Adaptive social distancing under variant-specific transmission dynamics in a spatial SEIIR model with reinforcement learning.

Epidemic control is inherently dynamic because viral transmissibility and human behavior co-evolve and vary across spatial scales. Consequently, identical intervention strategies can yield divergent outcomes depending on regional connectivity, temporal changes in transmission, and the relative costs of control. Many existing analytical and policy frameworks, however, assume fixed transmission rates or rely on static thresholds, limiting their ability to guide effective interventions in heterogeneous and evolving epidemic landscapes. We present a mathematical framework that couples a time-varying, multi-patch SEIIR model with reinforcement learning to generate adaptive, region-specific social-distancing strategies under varying cost scenarios. Using COVID-19 incidence and mobility data from 17 administrative regions in South Korea, we estimate time-varying transmission and construct a decision environment in which an agent observes epidemiological states, selects intervention intensities for each region, and receives rewards that integrate epidemiological and economic costs. In the early period, low intervention costs lead the learned policy to impose strong early actions in highly connected metropolitan regions, suppressing incidence after a single peak. Under high costs, sustained control is limited to Gyeonggi Province, allowing persistent circulation elsewhere. In the later period, cost considerations dominate, and high intervention costs suppress actions even during substantial epidemic waves. These results demonstrate that rapid temporal shifts in transmissibility can render strict suppression suboptimal and that optimal strategies may require tolerating ongoing transmission. They highlight the importance of adaptive, spatially explicit control frameworks that integrate mechanistic epidemic models with RL approaches.