In most of the papers I have read, calculation of weights is done by the following formula:

enter image description here

Where L is the dimensionality of my state and lambda is calculated by:

enter image description here

And the usual values for alpha, beta and kappa are given as 0.001, 2 (for Gaussian distributions) and 0 respectively.

My problem is that this would give a value of Lambda that is very close to - L, which drives the weights to huge negative values. This is affecting my covariance estimation, leading to wildly inaccurate predictions. This seems like a fundamental error in the way the weights are calculated, but I'm not sure if I'm missing some critical intuition here.

  • $\begingroup$ I usually set alpha around .8 $\endgroup$ – holmeski Jun 14 '19 at 18:24
  • $\begingroup$ I believe kappa is usually set to 3-L $\endgroup$ – holmeski Jun 14 '19 at 18:30
  • $\begingroup$ What dynamics are you using? Is there a toy problem you're working with? $\endgroup$ – holmeski Jun 14 '19 at 18:31
  • $\begingroup$ @holmeski, I am trying to predict the future motion of a body using noisy GPS measurements. Is there a basis for setting kappa? A lot of the sources I saw said a good default is 0 $\endgroup$ – saladboy97 Jun 15 '19 at 2:29
  • $\begingroup$ I was looking at the book Kalman Filtering and Neural Networks which cites A New Approach for Filtering Nonlinear Systems $\endgroup$ – holmeski Jun 15 '19 at 12:46

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.