search

MOSEK / Mosek.jl

Interface to the Mosek solver in Julia
MIT License
49 stars 28 forks source link

Conic duals fails Farkas test #56

Closed kaarthiksundar closed 9 years ago

kaarthiksundar commented 9 years ago

The attachment https://www.dropbox.com/sh/sjojrwff826zgn6/AABNrs_UFMlFSU1egPjzDVsDa?dl=0 contains a conic problem which I solve using Mosek. The julia program to solve the problem using Mosek is given here: https://gist.github.com/kaarthiksundar/10e645b8829a30a37d2e

The primal problem is infeasible and so I check the Farkas' certificate. Observe that the duals fail to satisfy the Farkas certificate.

CC: @mlubin

mlubin commented 9 years ago

ECOS gives a valid certificate for TOL=1e-4, but Mosek has a huge violation:

>> ||.|| = 8.434528581032842e8, rhs = 33.262782568756734
kaarthiksundar commented 9 years ago

There are cases where ECOS fails with that tolerance.

ulfworsoe commented 9 years ago

If I extract the underlying task and get the solution values directly, I get that ||A'y+slx-sux+snx|| ~ 1e-8 so it is probably a bug in the interface. I'll look at it.

ulfworsoe commented 9 years ago

Nope. That was a bug in MOSEK. I have passed it on to the interior-point guy.

erling-d-andersen commented 9 years ago

The problem is somewhat ill-posed. A lot of the conic constraints degenerate to one dimensional cones for instance which is a little bit strange but of course ok.

We are currently testing a fix that we plan to make public shortly.

kaarthiksundar commented 9 years ago

I ran the test instances with the fix (updated binaries) provided by mosek support. Preliminary tests indicate that the dual values pass the Farkas' test. I will check all the instances that failed the test previously and keep you guys posted by the end of the day.

Thanks for the prompt response.

kaarthiksundar commented 9 years ago

The new binaries passed the Farkas' test for all the instances that I generated. Hence, closing the issue.