Skip to content

Local uncertainty threshold is multiplied by force hyperparameter, but uncertainty is already normalized? #450

@gilbogilb

Description

@gilbogilb

https://github.com/mir-group/flare/blob/07a6ff51a26c382eb67d944aa0053d816392c29e/flare/learners/otf.py#L352C1-L359C18

Hi,
I was looking at the environment selection during an OTF step. In this section of code from the OTF learner, the environment selection is performed by is_std_in_bound for regular on-the-fly active learning. The function gets as argument the force noise hyperparameter and multiplies the threshold for its value (https://github.com/mir-group/flare/blob/07a6ff51a26c382eb67d944aa0053d816392c29e/flare/learners/utils.py#L50C1-L56C42) - however, if the variance_type is "local", the variance is already normalized wrt the kernel prefactor, and this would result in a (generally) lower threshold which is also depending on the hyperparameters - with, on a side note, no connection between the local uncertainty on environments and the force noise hyperparameter. Is this wanted/taken care of somewhere (e.g. should the thresholds always be set to negative values), or it is a problem in the code?
Thank you in advance.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions