Open tpr0p opened 1 year ago
I agree that it would be interesting to try this. I don’t know of any work that looks into this. People use very simple weighting functions. In the 3d_spots branch, we have a tanh test function that I think will be more stable. I suspect that people don’t use AL because balancing the weights is a very nonlinear function and it’s better to just go very slowly and use something that always works instead of tuning AL. But would be cool to try!
Add the Augmented Lagrangian method for increasing weights in experimental optimization.