[Ipopt] IPOPT and CUTEr

[Horand Gassmann]

Jonathan Hogg <jonathan.hogg at stfc.ac.uk> wrote:

> Dear all,
>
> Just wondering if anyone had successfully made IPOPT work with CUTEr
> recently?
>
> I've tried it with various linear solvers, but can't seem to get the
> "right" answer for several problems, even where almost every other
> solver under the sun seems to agree (of course some problems will have
> more than one local optimal, but I think I'm picking convex QPs mainly)
>
> One particular example where everyone else seems to agree is AUG2D,
> where I believe the correct answer is 1.6877257114E+06, but IPOPT/MA57
> is getting 3.3748235057934504e+06 (other solvers incl some reported by
> Andreas agree with the first answer at least in the first few sf).

I don't know anything about the CUTEr problems, but your mention of QP  
reminded me of a situation we encountered just a couple of weeks ago.  
Ipopt represents quadratic objective differently than many other  
solvers. Ipopt is "WYSIWYG", that is, the objective

x1^2 + 2 x1 x2 + 3 x2^2

is represented using three quadratic coefficients, 1, 2, and 3 as  
written. A lot of other solvers use matrix notation, often with an  
implicit factor of 1/2, so that the same objective is written as

1/2 (x' Q x), with x = (x1 x2), and Q =  [2  2; 2  6];

Since Q is usually assumed symmetric, you would represent this matrix  
with three coefficients as well, but they are rather different than  
before.

I can't say whether this is the source of your discrepancy, but it's  
worth checking.

Cheers

gus gassmann