?????
I´m using normalized Euclidean distance (weight = 1/stddev) with fast sampled deviation, which do change over time (k-NN for lazy people who don´t fine tune their bots).
I had the impression the changing weights were not being accounted for. My bot´s scoring dramatically decreased with the kd-tree. I suspect it is the cached coordinate limits inside each node. Or maybe I screwed up something while trying to use the library.
Changing weights is certainly intended to be accounted for, though I hadn't tested it as much as other features. I know it couldn't be the coordinate limits inside each node, because they are copying the min/max values from the points which don't change with weighting. Do you remember which version of the tree were you trying this with?