Some wireless networks have no fixed infrastructure and no reliable end-to-end path: phones carried by people, vehicles passing each other, devices in a disaster zone. In these networks, messages are passed hand to hand whenever two devices meet. The best protocols for this setting decide who should carry a message by looking at where each device has been and whom it has met. They then bet that this history predicts where the device will go next. These protocols are almost always tested in simulations lasting hours, on synthetic movement models whose statistics never change. Clinical research has a name for this failure: limited follow-up duration. A short trial can confirm an early effect while staying blind to decay, relapse, and whether the early measurement predicts the outcome that matters.
This note proposes to treat a routing protocol the way a trialist treats an intervention, following it long enough to see whether its early advantage lasts. Using a real 90-day city-scale mobility dataset that includes a 15-day emergency, we ask two questions. First, do protocol rankings measured over a short window predict rankings over weeks? Second, does the history these protocols depend on go stale, and can a simple aging rule repair it? The contribution is partly a measurement and partly a method. If short-horizon results turn out to be poor predictors of long-horizon results, much of the published comparison in this area rests on a weak proxy.
Core Claim
Geographic delay-tolerant routing uses node location plus stored movement history to pick relays in networks that are only intermittently connected. Protocols of this kind carry a hidden assumption: the encounter history they accumulate stays predictive. The clearest example is TBHGR in A Geographic Multicopy Routing Scheme for DTNs With Heterogeneous Mobility. It scores relays by their past encounter sets and by how often they visited the destination's area. It then switches from spreading multiple copies to single-copy forwarding near the destination. The paper shows TBHGR beating B-SaW, EBSR and GeoSpray on delivery ratio, latency and overhead. It does so only on short ONE-simulator runs with synthetic mobility, and it assumes no bandwidth or buffer contention.
Our claim is that on real mobility that drifts over weeks, three things happen. The advantage of history-heavy relay selection shrinks as history ages. It can reverse after an abrupt regime change such as an emergency. Short-window delivery ratio is a poor stand-in for long-horizon delivery ratio. The clinical parallel is direct. Sohal et al. found that six-month remodeling measures fail to predict long-term mortality after cardiac resynchronization. ADL+ found that several cognitive gains waned between the end of a 6-month intervention and 9-month follow-up. Both papers motivate treating the evaluation horizon as an experimental variable rather than a fixed convenience. We borrow the method, not the medicine.
Research Program
We will build a trace-driven evaluation from YJMob100K. The dataset contains 75 business-as-usual days and 15 emergency days for 100,000 people, recorded on 500 m grid cells in 30-minute bins. Its home locations correlate with census data at 0.781 at 1 km, so the spatial structure is realistic.
Two people are counted as "in contact" when they occupy the same cell in the same bin. This is far coarser than radio range, and it will overstate contacts. We therefore thin contacts: each co-location becomes a real contact only with probability $p$, swept from 0.01 to 0.5. We also cap each contact's transfer capacity, which relaxes TBHGR's no-contention assumption. The resulting contact traces are loaded into the ONE simulator, so every protocol runs on the same code base that TBHGR's authors used.
Each message has a destination node. Its "destination area" is that node's most-visited cells over the preceding window, which is how area-based protocols define it. Messages are generated at a constant rate across all 90 days, with a fixed time-to-live (the deadline after which an undelivered message is dropped). Each simulation continues for one full time-to-live after message generation stops, so no message is cut off before its fate is known.
The method component is deliberately simple. We give TBHGR-style relay scoring an aging rule for its stored history, at three levels of simplicity. The first is a sliding window of length $W$. The second is exponential forgetting, where an encounter or visit observed $\Delta t$ ago gets weight $e^{-\Delta t/\tau}$, with $\tau$ a time constant. The third is a change-triggered reset: history is discarded when a node's recent visit distribution departs from its stored one by more than a threshold in total-variation distance.
The analysis borrows from longitudinal epidemiology. We report delivery as a time-to-event curve per protocol, and we estimate "surrogate validity" in two ways. The first is the rank correlation (Kendall's $\tau_K$) between protocol rankings computed on the first $h$ days and rankings computed on the full 75 days, for $h$ from 1 to 14. The second is the change in each protocol's delivery ratio across the boundary between business-as-usual days and emergency days. Cheng et al. warn about time-related biases in observational studies. Here they constrain the design: a protocol must be scored only on messages it could have handled at the time, and history-building warm-up days must be excluded from its score.
Baselines
The naive reference point is history-free geographic forwarding. We will use GeoSpray, as run by Cao et al. against TBHGR. It spreads a fixed number of copies and hands them toward nodes currently closer to the destination area, with no memory. If aging history matters, this baseline is what history-based methods should converge toward once their history becomes useless. It is run on the identical contact traces, with the same copy budget, time-to-live and buffer caps.
Epidemic replication means copying every message to everyone met. It serves as a delivery-ratio ceiling and an overhead worst case, not as a competitor.
The strongest existing method is TBHGR itself, run with its full, never-forgotten history. That is effectively how short simulations use it. A Geographic Multicopy Routing Scheme for DTNs With Heterogeneous Mobility reports it outperforming B-SaW, EBSR and GeoSpray in the ONE simulator. However, there is no reported numerical result on Lacuna for TBHGR's delivery ratio, latency or overhead; only relative rankings are given. We will therefore first reproduce the relative ordering on a stationary synthetic run. This checks that our implementation is faithful before we move to real traces.
All protocols get the same tuning budget. That means a grid over copy count and time-to-live, selected on days 1–30, with the aging parameters $W$ and $\tau$ tuned on the same days only. Beating TBHGR means this: aged-history TBHGR achieves a higher delivery ratio at no greater overhead ratio on days 31–75 and on the emergency days, across at least three thinning levels $p$, with the difference outside seed-to-seed variation over ten seeds.
The key ablation isolates the aging ingredient. It compares frozen history (collected during a warm-up and never updated), full cumulative history, sliding window, exponential forgetting, and change-triggered reset, with everything else held fixed.
Why not just…?
Why not just run the existing synthetic simulations for longer? Longer runs of a stationary mobility model cannot reveal drift, because nothing drifts. The follow-up problem in clinical work is not calendar time as such. It is that the system changes over that time. A long run of a model whose statistics never change would only shrink confidence intervals around the same short-horizon answer. We include exactly this control: a 90-day stationary synthetic run. We predict its surrogate-validity correlation stays near 1, while the real trace's correlation drops.
Why not just add exponential forgetting and call the problem solved? This is the serious threat, and we treat it as an outcome rather than an objection. If a single tuned $\tau$ closes the gap on both routine and emergency days, the methodological contribution stands: short-horizon evaluation hid a real weakness. The algorithmic contribution then reduces to "use decay," and we will say so plainly. The change-triggered reset only earns its place if it beats tuned decay during the emergency period, where a smooth forgetting rate should adapt too slowly.
Why not rely on cellular or satellite backhaul and skip delay-tolerant routing entirely? In routine urban life that is often right, and we make no deployment claim for everyday settings. The regime where infrastructure-free relaying matters is disruption. YJMob100K's emergency period is included precisely to capture it. Navigo found that up to 40% of requests in its vehicular simulation could be served by nearby opportunistic carriers. We report emergency-period results separately so a reader can judge the regime of interest on its own.
Why not correct for truncated messages with a drain period instead of survival analysis? With a finite time-to-live, a drain period of one time-to-live does remove censoring, and we use it. The time-to-event framing is kept for a different reason. It lets us compare the whole shape of each protocol's delivery-over-time curve at a fixed horizon, not a single ratio.
Proposed Experiments
The first experiment calibrates the setup. It reproduces TBHGR's relative ordering over GeoSpray on a short stationary ONE run. The decisive experiment is the surrogate-validity curve. We will plot $\tau_K$ between the $h$-day and 75-day protocol rankings against $h$, on real traces and on the stationary control, at each thinning level $p$. A low $\tau_K$ at the 1–3 day horizons typical of published simulations would confirm the core claim. A value near 1 would refute it, and would itself be a useful negative result.
The second experiment follows each protocol's delivery ratio week by week. It asks whether full-history TBHGR's lead over GeoSpray shrinks with the age of its history, and whether it reverses on emergency days. The third experiment is the aging ablation, scored on days 31–75 and on the emergency window. It shows whether decay suffices or a change-triggered reset is needed.
All results will be reported across the contact-thinning and capacity sweeps. This ensures that no conclusion depends on the coarse grid being mistaken for radio range.
References
- A Geographic Multicopy Routing Scheme for DTNs With Heterogeneous Mobility
- A prospective evaluation of cardiovascular magnetic resonance measures of dyssynchrony in the prediction of response to cardiac resynchronization therapy
- ADL+: A Digital Toolkit for Multidomain Cognitive, Physical, and Nutritional Interventions to Prevent Cognitive Decline in Community-Dwelling Older Adults
- YJMob100K: City-scale and longitudinal dataset of anonymized human mobility trajectories
- Reply: Time‐Related Biases in Observational Studies of <scp>GLP</scp> ‐1 Agonists and Cancer Risk
- Navigo: Interest forwarding by geolocations in vehicular Named Data Networking