Best-arm identification with cascading bandits
share
We consider a variant of the problem of best arm identification in multi-arm bandits, where in each round, multiple arms are played in an ordered fashion until a nonzero reward is obtained. Since each round potentially provides information about more than one arm, the sample complexity can be much lower than in the standard formulation. We introduce a subroutine to perform uniform sampling, that allows us to adapt certain optimal algorithms for the standard version to this variant. As we prove, much of the analysis goes through as well.
READ FULL TEXT