Optimization with IPOPT, "memoization" technique seems to be overridden

[mranneberg]

I have, like probably a lot of real world applications, a function that calculates target and equality constraints and gradients of both in one function.
To make this work with the framework I want to use the "memoization trick", that is e.g. for the target function

# Helper function / memoization
function fgh_Target(I,x0)
    last_x = nothing
    L = 0.0
    dL = 0.0
    ddL = 0.0
    function update(x)
        if x != last_x || true # I need to use the "true" for ipopt
            #@info "Updating $x"
            x0[I] = x
            L,dLF,ddLF = eval_lagrange_design(x0)
            last_x = x
            dL = dLF[I]
            ddL = ddLF[I,I]
        end
        return
    end
    function f(x)
        update(x)
        return L
    end
    function ∇f(x)
        update(x)
        return dL
    end
    function Hf(x)
        update(x)
        return ddL
    end
    return f, ∇f, Hf
end

If I do not set the "if" in update(x)to be always true and call f(x) in julia with different arguments, it correctly updates all values for new x. But if I use it for IPOPT like this

    t, ∇t, ∇²t = fgh_Target(I,x,param,windspeed,altitude,Pnom,init_solution,optimize_plant,opt_weights)
    T = CustomHessianFunction(t, ∇t, ∇²t)
    model = Model(T)

it will not update and I have to set the if to be always true.

Before I go about defining a dummy problem and sharing this, does somebody already know what's going on? Maybe some internal optimization for function handles, or a specific behaviour of CustomHessianFunction?

[mohamed82008]

I will look into it. But would it be easier to define a dummy variable O, make that the objective and then add a new equality constraint O - obj(x) = 0 where obj is the objective function?

[mranneberg]

I will look into it. But would it be easier to define a dummy variable O, make that the objective and then add a new equality constraint O - obj(x) = 0 where obj is the objective function?

Thanks for the reply! Sorry, I was not very clear. The same thing happens with equality constraint functions, using the memoization "trick" fails for some reason when used with the "Custom..." Function in ipopt. Will make a small example next week.

[mranneberg]

Added example in #168

[mranneberg]

Hint:
I tried using IPOPT with NLPModelsIpopt and have exactly the same issue.

[mranneberg]

It works with
if x[:] != last_x[:]
and
last_x .= x

I saw the hints here: jump-dev/Ipopt.jl#34

This is not an issue with nonconvex or nlpmodels, but with the magic of rebinding, overwriting and the obtuse actions that we all do with writing a=b. Closing.

[mohamed82008]

Great to hear that this is resolved. Sorry I haven't been very responsive because I am in the middle of relocating.

[mranneberg]

No worries.
It's still weird that the issue only comes up when used within ipopt and not within Julia though.