Implement `gp_hedge` acquisition function
pfebrer opened this issue · 13 comments
Is your feature request related to a problem? Please describe.
I'm trying to migrate from scikit-optimize to this package because it seems more actively maintained. However, in scikit-optimize the default acquisition function was gp_hedge, a combination of the three acquisition functions implemented in this package, which is not implemented in this package.
Describe the solution you'd like
Would it be possible to add gp_hedge utility function to this package?
References or alternative approaches
The methodology is described in https://arxiv.org/pdf/1009.5419.pdf (Algorithm 2 in page 6).
And it is implemented in scikit-optimize loop here:
https://github.com/scikit-optimize/scikit-optimize/blob/a2369ddbc332d16d8ff173b12404b03fea472492/skopt/optimizer/optimizer.py#L538-L539
and here:
Additional context
I'm new to this field, so I might be missing some points like gp_hedge not being the optimal technique or being described initially in another paper.
Are you able and willing to implement this feature yourself and open a pull request?
- I'm not sure I would do it in the right way.
Hi @pfebrer,
this would seem like a good addition to the project, in my opinion. However, compared to other acquisition functions, this one seems to be non-trivial to integrate into the existing codebase, primarily because the process is stateful through the gains vector. In all likelihood this will mean that the resulting optimization process...
- ...will not be serializable.
- ...will have no clear separation between suggest/evaluate/register.
I currently don't have the bandwidth to implement this either. You could give it a try and see if you can work around these issues in a clever way.
PS: maybe I misread the paper -- I really just scanned the pseudocode on p.6 -- and this is not actually an issue.
Indeed that's the scary part about it, the acquisition function needs to keep a state.
But I don't know if it's really incompatible with the current code structure. You already define aquisition functions by initializing a python object, e.g.:
utility = UtilityFunction()So that object could keep a state that it's updated when you pass the utility function to the suggest method.
It's true that it could be stored in the UtilityFunction, but right now you can restore the state of the optimizer by simply registering the previous run's parameter/target value combinations (+constraint values, if applicable). To restore an optimizer that uses gp_hedge you would additionally have to store/restore the gain values (by all means, not a huge problem, just an added layer of complexity). If you want to serialize the optimizer via pickle, it will also not store the acquisition function, so you would have to deal with that separately.
Good point about suggest though, that should work without issue.
To restore an optimizer that uses gp_hedge you would additionally have to store/restore the gain values (by all means, not a huge problem, just an added layer of complexity). If you want to serialize the optimizer via pickle, it will also not store the acquisition function, so you would have to deal with that separately.
Hmm to be perfectly correct yes, but I don't think there would be a major problem with starting the utility from scratch, would there be?
If you restore the optimizer from a saved state (and you set the seeds to the same value), it should proceed as the uninterrupted optimization would, so I think this is a problem. You don't have reproducibility anymore.
It would also make the optimization worse -- the point of gp_hedge seems to be to derive which of the acquisition functions from the pool works best for a specific problem, this information will have been lost if you restart.
I really like this acquisition function, and I think it would be a great addition to the package but I don't think we can add it without some larger restructuring of the suggest step, serializability issues aside.
I would encourage you to try and implement it for use in suggest-evaluate-register (see here). If you can figure it out, let me know :)
Yes, ideally it should restart from the state of the gp_hedge, but I mean it's not a huge problem in practical terms if you just want to optimize a function. It might take a bit more time to relearn which utility function is better, but maybe it won't be that bad.
I don't know if I will have time to do it, because I'm just benchmarking different optimizers for my particular problem. Of course it is always nice to contribute to things, but I'm more of a user in this case 😅
Understandable. If I ever find the time, I would try to implement this, so feel free to leave the isssue open for now :)
@pfebrer I have redesigned the acquisition function API and implemented GPHedge. Would you be interested in reviewing and/or testing that code?
Cool!
I am not focused any more on the problem that brought me here, but I am sure I can find time to do some testing and benchmarking. Is it on the master branch already?
I'll push it to a separate branch when it's ready and ping you, if that works :)
Hi, this next couple weeks I'm swamped with work. I won't have the time to review it, I'm sorry 😅
No worries! Good luck with whatever's keeping you busy :)