Run cuPDLP on dual?

[ktp-max]

Hi all,

I am successfully running the GPU version of PDLP on a given program but I have several reasons to believe that it would be better to run the same method on the dual. Is there a simple interface within HiGHS which I could use to run directly cuPDLP on the dual without having to convert myself the complete program to its dual?

Thanks!

--Constantinos

[jajhall]

In short, no.

However, there is code within HiGHS to form the dual of a general LP - since simplex and IPM have a "dualize" strategy, so it shouldn't be hard to add. What's even easier, and maybe of wider use, is a facility to return the dual of an LP, so that users could pass that back to HiGHS and solve it as a primal problem.

Of course, if a primal problem has finite lower and upper bounds on a variable or constraint, then the dual problem would have more variables and constraints than the original problem, so deducing its solution would not be trivial.

[ktp-max]

Perfect, thanks! I will try to use this one. In the future, as you mentioned it would be nice to have such a dualize function publicly available. Thanks for the help!

--C

[jajhall]

Perfect, thanks! I will try to use this one. In the future, as you mentioned it would be nice to have such a dualize function publicly available. Thanks for the help!

--C

Can you build HiGHS from GitHub branches? If so, it's probably much quicker for me to write Highs::dualiseLp and Highs::getDualLpSolution than for you to interpret HEkk::dualize() and HEkk::undualize(). However, they probably wouldn't make it to the next HiGHS release - which we expect to be out by Tuesday.

[ktp-max]

Yes, I can definitely try to build from the branch. Thanks for the help!

[ktp-max]

Hi @jajhall

Any updates on this one? Would it be possible to have it in the next release? Thanks for the help

[jajhall]

Yes, I'll have this in the next release