More granular support of SOS1 and SOS2

[ValentinKaisermayer]

I have some solver-agnostic code where I already check:

MOI.supports_constraint(JuMP.backend(model), MOI.VectorOfVariables, MOI.SOS1{Float64})

But this does not work here, since Cbc does support them.

Why are SOS1 and SOS2 constraints supported

Cbc.jl/src/MOI_wrapper/MOI_wrapper.jl

Lines 483 to 489 in a7ea668

	function MOI.supports_constraint(
	::Optimizer,
	::Type{MOI.VectorOfVariables},
	::Type{<:Union{MOI.SOS1{Float64},MOI.SOS2{Float64}}},
	)
	return true
	end

when the solver states

Cbc.jl/src/MOI_wrapper/MOI_wrapper.jl

Lines 422 to 423 in a7ea668

	"There are known correctness issues using Cbc with SOS " *
	"constraints and no binary variables.",

Is there a way of not supporting them just for non-binary variables.

[blegat]

Unfortunately, the way supports communicates is constraint-wise so there is no way to communicate that a SOS constraint with non-binary variables is not supported. The only way would be to have different sets for SOS constraints with binary and non-binary. Alternatively, the Cbc solver could also correctly handle these cases

[odow]

Is there a way of not supporting them just for non-binary variables.

No. Just don't use Cbc. Use HiGHS instead.

[odow]

Closing as a duplicate of #151. I don't want to complicate MOI by adding more complexity to how we manage supports_constraint.

Our options are:

• Do nothing.
• Build and distribute un-released master version of Cbc and required dependencies
• Do something like adding a binary variable to the end of each SOS constraint.