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

Download:

Current browse context:

cs.LG

Change to browse by:

References & Citations

DBLP - CS Bibliography

Bookmark

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

Computer Science > Machine Learning

Title: Optimal Regularized Online Convex Allocation by Adaptive Re-Solving

Abstract: This paper introduces a dual-based algorithm framework for solving the regularized online resource allocation problems, which have cumulative convex rewards, hard resource constraints, and a non-separable regularizer. Under a strategy of adaptively updating the resource constraints, the proposed framework only requests an approximate solution to the empirical dual problem up to a certain accuracy, and yet delivers an optimal logarithmic regret under a locally strongly convex assumption. Surprisingly, a delicate analysis of dual objective function enables us to eliminate the notorious loglog factor in regret bound. The flexible framework renders renowned and computationally fast algorithms immediately applicable, e.g., dual gradient descent and stochastic gradient descent. A worst-case square-root regret lower bound is established if the resource constraints are not adaptively updated during dual optimization, which underscores the critical role of adaptive dual variable update. Comprehensive numerical experiments and real data application demonstrate the merits of proposed algorithm framework.
Subjects: Machine Learning (cs.LG); Information Theory (cs.IT); Optimization and Control (math.OC)
Cite as: arXiv:2209.00399 [cs.LG]
  (or arXiv:2209.00399v1 [cs.LG] for this version)

Submission history

From: Dong Xia [view email]
[v1] Thu, 1 Sep 2022 12:23:26 GMT (1966kb,D)

Link back to: arXiv, form interface, contact.