Hessian in TorsionTarget?

[mkrompiec]

I'm looking for ways to speed up torsion fitting. Is the Hessian of the objective function computed for the TorsionTarget if Newton-Raphson optimizer is used? Should BFGS be faster?

[leeping]

In my experience, the Newton-Raphson optimizer is usually faster because part of the objective function Hessian is calculated via the Gauss-Newton approximation. I don't have the theory for how to update the Hessian a la BFGS if the Gauss-Newton part is being recomputed at every step, but I think it may be possible.

[mkrompiec]

Thanks @leeping ! Do you have any tips for speeding up torsion fitting? We are fitting all rotatable torsions in polyaromatic molecules, 2-5 torsions each.

[leeping]

Have you tried the remote target feature? It allows all objective function components (target) to be computed in parallel across multiple machines, and it's something we use routinely to run the larger parameter optimizations.