Ensemble transform Kalman inversion #320
Conversation
|
The initial implementation seems to be working, but still needs tests, documentation, and examples. Note that the ETKI does not work with localization. The ETKI makes use of the inverse observation error covariance. Right now, the observation error covariance is inverted at every update time, but it may be worth storing in the process struct instead. |
|
Any developments here? It would be good maybe to just have the source code and unit tests in this PR, we can make another for examples & docs if you like |
There was a problem hiding this comment.
Hey! Thanks a lot for this!
- In what you have implemented I only tiny comments here - it all looks good otherwise!
- codecov has an issue, did you miss some testing out?
- One final test suggestion: could we perform one update with an increasingly high-dimensional output sized output space? e.g.
10^2, 10^3, 10^4, 10^5and time it? You could either just println the times, or you could compare loosely to theoretical scalings.
It would be good to have something like this to warn us in case future features corrupt the scalings for this method.
Once this is implemented, could you add the times to the PR description and I will be happy for merge!
329: Ensemble tranform Kalman inversion r=eviatarbach a=eviatarbach I somehow made some mistake with the branches and closed the last PR, #320, and couldn't reopen it. This should be ready to merge. Closes #324. ### Scaling test timings For Julia v1 tests outputs: ``` ETKI with 10 observations took 0.013366613000000005 seconds. ETKI with 100 observations took 0.026825424999999983 seconds. ETKI with 1000 observations took 0.17781986400000005 seconds. ETKI with 10000 observations took 1.249208766 seconds. ``` Co-authored-by: Eviatar Bach <[email protected]>
Closes #324