quadratic objective weirdness

[jump-dev]

quadratic objective weirdness

[IainNZ]

Just without even looking at things: sometimes Q is defined funny - if the diagonal should be 2 instead of 1 for some reason, you'd get this

[IainNZ]

e.g. some sort of 0.5 x^T Q x type thing

[mlubin]

I thought about that, but it doesn't really make sense given that gurobi is based on quadradic terms.

[IainNZ]

http://www.gurobi.com/documentation/5.5/reference-manual/node16#routine:GRBaddqpterms
Yeah documentation suggests it isn't like that - although I swear something I've seen is.
Does Gurobi.jl do something strange?

[IainNZ]

How does our current quadratic test not fail? Or is it actually testing for the wrong solution? (LOL)

[mlubin]

I wouldn't be surprised if we're testing for the wrong solution :)

I just tested the gurobi C++ interface, and it gave the right answer, so it's either Gurobi.jl or the Gurobi C interface. The binary search continues.

[mlubin]

See the fix. We weren't hitting it in the tests.

[IainNZ]

Ahh, subtle!