Managing unnecessary 0's in JuMP

Specific Domains Optimization (Mathematical)
1

I have a rather ugly quadratic objective function that requires a fair amount of manipulation of variables, and I suspect that one of the things that’s making this process incredibly slow is that I have lots of GenericAffExpr that are carrying around lots of unnecessary variables that have 0.0 coefficients. Is there an elegant way of dropping these?

2

May you post a MWE?

1 Like
3

Here’s the (incredibly hackish) code I wrote for doing this, I guess my question was whether there was an equivalent function already available in JuMP, which perhaps would have been better to use depending on the context

function _dropzeros(ex::T) where T<:JuMP.GenericAffExpr
    mask = Bool[x > eps(Float32) for x ∈ ex.ceoffs]
    T(ex.vars[mask], ex.coeffs[mask], ex.constant)
end

This seems to work quite well, by the way, and is much faster than I was expecting it to be.

4

I’m planning on changing AffExpr and QuadExpr to use a dictionary instead of a list of coefficients for JuMP 0.19 (discussion). That would resolve this issue.

1 Like
5

That sounds wonderful. I’ve been noticing that things start getting really ugly with quadratic expressions, it gets to the point that everything requires custom code to do efficiently.