Possible Bug with Maximization in Lower Level

[LukasBarner]

Hi all,
great package, it really makes life a lot easier!
I noticed some strange behavior when the lower level problem is a maximization. See attached a MWE illustrating this, it is based on: Siddiqui S, Gabriel SA (2013). An sos1-based approach for solving mpecs with a natural gas market applica-
tion. Networks and Spatial Economics 13(2):205–227.
Since, in the original publication there are two lower level players, I reformulated both lower level problems into a single one (with equivalent KKTs).
When minimizing the negative of the lower objective, results are as expected. However running the same code but switching to maximizing the actual objective gives wrong results.

The MWE is:

using JuMP, BilevelJuMP, Ipopt

F = [1,2]
c = Dict(1=>1, 2=>1)
C = 1
a = 13
b = 1

model = BilevelModel(Ipopt.Optimizer, mode = BilevelJuMP.ProductMode())

@variable(Lower(model), q[F] >= 0)
@variable(Upper(model), Q >= 0)

@objective(Upper(model), Max, ((a-b * (q[1] + q[2] + Q)) * Q - C*Q) )

@objective(Lower(model), Min, -((a-b * (q[1] + q[2] + Q)) * q[1] - C*q[1] + (a-b * (q[1] + q[2] + Q)) * q[2] - C*q[2] + b*q[1]*q[2]) )

optimize!(model)

@assert isapprox(value.(Q), 6; atol=1e-6)
@assert isapprox(value.(q).data, [2,2]; atol=1e-6) 

Edit: In an earlier version, there has been a copy/paste issue, this is now corrected

[joaquimg]

Can you try with a MIP solver and a MIP mode (sos, fortuny amat, indicator)?
Ipopt is a local solver só it might reach suboptimal solutions

[LukasBarner]

I can also reproduce this behavior with Gurobi and SOS1Mode().
See the code below. Indeed assertion fails for the lower level maximization problem, but works for a minimization of the negative objective.

using JuMP, BilevelJuMP, Gurobi

F = [1,2]
c = Dict(1=>1, 2=>1)
C = 1
a = 13
b = 1

model = BilevelModel(Gurobi.Optimizer, mode = BilevelJuMP.SOS1Mode())

@variable(Lower(model), q[F] >= 0)
@variable(Upper(model), Q >= 0)

@objective(Upper(model), Max, ((a-b * (q[1] + q[2] + Q)) * Q - C*Q) )

@objective(Lower(model), Max, ((a-b * (q[1] + q[2] + Q)) * q[1] - C*q[1] + (a-b * (q[1] + q[2] + Q)) * q[2] - C*q[2] + b*q[1]*q[2]) )
set_optimizer_attribute(model, "NonConvex", 2)

optimize!(model)

@assert isapprox(value.(Q), 6; atol=1e-3)
@assert isapprox(value.(q).data, [2,2]; atol=1e-3) 

[LukasBarner]

Managed to take a look at this now. For both Problems, the KKTs stated in BilevelJuMP are actually different.
I think this is related to a bug in Dualization.jl: jump-dev/Dualization.jl#142

[joaquimg]

Thanks! I will take a look.