search

lululxvi / deepxde

A library for scientific machine learning and physics-informed learning
https://deepxde.readthedocs.io
GNU Lesser General Public License v2.1
2.59k stars 733 forks source link

L-BFGS #1662

Open tanmay2798 opened 6 months ago

tanmay2798 commented 6 months ago

I wanted to ask that why after using L-BFGS with pytorch in https://i-systems.github.io/tutorial/KSNVE/220525/01_PINN.html significantly reduces the loss value?

praksharma commented 6 months ago

The question is fundamental to optimisers. While, second order optimisers are known for faster convergence, this is not generally true in cases where the loss lanscape contains saddle points.