Issues with Proximity about splinepy HOT 4 CLOSED

I implemented a crude version of the Levenberg-Marquardt-method, based on this. It significantly reduces the number of failures to:
[11 21 61 62 81 82 88] with the same example. It also uses a secondary metric from ILSB's numerics course (here as supposed to here), which seems to perform better.

There are two versions with a boolean flag here, where either the lambda * diag(J^TJ) is added or lambda * I, as shown in the wiki article. I did not observe any meaningful difference.

If only LM is used the failures are these:
[ 1 11 19 20 21 29 31 61 62 80 81 82 88 91]

@j042 : Is this worth further investigation?

from splinepy.

I implemented a crude version of the Levenberg-Marquardt-method, based on this. It significantly reduces the number of failures to: [11 21 61 62 81 82 88] with the same example. It also uses a secondary metric from ILSB's numerics course (here as supposed to here), which seems to perform better.

There are two versions with a boolean flag here, where either the lambda * diag(J^TJ) is added or lambda * I, as shown in the wiki article. I did not observe any meaningful difference.

If only LM is used the failures are these: [ 1 11 19 20 21 29 31 61 62 80 81 82 88 91]

@j042 : Is this worth further investigation?

Side-note. If we do not change the penalisation damping parameter, even if the coefficient rho is out of bounds, but only when the solution gets worse, it converges for all points. This is the latest version in the branch

from splinepy.

I implemented a crude version of the Levenberg-Marquardt-method, based on this. It significantly reduces the number of failures to: [11 21 61 62 81 82 88] with the same example. It also uses a secondary metric from ILSB's numerics course (here as supposed to here), which seems to perform better.
There are two versions with a boolean flag here, where either the lambda * diag(J^TJ) is added or lambda * I, as shown in the wiki article. I did not observe any meaningful difference.
If only LM is used the failures are these: [ 1 11 19 20 21 29 31 61 62 80 81 82 88 91]
@j042 : Is this worth further investigation?

Side-note. If we do not change the penalisation damping parameter, even if the coefficient rho is out of bounds, but only when the solution gets worse, it converges for all points. This is the latest version in the branch

Thanks a lot for implementing this, sounds great to me!

from splinepy.

I had a weird case where max_iterations value was set to 0 at ~60 query within a call. This happened at PySpline::Proximities. I wonder if that was compiler error or something that happens due to our implementation.

from splinepy.