Hessian with user-defined functions in MathOptNLSModel

Specific Domains Optimization (Mathematical)
, ,
1

I am trying to use user-defined functions with the MathOptNLSModel for solving nonlinear least squares problems. I use:

using JuMP                  
using NLPModelsJuMP         
using JSOSolvers             
model = Model()
g(x...) = 10*length(x) + sum(x[i].^2 .- 10*cos.(2π*x[i]) for i in 1:length(x))
register(model, :g, 5, g, autodiff = true)
x₀ = [1.0, 0.1, 0.2, -0.5, 1.0]
@variable(model, x[i=1:5], start = x₀[i])
@NLexpression(model, res[i in 1:5], g(x...))
nls = MathOptNLSModel(model, res, name = "NL")

(This is a simpler example of a more complicated problem, so I’m not looking to simplify anything down here.) This last line nls = MathOptNLSModel(model, res, name = "NL") gives the following stacktrace:

julia> nls = MathOptNLSModel(model, res, name = "NL")
ERROR: LoadError: Unsupported feature Hess
Stacktrace:
 [1] error(s::String)
   @ Base .\error.jl:33
 [2] initialize(d::NLPEvaluator, requested_features::Vector{Symbol})
   @ JuMP C:\Users\~\.julia\packages\JuMP\lnUbA\src\nlp.jl:427
 [3] parser_nonlinear_expression(cmodel::Model, nvar::Int64, F::Vector{NonlinearExpression})
   @ NLPModelsJuMP C:\Users\~\.julia\packages\NLPModelsJuMP\j5VnJ\src\utils.jl:371
 [4] MathOptNLSModel(cmodel::Model, F::Vector{NonlinearExpression}; name::String)
   @ NLPModelsJuMP C:\Users\~\.julia\packages\NLPModelsJuMP\j5VnJ\src\moi_nls_model.jl:29

This shows that the model does not support Hessians. I found similar issues like this online when googling the error message, but none seem to work, e.g. https://github.com/jump-dev/Pavito.jl/pull/12 and Unsupported feature Hess with user-defined functions using JuMP and Alpine, for this specific function.

How can I work around this issue while still supporting user-defined functions? Is there another interface I could use with my model and NLexpression res for solving this problem, if I can’t get around this error?

2

It seems to be the same error as in one of the links, so the problem is most likely with MathOptNLSModel - open an issue on github with them? They seem to check for hessians and hessians are not available:

image
image930×103 5.29 KB

As for solving the problem - I can only think about skipping the reformulation provided by NLPModelsJuMP and writing the sum of squares directly in JuMP. Something like:

@NLexpression(model, Summ, sum(res[i]*res[i] for i=1:5))
@NLobjective(model,Min,Summ)
set_optimizer(model,Ipopt) ### Add using Ipopt before that line - or any other solver that handles nonlinearities
optimize!(model)
2 Likes
3

Thanks @blob .
I posted the issue on their GitHub https://github.com/JuliaSmoothOptimizers/NLPModelsJuMP.jl/issues/94 and the problem is now fixed.