do sensitivities make sense in a tree?

[exaexa]

(might be better as dense vectors)

[exaexa]

this ain't super good but let's keep it open for comments. You can e.g. nicely get the image of sensitivity to stuff by looking at scalar products, like this:

st.parameters.something' * st.fluxes.someflux

[stelmo]

ah yeah, this should end up as some sort of nested tree: e.g.

sens_tree = (
  flux1 = (
    flux1 = 1,
    flux2 = 0.2,
    ...
  ),
  flux2 = (
    flux1 = -0.2,
    flux2 = 1.0,
    ...
)

[stelmo]

hmm maybe a matrix does actually make more sense, especially for gradient descent purposes... but there must be a better way of keeping track of which sensitivity a number in the matrix corresponds to?

[exaexa]

hmm maybe a matrix does actually make more sense, especially for gradient descent purposes... but there must be a better way of keeping track of which sensitivity a number in the matrix corresponds to?

the GD is doing precisely this:

parameters.XXX +=
  (sens_tree.parameters.XXX' * sens_tree.fluxes.someflux) *
  (measured.fluxes.someflux - solution.fluxes.someflux)

if you get the measured stuff in a tree, you can do this via C.zip.

(EDIT: as a bonus you can descend also on "indirect" variables for free, such as the total enzyme capacity etc)

[exaexa]

Design: The sensitivity tree would be better as dense, but pruned to just contain the interesting stuff (measured values etc).

[exaexa]

This now works nicely:

julia> sensitivity_tree.fluxes.TALA' * sensitivity_tree.parameters.PGM
0.007331400905132218

julia> sensitivity_tree.isozyme_forward_amounts.PGL.isozyme_1' * sensitivity_tree.parameters.GND
0.008737521507653744

we might actually want to soak the sensitivity tree creation into the package?