passing initial solution or warm start

[LucaVaresi]

Hi everyone,
I have a question regarding HiGHS and in particular highspy.
I have looked online but i have not found a way to specify an initial solution to the HiGHS solver is there a way to do that?

[jajhall]

It's possible to do it using Highs::setSolution, but the binding for this in highspy was forgotten in the last release. It will get into highspy soon - possibly this week as we address another issue

[LucaVaresi]

thank you!
Just another question on this topic, what type of object should be passed to Highs::setSolution?

[jajhall]

An instance of the HighsSolution struct with (at least) values for the variables - col_value. These must correspond to a feasible solution to be useful.

If you're solving a MIP, then (even) a partial assignment of the discrete variables will be exploited, so long as it corresponds to a feasible solution

[jajhall]

With https://pypi.org/project/highspy/1.7.1.dev1/ Highs::setSolution can now be used

[jajhall]

What value is None?

It should be sufficient to set to (say) 0.5 integer/binary variables that you don't want to warm start.

I'll be working on #1692 in the coming weeks, if you have any more experience to share.

[guifcoelho]

What if my integer variable bounds are 5 and 50? Would it be sufficient to put any value in the middle?

I believe None is the equivalent of NULL in C++ (sorry if I'm wrong, it's been centuries since the last time I programmed in C++). I was hoping that filling col_value with some Nones it would translate to NULLs to represent "empty" positions in the array. In my opinion, providing some way to say that this or that variable is not being set for warmstart could be easier to understand. As reference, I believe other solvers let you pass 2 arrays: a list of column indexes and a list of values for the variables.

[jajhall]

Yes, set them to any fractional value - it might be that an infeasible value is OK, actually.

I just look for the values that are within the integrality tolerance of an integer value. I fix their bounds to that value, and (currently) solve the resulting MIP. In practice I should just run the MIP solver until it has found an integer solution. The only danger with this approach is that fixing the user's chosen values might make the problem infeasible, and that could take ages to identify.

Interesting to learn about other solvers. That is a better way, even if it needs another setSolution method.

[guifcoelho]

This is how Xpress does it

I recommend writing a new method for this. Maybe keep setSolution the way it is and add a not method called setMipSolution or setHotStartSolution. In my opinion, this topic is relevent enough to get its own interface method but that's your call.

Also, would you mind opening this discussion again so it could be easier to find?

[jajhall]

Thanks. I hadn't meant to close the discussion. I'll reopen it once I've walked the dog