English

Tight Bounds for $\gamma$-Regret via the Decision-Estimation Coefficient

Machine Learning 2023-07-24 v2 Machine Learning

Abstract

In this work, we give a statistical characterization of the γ\gamma-regret for arbitrary structured bandit problems, the regret which arises when comparing against a benchmark that is γ\gamma times the optimal solution. The γ\gamma-regret emerges in structured bandit problems over a function class F\mathcal{F} where finding an exact optimum of fFf \in \mathcal{F} is intractable. Our characterization is given in terms of the γ\gamma-DEC, a statistical complexity parameter for the class F\mathcal{F}, which is a modification of the constrained Decision-Estimation Coefficient (DEC) of Foster et al., 2023 (and closely related to the original offset DEC of Foster et al., 2021). Our lower bound shows that the γ\gamma-DEC is a fundamental limit for any model class F\mathcal{F}: for any algorithm, there exists some fFf \in \mathcal{F} for which the γ\gamma-regret of that algorithm scales (nearly) with the γ\gamma-DEC of F\mathcal{F}. We provide an upper bound showing that there exists an algorithm attaining a nearly matching γ\gamma-regret. Due to significant challenges in applying the prior results on the DEC to the γ\gamma-regret case, both our lower and upper bounds require novel techniques and a new algorithm.

Keywords

Cite

@article{arxiv.2303.03327,
  title  = {Tight Bounds for $\gamma$-Regret via the Decision-Estimation Coefficient},
  author = {Margalit Glasgow and Alexander Rakhlin},
  journal= {arXiv preprint arXiv:2303.03327},
  year   = {2023}
}
R2 v1 2026-06-28T09:03:58.406Z