Greybox Objectives #3364
Greybox Objectives #3364
Conversation
There was a problem hiding this comment.
I haven't reviewed changes in _evaluate_hessian_if_necessary_and_cache yet, but this looks fine.
Just curious, what is the motivation for doing this instead of just using outputs and putting them in the objective? Is there some motivating use case where this works better?
| input_values | ||
| """ | ||
| raise NotImplementedError( | ||
| 'evaluate_objective called but not ' 'implemented in the derived class.' |
There was a problem hiding this comment.
Can you fix black's weird auto-formatting here?
| return self._pyomo_nlp.evaluate_objective() | ||
| obj = 0 | ||
| for nlp in self._nlps: | ||
| obj += nlp.evaluate_objective() | ||
| return obj | ||
|
|
||
| # overloaded from NLP | ||
| def evaluate_grad_objective(self, out=None): | ||
| return self._pyomo_nlp.evaluate_grad_objective(out=out) | ||
| ret = np.zeros(self.n_primals(), dtype=float) | ||
| for nlp in self._nlps: | ||
| ret += nlp.evaluate_grad_objective() | ||
|
|
||
| if out is not None: | ||
| ret.copyto(out) | ||
| return out | ||
|
|
||
| return ret |
There was a problem hiding this comment.
Do we want to sum multiple objectives? At first thought, it seems like we should only allow one objective to be implemented. If each grey box intends to implement an objective term, maybe the methods on ExternalGreyBoxModel should be called e.g. evaluate_objective_term.
There was a problem hiding this comment.
I agree with Robby in principle that I don't always want to sum my objectives. But am happy to accept this as a first pass
This is me. For applications I work with, it's common for people to just come with a black box and say "minimize". Clearly that approach is not ideal, but allowing this capability will enable us to reach out to people who are used to doing optimization that way and help us convert them to the light side of the force... |
Okay, so this is basically just for convenience.
I don't think it's clear at all. I'm usually a fan of eliminating auxiliary variables where possible, so this approach appeals to me. I was just curious if there was an example where it actually mattered for performance or reliability. |
It is not ideal in the context of it hides structure I could potentially implement in my optimization problem. IE, what if my black box is hiding a convex optimization problem? But yea, it is a matter of convenience not performance. |
|
@michaelbynum , @Robbybp - We are hoping to cut a patch release on Monday (the 14th). Any chance this is going to make it into that? |
Summary/Motivation:
This PR adds support for objectives in PyNumero grey box models.
Legal Acknowledgement
By contributing to this software project, I have read the contribution guide and agree to the following terms and conditions for my contribution: