Can you compare to MVT (Multi-Variate Testing)? From an admittedly surface reading of the post, this sounds like the somewhat common "A/B sucks, MVT is more accurate at optimizing and showing impact" with a newer optimization approach.
In MVT you're interesting in testing the interactions between different elements. If the elements are encountered in sequence the bandit analogue is reinforcement learning. If you have one page with different "slots" to fill then "bandit slate" algorithms might be appropriate. The key advantage is that all these approaches are online: they take advantage of information as it is received, and you can change things and the algorithms adapt. A/B testing and MVT
don't do either.
