Abstract
The virtualization of wireless networks enables new services to access network resources made available by the Network Operator (NO) through a Network Slicing market. The different service providers (SPs) have the opportunity to lease the network resources from the NO to constitute slices that address the demand of their specific network service. The goal of any SP is to maximize its service utility and minimize costs from leasing resources while facing uncertainties of the prices of the resources and the users' demand. In this paper, we propose a solution that allows the SP to decide its online reservation policy, which aims to maximize its service utility and minimize its cost of reservation simultaneously. We design the Optimistic Online Learning for Reservation (OOLR) solution, a decision algorithm built upon the Follow-the-Regularized Leader (FTRL), that incorporates key predictions to assist the decision-making process. Our solution achieves a $\mathcal{O}(\sqrt{T})$ regret bound where $T$ represents the horizon. We integrate a prediction model into the OOLR solution and we demonstrate through numerical results the efficacy of the combined models' solution against the FTRL baseline.
1. Introduction
In the landscape of 5G and beyond, Network Slicing has emerged as a critical paradigm, allowing Service Providers (SPs)—ranging from high-bandwidth streaming platforms to low-latency autonomous vehicle fleets—to lease virtualized resources such as bandwidth, CPU, and storage from a Network Operator (NO).
However, the market for these resources is inherently volatile. SPs must commit to resource reservations before observing actual user demand or the "spot price" of the resources. Traditional resource management often assumes a stationary environment where statistical properties remain constant over time. In contrast, real-world network traffic is notoriously non-stationary and unpredictable. This paper addresses the challenge of real-time decision-making for SPs who must maximize service utility and minimize leasing costs without the benefit of perfect foresight.
2. Methodology: The OOLR Algorithm
We frame the reservation problem as an Online Convex Optimization (OCO) problem. Our primary contribution is the Optimistic Online Learning for Reservation (OOLR) algorithm. OOLR extends the standard "Follow-the-Regularized-Leader" (FTRL) framework by introducing an "optimistic" component that anticipates future market states.
2.1 Optimistic FTRL
While standard OCO algorithms are purely reactive, looking only at past performance, OOLR incorporates a prediction of the gradient for the upcoming time step. By integrating this "hint" about the future, the algorithm can adjust its reservation strategy more rapidly in response to emerging trends. To maintain stability, the algorithm utilizes a sequence of proximal quadratic regularizers that adapt based on the accuracy of these predictions.
2.2 ARMA-OGD Prediction Module
To generate the necessary gradient predictions, we employ an Auto-Regressive Moving Average (ARMA) model updated via Online Gradient Descent (OGD). This approach offers several advantages:
- Efficiency: It is computationally lightweight, with a time complexity of $O(mq)$, making it suitable for real-time deployment at the network edge.
- Autonomy: It requires no expensive offline training phase, learning directly from the data stream as it arrives.
3. Theoretical Guarantees
A key feature of the OOLR framework is its mathematical robustness, often described as a "best of both worlds" guarantee. The algorithm's performance is measured by "regret"—the difference between the algorithm's cumulative cost and the cost of the best fixed strategy in hindsight.
- If the predictions are highly accurate, OOLR achieves a constant regret bound of $O(1)$.
- If the predictions are entirely incorrect, the algorithm gracefully degrades to the standard $O(\sqrt{T})$ regret bound.
This ensures that while the SP stands to gain significantly from accurate predictions, they are not penalized beyond the standard baseline if the environment becomes too chaotic to predict.
4. Experimental Evaluation
We evaluated OOLR using a combination of real-world and synthetic datasets to simulate a complex, non-stationary market.
- Real-World Traffic: We utilized aggregated traffic volumes from multiple base stations in Shanghai, recorded at 10-minute intervals over a one-month period.
- Synthetic Market Conditions: Resource prices were modeled using an AR(1) process (specifically an Ornstein-Uhlenbeck process), while resource utility was modeled using seasonal sine waves with stochastic noise to mimic fluctuating demand patterns.
4.1 Key Results
Numerical simulations demonstrate that the OOLR solution (specifically the version utilizing the ARMA-OGD predictor) consistently outperforms the standard FTRL baseline. The ARMA-OGD module successfully tracked the intricate, non-stationary signals within the Shanghai traffic data, allowing the SP to anticipate demand spikes effectively. Furthermore, OOLR proved robust in scenarios of "partial fulfillment," where the Network Operator could only meet a fraction of the Service Level Agreement (SLA) due to network congestion.
5. Conclusion and Future Work
OOLR provides a computationally efficient and theoretically sound framework for resource reservation in dynamic network slices. By combining optimistic online learning with lightweight ARMA-OGD predictions, SPs can navigate volatile markets with greater confidence.
Future research will explore the impact of strategic competition, where multiple SPs compete for the same pool of resources. Additionally, while ARMA-OGD offers excellent efficiency, we aim to investigate the integration of more complex neural network architectures into the OOLR framework to further enhance prediction accuracy in highly chaotic environments.
Summary updated