Should the learning rate multiplied by the number of rounds always equal to 1? My best out of sample results has been using an eta of 0.1 and 50 boosted rounds, but the total weightings add up to 5. Is that spurious or is there nothing special about the tree weights adding up to 1?
