[protocol] decide the reactive route-setup design on measurement: thesis single-route (A) vs source-parameterised band (B) vs the repo bound (C, shipped)

[danieljoppi]

Why this is open

#173 is fixed and shipped (PR #174), but with a bound that has no basis in either primary source. That was a deliberate, time-boxed choice: main was worse than v1.0.0 on large and loaded networks (large-scale 21.7% PDR, NRL 3071) and un-breaking it was separable from the fidelity question. This issue is the fidelity question, and it should be settled by measurement rather than by reading.

The trigger is that the 2007 thesis records the authors abandoning the mechanism #174 patches, for exactly the symptom #173 measured:

"…through experiments we experienced that it is hard to get a good balance between the number of routes that are obtained and the overhead that is created. High levels of overhead were often experienced. Therefore, we decided to restrict reactive route setup to the creation of just one single route, and to rely on proactive route maintenance to obtain multiple routes."
— 2007 thesis, lines 4676-4681

The three arms

	design	how the flood is bounded	source
A	enableMultipath = false: reactive setup lays one route; multipath comes from proactive maintenance	strict (src,seq) duplicate suppression	2007 thesis (the final design) — lines 3773-3778, 3785-3791
B	band-multipath parameterised as the authors ran it: a1 = 0.9 base, plus a2 = 2 for first-hop-disjoint ants	the band itself — at 0.9 it admits only ants better than the incumbent, so it is self-limiting	2007 thesis describing the older version — lines 4650-4671
C	band at factor 1.5 + per-(node, generation) broadcast bound	the repo-invented per-node bound	none — 1.5 is [1] §3.1; the bound is ours

C is what ships today. Note that we do not implement B's a2 first-hop-disjoint rule at all, and our single factor of 1.5 is looser than the authors' base 0.9 — that gap is the root of #173.

Evidence so far

Core testbench, one discovery on an 8-connected grid, reactive ants on the medium:

grid	nodes	A (multipath=off)	B (a1=0.9)	band at 1.5, unbounded	C (shipped)
5×5	25	28	28	420	94
7×7	49	54	54	30,090	236

Delivery was 17/17 in every arm, so none of these bounds costs reach at this scale.

ns-3 discrete taxonomy for C (30165318267), PDR / NRL:

scenario	v1.0.0	broken main	C (shipped)	AODV
paper-base	81.7 / 59.1	82.2 / 84.8	83.4 / 56.6	80.5 / 56.6
sparse-static	83.8 / 48.3	93.0 / 46.5	90.7 / 40.7	81.9 / 42.9
heavy-load	85.0 / 6.1	51.4 / 28.6	80.1 / 8.7	86.6 / 5.5
large-scale	75.8 / 99.9	21.7 / 3071.1	77.9 / 167.8	67.4 / 163.4

heavy-load is the open residual: C sits 4.9 pp below v1.0.0 and 6.5 pp below AODV. It is the scenario most sensitive to control overhead, so it is the one most likely to separate the three arms.

What to do

1. Implement A behind the existing gate (it is just the default flip) and B (needs a1/a2 and the first-hop-disjoint tracking — GenerationTracker already keys per generation, so this is where it goes).
2. Run the discrete taxonomy at runs=5 — not --quick's 2 — for A, B and C on identical seeds. The 2-run numbers above carry pdr_sd up to 6.43, which is wider than several of the differences being argued about.
3. Compare with sweep_summary.py --vs, whose baseline control will confirm the harness held still across arms.
4. Pick the shipped default on that evidence and record it in an ADR, since it supersedes #96.

Also blocked on this

• [protocol][fidelity] Reactive-ant multipath setup (1.5× filter, paper §3.1) — implemented, but regresses PDR in the ns-3 harness #96's evidence is invalid. It chose enableMultipath=true on an A/B showing PDR 92.3 vs 89.4 — measured with reactiveMaxBroadcasts = 2 in force, which truncated the flood at ~5 hops and therefore hid multipath's overhead cost. Whatever this issue concludes, [protocol][fidelity] Reactive-ant multipath setup (1.5× filter, paper §3.1) — implemented, but regresses PDR in the ns-3 harness #96's decision needs re-deriving.
• docs/configuration.md records antAcceptanceFactor = 1.5 as sourced to [1] §3.1. True but incomplete: the thesis supersedes it with the a1/a2 scheme. That row needs the thesis context regardless of which arm wins — the incompleteness is what let [protocol] reactive forward ants have no flood bound in dense graphs — PDR 75.8→21.7, NRL 99.9→3071 on large-scale (#169 regression) #173 hide.
• v1.1.0 should not be tagged mid-decision if the default is about to change again; either tag after this lands, or tag now and treat the outcome as the next minor.

Sources: 2007-ducatelle-thesis.pdf and 2004-ppsn-anthocnet.pdf in the private papers repo, read with the pdf-extract skill; line numbers index its extracted text.

[danieljoppi]

All three arms implemented and dispatched at runs=5

Discrete taxonomy, runs=5, time=300, propagation=range, 3.42-opt, identical seeds across arms:

arm	branch	configuration	run
A — thesis single-route	claude/arm-a-thesis-single-route	enableMultipath=false (strict (src,seq) dedup)	30168645100
C — shipped	main	band 1.5 + per-node bound 2	30168647546
B — source-parameterised band	claude/arm-b-source-band	a1=0.9, a2=2.0 (new first hop), reactiveMaxBroadcasts=-1	dispatched

Arm A is a pure default flip — the behaviour already exists behind the #96 gate — and make test is 22/22 including test_reactive_hop_reach, so single-route setup costs no reach. Arm B needed real work: a two-factor GenerationTracker with per-generation first-hop tracking (bounded at 8 first hops, on top of the existing FIFO cap), the new factor exposed as an ns-3 attribute in both m_config blocks, and a2 given its own field name rather than redefining antAcceptanceFactor — silently repurposing a field's meaning is the #173 mistake.

Arm B is deliberately run with no per-node bound, so the measurement isolates whether the band alone suffices.

Testbench cost, one corner-to-corner discovery on the 8-connected grid

side	nodes	A strict dedup	B a1=0.9/a2=2.0, unbounded	C band 1.5 + bound 2	band 1.5, unbounded
3×3	9	10	16	10	10
5×5	25	28	76	94	420
7×7	49	54	154	236	30,090

The band at the source's own parameterisation is self-limiting — 154 at 7×7 with no broadcast bound at all, ≈3.1n and flat in n, versus 30,090 for the same band at 1.5. That settles the mechanism question raised earlier: #173 was caused by the factor, not by the missing bound.

Two details worth carrying into the comparison:

• At 3×3, arm B costs more than A or C (16 vs 10). a2 deliberately admits disjoint-first-hop ants that a 1.5 band rejects in a tiny dense graph; the saving only appears once there is room to proliferate. So B is not uniformly cheaper — it trades a little early overhead for path diversity, which is exactly what it is for.
• With the per-node bound re-enabled at 2, arm B's grid costs 113 at 7×7 rather than main's 236 — so B and the fix(core): bound the reactive flood per (node, generation), not per ant path #174 bound compose, and that combination is a fourth configuration if the taxonomy makes it look attractive.

One correction to this issue's framing

The audit in #182 found that the thesis's own resolution of the overhead problem was arm A — deleting multipath reactive setup, not retuning its band (lines 4676-4681). So arm A is not merely "one option"; it is what the source actually shipped, and B is the design the source describes itself as having abandoned. That does not decide the measurement, but it sets the burden of proof: B or C should have to beat A on the numbers to justify shipping a superseded design.

Also relevant to interpreting whatever comes back: #179 (β exponents 20× off), #180 (proactive rate 10× slow + missing emission gate) and #181 (SINR metric absent) all affect this comparison, and #180 especially — arm A relies on proactive maintenance to build the path mesh, which is precisely the mechanism #180 shows is running 10× too slowly with no gate. Arm A is therefore being measured with its supporting mechanism handicapped, and a weak result for A should not be read as the thesis design failing until #180 is fixed.

---

Generated by Claude Code