Do automatic solution checking and validation

[whoburg]

Very low priority:

We could create GP.solution_check(self, sol). It would validate that a particular primal/dual solution pair solves the GP. Could be useful for sanity checks and trying out 'suspect' solution codes.

This is very very very low priority.

[bqpd]

Partway there!

[bqpd]

Currently validates primal feasibility, and that the sensitivities of free variables are close to 0.

[whoburg]

Todo: (optional) we could evaluate both the primal and dual objective functions and make sure that they match. I can do this and/or explain how.

[bqpd]

That would be great! I'm happy to implement it.

[whoburg]

To implement -- gpkit.GP.__run_solver already does some checks -- can add more there.

[bqpd]

Note to do this using the A matrix / matrix math, etc., not substitution.

[whoburg]

Will implement this once #226 is complete.

[bqpd]

Closing this will close half of the remaining issues! (#125 and #217) Go @whoburg go. :p

[whoburg]

where should this be implemented?

[bqpd]

In geometric_program.solve()

[bqpd]

Probably either just before or just after self.result = result: before means that no bad result is made available if these checks fail, after means that the bad result is made available but not returned.

[whoburg]

I should be able to work on this in a few hours.

[whoburg]

Sorry for delay -- still getting to this, but I want to investigate a possible bug first that I consider higher priority. More info soon.

[whoburg]

still psyched to implement this, but waiting until #338 and #335 are fixed.

[whoburg]

@bqpd how do I best get the slice of A corresponding to:

• the objective
• A specific posynomial

?

[bqpd]

I think gp.A.tocsr()[0,:] will get the objective, and >0 the constraints?

[whoburg]

TypeError: 'coo_matrix' object has no attribute '__getitem__'

Also, that's just the first monomial in the objective, but I can deal with that via p_idxs

[whoburg]

I think I have a solution where I don't actually need the answer to this, but I'm still curious...

[bqpd]

Sparse matrices are fun!

(yeah, you need a CSR sparse matrix to index, hence the tocsr above)

[whoburg]

doh, my fault.

[whoburg]

this is coming along -- hard at work