Revisiting Simple Regret Minimization in Multi-Armed Bandits
share
Simple regret is a natural and parameter-free performance criterion for identifying a good arm in multi-armed bandits yet is less popular than the probability of missing the best arm or an ϵ-good arm, perhaps due to lack of easy ways to characterize it. In this paper, we achieve improved simple regret upper bounds for both data-rich (T≥ n) and data-poor regime (T ≤ n) where n is the number of arms and T is the number of samples. At its heart is an improved analysis of the well-known Sequential Halving (SH) algorithm that bounds the probability of returning an arm whose mean reward is not within ϵ from the best (i.e., not ϵ-good) for any choice of ϵ>0, although ϵ is not an input to SH. We show that this directly implies an optimal simple regret bound of 𝒪(√(n/T)). Furthermore, our upper bound gets smaller as a function of the number of ϵ-good arms. This results in an accelerated rate for the (ϵ,δ)-PAC criterion, which closes the gap between the upper and lower bounds in prior art. For the more challenging data-poor regime, we propose Bracketing SH (BSH) that enjoys the same improvement even without sampling each arm at least once. Our empirical study shows that BSH outperforms existing methods on real-world tasks.
READ FULL TEXT
