Closed KalelR closed 1 year ago
Well sure why not do the PR that fixes the function?
seems to me you are correct and there isn't much debate to be had here :D
I can, but that test will fail. If you have an idea why, I could already fix it in the PR.
i don't understand which test you are referring to.
bump
The dummy bistable map in grouping_continuation.jl
tests fails with this change. It apparently leads to floating point errors that somehow screw up the matching. The reason for this is not obvious to me, but I thought could be for you. If not, I can take a look at this sometime in the next days :)
Can you open a PR that does this, and then I look at the failing test?
Hey
I just looked at the
_rescale_to_01
function, which we use to rescale each feature dimension onto the[0,1]
interval, and noticed it is clearly incorrect.Currently, it's
This does not rescale features to
[0,1]
.Instead, it should be
When I first wrote it, it was correct, I'm pretty sure. Somewhere in the middle of the refactoring we got this wrong. It might very well nott affect results, but we should correct it to do what it should.
I'm writing an issue because I implemented this and the dummy bistable test in
grouping_continuation.jl
got an error. I didn't have time to understand why this happens.. but I saw that this rescaling of features leads to small floating point errors (e.g. the rescaled features being0.49999999...2
instead of0.5
). Maybe the matching procedure does not match them because of the errors, thus creating new attractors? Or is the rescaling applied to the global pool (which would rescale the features a certain way) and then also applied to a local pool, for only attractor (whihc would rescale in a different way)?