We gratefully acknowledge support from
the Simons Foundation and member institutions.
Full-text links:


Current browse context:


Change to browse by:

References & Citations


(what is this?)
CiteULike logo BibSonomy logo Mendeley logo del.icio.us logo Digg logo Reddit logo ScienceWISE logo

Mathematics > Statistics Theory

Title: Refined Lower Bounds for Adversarial Bandits

Abstract: We provide new lower bounds on the regret that must be suffered by adversarial bandit algorithms. The new results show that recent upper bounds that either (a) hold with high-probability or (b) depend on the total lossof the best arm or (c) depend on the quadratic variation of the losses, are close to tight. Besides this we prove two impossibility results. First, the existence of a single arm that is optimal in every round cannot improve the regret in the worst case. Second, the regret cannot scale with the effective range of the losses. In contrast, both results are possible in the full-information setting.
Subjects: Statistics Theory (math.ST); Machine Learning (cs.LG); Machine Learning (stat.ML)
Journal reference: D. D. Lee; M. Sugiyama; U. V. Luxburg; I. Guyon; R. Garnett. NIPS 2016, Dec 2016, Barcelona, Spain. Curran Associates, Inc., pp.1198--1206, Advances in Neural Information Processing Systems 29
Cite as: arXiv:1605.07416 [math.ST]
  (or arXiv:1605.07416v2 [math.ST] for this version)

Submission history

From: Sebastien Gerchinovitz [view email]
[v1] Tue, 24 May 2016 12:36:47 GMT (22kb)
[v2] Mon, 27 Feb 2017 13:48:10 GMT (23kb)

Link back to: arXiv, form interface, contact.