0f34314d2dd0c1b9311cb8f40eb4f255-AuthorFeedback.pdf

We agree that finite-time regret is the2 performancemeasureofinterest. We significantly improved the regret guarantees w.r.t. The lower-bound remains the same, except that the KL divergence needs to be43 computed for some distribution in the sub-Gaussian family. This48 is very important in practice because it certifies that the algorithm effectively adapts to the problem's structure.