RoadRunner

<small><i>Beep Beep!</i></small>

RoadRunner dense-S MTP — a negative result, published as one

12,986-parameter learned A* heuristic + multi-token (multi-hop) draft heads for Western Australia road routing — with the full measured frontier, including everything that didn't work. If you came for "NN beats A*": it doesn't, and this card shows exactly where and why.

Bottom line

Locked objective was path ≤ 1.01× optimal and faster than vanilla single-query. That cell is still empty after dense/regime × fwd/bidir × w-sweep × admissibility-tune. Banked wins: 1.5× amortized throughput (many queries/target), 2.1× fewer expansions via guided-bidir.

Files

• roadrunner-dense-s-mtp.onnx (52KB, FP32, opset 18, dynamic batch) — inputs (xc[B,17] float32, cls[B] int64) → (r[B], mtp[B,3,3], q[B,16]). h = euc + r*(euc+300), euc_s = hav_km/110*3600. FP32, not BF16: trained BF16-autocast, but ORT CPU BF16 is spotty and at 13KB precision is not the bottleneck. ORT↔torch parity ~5e-5.
• samples.npz (256 val positions: feats, torch refs, truth next/hops, 1-hop + capped-8 2-hop neighbourhood, projection weights) + meta.json.
• demo_mtp.py — numpy+ORT only. Replays pointer + MTP drafts on the bundle.

Quickstart

pip install onnxruntime numpy
python demo_mtp.py   # needs only the .onnx + samples.npz in this folder

How MTP works here (and its honest limits)

Heads k=2,3,4 predict polar (sin,cos,logdist) of nodes 2–4 ahead, trained discriminatively (CE over the base node's real neighbours by polar distance — not global regression). Measured on the bundle: pointer acc 0.80, k2 snap-accept 137/256 forced vs 114/206 chained, mean accept 1.44 teacher-forced (repo sim_mtp.py).

Two findings worth knowing before you build on this:

1. The heads are distance-uncalibrated by design. CE only needs ranking, so logdist parks at ~11 (≈90,000 km vs ~0.5 km truth). All signal is in the angle (30° mean err vs 90° chance); the snap decode ranks by angular alignment. Do not read the vectors as displacements.
2. Teacher-forced accept ≠ rollout stability. 0.75 pointer acc compounds (0.75^174 ≈ 0): greedy rollout reached the target on 0/195 val ODs (repo rollout.py). MTP's remaining form is subgoal jumps with exact fill, not free rollout.

Reproduce everything

Code + docs repo runs the full chain (parse → simplify → engine checks → 100k-OD teacher → train → val): see ARCHITECTURE.md there. Key scripts: data/export_onnx.py (this bundle), data/rollout.py, data/bench_bidir.py, src/train.py --freeze-base (MTP-head-only training that kept the crown).

Provenance / license note

Weights + demo + card in this repo: MIT. Road topology: Geofabrik WA OSM extract (ODbL — attribution retained above; derived graph/teacher data inherits share-alike, see source repo). Attribute side-data: Main Roads WA (CC-BY-4.0). Model code: see repo license.