Unable to create vector out of AffExpr in some cases

Specific Domains Optimization (Mathematical)
1

MWE:

using JuMP
model = Model()

x1 = @variable(model, base_name = "x1")
x2 = x1 + 1
x3::AffExpr = x1 

# Works: x1 is of type VariableRef
v1 = [sin(x1), x1]
# Works: x2 is of type AffExpr
v2 = [sin(x1), x2]
# Does not work: x3 is also of type AffExpr
v3 = [sin(x1), x3]

Hits error in Base.vect. So not sure if this is a JuMP issue or a Base issue.

1 Like
2

That sure looks like a bug. I think the bug is in nlp_expr.jl in JuMP:

function Base.convert(
    ::Type{<:GenericNonlinearExpr},
    x::GenericAffExpr{C,V},
) where {C,V}
    args = Any[]
    for (variable, coef) in x.terms
        if isone(coef)
            push!(args, variable)
        elseif !iszero(coef)
            push!(args, GenericNonlinearExpr{V}(:*, coef, variable))
        end
    end
    if !iszero(x.constant) || isempty(args)
        push!(args, x.constant)
    end
    if length(args) == 1   # <--- BUG is here
        return args[1]
    end
    return GenericNonlinearExpr{V}(:+, args)
end

When args is length 1, the return value does not go through the GenericNonlinearExpr constructor and can be of the wrong type.

When removing the if length(args) == 1 statement, the example in the post goes through. This may also be the fix, as this seems to mainly function as an optimization.

2 Likes
4

Thanks a lot. I have temporarily copied your version to my file and it works as expected.

5

Another work-around is v3 = Any[sin(x1), x3].

I’ll take a look at fixing this.

1 Like