search

SciML / DataDrivenDiffEq.jl

Data driven modeling and automated discovery of dynamical systems for the SciML Scientific Machine Learning organization
https://docs.sciml.ai/DataDrivenDiffEq/stable/
MIT License
403 stars 57 forks source link

Measurement timepoints of different states not aligned? #312

Open thomvet opened 2 years ago

thomvet commented 2 years ago

This might be more of a question/feature request than an issue (depending on the answer): Is it currently possible to infer systems for which the measurement timepoints are different for different states? Say, I am measuring x1 at t = 0, 3, 5, 7, 9 and x2 at t = 0, 2, 4, 6, 8, 10.

If yes, how? I am happy to create a simplistic tutorial for the docs as payment for meaningful pointers! ;)

(my real data is not that regularly spaced and I don't have the same number of measurements for my different states - one of them is more costly/time-consuming to measure, so gets measured less often)

AlCap23 commented 2 years ago

You're right. Currently this is not supported.

I assume your states are entangled in such a way that separate discovery is not possible. Otherwise you could divide the recovery process up to du = f(u) and dv = (v), u and v being the states.

If you can use an interpolation, try to provide the missing measurement points in advance. This would be similar to the example from the latest iteration of the universal differential equations paper. Otherwise this resembles a different optimization problem with additional variables for the missing states.