How does Optimization.jl relate to JuMP.jl

[dehann]

Hi, perhaps a silly question, but I was not able to find a direct answer. The JuMP.jl package provides a common interface to many of the same optimizers listed in the documentation for Optimization.jl. Sure there is a good description of the complementary aspects between JuMP.jl and Optimization.jl but I was not able to easily find it?

Thought it might have something to do with discrete vs. continuous, but as far I know JuMP covers both of those pretty well. JuMP being domain specific might be more tailored for operations research, is that perhaps part of the difference?

PS, JuMP replaced previous JuliaOpt

[ChrisRackauckas]

A nice overview is https://discourse.julialang.org/t/ann-upcoming-refactoring-of-jumps-nonlinear-api/83052/9 .

JuMP focuses on a DSL. That is both good and bad. It's good because there's a lot of extra optimizations that JuMP can do, if a code fits into its DSL. We in theory can get there with better versions of Enzyme, but we're not there yet and it's inevitably harder to do in our case because we allow for any general Julia code. But, the JuMP DSL is very limiting. For one, it's hard to mix it with very general Julia code, so something like an imperative while loop just won't fit into it. Another thing is that its format requires splatting arguments, i.e. you cannot define your own operators with f(x::Array), so you have to do f(x,y,z,...) (which is what is meant by https://discourse.julialang.org/t/ann-upcoming-refactoring-of-jumps-nonlinear-api/83052/5?u=chrisrackauckas). For some use cases this is fine, but for some others, like Neural ODEs, you really need to use arrays etc.

So the two very much compliment each other. Over time, ModelingToolkit's symbolic support will continue to grow in a way that will cover this semi-DSL space and optimizations more, but it's not there yet, and so JuMP's nonlinear DSL optimizes better for the set of codes which fit into it.

[dehann]

Hi Chris, got it thanks. My internet searches did not find the Discourse posts so hopefully this will help discover them.