Is it possible to personalize the observation function in python?

[Wenbo11]

Hey,

The GCNN branching tutorial samples observation by the build-in functions: ecole.observation.StrongBranchingScores() and ecole.observation.Pseudocosts(). Is it possible to personalize the observation function (branching policy) for ecole in python? Thank you so much!

Best,
Wenbo

[AntoinePrv]

Hi,

The observation function is an object that extract data from SCIP, typically to get the data you would feed to the ML model.
It can be done in Python or in C++, either by inheriting of an existing one (if you need some of its data), or from scratch. You can see more details here: https://doc.ecole.ai/master/howto/create-functions.html

Ecole implements a few observation functions that we deemed valuable to many users (hopefully more to come).
In the case of Pseudocosts and StrongBranchingScores, they are observation functions.
As such, they do not perform any branching, only extract data, even if they are related to a branching policy.
They are useful to get "expert" data and reproduce such branching rules (by taking the argmax of the scores and passing it to env.step).

Hope this answer your question. Let us know if we can elaborate on anything.

[Wenbo11]

Thanks for your response! It helps a lot. I have a small suggestion for the documents of **dynamics_kwargs. I guess it could be helpful to detail it in the document. For instance, if I want to query the pseudobranching candidates rather than the LPbranching candidates, I can directly pass pseudo_candidates=True when initializing the environment.

[AntoinePrv]

Thanks for the feedback.
We documented this in the Environment parameters section

Each environment can be given a set of parameters at construction, in order to further customize the task being solved. For instance, the Branching environment takes a pseudo_candidates boolean parameter, to decide whether branching candidates should include all non fixed integral variables, or only the fractional ones.

In your opinion, is there a way we can improve on this?
Perhaps we could document in the Creating environment section that Dynamics constructor argument are forwarded from Environmnents extra arguments

[Wenbo11]

Yes, I personally believe it can be helpful to document it in the Creating environment section.

[AntoinePrv]

I have added another section to the doc