Convex solver problems

[pavelkomarov]

There are several places in the code where the convex solver is given explicitly as MOSEK, but I'm having trouble running MOSEK, because there doesn't appear to be a version compatible with my python installation. Should we loosen the requirements and let cvxpy use whichever solver it wants?

[pavelkomarov]

I can run the unit tests on lineardiff using cvxopt, but I see this result (no points at domain edges). Not sure whether this is a not-having-MOSEK issue or whether it's more foundational to lineardiff.

[pavelkomarov]

I think this might be part of a larger problem where the convex solver just doesn't work all the time. For instance, when testing velocity against $x(t) = 1$, I get Solver 'CVXOPT' failed. We need to have checks in place to make the TVR methods robust.

[pavelkomarov]

In #106 today I've added some code to the top of the TVR module to try importing MOSEK and warn if it can't be. All methods get MOSEK as their default solver if it can be imported, but they otherwise get None, which allows CVXPY to do whatever it wants, which appears to be defulting to ECOS, see the _solve method. Hopefully that is robust enough to work better than CVXOPT was.

[pavelkomarov]

Indeed, this is the way. CVXPY is able to choose a solver that works for lineardiff if I pass the solver as None. I actually don't have ECOS installed, so it must be choosing something else.

[pavelkomarov]

I've made the default solver None across the codebase so I can stop raising warnings about users not having MOSEK. CVXPY is pretty smart about choosing solvers, so I think it's best to let it do its thing. I've kept comments throughout the documentation that we found MOSEK to be most robust, although I'm not sure that's current information. I was finding CLARABEL and OSQP also worked for me, although OSQP fails with Lineardiff in the CI run, so that unittest now explicitly sets the solver.