simultaneously evaluated objective and constraint

[tpapp]

Thanks for this great package, I find it very useful for structural estimation in macroeconomics.

A lot of problems I work with calculate the objective and the constraint simultaneously. Eg (stylized)

1. given a set of structural parameters θ and prices p,
2. agents solve their optimization problem to obtain an policy k = f(θ, p),
3. mass distributions d = g(θ, p, k) are calculated, which is the most costly step, d is large
4. the objective is a moment distance m(θ, p, d), while the constraint (market clearing) is c(θ, p, d) = 0.

Eg typically θ and p have a few ten-hundred elements, d thousands or millions.

Currently the way NLPModel and ADNLPModel are set up, f and g is evaluated twice. My understanding of augmented Lagrangian methods suggest that for an inner objective, they objective and the constrained are combined, so a function that returns both in a single pass could be used.

Sorry if this can be done in an obvious way with the current setup already, I could not figure it out.

[dpo]

The NLPModels API makes provisions for an objcons() method but it isn’t used in the current implementation of Percival. We’ll have to think how to work it in.

Sorry if this isn’t a permanent solution but for the time being, perhaps a simple caching mechanism could save you some computations?!

[tpapp]

Where is objcons documented?

[tmigot]

Hi @tpapp
objcons as well as all the functions used to access the NLPModel in general, are defined in NLPModels.jl. So, you can do:

using NLPModels
? objcons

or visit directly NLPModels.jl's doc