Patrick Kofod Mogensen

Testing against SciPy to assess relative performance

1

Shameless repost of https://github.com/JuliaOpt/Optim.jl/issues/298 @ChrisRackauckas wrote: *Are there any tests against the SciPy LM implementation to see if this implementation is as robust as possible (assuming the SciPy implementation is...

Add tools to extract x, F, and J from the trace.

1

Fixes #246 by adding `x_trace`, `F_trace`and `J_trace`. `extended_trace` has to be set to store arrays and not just norms.

Variants of Anderson Acceleration

3

see https://github.com/JuliaNLSolvers/NLsolve.jl/issues/128#issuecomment-522333787 https://stanford.edu/~boyd/papers/pdf/scs_2.0_v_global.pdf

Types as API for choosing methods

7

I do think that we need to adapt this here, just as we did in Optim. Is anyone strongly against?

Make a better interface to construct the merit function (use `only_fg!` for example)

https://github.com/JuliaNLSolvers/Optim.jl/issues/802

TODO: New LineSearches

1

ref changes in https://github.com/JuliaNLSolvers/LineSearches.jl/pull/80 - [ ] Don't allocate vectors. We might as well have them as scalars to avoid allocations - [x] We should also free the deepest layer...

Uniformly check conditions

5

like ```julia if vecnorm(s) == 0 Base.error("Step direction is zero.") end ```

Move some removed tests

1

- [ ] Counter: Find a situation where all the (non-trivial) linesearches require at least two evals (e.g. Himmelblau above?) - [ ] Optim usage: This should be covered by...