Phase 3b: two hypotheses fix the course, not the speed
Phase 3 found a network that beats a tuned IMM tracker after a target turns, but that averages "turned" and "not turned" on targets that hold course. Phase 3b gave the network a way out. It outputs two estimates and the probability that the target has turned, and reports the more probable estimate.
It half worked. On the same validation split:
| Measure | Phase 3b | Phase 3a | IMM tracker | Limit |
|---|---|---|---|---|
| After a turn: median range error | 10.3% | 10.5% | 13.6% | below the IMM |
| After a turn: median course error | 14.5° | 14.4° | 24.1° | below the IMM |
| Held course: course error | 6.1° | 11.1° | 5.0° | 10° |
| Held course: speed error | 1.15 kt | 1.60 kt | 0.71 kt | 1 kt |
The speed interval, [1.05, 1.25] kt, lies wholly above the limit. The pre-registered rule therefore stopped phase 3b at validation, and its test split was never scored. Under SRD Annex B, this was the last phase-3 attempt.
Records: results and the validation record.
What was tested
Only the network's output changed. The data, the frozen IMM reference and the gate were all as in phase 3.
- The network gives a held-course estimate, a turned estimate, and P(turned).
- The flag's training. The training label is the truth: has the target's course changed yet?
- The forecast in the decision record, written before training, was held-course course about 6° and speed about 0.95 kt (0.8–1.2), with a 40% chance of reaching the test split.
What happened
- Course: fixed. Held-course course error fell from 11.1° to 6.1°, as forecast. It no longer grows with time.
- After a turn: unchanged. The network still beats the IMM by about a quarter on range and 40% on course.
- Speed: not fixed. Speed fell from 1.60 to 1.15 kt, inside the forecast band but above the limit.
The flag explains the speed failure:
- It detects real turns well. The median delay is 70 s, and it misses 0.5%.
- It also false-alarms. On 29–31% of held-course rows five and ten minutes after the zig, it calls "turned". In this data, those minutes are when targets usually start to turn.
- A false alarm is costly. The turned estimate is biased slow, so each one costs about 1.8 kt.
- Even without false alarms, speed sits at the limit. Where the network says "held", its speed error is 1.00 kt, so a perfect flag would leave it there too. The IMM manages 0.71 kt and batch least squares 0.31 kt.
Reading it critically
- Switching moves the failure rather than removing it. An averaged output blurs every estimate a little. A switched output is sharp, but sometimes sharply wrong: held-course course P90 rose from 34° to 51°.
- No threshold tuning. A stricter threshold than 0.5 would cut false alarms but slow detection. Choosing it on this split would make validation a training signal, so it was not done.
- Two changes at once. The output form and its supervision changed together, as declared, so this result cannot separate them.
- The finding is consistent. Two different network designs beat a tuned tracker after a turn, and both give back accuracy on targets that hold course. Neither dominates the tracker, and dominating it is what gate G4 asks.
What next
SRD Annex B §B.5's second branch applies, as it did for phase 3:
- publish the negative result, which this post does;
- the owner then decides separately whether an end-to-end learned tracker is a new project.
Real or recorded data stay deferred either way.