PDF version of this document

Optimisation Issues

The optimisation regime needed to be changed as it often got stuck without any constructive paths to a solution being found. It was at first unknown whether the objective function was malformed, or perhaps its handling by the general-purpose Nelder-Mead optimiser was not effective. It could be seen that there was no justification in believing that the function was truly stuck. Sharp drops in value could occasionally be observed, meaning that good warps were finally found and applied to the data. Figure makes it rather crystal-clear.

Figure: Drops which illustrate the problems with optimisation.

Even at present some moderate drops are observed at times. With a less stochastic choice of warps, that unwanted effect might vanish.