problem is framed as a Markov Decision Process (MDP) implemented within a custom OpenAI Gym environment, campus-v0.
State and Action Spaces
To maintain computational efficiency, the state space tracks the percentage of infected students per course and broader community risk levels, discretized into three tiers (0–33%, 33–66%, and 66–100%). The agent's action space consists of proposing attendance levels for each class: 0% (fully online), 50%, or 100%.
Transmission Model and Reward Function
State transitions are governed by an equation based on the SIR (Susceptible-Infectious-Recovered) model, accounting for both community spread and peer-to-peer classroom transmission. The agent's objective is defined by a weighted reward function: $$R(n) = \alpha \sum (\text{Allowed Students}) - (1 - \alpha) \sum (\text{Infected Students})$$ The parameter $\alpha$ (ranging from 0 to 1) allows administrators to weight the relative importance of attendance versus infection control.
Results
The Q-learning agent was evaluated over 5,000 training episodes in the simulated environment.
Learning Performance
The agent demonstrated a consistent ability to optimize its policy over time. As shown in the performance analysis, the expected returns increased steadily as the agent explored the state-action space and converged on more effective strategies.
Policy Trade-offs
By varying the reward weight $\alpha$, the researchers mapped a Pareto-optimal frontier between attendance and public health. When $\alpha$ is low, the agent prioritizes safety, maintaining low infection rates by restricting attendance. Conversely, higher $\alpha$ values lead the agent to allow more students on campus, accepting a predictable increase in infections to maximize educational output.
Discussion and Future Work
This proof-of-concept demonstrates that RL can effectively navigate the complex trade-offs of pandemic-era campus management. The discretization of states and actions makes the problem tractable for standard Q-learning, providing an interpretable alternative to deep learning models.
However, several limitations remain. The current SIR-based transition model is a high-level approximation; future iterations should incorporate granular indoor epidemic spread models that consider room ventilation and physical layouts. Additionally, the model currently assumes a homogeneous population. Future work will aim to differentiate risks based on activity types (e.g., lectures vs. laboratories) and refine the discretization for more nuanced planning.
Summary updated