Online Learning for Network Resource Allocation

# 168





Abstract

The proliferation of connected devices has fostered a diverse landscape of network applications, ranging from content delivery and interpersonal communication to intervehicular networks. Emerging use cases, such as autonomous driving, augmented reality, and tactile internet, demand stringent latency and bandwidth requirements from network resources. Network resource allocation (NRA) plays a significant role in catering to these demands while minimizing computational or communication overhead. By optimizing resource placement across the network, NRA aims to expedite service delivery and alleviate network congestion. However, NRA poses significant challenges in practical scenarios where network parameters, such as latency and operating costs, exhibit temporal variations and demands arrive unpredictably in a sequential manner. In small-cell mobile networks, for instance, the high and fluctuating user churn rate impedes efficient spectrum allocation to cells. Similarly, balancing latency gains across served areas by placing content files at edge caches is a complex task due to non-stationary and rapidly changing request patterns. Additionally, the increasing virtualization of network systems, as evidenced by extensive measurement studies, introduces cost and performance volatility. This uncertainty is exacerbated for services that process user-generated data, such as streaming data applications, where performance metrics, such as inference accuracy, hinge on a priori unknown input data and dynamically selected machine learning libraries. This seminar will explore various instances of the NRA problem, including exact caching, similarity caching, and inference delivery networks. Additionally, it will address the fairness dimension of NRA. By leveraging the powerful techniques of online learning and adversarial analysis, we will identify online policies that deliver robust performance guarantees even in unpredictable environments devoid of statistical regularity.

Dr. Tareq Si Salem, Northeastern University, Boston.

Dr. Tareq Si Salem is currently a Postdoctoral Research Associate at Northeastern University in Boston Massachusetts. He holds a Master of Science (M.Sc.) and a Doctor of Philosophy (Ph.D.) in Computer Science both obtained from Université Côte d'Azur in France in 2019 and 2022 respectively. During his doctoral studies, he was hosted by Inria Sophia Antipolis. Additionally, he held a visiting researcher position at Delft University of Technology (TU Delft) in 2022.       His research interests lie at the intersection of machine learning, networking, and modeling, with a focus on learning algorithms that satisfy system constraints such as privacy, safety, fairness, memory, and communication. His work has been published in top academic conferences and journals including the IEEE/ACM ToN, ACM ToMPECS, ACM SIGMETRICS, ITC, and IEEE INFOCOM. He also received the Best Paper Award at the 33rd International Teletraffic Congress (ITC'33) in 2021.