Skip to content

ACO routing on satellite constellations — prior art and the deterministic-topology objection

Status: web/abstract-level survey, 2026-07-26, with §2 since checked against a full-text source (see §2). Answers the gating question on #202 for the satellite track (#192). Everything outside §2 remains abstract-level — see §8 Limits for what still needs a human pass.

Companion page: network-regimes.md draws the two regimes and tabulates how they differ. Read that first if the question is "why doesn't my MANET intuition transfer?"; read this one for "what has already been published, and what may we claim?"

1. The question this answers

Before spending anything on a satellite substrate (#193) or a satellite module (#195), the track needs an answer to the objection a reviewer raises in one sentence:

LEO topology is deterministic. The inter-satellite-link graph for any second of the next decade follows from the TLEs. AntHocNet exists because MANET topology is unknown and stochastic. Why would you route a predictable network with an algorithm designed for an unpredictable one — when you can precompute shortest paths?

The objection is sound: the research substrate the track was considering (Hypatia) precomputes per-timestep forwarding tables rather than running a routing protocol at all.

Correction, 2026-07-26. An earlier revision of this document also asserted that "Iridium runs virtual-topology-based snapshot routing operationally". That claim came from a web-search snippet and is not supported by the full-text source since checked (Westphal et al. 2023, below), whose only mention of Iridium is historical — the late-1990s/early-2000s proposal wave. The claim may still be true, but it is unattributed and must not be relied on until a source is found. It is withdrawn here rather than quietly softened, because "an operational system does X" is exactly the kind of load-bearing fact a reviewer would check.

2. What the mainstream actually does

The LEO routing literature splits into two families:

§2 is the one section checked against a full-text source: Westphal, Han & Li, LEO Satellite Networking Relaunched: Survey and Current Research Challenges, ITU Journal / arXiv 2310.07646 (2023). Verdicts below are marked accordingly.

Family Mechanism Evidence
Static / topology-exploiting Snapshot routing — the topology is treated as static for a snapshot's duration, routing is computed on it, and a new table is populated as the topology evolves ✅ Confirmed by Westphal et al. §1, which splits satellite routing into exactly two categories, snapshot and dynamic, and notes the tables are typically precomputed and uploaded or stored. Its §4.1 adds that each satellite holds a few snapshots' tables, uploaded periodically from ground stations
Dynamic On-demand and traffic-aware routing; the protocol must adapt as the topology changes, and a snapshot's stable lifetime is short ✅ Confirmed — the other half of the same split
Named taxonomy: virtual topology (VT), virtual node (VN), time-expanded graph (TEG) Finer-grained modelling of the static family ⚠️ Not attributable to Westphal et al. — none of the three terms appears in it. The taxonomy comes from other (abstract-level) sources and may well be right, but this document cannot cite that survey for it

This is the bar. Any claim we make is measured against a precomputed shortest-path control, not against AODV — which is exactly the framing already written into #196 / #216, and this survey confirms it is not a self-imposed handicap but the field's actual convention.

2.1 What protocols actually run on satellites

Checked against the same source (§4.2), because "what is the baseline" is a question about deployed and proposed practice, not only about the ACO corner.

The starting point is the terrestrial suite, and it does not work. Westphal et al. record that TCP/IP, IPv6 for addressing, BGP to share routing information and OSPF to compute routes are commonly used, then state directly that "these are inadequate in the high mobility, highly dynamic environment of satellite constellations" (§4.2). They cite an analysis finding network usability falls below 20% if IP protocols are applied as-is.

So the real landscape is adaptations of the terrestrial stack:

Family Named examples What it does
Snapshot / precomputed virtual topology, per-snapshot Dijkstra Treat the topology as static per snapshot; tables typically precomputed and uploaded from the ground
Link-state, predictability-aware OPSPF (Orbit Prediction Shortest Path First), ASER (Area-based Satellite Routing, hierarchical) OSPF adapted to the orbit's periodicity. OPSPF exists to fix "endless route convergence with OSPF that consumes expensive inter-satellite link bandwidth" — i.e. plain OSPF's reconvergence is the problem being solved
Shortest-path refinements StableRoute Among equal-cost paths, prefer the one requiring fewest route updates — churn, not distance, is the cost being minimised
BGP adaptations — Studied for link intermittency on the space/ground link and its effect on peering sessions; the survey states BGP "needs to be modified as well"
Traffic engineering segment routing (SR), LCRA Zone-based path selection to spread load across the constellation
Disruption-tolerant DTN / Bundle Protocol (IETF WG) From deep-space work; addresses disruption rather than constellation ISL routing
Bio-inspired ACO variants (§3) Load balancing

Not covered by this source, and flagged as such: the access/link layer (DVB-S2/RCS2 for GEO — what SNS3 models; CCSDS for space links; 3GPP 5G NTN integration, which the survey does cite) and transport-layer work (PEPs, TCP variants — though the survey notes routing being optimised to stay compatible with TCP's timer granularity, which inverts the usual layering).

What operators actually run is largely undisclosed. Starlink's inter-satellite routing is proprietary. This document deliberately makes no claim about any operator's deployed protocol — see the withdrawn Iridium assertion in §1 for why.

2.2 Consequence for our baselines

Two things follow, and both change #216:

  1. AODV / OLSR / DSDV are not satellite protocols. Nobody proposes them for a constellation. Comparing AntHocNet only against them on an ISL grid would be a strawman, and a reviewer would say so. They stay useful as a sanity row — "here is what a naive terrestrial protocol does here" — never as the claim.
  2. The realistic incumbent is predictability-aware link state, not just "precomputed shortest path". OPSPF is the sharper control: it is what the field proposes because plain OSPF reconverges badly, so beating a precomputed table is a weaker result than beating OPSPF. A precomputed control remains the right first implementation (it is simpler and bounds the best case), but the ceiling it represents should be named honestly.

3. Prior art: ACO on LEO constellations

The field is substantial and converged. Every ACO-on-LEO paper found frames its contribution as load balancing / congestion, not as topology discovery.

Work What it optimises Baseline(s) Platform
Deng et al., An ACO-Based Routing Algorithm for Load Balancing in LEO Satellite Networks, Wireless Communications and Mobile Computing, 2022 — ACORA-WR Load balancing; ant movement confined to a window; combines path distance, transmission direction, link load LBRA-CP, SPR (shortest-path routing), LCRA not stated in abstract
Zhi et al., Load Balancing Routing Algorithm for LEO Satellite Networks Based on ACO, Internet Technology Letters, 2025 — LBRA-ACO Globally optimal paths + real-time link monitoring + periodic rerouting of degraded paths not stated in abstract not stated in abstract
Routing Optimization of LEO Satellite Network Based on Genetic Ant Colony Algorithm, 2024 GA+ACO hybrid for dynamic load balancing reports gains in optimal-path search rate, RTT, packet loss not stated in abstract
An Ant Colony Based Approach to Multi-Constraint Routing, IEEE, 2023 Multi-constraint (QoS) routing — —
A distributed QoS routing based on ant algorithm for LEO satellite network, J. Electronics (China), 2006 Distributed QoS routing — —
Application of ACO to Adaptive Routing in LEO Telecommunications Satellite Network (early) Adaptive routing compared against link-state algorithms custom C++ + MATLAB frontend
ACO-PSO hybrid, IJSDR, 2025 Optimal path selection — —

Independently corroborated: Westphal et al. (2023) — a survey with no stake in ACO — also references a LEO work applying ant colony optimization as a learning mechanism to optimize end-to-end paths under fragment routing. So the presence of ACO in this literature is not an artifact of searching for it.

Two facts from this table drive everything below.

(a) The congestion framing is the field's answer to the objection — and it is the same answer this repo hypothesised independently. The stated motivation is consistent across the literature: bursty Internet traffic plus distributed satellite links means traffic-intensive regions congest while other links sit idle. Orbital mechanics predict the topology; they do not predict where the traffic goes. That is the gap an adaptive protocol occupies, and it is precisely the mechanism AntHocNet already has (multipath + delay-weighted pheromone + the A2 congestion metric, #55/#67).

(b) Shortest-path controls are already standard here. ACORA-WR is compared against SPR among others. So a submission without that control would be rejected, and one with it is merely meeting the existing bar — not exceeding it.

4. Where the actual gap is

The gap is not algorithmic. It is in evaluation:

  • The ACO-on-LEO work evaluates on NS-2 and on custom C++ simulators with a MATLAB frontend. No reproducible artifact, no shared scenario definition, no common baseline implementation — the results are not comparable to each other, let alone reproducible by a third party.
  • Meanwhile the satellite-simulation infrastructure has moved on: Hypatia, ns3-leo, xeoverse, SNS3 all exist on ns-3, and recent work builds trace-driven emulation on top of Hypatia.
  • No implementation of AntHocNet for satellite topologies on ns-3 was found.

Crucially, ns3-leo states our exact use case as its design goal: a simulation environment where existing routing protocols from the WSN and MANET areas can be evaluated in satellite networks. That is a direct hit on #198 and a strong signal for the substrate decision — and it further weakens the case for building our own module (#195).

Two substrate facts recovered from the Westphal et al. full text, both new to this track:

  • ns3-leo is peer-reviewed, not merely a GitHub repository — Schubert, Wolf & Kulau, ns-3-leo: Evaluation Tool for Satellite Swarm Communication Protocols, IEEE Access 10 (2022), 11527–11537, doi:10.1109/ACCESS.2022.3146770. Worth knowing before betting a substrate decision on it, and citable.
  • A candidate the track had missed: Pedro Silva, Satellite Mobility Model for ns-3 Simulator (https://gitlab.inesctec.pt/pmms/ns3-satellite) — and Westphal et al. state that Hypatia is built on it. That makes it the layer beneath the option #197 evaluates: potentially the mobility half of a substrate without Hypatia's precomputed-forwarding baggage, and a plausible ready-made answer to the roll-your-own floor in #200.

5. The defensible claim

There is one, and it is not "ACO is novel for LEO". That space is crowded and converged; proposing another ACO variant would be the eighth such paper and would have to beat tuned, published load-balancing algorithms on their own ground.

The defensible claim is the repo's existing positioning (#31, #109) applied to a new domain: an open, tested, simulator-agnostic implementation of a canonical ACO routing protocol, evaluated on a current, reproducible ns-3 satellite substrate against a precomputed shortest-path control — the comparison the existing ACO-on-LEO literature asserts but does not make reproducible. The contribution is "here is what an unmodified, paper-faithful ACO protocol actually does on a constellation, measured against the control that exploits determinism, with an artifact anyone can re-run", not "here is a better algorithm".

That claim is worth making only if the measurement is honest about the likely outcome: on a quiet, fully-predicted constellation the control should win, and the interesting cells are the ones where its assumptions fail.

5.1 The competitor set is wider than the ACO papers

A correction to this document's own earlier framing, from the §2.1 protocol check. §3 established that the ACO-on-LEO field converged on load balancing, and §4 concluded the gap is reproducibility rather than mechanism. But the congestion niche is not contested only by ACO papers:

Segment routing already targets the same problem, is standardised, runs on real hardware, and needs no per-packet stochastic decision — Westphal et al. record SR being used to divide the constellation into zones and spread traffic across them, alongside LCRA as a low-complexity alternative.

That does not overturn §5, but it sharpens what must be argued. Our claim cannot be "adaptive routing addresses congestion that precomputed paths cannot see" — so does SR-based traffic engineering, deployably. It has to be narrower, e.g.:

  • distributed and reactive without a controller — SR zone assignment is computed centrally from a traffic view; ACO's pheromone is local and needs no such view, which matters exactly when the controller's view is stale or unreachable (the disrupted regimes in §6);
  • no separate traffic-engineering plane — the same mechanism that discovers routes also balances them.

Both are testable, and neither is tested by beating a precomputed shortest path control alone. Anyone taking this to publication should expect "why not segment routing?" as a reviewer question and have an answer that is not "we didn't compare".

6. What the benchmark must create

Directly actionable for #216 — these are the regimes where the deterministic control is blind:

  1. Congestion the control cannot see. Asymmetric offered load over paths of equal length; the literature's own motivation (bursty traffic, idle links elsewhere). Requires the non-wifi ILinkState signal (#206) or AntHocNet's congestion mechanism is inert on ISLs.
  2. Failures it did not predict. Unscheduled ISL loss, degraded or jammed satellites — as distinct from the scheduled topology changes the control already has in its tables.
  3. Endpoint churn. Ground-station handover (#211) forces the control to recompute whole paths; traffic endpoints should therefore be ground stations, not satellites.

A benchmark that cannot produce these cells cannot falsify the claim, which makes building them part of #216 rather than a follow-up.

7. Consequences for the track (see also §8 for what this does not establish)

Ticket Effect of this survey
#192 Track is not closed as a negative result — a defensible claim exists, but it is an evaluation claim, not an algorithmic one
#193 Substrate must be a current, reproducible ns-3 base; that is now a selection criterion, not a nice-to-have
#198 Promoted — ns3-leo's stated goal is literally this use case, and it is peer-reviewed (IEEE Access 2022) rather than only a repository
#200 New candidate: Silva's Satellite Mobility Model for ns-3 — Hypatia is built on it, so it may already be the roll-your-own floor this spike was going to estimate
#197 Assess Hypatia's mobility layer separately from its precomputed forwarding — the former is Silva's model, and may be usable without the latter
#195 Weakened — the infrastructure exists; building our own is harder to justify
#206 Confirmed load-bearing — congestion is the claim, and the signal is wifi-only today
#216 Control is mandatory and matches field convention; §6 defines the cells. Updated by §2.1/§5.1: AODV/OLSR/DSDV are a sanity row, not the claim; OPSPF is the sharper realistic incumbent; and segment routing contests the same congestion niche, so "why not SR?" needs an answer

8. Limits of this survey

Stated plainly so nobody mistakes its reach:

  • Abstract-level, except §2. This began as a web search over titles, abstracts and landing pages. §2 has since been checked against the full text of Westphal et al. (2023), supplied by the maintainer on 2026-07-26 — which is how the Iridium claim was caught and withdrawn. Every other section is still abstract-level. The environment cannot fetch academic PDFs itself (proxy 403), the same constraint that gated the 2007 thesis work (#58/#88).
  • One claim has already been withdrawn on contact with a source (§2, Iridium). Treat that as a base rate, not a one-off: the unchecked sections were written the same way.
  • Baseline columns are largely unfilled, which is the single most important column in §3. Only ACORA-WR's baselines were recoverable from the abstract. Whether the other works use a shortest-path control decides how strong the "not reproducible / not comparable" claim in §4 really is.
  • Simulation platforms are mostly unstated in abstracts. The NS-2 / custom C++ finding comes from two of seven works; treat §4's first bullet as indicative, not established.
  • No claim is made about result quality in any surveyed paper.

Worth a human pass, in priority order

  1. Deng et al. 2022 (ACORA-WR) — the most directly comparable work, and the one confirmed to use an SPR baseline. Full text would settle the bar we must clear and the scenario parameters to mirror.
  2. Zhi et al. 2025 (LBRA-ACO) — the most recent; establishes current state of the art.
  3. ~~The ITU/arXiv survey~~ ✅ obtained and checked, 2026-07-26. Results are folded into §2, §3 and §4 above; it confirmed the snapshot/dynamic split, refuted the Iridium claim, could not support the VT/VN/TEG taxonomy, corroborated ACO's presence in the literature, and supplied the ns3-leo citation plus a substrate candidate the track had missed.

Land sources in the private papers repo under AntHocNet/library/ and grep them with the pdf-extract skill rather than reading them into context. Note that this (public) repo keeps digests and citations only — never extracted text.

Sources