robbert-harms
commented
3 years ago Hi Stefania,
The patience is indeed the number of iterations for the fit process, but slightly different from how you interpret it. Every optimization algorithm has, internally, a loop. For example, a simple Gauss-Newton search will require a few iterations to converge to a good estimate. Precisely this iteration counter is controlled (indirectly) with patience.
The idea of patience comes from the problem that every optimization algorithm does a different amount of work per iteration. Some algorithms require only 5 iterations to reach convergence, while others do less work per iteration and as such require 100 iterations. The idea of patience is to map these values to a common denominator, i.e. for the first algorithm we would set nmr_iterations = 20 * patience while for the other we would set nmr_iterations = patience. Or something to that effect. This abstraction is not a very accurate though and one often finds himself tuning patience per algorithm.
For the rest, MDT does not do any stochastic parameter selection before or during optimization.
Best wishes,
Robbert Harms
StefaniaOliviero
Hi, my issue regards the option "patience" for the fit processing: as I understood, it should be used to set the number of iterations of the fit process. The question is "What is the difference between the different iterations? The starting points for the non-initialized parameters?" "If yes, which is the stochastic method used to select the starting points? Is there a kind of perturbation to the first set of starting-points?"
Regards Stefania Oliviero