← All Cracked Math projects
PM-I07NOT STARTEDINTERMEDIATEMATHPYTHONIMPACT ★★☆☆☆

Kalman filters that pass their own consistency test

Low RMSE is not enough: your filter has to be statistically honest about its own uncertainty.

EST.12 h
REWARD+300 XP
AFTERRoger Labbe + 2

ML-ARENA BENCHMARKS

PROOF: BENCH
METRICINDUSTRYMINE
Position RMSE; fraction of runs whose average NEES (normalised estimation error squared) falls inside the 95% chi-square band
GOAL · RMSE equal to filterpy within 1% on (a). On (b), UKF RMSE <= EKF and NEES consistent in >= 90% of runs. The RTS smoother cuts RMSE by >= 30% vs the forward filter.
filterpy (Labbe) · measured locally—
no results yet
YOUR RESULTS

Your recorded results appear here.

Brief

A Kalman filter tracks a moving object from noisy measurements and reports both an estimate and its uncertainty (a covariance matrix). Implement KF (the linear Kalman filter), EKF (the extended version, which linearises a nonlinear model with Jacobians), UKF (the unscented version, which pushes a few chosen sample points through the nonlinear model instead) and the Rauch-Tung-Striebel smoother (a backward pass that improves past estimates using later data). Apply them to (a) a 2D constant-velocity tracker and (b) bearings-only tracking, where you observe only the direction to the target, not its distance. Tune the noise covariances by maximising the innovation likelihood (the probability of the filter's observed prediction errors under its own model). Verify the filters with NEES/NIS chi-square tests over 500 Monte Carlo runs; these check whether the actual errors are as large as the filter claims, no larger and no smaller.

VICTORY CONDITIONS

  1. RMSE equal to filterpy within 1% on (a). On (b), UKF RMSE <= EKF and NEES consistent in >= 90% of runs. The RTS smoother cuts RMSE by >= 30% vs the forward filter. (Position RMSE; fraction of runs whose average NEES (normalised estimation error squared) falls inside the 95% chi-square band).
  2. A public repository with a README that shows your result and how to reproduce it.

Terms in plain words

Kalman filter
An algorithm that combines a motion model with noisy measurements to give the best estimate of a hidden state and its uncertainty, step by step.
EKF / UKF
Two ways to apply the Kalman filter to nonlinear problems: the EKF linearises with Jacobians, the UKF propagates a small set of sample points.
RTS smoother
A backward pass after filtering that uses future measurements to refine every past estimate.
Innovation
The difference between an actual measurement and the filter's prediction of it.
NEES / NIS
Normalised squared errors of the state and of the innovations; for an honest filter they follow a known chi-square distribution.

Why it rates ★★☆☆☆

Well-trodden ground. The NEES consistency checks are the non-standard, credible part.