search

SciML / DiffEqParamEstim.jl

Easy scientific machine learning (SciML) parameter estimation with pre-built loss functions
https://docs.sciml.ai/DiffEqParamEstim/stable/
Other
60 stars 34 forks source link

SDE-specific estimation methods #41

Open ChrisRackauckas opened 7 years ago

ChrisRackauckas commented 7 years ago

Here's an example from @mschauer

Chris Rackauckas [1:45 PM] 
If I'm getting it correct
[1:46] 
at a high level it's like
[1:46] 
1) Get an SDE from the user, but build a slightly transformed SDE using the data
2) Solve that transformed SDE using whatever discretization
3) The stuff with Girsonov's theorem etc. says how to change the parameters and resolve.

Is that right?

Moritz Schauer [1:49 PM] 
yes, add this into the overarching gibbs sampler
- parameters given full path and prior
- "slightly transformed" full path given data (plus accept/reject step with girsanov)

For now, method of moments and Monte Carlo likelihood estimations will do fine, but this may be more efficient. This is a Bayesian method so it may need to go to the other repo, but keeping it here for reference for now.

ChrisRackauckas commented 7 years ago

@mschauer has a goldmine of papers from which we can pull from:

http://pub.math.leidenuniv.nl/~schauermr/

ChrisRackauckas commented 6 years ago

https://www.tandfonline.com/doi/full/10.1080/17442508.2017.1381097