Open bischtob opened 2 years ago
Currently EKI can only minimize nonlinear least squares problems, i.e.,
$\min_\theta \|\mathcal{G}(\theta) - y\|^2$
so it couldn't be added to a general-purpose optimization package.
However, one way to transform the nonlinear least squares problem into a general one would be to take $\mathcal{G}(\theta)\to\exp{\mathcal{G}(\theta)}$ and $y\to 0$. I don't know how well this would work but might be worth trying.
Can we hook EKP.jl into that project? We could then calibrate using DiffEqFlux.jl.