Skip to main content

Naive baseline

Source: packages/classical/src/naive.ts. Requirements: SRD §5 baseline 1, spec 001 FR-013.

The naive baseline ignores the bearings entirely and predicts training-set means, so any model that cannot beat it has learned nothing about geometry.

Fit (train split only)​

For each training label (the last sample of each scenario for final; every sample after zig_index for per_timestep):

  • mean of ln(range / 1 km)
  • mean course vector (mean sin θ, mean cos θ)
  • mean speed

Predict​

Every scenario receives the same estimate: range exp(mean ln r) km, course atan2(mean sin, mean cos), mean speed. The reported position uses the last valid bearing up to that sample, so position error reflects range error only.

Because target courses are drawn uniformly, the mean course vector is short and the predicted course is essentially arbitrary: the naive course error is ≈ 90° median by construction.

Reproducibility​

The fit is a deterministic pass over the training file; G0 fits it twice independently and requires identical metrics (SC-005).