-
Notifications
You must be signed in to change notification settings - Fork 17
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
EKP for general likelihoods? #262
Comments
I've had some discussion with @eviatarbach about this. The key point is that, in the notation of the paper linked, they view a forward map In our code, we view the EKP forward map If you do so, with perfect observation The only non-default option you therefore require if you wanted to do something like this, is to be able to observe the non-deterministic map perfectly. For this we have a keyword argument
Quick resolutionYou should be able to run the code with the following alteration
Let us know how you get on! Future actions
|
Thanks @odunbar, this is super helpful. I have to deal with count (integer) data (eg number of new infections per day for an infectious disease), that I normally transform via log(x+1). In the standard framework, I get biased parameter estimates (populations are small, so my statistics are far from being Gaussian), though for an initial condition for another scheme, your EKP library is super useful. |
Thanks! I should caveat that EKP still only has good theory for linear & Gaussian problems. That is to say, although our parameter distribution setups and the algorithms do allow for more interesting priors and likelihoods by pushing nonlinearities/non-Gaussianities into the forward map, this can still yield biases, as EKP still only evolve a Gaussian approximation to perform optimization. In practice though it is has shown effective in many cases. PS RE the count data, I am by no means an expert, but we do have some work in a filtering context, where we converted infection test-results into PPV/FOR data for data assimilation (c.f. "Synthetic Data" section). This is a different transformation that puts the data into [0,1], and might be of interest to you, though the boundaries can still prove challenging. |
Awesome! Thanks for the write-up, I'm pleased it worked out for you, it's really great to hear. If you ever work with a model you can't run >10^4 times for EKS, we do also have an ML-based pipeline [Please let me know if you are interested in trying this out too. It should hopefully just require a short script taking the priors and EKI ensemble as input] |
Closing as resolved. |
Is it straightforward to implement the algorithm described in https://www.sciencedirect.com/science/article/pii/S0167715222000967?
The text was updated successfully, but these errors were encountered: