Extended Kalman Filters

Tutorials
1

I am learning about Kalman filters and trying to extend the tutorial a little bit. I have added a few notes and started playing with parameters to try to understand what is going on better. Here is my plot so far:

50%20AM
Screen Shot 2019-08-25 at 10.43.50 AM.png1500×1002 68.3 KB

  1. My first question is, why is the code used to generate the plot not included? Having this code would go a long way in helping beginners play around with the tutorial.

  2. The number of dimensions looks like it is set at 4. I’m not sure why this is because I think we are just plotting using 2 dimensions. The author also claims there are 100 time steps yet it is set at 10.

  3. In my plot, it is hard to tell the difference between noisy signal and the predictions from the filter. As you can see, the measurements and state predictions overlap. I replaced one of the measurements with an outlier: zs[3] = torch.tensor([ 0.0460, 0.033]), yet the filter is unaffected in predicting state. The optimizer also converges in a few iterations. Enabling a bit more noise in the predictions would help show how the Kalman Filter is modeling the measurement uncertainty

  4. In general, in all of these tutorials, I would like to see the tuning knobs explicitly set at the top of the notebook and maybe a short description as to how they effect the sampling distributions or parameters.

2

great feedback! we’ve been meaning to rewrite that tutorial for a while, but still haven’t gotten around to it :slight_smile: to address some of your questions:

  • the plotting code wasnt included because it (and the KF implementation) was repurposed from another library and a bit messy to include. we do try to include the plotting code when possible. in this example, the plotting code can be found in Standard Cognition’s open source example
  • the position vector has dim 2 and velocity vector has dim 2

if you come up with a more compelling example than the simplistic one in the tutorial, we’d welcome a PR!

3

Awesome! I think I was pretty close with my code. Just couldn’t figure out what series to use for ellipses. I will work on the tutorial code some more.