Todo list

[mschauer]

Now that the test show that this is all fine 🎉, some next steps

• Example: for example linear and non-linear continuous discrete smoother (van der Pol?) (Moritz)
• Consistency between argument names/order in Mitosis and MitosisStochasticDiffEq (Moritz)
• Merge p and pest? (Moritz)
• Use WGaussian{(:μ, :Σ, :c)} from Mitosis (Moritz)
• Ensemble version (Frank)
• Doc-Strings (Moritz, Frank)
• Implement Proposition 5.6. of https://arxiv.org/pdf/1712.03807.pdf (Frank)
• StaticArrays (Frank)
• StaticArrays for guided solution (Frank)
• Interpolating option for forward guiding (Frank)
• AD for fjac! in Regression (Frank)
• Regression could also accept a SDEProblem, not only a SDEKernel
• Regression could work with in-place versions f!, g!
• For non-centred reparametrizations we need to allow to give a fix Wiener process to the sampler.
• If sigma is invertible, is is convenient to have a function which takes a trajectory and solves for the innovations/driving Wiener process.

[frankschae]

For non-centred reparametrizations we need to allow to give a fix Wiener process to the sampler.

This is just for the forward guiding, isn't it? Thus, by feeding a NoiseProcess from the DiffEqNoiseProcess package to forwardguiding, this is now implemented, too. For example, we can use a NoiseGrid in a MH sampler:

# random walk proposal for innovations
Zᵒ = ρ*Z + sqrt(1 - ρ^2)*randn(length(s))
Z1 = cumsum([0;sqrt(dt)*Zᵒ[1:end-1]])
NG = NoiseGrid(s,[[Z] for Z in Z1])

See:

MitosisStochasticDiffEq.jl/src/guiding.jl

Line 112 in b66dcd6

prob = SDEProblem{inplace}(guided_f, guided_g, u0, get_tspan(trange), p, noise=Z)

and
https://github.com/mschauer/MitosisStochasticDiffEq.jl/blob/main/test/guiding_test.jl#L319-L325