Oscar Dowson

Here's where we do it: https://github.com/jump-dev/HiGHS.jl/blob/c736c0260d9bf74dcb544e4e76ebf05a00ac0593/src/MOI_wrapper.jl#L2107-L2193 * We don't include the offset if there is a ray * If there is a primal solution we pick the closest bound *...

@bernardokp I think you should write the model like this. It's far more explicit and readable. I still don't really get what the `x` and `y` in your model are...

Okay, now I have  We do't get the analytic solution because of the SAA.

@bernardokp how about now? There is ambiguity in the infinite horizon policy: does the agent get to make an initial "buy" decision before observing the first realization of demand? Or...

Cool that's the same model then. The `L(y)` formulation isn't any nicer, because you need to introduce new variables to represent the positive and negative components of `y`.

Nope. `ret` is a vector, with the non-zeros in the same order as `jacobian_structure` and `hessian_lagrangian_structure`.

Oracle meaning something you can ask a question of and get back an answer. We could have named this "VectorNonlinearCallback" but oracle is a fairly (?) standard term-of-art (if hard...

Sometimes things go the other way. In the stochastic programming world, they use terms like "hazard-decision" to mean "observe the uncertainty first, then make a decision", but the term is...

I think no. There are many problems with a multithreaded callback that I don't see us resolving anytime soon.

> Why did you choose to store the Hessian as a vector + tuple of indices Because it was simple, and because that's the format that MOI requires. > It...