mljs / levenberg-marquardt

Curve fitting method in JavaScript
MIT License
70 stars 15 forks source link

Implement any initial approaching using a Newtons method #27

Open andcastillo opened 5 years ago

andcastillo commented 5 years ago

We should implement a globalization part. It could put the method close enough to the optimum before starting with the gradient descent method.

jacobq commented 5 years ago

Hi @andcastillo and sorry I haven't been helping with this library lately (I had planned to do so months ago). Would you mind elaborating on how the current implementation differs from the algorithm described by the papers/articles linked in the README? I thought the "damping factor" basically allowed LM to work as a blend of those two modes. If it's easier for you to explain with an example / PR feel free to do it that way.

andcastillo commented 5 years ago

I will elaborate later, but up to I know, the damping value is adjusted within the optimization loop, based on the update speed. If the algorithm is changing too slowly, the damping value is increased.