Mobile clinics, such as vaccination vans, screening trucks and maternal-care units, are usually planned from a single map of where "the population" is and where it goes. That map averages very different groups together. In Virginia, the 25 best vaccine-van locations for Black residents and the 25 best for Latinx residents shared only 2 sites (Mehrab et al.). A plan built on the average serves neither group well. This proposal studies a second, less-noticed consequence of that averaging. Real fleets lose vans to breakdowns, lose staff, and lose radio contact. When that happens, a planner that reasons over aggregate demand will tend to move the remaining vans toward the dense majority footprint. The groups whose footprints were never visible in the aggregate then lose coverage first and lose the most.
The core claim is a measurable mechanism. We expect equity loss from fleet disruption to be concentrated in demographic groups whose movement footprints do not overlap the aggregate peak. We further expect that giving the planner separate demand layers for each group, each discounted by that group's measured access friction, removes most of this concentrated loss, regardless of whether the planner is a classical optimizer or a learned multi-vehicle policy. Learning-based coordination is tested as a secondary question with a specific scope. We ask whether a policy that handles any number of vans can keep coverage equitable when vans must decide locally because the central solver is unreachable. If re-optimization matches it everywhere else, we will say so.
Core Claim and What Is New
Two existing lines of work meet here without having been joined. The mobility line shows that splitting demand by community changes what you see. Ada-MSTNet forecasts inflow and outflow separately for 25 demographic communities over 784 one-kilometer cells in Beijing and beats aggregate forecasters by 15.77% to 22.78% in RMSE (Liu et al.). Mehrab et al. show that separate demographic footprints matter for clinic placement. But those forecasts were never connected to a dispatch decision, and Mehrab's pipeline is a weekly ranking that treats each van as an independent point. The access-disparity line supplies the objective. Modeling Access Differences to Reduce Disparity in Resource Allocation models disadvantaged groups as acquiring a resource at a slower rate, scaled by an access-gap factor $\eta \in (0,1]$. It proves that proportional or volume-seeking allocation lets advantaged groups saturate supply first. Its formulation, however, is one-shot and static. Guo et al. document the "double burden" in 954 US counties, where low-income residents are both farther from care and less likely to have transport (Income disparities in driving distance).
What no paper in the file does is ask how disaggregated demand and access-gap weighting behave when the fleet itself changes over time. The file's own framing of the gap, "put MARL on mobile clinics with an equity reward," is not what we propose. A skeptic would correctly answer that a rolling re-optimization would do as well. We instead make the disaggregation-under-disruption effect the object of study and treat the choice of planner as a variable.
Model
Demand is represented as $G$ layers. Layer $g$ gives, for each cell $i$ and hour $t$, the expected number of people $d_{g,i,t}$ from group $g$ present there, estimated from foot-traffic or query data. A van stationed at site $s$ produces expected uptake for group $g$ of
$$u_{g,t}(s) = \eta_g \sum_i d_{g,i,t}\, e^{-\beta\, c_{i,s}},$$
where $c_{i,s}$ is road-network travel time from cell $i$ to site $s$, computed on OpenStreetMap as in Xu et al., not straight-line distance. The term $e^{-\beta c}$ is the standard distance-decay assumption. $\beta$ is the decay rate, which we fit from Mehrab's dose-versus-distance data around the 147 deployed sites. $\eta_g$ is the group's access gap.
The fleet objective is total uptake minus a penalty on the gap between the best- and worst-served group's coverage fraction, weighted by $\lambda$. This places the volume-versus-parity tension the file identifies directly in one tunable term. The aggregate planner is the same planner given a single layer with $\eta = 1$.
Disruption enters through a stochastic process. Vans fail and return, staff with different skills join and leave, and communication with the central planner drops for stretches of hours. Vans may carry vaccination, point-of-care testing, or imaging skills, following the heterogeneous assets described by Liu et al. and Shao et al..
Research Program
We will build an open simulator from three grounded settings. The first is Virginia, using SafeGraph-style foot traffic, Census Block Group demographics, and the vaccination records used by Mehrab et al. It is the only setting with real deployed mobile sites to validate against. Data access for SafeGraph may be restricted, and we will report what can be released. The second is Beijing, using Liu et al.'s 25-community demand tensors, which give the richest demographic layering but no health outcome. The third is rural Kebbi State, Nigeria, using the facility and population grids from MamaCare AI, with OpenStreetMap travel times replacing its Haversine distances.
On these settings we compare planners that differ along two independent axes. The first axis is aggregate versus layered demand. The second is the planner type: rolling re-optimization or a learned policy. The learned policy uses attention over a variable set of van tokens and demand-layer tokens, so a van dropping out does not break the input. It is trained on randomly sampled sub-fleets, following the entity-subgroup training of REFIL. We borrow COPA's coach-and-player structure, in which a central coach broadcasts a compact strategy vector only when it changes enough (Coach-Player MARL). This lets vans keep acting on a stale strategy during communication outages.
We deliberately do not apply MIPI-style penalties that make each agent ignore its teammates (Wang et al.). Such penalties guard against brittle teamwork, but they would also erase the necessary pairing between, say, an imaging van and its technician. Whether a weak version helps is an ablation, not a design commitment.
Baselines
The naive reference point is population-proportional allocation: vans are assigned to districts in proportion to headcount and stationed at each district's densest site. Modeling Access Differences predicts this will widen disparity whenever $\eta_g < 1$. It anchors the bottom of the scale.
The deployed-practice baseline is Mehrab et al.'s weekly demographic cell ranking, run on the same data, with vans assigned to the top-ranked cells. Mehrab et al. report the 2-of-25 overlap and that sites within 1 km of recommendations gave more doses. No exact dose figure is given, and no reported result was found on Lacuna for disparity or coverage under fleet disruption for this method. We will also run it re-ranked daily rather than weekly, so that update frequency is not confounded with method.
The strongest competitor is rolling-horizon re-optimization. This is a p-median style integer program over road-network travel times, in the spirit of Xu et al., with the uptake objective above. It is re-solved every planning period and immediately after every disruption event, using exactly the same layered demand forecasts and $\eta_g$ values as the learned policy. Xu et al. report that optimal network placement roughly halves average travel distance relative to existing layouts. That figure concerns static facilities, and no reported result was found on Lacuna for a rolling re-solve applied to mobile fleets. For fairness, the optimizer gets a wall-clock budget per solve equal to what the learned policy's training amortizes over the evaluation horizon. Both use the same forecasts, and both are tuned on the same validation weeks.
The learned policy counts as beating it only if it does one of two things on held-out weeks. It must lower the worst-group coverage loss after disruptions by a margin larger than seed variance, or match it under connectivity while clearly beating it during communication outages. In those outages, the optimizer can only execute its last plan.
The key ablation removes the demographic layers: every planner is given aggregate demand with $\eta = 1$, everything else unchanged. A second ablation keeps the layers but sets every $\eta_g = 1$. This separates the effect of seeing separate footprints from the effect of weighting groups by access friction. A flat MAPPO policy with fixed-size inputs is included only to confirm the known failure of fixed-input methods when roster size changes, which VAST and MIPI already document on games (Phan et al.).
Why not just…?
Why not just re-solve an optimization problem every time something changes? This is the real threat, and the design treats it as such. Given the same layered forecasts, rolling re-optimization may capture almost all of the equity benefit. If it does, the result is still useful: it would show that the demographic layers, not the learning, do the work. The learned policy would then be reported as unnecessary when the fleet stays connected. Its remaining claim would rest only on the communication-outage condition, where a central solver cannot be consulted. COPA's finding that coordination survives with communication cut to 13%–15% of steps suggests a coached policy might degrade gracefully there, but that result comes from a toy fire grid, so the claim is open.
Why not just run Mehrab's ranking more often? The daily-ranking baseline tests this directly. We expect it to fall short for a structural reason. It scores cells independently, so after a van drops out it cannot shift a neighboring van to cover the orphaned footprint. Each remaining van keeps its own top cell. If daily ranking nearly matches the joint planners, coordination between vans matters less than we think. We would report that as a finding.
Why not just forecast demand better and plug the forecast into any planner? That is half of our hypothesis, not an alternative to it. Ada-MSTNet's 15.77%–22.78% RMSE gain shows layered forecasts are more accurate. What is unknown is whether accuracy translates into equitable coverage once disruptions occur. The layer ablation answers this by holding the planner fixed and varying only what demand it sees.
Why not just allocate with the access-gap optimizer from Modeling Access Differences? That method decides how much supply each zone gets, once and for an infinite horizon. It says nothing about where vans sit hour by hour or how to recover from a breakdown. We reuse its $\eta$ weighting inside every planner, and we compare against its static allocation executed without re-planning.
Proposed Experiments
The decisive check is a disruption-response experiment run in all three settings. We randomly remove 10% to 40% of vans at random times and measure each group's coverage fraction before and after. The main quantity is the ratio of the worst-off group's coverage loss to the average group's loss. Our hypothesis predicts this ratio stays well above one for aggregate planners and falls toward one for layered planners. If layered and aggregate planners show the same ratio, the central claim is wrong.
A retrospective check in Virginia asks whether layered planners recommend sites closer to the deployed sites that achieved high dose counts, and specifically closer to those serving minority Block Groups. This is the only test against real outcomes. Its weakness is that deployed sites were chosen partly by Mehrab's own method, so we report it as consistency evidence rather than proof.
The outage experiment cuts contact with the central planner for windows of 2 to 24 simulated hours. It compares the coached learned policy against the optimizer executing its last plan. Across all experiments we report total uptake alongside the disparity measures and sweep $\lambda$ to trace the volume-versus-parity frontier. This makes visible how much throughput any equity gain costs, which the throughput-maximizing approach of Mehrab et al. and the parity results of the access-gap paper leave unreconciled.
References
- Data-Driven Real-Time Strategic Placement of Mobile Vaccine Distribution Sites
- Community-Aware Multi-Task Transportation Demand Prediction
- Modeling Access Differences to Reduce Disparity in Resource Allocation
- Income disparities in driving distance to health care infrastructure in the United States: a geographic information systems analysis
- Deconstructing laws of accessibility and facility distribution in cities
- Autonomous Mobile Clinics: Empowering Affordable Anywhere Anytime Healthcare Access
- Deep Learning Empowers Lung Cancer Screening Based on Mobile Low-Dose Computed Tomography in Resource-Constrained Sites
- MamaCare AI: A GeoAI Framework for Maternal Healthcare Accessibility Analysis in Underserved Communities
- Randomized Entity-wise Factorization for Multi-Agent Reinforcement Learning
- Coach-Player Multi-agent Reinforcement Learning for Dynamic Team Composition
- Mutual-Information Regularized Multi-Agent Policy Iteration
- VAST: Value Function Factorization with Variable Agent Sub-Teams