Phase 4c: one track, a tracker that listens to a network's range
· 3 min read
Phase 4b's hybrid passed G5, but it was two estimators stitched together: T2's range beside the IMM's velocity. Phase 4c keeps one filter. The frozen IMM takes the frozen network T2's range as an extra measurement, weighted by T2's own stated uncertainty and inflated four-fold, once a minute. Nothing was trained.
It passed G5 on a fresh 40,000-scenario test split (seed 214), decided by the agent review panel:
| Fused tracker / IMM, 97.5% interval | Limit | Result |
|---|---|---|
| Range, all gate rows | < 1 | 0.82 [0.81, 0.84] |
| Course, all gate rows | < 1 | 0.95 [0.94, 0.96] |
| Range, settled rows | < 1.1 | 0.86 [0.84, 0.89] |
| Course, settled rows | < 1.1 | 0.95 [0.93, 0.97] |
| Speed, settled rows | < 1.1 | 0.90 [0.87, 0.92] |
Records: the gate record and the panel review.
What improved on phase 4b
- One state. Range, course and speed come from one filter. Projected forward, its bearing stays as close to the truth as the IMM's own (1.02° after 5 min against 1.06°), where the hybrid's drifted to 1.49°.
- Settled rows improved because of estimation, not copying. Settled range, course and speed are all better than the IMM's and the hybrid's. The hybrid's settled velocity was mostly the IMM's.
- The hybrid's worst case is gone. Before any target manoeuvre the hybrid's range was 1.43 × the IMM's. The fused tracker's is 1.03 × [0.99, 1.07]: parity.
What it does not show
- It does not beat the IMM before a target manoeuvres. On validation it seemed to (0.94 ×), but on the fresh seed it only matches it. The gains come after manoeuvres: 0.74 × the IMM's range error there.
- The EKF and batch least squares are still better before any manoeuvre. Against the EKF it would fail settled course (1.41 ×). Overall it beats the EKF by a wide margin (range 0.80 ×, course 0.78 ×).
- The hybrid is still better on overall course (fused 1.04 × the hybrid).
- The settings were chosen in-sample, from 18 variants on validation. The decision record says 12; the panel found the error after scoring, and it is corrected around the frozen record. Only the fresh seed decides, and it did.
- Its uncertainty is not calibrated. Like the IMM's, its covariance is too wide on typical rows and too narrow in a small tail.
- It is not a learned tracker. It is a filter that listens to a frozen network. Phases 4b and 4c use the same T2 and IMM, so they are not independent replications. Both show that T2's range helps this IMM. Neither shows that a network alone tracks better.
The long view
After G0 to G5 and three phase-4 attempts:
- What networks add. A learned range estimate is genuinely useful to a tuned tracker, mostly after the target manoeuvres. Folded into the filter as a measurement, it gives one consistent track that beats the tuned IMM on every G5 check.
- What they do not do. On clean, straight-leg geometry the classical tools (an EKF, a batch fit) remain better.
Real or recorded data are still the untested step.