|
|
|
|
|
by tednaleid
5756 days ago
|
|
It doesn't need to be that random on your side of things. It just needs to be unexpected/unpredictable by the subject getting the reward, but in some way paired with the performance you want to elicit more of. Think of gambling, the reward is very predictable if you know the seed values and underlying algorithms, but unpredictable from the gamblers perspective. This is a classic Variable Ratio (VR) schedule of reinforcement (http://en.wikipedia.org/wiki/Reinforcement#Schedules_of_rein...). To get the highest rates of responding, you don't want to tie it to a known schedule, that would instead be some sort of fixed or variable interval schedule, which isn't anywhere near as reinforcing as a variable ratio one. It should be possible to get multiple bonuses within a short period of time, as long as the desired metrics are being hit. That way, you'll continue to get strong behavior throughout the entire period rather than bursts of behavior as the end of the period nears (rather than the scalloping that you see on the wikipedia graphs). |
|
Now, the first question I need to ask is "is this crossing a line" - is it too manipulative... then, if it's not, I need to come up with a bonus scheme.
I was thinking about basing the bonuses on how much revenue has grown past our previous high. (with the number of employees, one guy can really make a big difference, and if we are operating properly, we grow by around 10-15% a month.)
hm. It seems I could pay out on random days on a month that beats the previous months receipts? I think I'd want it to be a 'peak month' rather than a 'peak day' to discourage gaming, I mean, I see how you could queue up customers for a few days... doing this over months could be bad, especially if when the months begin changes randomly.
But yeah, a formula that people can know, with, perhaps, random numbers that aren't known until the period end sounds like the way to go. I think people would mind the manipulation less if it was clear what I was doing and why.