When many radios share a frequency band and nobody coordinates them, each radio raises its transmit power to be heard. That extra power becomes noise for its neighbours, who raise their power in turn. Wireless game theory has studied this "your signal is my noise" spiral for decades and usually scores it by total throughput or total energy. This proposal asks a different question: who loses the spiral when the radios belong to rural clinics. The radios worst placed to win a power race are the ones farthest from a receiving tower. In many regions those are also the clinics farthest from a hospital. If that holds, uncoordinated power competition adds a hidden communication penalty on top of existing geographic health disparity. A network can look efficient in aggregate while systematically failing the patients who most need remote diagnosis.
We propose to measure this effect on real health geography and then remove it. The measurement uses facility and population data from Kebbi State, Nigeria, together with diagnostic payloads from an existing rural screening system. The remedy is a decentralized power-allocation game with two changes. First, each clinic's goal is to deliver a specific diagnostic case before a clinical deadline at minimum battery cost, not to maximize data rate. Second, a lightweight hierarchy decides who yields to whom, ordered by clinical urgency and remoteness. The central test is whether this beats a much simpler fix: rate maximization reweighted toward fairness. If it does not, we will say so.
Core Claim
We make two claims and treat them separately because either could fail on its own.
The first claim is empirical. In decentralized rural uplinks, the Nash equilibrium of selfish power control (the operating point where no radio gains by changing its own power alone) concentrates missed clinical deadlines among the most geographically remote facilities. The failure rate is worse than distance alone would predict. It is amplified by competition from nearer nodes.
The second claim is about design. Most of this amplification can be removed without a central controller and without global channel knowledge. Two things are needed. Each node's utility must reflect what it actually has to send: a roughly 6.5 KB structured finding most of the time, and occasionally a multi-megabyte raw image for specialist escalation. And decoding or transmission priority must be ordered by urgency and remoteness rather than by channel strength.
This scoping matters, and one finding constrains it directly. On the Base Station Selection and Base Station Sharing in Self-Configuring Networks analyzes decentralized base-station association as a potential game. It reports a Price of Anarchy (the ratio of the coordinated optimum to the selfish outcome) "remarkably close to 1." Selfish play may therefore cost little in total efficiency. Our claim does not contest that. It is about how the loss is distributed across geography, which an aggregate ratio near one can hide entirely.
We also restrict the setting deliberately. Licensed cellular uplinks already run closed-loop power control from the base station, so they are not a non-cooperative game. The regime here is the one where no operator coordinates the radios. That includes community long-range Wi-Fi or TV-band backhaul, mesh relays between health posts, and mobile clinic units sharing unlicensed spectrum.
Where the Pieces Come From
The health side supplies the motivation and the data. MamaCare AI and its companion challenge paper GeoAI for Maternal Healthcare Accessibility map maternal-care deserts in Kebbi State. They use 1,207 GRID3 facilities and WorldPop population grids, and find that 62% of women of reproductive age live more than 5 km from a facility. They then plan mobile-clinic phases and SMS-based triage. Both treat the communication link as always available. We reuse their geography, but put an actual shared radio channel underneath it.
A Cloud-Edge System for Multimodal Clinical Screening in Resource-Constrained Rural Settings runs 17 diagnostic models locally. It sends only about 6.5 KB of structured evidence to a cloud language model, against 2–7 MB for raw uploads. It reports stable 25–38 second end-to-end latency on simulated single links from 500 kbps to 5 Mbps. It supplies our traffic model. We also use it as a constraint, since it implies that ordinary payloads are small. Its links are isolated pipes with no competing users, and that is exactly the assumption we relax.
The wireless side supplies the methods. Each of the following papers optimizes a symmetric, utilitarian objective with no notion of whose packet matters more.
- Learning Non-myopic Power Allocation in Constrained Scenarios provides NMPA. A graph neural network trained with reinforcement learning scales the output of an unrolled per-step power solver so that finite batteries last a whole episode. We reuse its two-level architecture almost directly.
- Introducing hierarchy in energy games shows that adding hierarchy improves energy efficiency for leader and followers alike. Hierarchy here means either a Stackelberg leader that moves first or successive interference cancellation, in which the receiver decodes users one at a time and subtracts each before the next. It also shows that the decoding order determines who benefits. That order is the knob we reassign to clinical criteria.
- Conjectural Equilibrium in Water-filling Games shows how a node can estimate locally how interference responds to its own power. It does this with a linear belief fitted by small power perturbations, which avoids the need for global channel information.
- Dynamic Power Allocation Games in Parallel Multiple Access Channels gives convergence guarantees under single-user decoding.
The hierarchy paper also flags a caveat we inherit: successive interference cancellation breaks exact potential structure, so those guarantees do not carry over.
Research Program
The geographic interference testbed. We place Kebbi State GRID3 facilities and planned mobile-clinic routes on terrain and the OpenStreetMap road network. We compute path loss to candidate receiving points and attach WorldPop female-population counts to each facility's catchment area.
The file gives no tower-location dataset. Receiver placement is therefore a stated assumption, varied across three scenarios: towers at LGA (local government area) headquarters, towers at secondary hospitals, and a sparse random placement. Results must hold across all three to count.
The workload. Cases arrive at each node as a stochastic stream. Payload sizes and deadlines come from the 100-case multimodal benchmark of the cloud-edge paper. Most cases send the 6.5 KB structured finding. A tunable fraction escalate to raw-image uploads, and each case carries an urgency class with a deadline.
The utility. Node $i$ chooses a transmit power $p_i$. Its utility is
$$u_i(p_i, p_{-i}) = w_i \Pr\left[T_i(p_i, p_{-i}) \le D_i\right] - c_i(b_i) p_i,$$
where:
- $T_i$ is the time to deliver the node's current payload, given the signal-to-interference-plus-noise ratio produced by everyone's powers $p_{-i}$;
- $D_i$ is the clinical deadline;
- $w_i$ is an urgency weight;
- $c_i(b_i)$ is an energy price that rises as the remaining battery $b_i$ falls. Remaining battery is replenished by solar charging during the day.
This objective has a ceiling that rate objectives lack. Once a node will deliver on time with high probability, extra power buys nothing. That ceiling is the mechanism we expect to stop the power spiral.
The hierarchy and the learner. Priority among nodes contending for the same receiver comes from a public score combining urgency with travel time to the nearest hospital. Under successive interference cancellation, this score sets the decoding order. Without it, it sets who acts as the Stackelberg leader.
Each node learns its power with an NMPA-style policy. The neural network's inputs are extended with urgency, deadline slack, battery state and the conjectural interference estimate. Its action space is extended with one binary choice: send the structured finding now, or attempt the raw escalation.
We will prove what we can, including existence of equilibria and convergence in the single-user-decoding special case. Where potential structure breaks, we will report convergence empirically and not claim a guarantee.
Baselines
The naive reference is fixed equal power at the regulatory maximum with no adaptation. It shows the raw effect of distance with no competition dynamics, so it isolates how much of the remote-clinic deficit is pure geography.
Selfish SINR-target power control. Each node independently raises power until it reaches a target signal-to-interference-plus-noise ratio, in the style of A General Theory for SIR Balancing. This is the escalating regime whose effect on disparity we want to measure.
NMPA unmodified. This is the strongest existing learned method for battery-constrained interference networks, and the closest in mechanism to ours. Its reported result is a 15–20% improvement in cumulative episodic sum-rate over myopic allocation, with near-zero battery-constraint violations (Learning Non-myopic Power Allocation in Constrained Scenarios). That number comes from 10 transceiver pairs in a synthetic 60×60 area. We will first reproduce it in that setting to validate our implementation, and only then run it on the Kebbi testbed.
NMPA with alpha-fair rewards. This variant keeps NMPA's architecture but replaces sum-rate with a fairness-weighted rate (alpha = 1 for proportional fairness, alpha = 2, and the max-min limit). It is the most dangerous competitor, because it targets disparity directly with no clinical modelling. We found no reported result on Lacuna for alpha-fair power control in this or any comparable setting.
Centralized optimum. We compute a centralized optimum of our own objective as an upper bound.
Fair comparison. All learned methods share:
- the same geography, channel draws and case streams;
- the same network size and training steps;
- an equal hyperparameter search budget of 30 trials per method.
What counts as beating the baselines. Beating alpha-fair NMPA means a statistically significant gain on the primary metric: the population-weighted fraction of urgent cases delivered on time in the most remote quintile of facilities. The gain must hold across all three tower scenarios, with no worse than 5% loss in total on-time delivery and no increase in energy per delivered case.
Ablations. Two ablations isolate the key ingredients. The first keeps the deadline utility and removes the urgency–remoteness hierarchy. The second keeps the hierarchy and returns to a rate utility.
Why not just…?
Why not just compress everything at the edge, since 6.5 KB barely needs a channel? This is a real threat, and it is the first thing we test. If almost every case is a 6.5 KB finding and deadlines are minutes, contention may never matter. The cloud-edge paper's 25–38 second latency is likely dominated by language-model reasoning rather than transmission.
We will sweep two quantities: the escalation fraction, from 0% to 20% raw uploads, and node density. If the remote-quintile deficit stays negligible across this range at selfish equilibrium, the first claim is false. We would then report edge compression as sufficient.
Two arguments suggest it will not be sufficient. First, under interference the binding constraint at the cell edge is whether the signal clears the decoding threshold at all, not payload size. A node that cannot be decoded cannot send 6.5 KB either. Second, battery cost from repeated failed attempts accumulates regardless of payload size.
Why not just reweight NMPA with a standard fairness objective? Alpha-fair rates equalize throughput, not deadlines. Throughput is the wrong currency when payloads are small and bursty and urgency varies across cases. Alpha-fairness has no way to prefer an urgent obstetric case at a mid-distance clinic over a routine upload at a remote one.
This is still an empirical question, and the alpha-fair baseline and the hierarchy ablation exist to answer it. If alpha-fair NMPA matches us on the remote-quintile metric, we will conclude that the clinical utility is unnecessary and that the contribution reduces to the measurement study.
Why not just let the base station or a central server schedule everyone? Where an operator controls the uplink, that is the right answer, and our setting excludes it. The open question is whether decentralized networks can approach the centralized result. The gap to the centralized optimum is reported for every method.
Coordination overhead also matters. Classical hierarchical solutions assume the leader knows every channel gain, which Introducing hierarchy in energy games does not avoid. Our nodes use only a public priority score and local conjectures, so we also report signalling bytes per case.
Decisive Checks
Precondition. Before any learning, we compute the Spearman correlation between each facility's travel time to the nearest hospital and its path loss to the nearest receiver, under each tower scenario. If the correlation is weak, distance to towers and distance to hospitals are decoupled. The "compounding disparity" story would then collapse, and we would report that instead.
First claim. At selfish equilibrium we measure the remote-to-central ratio of missed urgent deadlines. We compare it with the same ratio under fixed equal power. Amplification means the equilibrium ratio significantly exceeds the fixed-power ratio. This must hold across density and escalation sweeps, with confidence intervals over 50 channel and arrival seeds.
Second claim. The test is the comparison against alpha-fair NMPA and the two ablations described above. Each result is translated into population terms using WorldPop catchments, as the number of women whose urgent case would reach a remote specialist in time.
Generalization. We train on two of the three tower scenarios and test on the third, and report whether the learned policy still beats the baselines. This guards against a policy that has memorized one map.
References
- On the base station selection and base station sharing in self-configuring networks
- MamaCare AI: A GeoAI Framework for Maternal Healthcare Accessibility Analysis in Underserved Communities
- A Cloud-Edge System for Multimodal Clinical Screening in Resource-Constrained Rural Settings
- Learning Non-myopic Power Allocation in Constrained Scenarios
- Introducing Hierarchy in Energy Games
- Conjectural Equilibrium in Water-Filling Games
- Dynamic power allocation games in parallel multiple access channels
- A General Theory for SIR Balancing