Variable index allocation problem for silent variables about gams.jl HOT 9 CLOSED

Any solution is correct. 0 is just the least worst option, and most often the one the user intended.

from gams.jl.

I still doubt that 0 is "the correct answer", but

Getting NaN solutions is going to cause no-end of issues.

is a very good point. I will change the behaviour to return 0 with the next release. Thanks a lot for pointing this out!

from gams.jl.

Thanks for your quick responses and for fixing the bug.

from gams.jl.

Just chiming in here: I don't know about the GAMS issue, but please talk to me if you have development plans for SDDiP odow/SDDP.jl#246!

from gams.jl.

Thanks very much for reporting! I can confirm that the used GAMS savepoint option only includes the symbols that are present in the model, which results in the incorrect behaviour reported by you.

from gams.jl.

This has been fixed with GAMS.jl v0.1.6. The output of the above program will be:

8-element Array{Float64,1}:
 NaN
 NaN
   1.0
   0.0
   0.0
   0.0
   0.0
 NaN

I thought about using missing instead of NaN, but that resulted in errors with MathOptInterface. missing would have been nice, but NaN works as well. Thanks again for reporting the issue!

from gams.jl.

Isn't the correct answer 0, not NaN?

from gams.jl.

Good question. I have thought about this for a while and I think that missing or NaN is best here as it indicates the value could have been set to anything since it wasn't part of the optimization. If I set it to something, e.g. 0, this is unclear to the user. But I am happy to hear your thoughts about it. Is there any convention in MathOptInterface to set the value to 0 in this case?

from gams.jl.

I think the convention in most (every?) solver is that unspecified variables default to 0. That's what pure-Gurobi does, so it makes sense that is what GAMS-Gurobi should do:

using JuMP, Gurobi
approx_model = Model(Gurobi.Optimizer)
@variables(approx_model, begin
    θ <= 100
    x[1:8]
end)
@objective(approx_model, Max, θ)
@constraint(approx_model, θ - 0.65 * x[3] + x[4] - x[5] + x[6] + x[7] <= 36)
optimize!(approx_model)
julia> value.(x)
8-element Array{Float64,1}:
  0.0
  0.0
 98.46153846153845
  0.0
  0.0
  0.0
  0.0
  0.0

For SDDP, we take the solution to the Lagrangian dual, and use the optimal x as an input to a different LP. Getting NaN solutions is going to cause no-end of issues.

from gams.jl.