Open yangtzech opened 3 years ago
Hi yangtzech,
sorry for the delay. This code was implemented when I was studying kalman filters more or less 5 yeas ago. At that time I was working with speech processing and I tried to use DEKF for speech filtering and also source separation. The results were not so good and we chose to continue using others techinques.
Well, I told you all this story because I really don't remember why I used the feedforward neural net. When I review the source code this morning, I saw that some parts was based on https://github.com/ziyangli/ekf_tools/blob/master/matlab/ekf/nnekf.m. But I'm not sure what was modified. Maybe looking at ziyang github helps you.
I hope it helps you! Eduardo
Thanks for your help and the explanation. I'll go check it first.
Hi jcsilva,
Sorry to bother, but thanks for your code. And I learned a lot from it.
As far as I know, DEKF uses two Kalman filter to estimate parameters and states of dynamics systems simultaneously. And I think u replaced the dynamic system's model with a two-layer FNN. So, there is no need to calculate dF/dx and dF/dw with some complicated matrix calculation concerning pri/pos value of x/w.
And I still have some questions. Am I right about it? And is it possible u can tell me where u get to learn this, or it's just your brilliant idea?
Yours sincerely, yangtzech