Jump.jl: Parameters for linear/quadratic problems

[joehuchette]

Ref #304, there's still a mismatch between the support for parameters in nonlinear and linear/quadratic problems. Potentially we could store the A matrix in the model explicitly and update coefficients for different parameter values, reusing the same sparsity pattern. I think the main hurdle is building up expressions or constraints where AffExprs and parameters are co-mingled.

[mlubin]

Just remembered the reason why we can't do this. Let's say N is a parameter, then you do:

@defVar(m, x[1:N])
@addConstraint(m, sum{x[i],i=1:N} <= 1)

What happens when you change N? It's not just about updating the coefficients in the constraint matrix.

[joehuchette]

Well we can never have that, but there's still the possibility of supporting "coefficient" parameters.

[mlubin]

Can they be used inside of filter conditions?

[joehuchette]

No, it would only be parameters for variable coefficients or constant terms in affine/quadratic constraints.

@defParam(model, p)
@addConstraint(model, p*x == p^2-1)

By no means as general as the nonlinear parameters, but better than nothing.

[mlubin]

The ratio of increased complexity of _everything_ in JuMP to usefulness is pretty high here

[joehuchette]

Wait, do nonlinear parameters work in sums or filter conditions right now?

[mlubin]

No, just inside expressions

[joehuchette]

So

@addNLConstraint(m, sum{x[i],i=1:N} <= 1)

doesn't work either when N is a parameter?

[mlubin]

No, although AFAIK it does in AMPL

[cossio]

+1, would really like to have this feature

[mlubin]

This is entirely speculative and out of scope for JuMP 1.0, so I'm tempted to close.

[odow]

Yeah. I have an application for this, but I don't think the speed improvement gained by poking the constraint matrix over rebuilding the model outweighs the additional complexity in JuMP.

[mlubin]

If anyone would like to rewrite JuMP to support this, get in touch. No use keeping the issue open though.