As I found out in my previous question Where to get transtion matrix for Kalman filter?, I need a model for correct usage of Kalman filter.

In paper On the Intrinsic Relationship Between the Least Mean Square and Kalman Filters there is model-less Kalman filter - The Kalman filter for deterministic states:

\begin{align} \textbf{g}(k) &= \textbf{P}(k)\textbf{x}(k) / (\textbf{x}^T(k)\textbf{P}(k)\textbf{x}(k) + r)\\ y(k) &= \textbf{x}^T(k)\textbf{w}(k)\\ e(k) &= d(k) - y(k)\\ \textbf{w}(k) &= \textbf{w}(k-1) + \textbf{g}(k)e(k)\\ \textbf{P}(k) &= \textbf{P}(k-1) - \textbf{g}(k)\textbf{x}(k)\textbf{P}(k-1) \end{align}

where $\textbf{g}(k)$ is gain vector $\textbf{P}(k)$ is covariance matrix, $e(k)$ is error, $\textbf{w}(k)$ adaptive filter weights and $\textbf{x}(k)$ are inputs (not really a states here), $d(k)$ is target, $y(k)$ is filter output.

Nice thing is that $\textbf{x}^T(k)\textbf{P}(k)\textbf{x}(k)$ is scalar, so no matrix inversion is needed.

  • What is difference of this in comparison with RLS filter?
  • What should be better?
  • Is it case specific or is there some obvious difference what I am missing?
  • Is it even a good idea to use this Kalman filter over RLS or NLMS?


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Browse other questions tagged or ask your own question.