Nothing much to see. The main improvement is the switch from pure gradient descent to a two-stage potential reduction algorithm.
The first stage is to perform random transformations on the system and discard the transformations that increase the potential. The second stage is the classic gradient descent.
Nothing much to see. The main improvement is the switch from pure gradient descent to a two-stage potential reduction algorithm.
The first stage is to perform random transformations on the system and discard the transformations that increase the potential. The second stage is the classic gradient descent.
So yeah, huge pre-simulation speedups.