Feedback Control for Epidemic Mitigation

29 concepts20 papers1 proposal

Traditional epidemiology is primarily descriptive. Models like the classic SIR (Susceptible-Infectious-Recovered) framework are designed to simulate and observe how a disease spreads through a population under fixed parameters. When public health officials want to evaluate an intervention—such as social distancing, testing campaigns, or quarantines—they typically run "what-if" simulations, manually tweaking parameters to see the resulting curves.

As noted by Sebin Gracy et al. and Alain Y. Kibangou et al., this open-loop approach lacks integrated mechanisms for active, automated intervention. If the initial estimates of viral transmissibility are slightly wrong, or if human behavior changes unexpectedly, static policies can fail catastrophically, leading to either overwhelmed healthcare systems or unnecessary economic stagnation.

To solve this, a growing body of research reframes epidemiological models as dynamic control systems. By treating public health interventions as continuous control inputs, researchers apply rigorous control theory to automatically calculate, adapt, and guarantee optimal mitigation strategies in real time.

The Epidemic as a Control System

To understand this paradigm shift, consider how a standard compartmental model changes when viewed through the lens of control theory. In a descriptive model, the transition rate between compartments (e.g., from Susceptible to Infected) is dictated by a fixed biological parameter, $\beta$.

In a controllable epidemiological model, we introduce a time-varying control input, $u(t)$, which represents the intensity of active interventions. For instance, Aaron Ames et al. and S. Z. Sayed Hassen et al. augment standard models by letting $u(t)$ scale the transmission rate, effectively transforming $\beta$ into $\beta(1 - u(t))$, where $u(t) = 0$ means business-as-usual and $u(t) = 1$ represents complete isolation. Similarly, Alain Y. Kibangou et al. treat the daily testing rate as a dynamic control input that actively shunts individuals from an undiagnosed infectious compartment to a diagnosed, isolated compartment.

Once the system is framed this way, the goal changes from predicting the future to steering the system's state vector toward a safe target, such as eradicating the disease or flattening the infection curve to match hospital capacity.

Closing the Loop: Why Feedback Beats Prediction

Why do we need closed-loop feedback instead of just planning a schedule of lockdowns in advance? The answer lies in uncertainty and robustness.

An open-loop policy relies entirely on the accuracy of the model. If a virus mutates or contact networks change, the planned policy fails. A closed-loop feedback mechanism, however, continuously measures the actual state of the population (e.g., daily new hospitalizations) and adjusts the intervention level based on the error between the observed state and the target state.

       +-----------------------------------------+
       |                                         |
       v                                         |
[Feedback Controller] -> [Intervention u(t)] -> [Epidemic Dynamics] -> [Observed Cases/Hospitalizations]

Francesco Di Lauro et al. demonstrated that a nonlinear feedback controller can successfully keep infection peaks exactly at healthcare capacity limits, even when the underlying epidemiological parameters are poorly understood. This self-correcting behavior is highly cost-effective. As shown by Shouhuai Xu et al., state-feedback control strategies can maintain infection levels within a strict target range without prior knowledge of attack parameters, preventing over-investment in restrictive measures during low-risk periods while guaranteeing containment when infections spike.

Enforcing Hard Bounds via Safety-Critical Control

In many control applications, minimizing a total cost over time is secondary to enforcing absolute safety boundaries. In a pandemic, the ultimate safety boundary is healthcare capacity: if hospitalizations exceed available ICU beds, mortality rates skyrocket.

To enforce these hard constraints, researchers leverage Control Barrier Functions (CBFs), a tool from safety-critical control theory. Instead of optimizing for a vague balance of health and economics, a CBF defines a mathematically "safe set" of states (e.g., all states where hospitalizations remain below capacity) and ensures that the control input $u(t)$ grows strong enough to deflect the system's trajectory whenever it approaches the boundary of that set.

Aaron Ames et al. extended this framework to handle a major real-world complication: measurement delays. Because of incubation periods and testing bottlenecks, the data public health officials see today reflects infections that occurred a week ago. By combining safety-critical control with predictor feedback, the authors demonstrated that it is possible to formally guarantee that hospitalizations stay within safe limits despite these destabilizing delays.

Rolling Horizons and Predictive Optimization

When public health policies must balance competing objectives—such as minimizing deaths while minimizing the socio-economic cost of lockdowns—Model Predictive Control (MPC) becomes the tool of choice.

MPC operates on a rolling horizon. At each time step (e.g., every day), the controller uses the epidemiological model to predict the trajectory of the disease over a short future window (e.g., 30 days). It solves an optimization problem to find the best sequence of interventions over that window, balances the trade-offs, executes only the first day's intervention, and then discards the rest. The next day, the process repeats with fresh data.

Today (t=0): Solve optimal 30-day plan -> Execute Day 1 policy.
Tomorrow (t=1): Receive new data -> Solve new optimal 30-day plan -> Execute Day 2 policy.

This framework, utilized by Nicholas J. Watkins et al. and S. Z. Sayed Hassen et al., allows policies to be highly adaptive. By framing the problem as a chance-constrained optimization, the system automatically updates its strategy as new data arrives. Zhe Xu et al. showed that even complex, non-linear epidemic control problems can be solved with high computational efficiency, making daily online policy adjustments fully viable.

To ensure these optimization routines are grounded in reality, they must be paired with robust state estimation. Because many cases are asymptomatic or unrecorded, M. Umar B. Niazi et al. designed specialized observers to reliably reconstruct the true hidden state of an epidemic from noisy, incomplete data, proving that optimal control policies derived from these robust estimates significantly reduce adverse outcomes like fatalities.

Granular and Networked Control

While macro-level compartmental models assume a well-mixed population, real-world diseases spread across complex networks of individuals and geographic regions. Control theory scales to these granular structures through several distinct mechanisms:

  • Spatial and Mean-Field Control: Instead of treating population movement as a random variable, Siting Liu 0003 et al. utilized a mean-field control framework that treats population velocity and displacement as controllable variables, optimizing spatial movement to isolate outbreaks while minimizing societal disruption.
  • Network Micro-Interventions: Rather than applying blanket restrictions, Lars Lorch et al. modeled disease spread using stochastic differential equations with jumps to determine exactly who to treat or isolate and when. This dynamic, state-dependent network control consistently outperforms static graph heuristics (like targeting high-degree nodes).
  • Heterogeneous Strategies: Nodes in a network have different characteristics. Shanshan Chen et al. demonstrated that tailoring asymmetric, node-specific immunization and quarantine rates is far more cost-effective than uniform, population-wide mandates.
  • Incentive and Multi-Agent Modeling: When central authorities cannot mandate behavior directly, they must rely on incentives. Régis Sabbadin et al. framed this as a Leader-Follower Markov Decision Process (LF-MDP), computing optimal, state-dependent incentive policies that encourage independent actors (such as individual farmers managing livestock disease) to take actions that collectively contain an epidemic.

As the complexity of these environments grows, traditional control derivations can become mathematically intractable or limited by state-space explosion, as explored by Marco Roveri et al.. To bridge this gap, toolboxes like the one developed by Cédric Colas et al. wrap epidemiological models in reinforcement learning interfaces, allowing model-free agents to discover these highly effective, closed-loop feedback policies automatically through simulation.

Key Foundational Work

Adaptive Epidemic Dynamics in Networks: Thresholds and Control (2013) — contributed 2 concept elements to this direction.
Passivity framework for composition and mitigation of multi-virus propagation in networked systems (2015) — contributed 2 concept elements to this direction.
A Tractable Leader-Follower MDP Model for Animal Disease Management (2013) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_epidemiological_control_systems_75131bbd,
  title = {epidemiological control systems},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/epidemiological-control-systems/rdr_75131bbdd5a8406cbcd3219ba5a99f77},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_epidemiological_control_systems_75131bbd_v1,
  title = {epidemiological control systems},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/epidemiological-control-systems/rdr_75131bbdd5a8406cbcd3219ba5a99f77/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake