Can OptimizationMOI use the common maxiters interface?

[SebastianM-C]

Question❓

I was wondering if there's something that prevents OptimizationMOI from using the globally passed maxiters. Currently (Optimization v4.0.3, OptimizationMOI v0.5.0), when I pass that, I get

┌ Warning: common maxiters argument is currently not used by Feasibility
│ 
│ Subject to:
│ 
│ Nonlinear
│ . Set number of iterations via optimizer specific keyword arguments.
└ @ OptimizationMOI ~/.julia/packages/OptimizationMOI/tpWKG/src/OptimizationMOI.jl:87

[Vaibhavdixit02]

The issue is that specific solvers need differently named arguments for this so there isn't an obvious mapping

[SebastianM-C]

I see...
I was wondering where can I look into the details for Ipopt in particular, as I'm using "max_iter" => x for setting the maxiters, but that doesn't correspond to what opt_state.iter considers as iterations. I think that in the optimization state we get some internal iterations too, so I was wondering if I'm using the wrong argument or if there's a bug / different convention used in Optimization.jl

[ChrisRackauckas]

maxiters needs should just do the natural thing with each solver. Of course some solvers will have outer and inner iterations etc. but nobody will look into it. We need to stand by the same principle everywhere: the simple things should work, and the complex things should be possible.

[SebastianM-C]

I was wondering where can I look into the details for Ipopt in particular, as I'm using "max_iter" => x for setting the maxiters, but that doesn't correspond to what opt_state.iter considers as iterations.

I realized that the issue here is that the OptimizationMOI wrapper uses the objective function for the callbacks. In the case of Ipopt at least, we need to use https://github.com/jump-dev/Ipopt.jl?tab=readme-ov-file#solver-specific-callbacks since we would want to get the callback called after a step is accepted and not on every internal iteration. I tried this and it does indeed match with what max_iter is setting. Would it make sense to have a OptimizationIpopt or some more specific handling for Ipopt?

[ChrisRackauckas]

OptimizationIpopt would be a good library. Yes, if there's special things we can do for IPOPT let's do it, it is used often enough that making specific improvements to it is a good idea.