[bug #22033] Not possible to minimise with "Levenburg-Marquardt". -- May 11, 2014 - 20:42

Follow-up Comment #2, bug #22033 (project relax):

There used to be a catch for this with a RelaxError.  The problem is that we
do not have gradients (or Hessians) derived and hardcoded into the dispersion
target function code.  Therefore Levenberg-Marquardt will not work, nor any of
the more advanced algorithms (Newton optimisation for example).  Numerical
gradients could be calculated in minfx https://gna.org/projects/minfx/, but
this has not been implemented.  Numerical gradients are hugely computationally
expensive and can increase the computation time by an order of magnitude. 
Anyway, feel free to try to reintroduce the RelaxError if the Nelder-Mead
simplex algorithm is not selected.

    _______________________________________________________

Reply to this item at:

  <http://gna.org/bugs/?22033>

_______________________________________________
  Message sent via/by Gna!
  http://gna.org/