Closed matsen closed 11 years ago
Well, the full optimization is going to be a bit gnarly, and overkill because we won't be putting those actual sequences into the tree. Would it be possible to do a variant of the greedy-Voronoi algorithm that doesn't actually add the sequences into the tree? That is, voronoi regions can be made from either leaves or placements. We recur through the tree, and score each placement by the amount by which it would reduce the ECLD, select the max one, redo voronoi regions, etc. I think this would be great.
Hm, meh. Someone is going to have to want to pick up the baton here.
Given a collection of placements on a tree and a number k, find a subset of k of those placements that most decreases the ECLD if they were to be added as leaves (with the ML branch length of the best placement). The rest of the placements "act" as they do in the cutting algorithm -- as bits of mass on the edges of the reference tree.