search

aaronkl / OptSizeChooser

0 stars 1 forks source link

cost function kernel #2

Open simbartonels opened 10 years ago

simbartonels commented 10 years ago

Currently it is possible that predicted costs are negative (which happened for small datasets and that's why the chooser avoided those). A solution is to put the durations on a log scale for modelling and to exponentiate later. However, this might render the polynomial kernel useless. What are our options?

simbartonels commented 10 years ago

A possible solution is to exponentiate before dividing the KL divergence.