<p>Hidden transmission and network heterogeneity have emerged as major challenges in epidemic control, as evidenced during the COVID-19 pandemic. Asymptomatic and pre-symptomatic spread, combined with uneven human contact patterns, contribute to super-spreading events that are difficult to detect and manage in real-time. Most existing models assume full observability of infections–an unrealistic premise in practical settings. We present a data-informed decision-support framework that integrates individual-based modeling (IBM) with reinforcement learning (RL) to optimize testing and quarantine policies under partial observability. Our IBM captures person-level disease dynamics over a scale-free contact network and is calibrated with COVID-19 data from South Korea. The RL agent is trained under surveillance regimes where only symptomatic cases are visible. Even with limited information, the agent learns policies that uncover hidden state dynamics, balancing disease suppression with economic costs. Our approach underscores the importance of realistic observability assumptions and provides a tool for managing outbreaks under uncertainty.</p>