Cardiologists learned an expensive lesson: how narrow an artery looks on a scan is a poor guide to whether it actually starves the heart of blood. Roughly a third of intermediate narrowings are misjudged by eye (Lee et al.). The fix was to measure function under stress instead. The standard measurement is fractional flow reserve (FFR), the pressure ratio across a narrowing while drugs force the heart's peak blood demand. Clinicians also learned to use it as a triage tool. A cheap simulated FFR computed from a CT scan settles most cases. An invasive pressure wire is reserved for the borderline "gray zone," where simulations are least reliable (Ding et al.). Lesions that pass the functional test are left alone.
Proactive network congestion control is still at the "looks narrow on the scan" stage. Current controllers forecast each link's utilization and reroute traffic before a link saturates (Labonne et al.). A link that is forecast to be busy is not necessarily one whose saturation hurts end-to-end service. Its traffic may have slack elsewhere, or demand downstream of it may be small. We propose to carry the clinical decision structure over intact. First, define a per-link functional reserve: the fraction of its flows' forecast peak demand that the network can still deliver when the link is at capacity. Second, compute it cheaply with a learned graph flow solver run on a stress-level forecast. Third, in the gray zone only, confirm it with active probing, the network's pressure wire. Fourth, intervene only on links that fail the test. The claim is that this matches or beats the service quality of utilization-triggered proactive control while making far fewer reconfigurations and sending far less measurement traffic.
Core Claim
The utilization-based triggers used in proactive traffic engineering are the networking analogue of anatomy-guided treatment. They will show a substantial discordance with functional significance, concentrated in an intermediate band of utilization. A function-guided policy with gray-zone triage turns that discordance into fewer unnecessary interventions at equal or better service quality.
The functional reserve of link $e$ is defined as
$$R_e = \frac{\Phi_e(\hat D_q,\, c_e)}{\Phi_e(\hat D_q,\, \infty)}$$
where $\hat D_q$ is the $q$-quantile ("stress") forecast of the traffic matrix, $c_e$ is the link's nominal capacity, and $\Phi_e(\cdot, c)$ is the total demand delivered for flows routed through $e$ when the network re-optimizes routing with $e$'s capacity set to $c$. This mirrors the physiological definition of FFR: maximal flow with the narrowing divided by maximal flow without it. It makes three ingredients explicit that clinicians found essential. Severity is measured under peak demand ($q$ high, analogous to drug-induced peak flow). It depends on downstream demand, the way an identical narrowing matters more in an artery feeding a large muscle mass (Lee et al.). And it accounts for whether alternative paths can absorb the load.
What is new is the decision procedure and its evaluation, not the existence of capacity-aware optimization. That procedure has three parts: a stress-forecast functional index, a gray zone where cheap estimates defer to measurement, and deferral of links that pass. It is evaluated with clinical-style discordance, negative-predictive-value, and deferral-safety analyses that networking work has not applied.
How the Two Directions Are Used
The vascular literature plays two distinct roles here. It motivates the gap: Lee et al. quantify discordance between visual and functional assessment. Ding et al. show a large accuracy jump from anatomical CT reading (AUC 0.64) to simulated FFR (AUC 0.97). Westra et al. show the same pattern for angiography-derived flow ratios (AUC 0.92 versus 0.64 for 2D anatomical measurement). Boerhout et al. show that routine functional testing changes outcomes even when anatomy looks benign. It also supplies method. Graflow is a neural solver that predicts pressures and flows on a vessel graph, trained with mass-conservation losses rather than labels. We reuse its physics-informed loss on a fluid relaxation of the network flow problem to compute $R_e$ quickly. Graflow is static, steady-state, and was tested only on synthetic trees. We apply it to forecast-driven, time-varying demand on communication graphs.
The finding that the cheap index is least reliable in a middle band (Ding et al., worst agreement for simulated FFR between 0.70 and 0.80) constrains the design. It is why we do not trust the surrogate everywhere and spend measurement budget only there. The pure-image deep learning result of Oliveira et al. is a second constraint: 72% accuracy and 48% positive predictive value without physics, but 90% negative predictive value. It suggests that physics-free predictors may be safe for deferral yet poor at deciding when to act. Our evaluation separates these two error types.
The networking side supplies the application regime and the competitors. Labonne et al. forecast single-link utilization 15 seconds ahead with an LSTM and trigger load balancing before saturation. That is precisely the "anatomical" trigger we contest, and its 115-feature telemetry burden is the overhead we aim to cut. NeuRoute pairs a traffic-matrix forecaster with a neural network that imitates an expensive multicommodity routing heuristic on a fixed topology. It re-optimizes everything every interval and does not ask which interventions are worth making. Liaskos et al. use predictive backpressure, which moves traffic down queue-length gradients, with loop filters on Internet backbones. It is a local, queue-driven alternative. Pozanco et al. predict downstream intersection queues from upstream densities in road traffic. They give us a second, non-packet flow domain in which to test generality.
Research Program
We will build three components and one evaluation protocol.
The first component is a stress forecaster. A quantile traffic-matrix predictor, an LSTM like those in Labonne et al. and NeuRoute but trained with a quantile loss, outputs $\hat D_q$ for $q$ in {0.5, 0.9, 0.99}.
The second component is a functional-reserve solver. The reference solver computes $R_e$ exactly from two linear programs per candidate link. This is affordable offline but too slow online. The fast surrogate is a graph neural network over the topology. It takes forecast demands and capacities and is trained with a Graflow-style loss: flow conservation at nodes, capacity limits on edges, and a small supervised term on LP-computed reserves. It outputs all $R_e$ in one forward pass with an uncertainty estimate. Because the conservation loss is topology-agnostic, we test whether it transfers across topologies, which is a stated weakness of NeuRoute's fixed-graph training.
The third component is a triage policy. Links with surrogate $R_e$ clearly above a threshold $\tau$ are deferred. Links clearly below $\tau$ trigger proactive rerouting of their flows. Links within a gray band $[\tau - \delta, \tau + \delta]$, or with high surrogate uncertainty, get an active probe. A short paced test burst along affected paths measures delivered throughput directly, standing in for the pressure wire. The band width $\delta$ sets the trade-off between measurement traffic and decision accuracy and is swept explicitly.
The evaluation protocol borrows clinical methodology. Ground truth for each link and interval is whether leaving the link alone causes a service-level violation in the realized future, determined by counterfactual simulation. Against this ground truth we report the discordance rate between utilization-triggered and function-triggered decisions, broken down by utilization band. We also report sensitivity, specificity, and especially negative predictive value (how often a deferred link turns out to be safe), and a closed-loop comparison.
Baselines
The naive reference point is reactive control. It reroutes only after a link exceeds a utilization threshold in measured telemetry, which is the regime Labonne et al. describe as arriving too late. We also include static shortest-path routing with no intervention. Together these anchor how much proactivity buys at all.
The anatomical proactive baseline is the LSTM utilization-forecast trigger of Labonne et al.. It is retrained on our traffic and topologies with the same forecasting backbone and data as our forecaster, and its trigger threshold is tuned on validation traffic over the same grid size we use to tune $\tau$ and $\delta$. That paper reports an average RMSE of 0.0122 for single-link utilization forecasts on its VirtualBox/iPerf3 testbed. Beating it means fewer service-level-violation minutes, or equal violation minutes with significantly fewer reconfigurations, on held-out traffic.
The strongest existing method is full predictive re-optimization in the style of NeuRoute. The traffic matrix is forecast and the whole routing is re-solved every interval. We run the exact heuristic as an upper bound on routing quality and a NeuRoute-style imitation network for the latency-matched version, both on GÉANT2 with identical forecasts. NeuRoute reports under 0.05% error relative to its heuristic, with inference of 30 ms versus 120 ms for the heuristic. The file has no reported result on Lacuna for service-level-violation rates or reconfiguration counts of NeuRoute on GÉANT2, so we will measure them ourselves. This baseline should do about as well as ours on raw service quality. Beating it means matching its violation minutes within a confidence interval while cutting rule changes and telemetry bytes substantially. If it also wins on violations, our selective approach adds nothing but savings, and we will report that.
For backbone experiments we include predictive backpressure with STITCH loop filtering from Liaskos et al.. It is run on the CAIDA topology with ITMGen traffic as in that paper, which reports roughly 100% average throughput gain over shortest-path routing. For road traffic, the PELEA planner of Pozanco et al. is the baseline. It reports average waiting time falling from 582 s under static control to 121 s on the Houston grid in SUMO.
Three ablations isolate the key ingredients. The first computes $R_e$ from the median forecast ($q=0.5$) instead of the stress forecast, testing whether peak-demand evaluation matters. The second removes the gray-zone probe, trusting the surrogate everywhere. The third replaces $R_e$ with the linear program's shadow price on each link's capacity, a cheap marginal-sensitivity score that needs no new index.
Why not just…?
Why not just re-optimize the whole network every interval on the forecast, as NeuRoute does? This is the most serious threat. A full re-solve implicitly accounts for function, so on service quality it may tie with us. The argument for selectivity is cost, not accuracy. Every re-solve can churn forwarding rules across the network and requires full-matrix telemetry, which Labonne et al. already identify as a burden. The closed-loop comparison reports violation minutes, rule changes per hour, and control-channel bytes side by side. If full re-optimization achieves similar churn because the optimizer is naturally stable, the proposal's practical case collapses to the diagnostic finding alone.
Why not just lower the utilization trigger threshold? A more conservative threshold catches more true bottlenecks but also flags more harmless ones. This is the anatomical-grading trade-off that left anatomical CT reading at 62% specificity in Ding et al.. We sweep the threshold to trace its full trade-off curve between violations and interventions and compare it against ours. If any point on the utilization curve dominates our policy, the functional index is unnecessary.
Why not just use the linear program's shadow prices, or a physics-free graph network that predicts path delay directly? Shadow prices measure marginal sensitivity at the current solution. They can be zero for a link that becomes binding only under the stress forecast, which is exactly the non-marginal, peak-demand behavior $R_e$ captures. The shadow-price ablation tests whether that distinction matters in practice. A physics-free predictor is the analogue of Oliveira et al., whose high negative predictive value but 48% positive predictive value suggests it would defer safely but over-trigger. We train one with identical capacity and data and compare both error types and cross-topology transfer.
Proposed Experiments
The first experiment is a diagnostic study with no control loop. On GÉANT2 (23 nodes, 38 links) and on CAIDA backbone subgraphs with ITMGen traffic, we replay forecast traffic and label each link-interval by counterfactual violation. We then measure how often utilization-forecast triggers and $R_e$ disagree, and where. The decisive check is whether discordance is substantial and concentrated in intermediate utilization. If utilization and $R_e$ agree on nearly every link, the networking gap does not exist in these regimes and we stop.
The second experiment is closed-loop control. All baselines run in packet-level simulation over the same traffic traces, including injected surges and micro-bursts. We report violation minutes, loss, reconfigurations, and measurement traffic, sweeping the gray-band width $\delta$ to trace measurement cost against accuracy. The decisive checks are two. The first is negative predictive value of deferral: it must be high, or the policy is unsafe. The second is whether the gray-zone probe recovers most of the gap between surrogate-only decisions and exact-LP decisions while probing a small fraction of links.
The third experiment tests transfer and generality. We train the surrogate on GÉANT2 and test on CAIDA subgraphs and on GÉANT2 with links removed, against a NeuRoute-style network that must be retrained. We then apply the same index, with delivered vehicles in place of delivered packets, to the SUMO Houston grid against PELEA. This tests whether the function-versus-structure gap appears in a physical flow network as well as a packet network.
References
- Discordance Between Angiographic Assessment and Fractional Flow Reserve or Intravascular Ultrasound in Intermediate Coronary Lesions: A Post-hoc Analysis of the FLAVOUR Trial
- Diagnostic accuracy of noninvasive fractional flow reserve derived from computed tomography angiography in ischemia-specific coronary artery stenosis and indeterminate lesions: results from a multicenter study in China
- Predicting Bandwidth Utilization on Network Links Using Machine Learning
- Diagnostic Performance of In‐Procedure Angiography‐Derived Quantitative Flow Reserve Compared to Pressure‐Derived Fractional Flow Reserve: The FAVOR II Europe‐Japan Study
- Coronary function testing vs angiography alone to guide treatment of angina with non-obstructive coronary arteries: the ILIAS ANOCA trial
- Graflow: Neural Blood Flow Solver for Vascular Graph
- Non-invasive derivation of instantaneous free-wave ratio from invasive coronary angiography using a new deep learning artificial intelligence model and comparison with human operators’ performance
- NeuRoute: Predictive dynamic routing for software-defined networks
- Backpressure on the Backbone: A Lightweight, Non-Intrusive Traffic Engineering Approach
- Urban Traffic Control Assisted by AI Planning and Relational Learning