However I have been thinking about it since, and it is possible to design a multi-armed bandit approach with logarithmic regret (though higher by a constant factor than a traditional approach), that can handle the varying performance ratio. It also would allow you to add variations at any time.
There remain operational differences, but this problem is fixable.
However I have been thinking about it since, and it is possible to design a multi-armed bandit approach with logarithmic regret (though higher by a constant factor than a traditional approach), that can handle the varying performance ratio. It also would allow you to add variations at any time.
There remain operational differences, but this problem is fixable.