Optimizing Place-Level Closure Decisions for Strict Epidemic Control: An Agent-Based Simulation and Reinforcement Learning Approach

The COVID-19 pandemic highlighted the challenge of designing intervention policies that suppress transmission while limiting economic disruption. Although lockdowns effectively reduce infections, prolonged restrictions impose substantial social and economic costs. Balancing these competing objectives is therefore a central problem in epidemic policy design. This study proposes a reinforcement learning framework for the selective closure of community places, treating each place as an individual decision unit. We develop a spatial agent-based simulation model in which infection spreads through place-based contacts and individual mobility follows a gravity-based formulation that captures place attractiveness and locality. The epidemic state is represented as spatial maps and processed by a deep neural network to learn closure policies that minimize the combined health and economic costs of epidemic control. We evaluate the learned policies under a cost formulation that explicitly balances infection burden against economic losses caused by place closures, weighted by the health and economic cost coefficients ch and ce, respectively. Simulation experiments were conducted in a small-scale virtual city under a strict epidemic-control scenario. The learned policy reduces total cost relative to threshold-based lockdown policies. The reduction is statistically significant when economic losses receive moderate or greater weight in the objective (ce = 0.04 and 0.06), whereas the difference is not statistically significant at ce = 0.02 (p = 0.12). Beyond the conventional emphasis on high-risk places, the results identify locality as an important determinant of efficient closure decisions. These findings suggest that spatially informed, place-level interventions can improve the efficiency of epidemic control while mitigating economic losses.