Butterfly effect

[lvdmaaten]

Butterfly effect

[lvdmaaten]

Given the nature of the Barnes-Hut approximation, this is not so surprising, is it? A small change in a single data point may push it from one quadrant to another, thereby changing the output of computeNonEdgeForces() because both some summary statistics as well as the pruning may change. This change then starts propagating for potentially many hundreds of iterations, which may lead to a different solution. I'm not quite sure what you want me to do about this?

[NikTuzov]

Does it make sense to apply some averaging like it is done in random forest, or you can always say that of two or more solutions one is the best, so there is no point in averaging them?

[lvdmaaten]

You could try, but it would substantially increase the computational costs at potentially limited gains.