Skip to main content

Phase 4 plan: two learned trackers, judged on manoeuvring targets

· 3 min read

Phases 3 and 3b ended the same way. A network beats a tuned IMM tracker after a target turns, but gives that back on targets that hold course. SRD Annex B left the next step to the owner, who chose the separate project it names: an end-to-end learned tracker.

SRD Annex C fixes the design before any code is written.

Two trackers, one gate​

  • A learned-gain filter (KalmanNet-style). The bearing and motion physics stay explicit, and a recurrent network learns the gain: how fast to adapt when the bearings stop fitting a straight track. The hope is filter accuracy on straight legs and network agility after a manoeuvre.
  • A pure recurrent tracker. The LSTM family we already have, outputting position, velocity and its own uncertainty at every step. It is the simpler bet, carrying the known weakness: on straight-leg speed it has stayed at about 3 × classical for six iterations.

Both are trained from scratch on the same data and judged on the same rows.

A harder problem​

Generator 0.5.0 gives the target 0–3 course changes and, in 30% of scenarios, a speed change. Manoeuvres become the norm, which is where a constant-velocity model should struggle most. The IMM and EKF are re-tuned for it, and the better one is frozen before training.

Gate G5​

A tracker passes if:

  • it is better overall: its median range and course error, over rows at 5, 10, 15 and 20 minutes after the zig, is below the reference's;
  • and it does not lose much once things settle: on rows more than 5 minutes after the last manoeuvre, it is within 1.1 × the reference on range, course and speed.

The second condition stops phase 3's trade, winning after turns by losing on straight legs, from counting as a pass.

The statistics are stricter than before:

  • The bootstrap resamples whole scenarios, because one scenario contributes several rows.
  • Each tracker's intervals are 97.5%, because two trackers get a shot at the gate.

Reading it critically​

  • What follows a failure is not fixed in advance. If G5 fails, the owner decides what follows then. That is weaker than Annex B, where both branches were set before the result. The write-up will say so.
  • The speed bar is demanding. The settled-speed check asks the trackers to come within 10% of a reference that is already about 2 × least squares on straight-leg speed.
  • The learned-gain filter is a risk. Backpropagation through about 300 filter steps may be slow or unstable in tfjs. Its cost is measured before the decision record commits to it.