Online Reusable Resource Allocation with Adversarial Requests

Yaolong Yu, Hanrui Zhang, Zeyu Zheng, Sinno Jialin Pan

Keywords: Reusable resource allocation, online learing, adversarial environment, cloud computing service

Abstract:

We study an online reusable resource allocation problem with adversarial inputs, where a platform must decide, over a horizon $T$, whether to accept incoming job requests with adversarial resource demands and durations. The goal is to maximize cumulative revenue subject to resource budget constraints.
We propose a class of \textit{online dual dynamic learning algorithms} and \textit{learning from pricing experts algorithms} that achieve asymptotically optimal competitive ratios, remain computationally efficient, and further improve performances across different regimes of maximum job duration and resource demand. We further extend the model to allow flexible resource allocations, where neither the demand nor the duration is fixed; instead, the amount of allocated resource influences job duration. To address this more general setting, we reduce the problem to the binary allocation case via resource discretization. We prove that the resulting loss is bounded by a constant depending only on the total budget, and this bound is nearly optimal.

Primary Area: learning theory

Submission Number: 4204