Glossary

TMA and ML terms as used in this project, alphabetical. Algorithm details are linked where they exist.

Across the line of sight — A zig that puts the target on ownship's other bow, judged from the bearing at the start of the turn. It reverses ownship's velocity component across the line of sight, the change bearings respond to most. It is iteration 3's rule for both turns (zig_rule: across_los; see generator).

Bearing — True direction from ownship to the target, clockwise from north, in radians in [0, 2π). The only measurement the estimators receive about the target.

Bearing error, realistic — Generator 0.6.0's added bearing error (spec 019, SRD Annex C §C.14). It is a correlated Gauss–Markov term (sd 1.0°, τ 30 s) plus a drifting bias (sd 0.5°, τ 20 min), on top of the white noise σ. It is applied to σ > 0 scenarios only. Under it, G5 passed for the fused tracker against the frozen IMM (phase 7), but T2 alone had lower range error.

Bearing-line intercept — c = x_o cos b − y_o sin b for ownship at (x_o, y_o) and bearing b: every point (x, y) on the line of bearing satisfies x cos b − y sin b = c. A derived input channel (in km, scaled). See feature layout.

Bearing rate — How fast the bearing changes (rad/s). The derived feature set estimates it at each row as the least-squares slope of the bearings in the trailing 3 min (≥ 3 valid samples), using no later samples. See feature layout.

Bearings-only TMA — Estimating a target's position and velocity from a time series of bearings, with ownship's own track known. See least-squares solver.

Causal — A model whose output at time t uses only inputs up to t. The instantaneous variant is causal.

CUSUM — Cumulative-sum change detector: it accumulates evidence step by step, never below zero, and alarms when the sum crosses a threshold. Phase 5's baseline is a clipped CUSUM on the reference IMM's innovations (spec 017).

Classical baseline — The batch least-squares bearings-only solver; the yardstick a model must approach for G3. See least-squares solver.

Confidence interval (CI) — The range a metric would plausibly take on another test set of the same size. Here a percentile bootstrap: the test scenarios are resampled with replacement 2,000 times (seeded), and the 2.5th and 97.5th percentiles of the recomputed median are reported. Ratio intervals (model/naive) resample both estimators' scenarios together (paired). See evaluation.

Cramér–Rao bound — The smallest variance any unbiased estimator can achieve from given measurements and noise. For a constant-velocity target it says how well the bearings themselves pin down range, course and speed. At G2 the classical solver's speed error (0.62 kt) is at it.

Control model — A deliberately simpler model trained on the same data as a sequence model (a causal CNN for the instantaneous LSTM, an MLP for the BiLSTM). If the sequence model cannot beat its control, its result says little about learning the geometry. See models.

Dropout — A missing bearing sample (valid = false). Drawn independently per sample.

EKF (extended Kalman filter) — A recursive estimator that predicts the target state under a constant-velocity model and corrects it with each bearing; its process noise σ_a sets how fast it can follow a turn. Reported in phase 3; in Cartesian coordinates it drifts after a target turn (Aidala's instability), so the gated reference is the IMM.

Early stopping — Ending training when validation loss has not improved for patience epochs; the best epoch's weights are kept.

Fused tracker — The reference IMM with a learned tracker's range added as a second, log-range measurement whose sd is the network's own stated sd × k, applied every M samples (spec 014, note). Unlike the tracker hybrid, one state carries range and velocity, though the network's range is correlated with the same bearings. Judged at G5 as a combination of a filter and a frozen network, not as a network alone.

Gate — A pass/fail decision at the end of an iteration (G0–G3), judged on fixed metrics. See evaluation.

Gross error — A scenario whose range error exceeds 50%, or whose estimator did not converge. Reported as a rate because TMA failures are heavy-tailed.

G4 — Phase 3's gate (SRD Annex B §B.4): on manoeuvring scenarios five minutes after the target's turn, the network alone must beat the better of the tuned IMM and EKF on median range and course error, not exceed its gross-error rate, and not regress from G2 on held-course scenarios.

G5 — Phase 4's gate (SRD Annex C §C.5): an end-to-end learned tracker must beat the re-tuned reference on median range and course error over all σ = 1° rows at zig + 5, 10, 15 and 20 min, and stay within 1.1 × the reference on range, course and speed on settled rows. Two trackers are judged, each with 97.5% intervals.

G6 — Phase 5's gate (SRD Annex C §C.12): a manoeuvre detector built from frozen models must detect target course changes faster than a CUSUM on the reference IMM's innovations, at the same false-alarm rate (0.5 per quiet hour). Not reached: phase 5 stopped at validation.

G7 — Phase 6's gate (SRD Annex C §C.13): the fused tracker's post-hoc recalibrated 50% and 90% intervals for range, course and speed must cover within ±2 points of nominal and be narrower than the reference IMM's. Inconclusive on seed 215; passed on the one re-score (seed 216), on the pooled rows only.

Held-out geometry — A test set drawn from parameter ranges outside training (for example 20–30 km initial range), used to measure extrapolation.

Hybrid estimator — Classical's least-squares estimate, replaced by a network's where classical fails or the two ranges differ by more than 50% (spec 009, note). Its accuracy is classical's; it is judged at G3 as a hybrid, not as a network alone.

IMM (interacting multiple model) — A tracker that runs a constant-velocity and a coordinated-turn EKF side by side and mixes them by how well each explains the bearings. Phase 3's gated reference, tuned on validation and frozen (note).

Learned-gain filter — A Kalman-style filter whose motion and bearing models stay explicit but whose gain is computed each step by a recurrent network, trained end to end through the recursion (KalmanNet, Revach et al. 2022). Phase 4's tracker T1 (SRD Annex C §C.4).

Inconclusive — A gate check whose 95% interval contains its threshold: the data cannot say which side the true value is on. A gate with an inconclusive check, and no failing one, is inconclusive; the agent review panel decides it under a pre-registered rule.

Instantaneous variant — The model that gives the operator's running solution: at each time, range, course and speed from bearings so far. Scored as a curve of error against time since the zig.

Leg — A period of constant ownship course and speed. Scenarios have two legs (joined by a rate-limited turn from iteration 2), or three from iteration 3 (a second zig).

Manoeuvre (target) — From phase 3 (generator 0.4.0), one rate-limited target course change of 30–120° in 80% of scenarios; manoeuvre_index is the first sample after it completes, and gate G4 counts time from there, as earlier gates count from zig_index.

Naive baseline — Predicts training-set means and ignores the bearings. Anything that cannot beat it has learned nothing. See naive baseline.

NEES (normalised estimation error squared) — The state error against the truth, weighted by the filter's stated covariance: eᵀP⁻¹e. For a 4-element state with an honest covariance its mean is 4 and 95% of values fall below 9.49. Used to check whether a filter's uncertainty can be taken at face value.

Noise σ — Standard deviation of the Gaussian error added to each bearing (0.5°, 1° or 2°).

Observability — Whether the bearings determine the target state. From a single straight leg they do not determine range; the zig makes the problem observable. Small zigs and long ranges make it weakly observable.

Ownship — The observing platform, whose track is known exactly.

Provenance — The record of what produced a file: a dataset header names the commit of the generator code and whether the working tree was dirty; features files carry that header forward, and training refuses data whose provenance is missing, dirty or from a generator other than the one G0 verified. See generator.

Pseudo-linear measurement — A bearing rewritten as an equation linear in the target's position: x cos b − y sin b = x_o cos b − y_o sin b. With constant target velocity, each bearing gives one linear equation in the target's initial position and velocity, the basis of pseudo-linear TMA estimators and of the bearing-line intercept channel.

Rate-limited turn — Ownship's zig flown as a constant-rate turn at constant speed (iteration 2: 0.5–3°/s, so 90° takes 30–180 s) instead of an instantaneous course change. zig_index is the first sample after the turn completes. See generator.

Scenario — One ownship-versus-target run over a window: two tracks, the bearings and the metadata that generated them. See the data dictionary.

Settled row — A G5 row more than 5 min after the target's last manoeuvre ended, or in a scenario with no manoeuvre: where the constant-velocity model is right again (SRD Annex C §C.5).

Shuffled-label test — A leakage check: the instantaneous model trained with each scenario's targets swapped for another's. It should do no better than the naive baseline; if it does, the inputs leak the answer.

Two-hypothesis output — Phase 3b's network output (spec 011): an estimate assuming the target has held course, one assuming it has turned, and the probability that it has turned. The more probable estimate is reported, so the network does not average the two (as phase 3's single output did). See models.

Time-late variant — The model that reconstructs the target track after the fact, using the whole window (non-causal). It has strictly more information than the instantaneous variant, so it must score at least as well.

Velocity outputs — A model output parameterisation (iteration 4): the network predicts the target's east and north velocity, scaled by the top of the target speed range, and course and speed are derived from them. Earlier models predicted course as sine and cosine and speed on its own. See models.

Window — The span of bearings a model sees: 10–40 minutes at one bearing every 10 s, padded to 240 rows.

Zig — Ownship's course change part-way through the window (30–90° at 40–60% of the window; instantaneous in iterations 0–1, a rate-limited turn from iteration 2). Iteration 3 adds a second zig 3–5 minutes into leg 2, and both turns are across the line of sight.