Help with JuMP variables

Specific Domains Optimization (Mathematical)
,
1

I am trying to solve a constrained optimization problem using JuMP, but there is apparently a conflict between how I define the variables and one of the packages I call in the objective function. I just started with JuMP so it’s probably a very minor thing that I do not understand. The following is a MWE to illustrate the issue (the example itself makes no sense, but it is the same error I get in my code):

using QuantEcon  # loads MarkovChain
using GLPK

function f_matrix(p_m)
    MC = MarkovChain(p_m)
    index_1 = stationary_distributions(MC)[1]
    return index_1[1]
end

model = Model(GLPK.Optimizer)
@variable(model, 0 <= p_matrix[1:3,1:3]  <= 1)
@objective(model, Min, f_matrix(p_matrix))

When running @objective(model, Min, f_matrix(p_matrix)) it gives the following error:

ERROR: MethodError: no method matching isless(::VariableRef, ::VariableRef)
2

JuMP is a package for structured optimization and has its strengths and weaknesses. It isn’t well suited for black-box nonlinear optimization problems like you have there.

See, “Should I use JuMP?”.

You may be better off with something like Optim.jl (https://github.com/JuliaNLSolvers/Optim.jl).

Note that JuMP does support user-defined nonlinear functions, Nonlinear Modeling · JuMP. That would look something like the following (I didn’t try, there might be typos, etc. It also needs stationary_distributions to be differentiable using ForwardDiff. Not sure if it is.).

using QuantEcon
using JuMP
import Ipopt

function f_matrix(x...)
    p_m = reshape(collect(x), (3, 3))
    MC = MarkovChain(p_m)
    index_1 = stationary_distributions(MC)[1]
    return index_1[1]
end

model = Model(Ipopt.Optimizer)
@variable(model, 0 <= p_matrix[1:3, 1:3] <= 1)
register(model, :f_matrix, 9, f_matrix; autodiff = true)
@NLobjective(model, Min, f_matrix(p_matrix...))

(Note that you need Ipopt, a nonlinear optimizer, instead of GLPK, which is a linear optimizer.)

I would still start with Optim instead, however.

1 Like