A minimal NLP interface to MOI?

[jd-foster]

This is more a question than an issue...
There was a discussion with @mlubin at the NLopt repo around a designing an simpler API for nonlinear optimization based on MOI here: jump-dev/NLopt.jl#128 (comment)

How much interest do you think there would be for contributing/developing this? I can imagine there is a reasonably common use case where a non-expert user of optimization has a list of functions/vectors that form their objective and constraints, and they 'just' want to call a Julia function that 'solves' the problem using an MOI wrapped solver, without translating the problem directly into say JuMP. Mostly curious for perspectives; this functionality may also encourage solver/algorithm developers to hook into MOI.

[mohamed82008]

Check out https://github.com/mohamed82008/Nonconvex.jl.

[odow]

How much interest do you think there would be for contributing/developing this?

I'm sure there would be interest. But this seems like a case for a package other than MathOptInterface.

Nonconvex.jl is an excellent option.

Other options are:

• https://github.com/SciML/GalacticOptim.jl
• https://github.com/JuliaSmoothOptimizers/NLPModels.jl

Presumably one could also write an Optim interface to Ipopt.

There will be a section on nonlinear programming at the upcoming JuMP-dev. It should be a good starting point to get a lot of people interested in similar things together.

[jd-foster]

Great links, thanks fellow Oceania optimisers. 😀

How "automatic" is the Nonconvex.jl link to MathOptInterface, in the sense that a new wrapped solver for MathOptInterface can be dropped in to a solve using Nonconvex?

[odow]

How "automatic" is the Nonconvex.jl link to MathOptInterface

That might be something better discussed in the Nonconvex.jl repo.

I'm going to close this issue as out-of-scope for MathOptInterface. We should focus effort on this front into existing packages.

[jd-foster]

No worries, I'll have a closer look at the other repos. Thanks again.