Skip to main content

Hybrid estimator (classical with a network fallback)

Source: packages/classical/src/hybrid.ts (hybridSwitch), CLIs tma-hybrid and tma-g3 (packages/eval). Requirements: SRD §5 (G3), Annex A lever B.2.6, spec 009.

Why a hybrid​

After G2, the models were 2.1–3.5 × the least-squares solver on range, course and speed. They made fewer gross errors early in the track. Spec 009's research looked at how the two could be combined:

  • Starting the solver from the network changes almost nothing. It reproduces classical in all but 0.13% of rows.
  • Classical's gross errors are not convergence failures. They are the least-squares optimum itself lying far from the truth. On seed 205's validation split, the true state had the lower cost in 1 of 2,379 gross rows.
  • The network is still right there. Its range was within 50% of the truth in 91% of those rows.

So the hybrid keeps classical's estimate and uses the network only where the two disagree.

Rule​

At each timestep after the zig, with classical row c and network row n (both causal):

  • Classical failed: use n if c did not converge or is implausible (classical then reports converged = false).
  • Large disagreement: use n if |r_c − r_n| / r_n > 0.5, where 0.5 is SWITCH_RANGE_REL, the same bound as a gross error.
  • Otherwise: use c.

The network is the frozen iteration-4 instantaneous LSTM. The rule has no fitted parameters.

What it claims​

It is a hybrid result, not "a network alone":

  • its accuracy is classical's;
  • its contribution is a lower gross-error rate where least squares lands far from the truth.

G3 (SRD §5) is judged for it against classical by tma-g3, on the σ = 1° slice, with paired bootstrap 95% intervals and FR-015:

  • the median and P90 ratios of range and course error at zig + 300 s, and of speed error at zig + 600 s, against 1.2;
  • the gross-error-rate difference at zig + 300 s, against 0.

The tracker hybrid (phase 4b, spec 013)​

Source: trackerHybrid in packages/classical/src/hybrid.ts; CLI tma-hybrid-track. Requirements: SRD Annex C §C.10 (owner decision 30).

Phase 4's recurrent tracker T2 beats the tuned IMM on range overall and on velocity after a manoeuvre, but loses on velocity on straight segments (and on range before any manoeuvre). The tracker hybrid keeps each one's strength:

  • Range is always T2's.
  • Course and speed are T2's only where the two estimators' courses differ by more than θ = 45°; elsewhere they are the IMM's. The IMM lags a manoeuvre, so a large disagreement means a manoeuvre is in progress or recent.
  • Causal. The rule reads only both estimators' outputs at the same row.
  • θ was chosen on seed-211 validation by a criterion declared in advance (the largest smallest margin over G5's five checks) and frozen; it is part of the estimator name (hybrid-<network>-<tracker>-c45).
  • Uncertainty. The hybrid keeps T2's stated range sd; its velocity sd is reported only where T2's velocity is used.

On validation (in-sample) it is 0.85 × the IMM on overall range and 0.90 × on course, and 0.97–0.99 × on settled range, course and speed (spec 013 R2). It passed G5 on fresh seed 213 (experiments/20261101-p4h-g5/).

Provenance (schema 0.17.0). A network's prediction file names the SHA-256 of the weights it was made with (model_weights_sha256), and the hybrid's file names the SHA-256 of both input files (inputs). On a test split, tma-g5 refuses a carried file unless it traces to a --model's weights, directly or through inputs that are, byte for byte, files given to the gate.

Not one track state. Range (with position) comes from T2 and velocity mostly from the IMM, so the bearing rate they imply is neither estimator's. On seed 213's G5 rows (bearing-rate.json, from scripts/research/spec013-bearing-rate.mjs), the error of the implied bearing rate against truth is a median of 0.24°/min for the hybrid, against 0.20 for the IMM and 0.23 for T2. On settled rows it is 0.13, against 0.10 and 0.14. Propagated 5 min at constant velocity, the hybrid's settled bearing is off by a median of 1.5°, against 1.0° for the IMM, 0.9° for the EKF and 1.6° for T2. Mixing the two estimators adds little inconsistency beyond T2's own. On settled rows, though, its state is less consistent with the bearings than either tracker's.