NonlinearSolve extension ? #2
Description
Awesome package by the way, very versatile !
The ECA example shown in the README is great,
but my model is much more expensive than gaussians, so it took too long (about 3 min and 20_000 function evaluations).
With NonlinearSolve.jl's NonlinearLeastSquaresProblem + LevenbergMarquardt solver
it was much faster (with relaxed tolerances, which I did not yet tried with ECA, but I'd expect a local algorithm to be faster),
although not nearly as robust of course (ECA is great for that).
EDIT: the ECA() results were already satisfying when forced to stop after 2 s, like that:
sol = solve(optimization_problem, ECA(), optimization_model;
maxtime = 2.0, # stop after maxtime s
)
The NonlinearSolve approach was found on discourse and is based on simple residuals rather than distributions/likelihood.
Would a PR for an extension be welcome ?
Or would you prefer to see how the code would look applied to the README example first ?