Computer Science > Computer Science and Game Theory
[Submitted on 23 Feb 2022 (this version), latest version 17 Oct 2022 (v2)]
Title:Nonstationary Dual Averaging and Online Fair Allocation
View PDFAbstract:We consider the problem of fairly allocating items to a set of individuals, when the items are arriving online. A central solution concept in fair allocation is competitive equilibrium: every individual is endowed with a budget of faux currency, and the resulting competitive equilibrium is used to allocate. For the online fair allocation context, the PACE algorithm of Gao et al. [2021] leverages the dual averaging algorithm to approximate competitive equilibria. The authors show that, when items arrive i.i.d, the algorithm asymptotically achieves the fairness and efficiency guarantees of the offline competitive equilibrium allocation. However, real-world data is typically not stationary. One could instead model the data as adversarial, but this is often too pessimistic in practice. Motivated by this consideration, we study an online fair allocation setting with nonstationary item arrivals. To address this setting, we first develop new online learning results for the dual averaging algorithm under nonstationary input models. We show that the dual averaging iterates converge in mean square to both the underlying optimal solution of the ``true'' stochastic optimization problem as well as the ``hindsight'' optimal solution of the finite-sum problem given by the sample path. Our results apply to several nonstationary input models: adversarial corruption, ergodic input, and block-independent (including periodic) input. Here, the bound on the mean square error depends on a nonstationarity measure of the input. We recover the classical bound when the input data is i.i.d. We then show that our dual averaging results imply that the PACE algorithm for online fair allocation simultaneously achieves ``best of both worlds'' guarantees against any of these input models. Finally, we conduct numerical experiments which show strong empirical performance against nonstationary inputs.
Submission history
From: Luofeng Liao [view email][v1] Wed, 23 Feb 2022 16:50:24 UTC (227 KB)
[v2] Mon, 17 Oct 2022 16:53:44 UTC (1,418 KB)
References & Citations
Bibliographic and Citation Tools
Bibliographic Explorer (What is the Explorer?)
Litmaps (What is Litmaps?)
scite Smart Citations (What are Smart Citations?)
Code, Data and Media Associated with this Article
CatalyzeX Code Finder for Papers (What is CatalyzeX?)
DagsHub (What is DagsHub?)
Gotit.pub (What is GotitPub?)
Papers with Code (What is Papers with Code?)
ScienceCast (What is ScienceCast?)
Demos
Recommenders and Search Tools
Influence Flower (What are Influence Flowers?)
Connected Papers (What is Connected Papers?)
CORE Recommender (What is CORE?)
arXivLabs: experimental projects with community collaborators
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.