PDF version of this entire document
A mixture/hybrid approach does not work quite so well just yet. That would be because there is an impact on what GMDS is measuring when the resolution gets increased. We need squash a bug related to how GMDS behaves at high resolutions sometimes (which is the cause for the previous large experiments getting GMDS results wrong, and thus rendered unusable).
The method is now in a state where, if a difference is noticeable enough, then it is easy to correctly classify in almost all cases, but if the difference is not very small either, then the decision is left to be made in another way.
As expected, lowering the resolution gives poor results, so it misses the point somewhat except when the effect of different parameters gets learned.
|
|