Enforce the constraint, but less accurate solution?

[jiu...@gmail.com]

How to configure the settings so that: 1) must satisfy the constraint (or very very close), and 2) the solution can be less accurate, because we want the solution very fast.

I tried to change the eps and polish, but it either slowdown the solver or violate the constraint badly.

Thanks. jc

[Goran Banjac]

Please have a look at the following link for a recent discussion on this topic:
https://github.com/oxfordcontrol/osqp/issues/97

Hope you find it helpful.

Goran

[jiu...@gmail.com]

Hi Goran,

Thanks for the reply, it is helpful.

OSQP is pretty slow for our problem of ~3000 variables, it takes 1.7 seconds compare to Mosek's 0.1 second. Maybe the default linear solver is slow. So I installed IntelSWTools as directed on the osqp.org. However i could not make it work with OSQP R package. Here is the error:
ERROR in lh_load_lib: Windows error while loading dynamic library mkl_rt.dll, error = 126
ERROR in osqp_setup: mkl pardiso linear system solver not available.
Tried to obtain it from shared library

I added the lib and include folder under C:\Program Files (x86)\IntelSWTools\compilers_and_libraries_2019.0.117\windows\mkl to the system path. but I still get that error.

Do you have any suggestions? Thanks.

jc

[Bartolomeo Stellato]

Hi,

You should add the directory where the `mkl_rt.dll` file is. It should be in a folder like the following one (not just `windows\mkl`). 

C:\Program Files (x86)\IntelSWTools\compilers_and_libraries_2019.*.***\windows\mkl\lib\<arch>_win

However I prefer the instructions we provided in the docs to set it up. Did your run the following command?

%MKLROOT%\mklvars.bat intel64

It should set the paths for you. It is strange your problem solution is so slow. How many iterations does it take? It might be related to scaling or wrong rho value.

Would you be able to send us the problem data? 

Bartolomeo