Comments

[dpo]

I haven't yet tried to run your code but here are a few comments and suggestions from a first reading:

• obj() allocates a new vector every time for the constraint values; why not preallocate a vector of size ncon inside the AugLagModel and use cons!() instead?
• obj() and grad() both evaluate the constraints, so you probably want to implement objgrad!() and save one constraint evaluation;
• we should probably implement memoization so grad() does not reevaluate the constraints where obj() just evaluated them;
• hess_structure!() doesn't look right; the J'J term could fill in the Hessian substantially. It's not hard to figure out the sparsity pattern of J'J from that of J;
• hess_coord!() doesn't look right either. Also it should call hess_coord!() and jac_coord() on the original nlp;
• there is no need to implement grad(), hess(), hess_structure(), hess_coord() and hprod() because you already implement the ! version of those methods;
• there is no need to implement hess_op() and hess_op!() because you already implement hprod!().

Note that because you use TRON as subproblem solver, you don't need hess_structure!() or hess_coord!(). You could leave those out for now.

[Egmara]

• Unnecessary functions have been removed;

• here replaces cons function #15 I used the pre-allocation for constraint values in AugLagModel and cons!() instead cons() in obj(), grad!() and hprod!() functions. But, I dont implement memoization yet;

• The objgrad!() function for AugLagModel is not required in the inner solver (tron), do I need to implement this?

[dpo]

Is it feasible to modify TRON to use objgrad()? It seems like we would save quite a bit here.

[abelsiqueira]

I've created JuliaSmoothOptimizers/JSOSolvers.jl#30 to handle that.