REU Supplement: CRII: CSR: Federated Resource Management in Mobile Edge Computing

Grants and Contracts Details


CSR: REU Supplement to “CRII: CSR: Federated Resource Management in Mobile Edge Computing” PI: Dr. Hana Khamfroush Department of Computer Science, University of Kentucky 1 Summary of NSF-Funded Project The recently awarded grant proposal, entitled “CRII: CSR: Federated Resource Management in Mobile Edge Computing” (Federal Award ID Number: 1948387), proposed the incorporation of federated learning (FL) technique to resource management in mobile edge computing (MEC) systems to provide a better trade- off between user quality-of-service (QoS) and the system’s costs. Brie?y, MEC is a framework to provide a second layer of computational devices deployed at the edge of a computer network rather than solely rely- ing on a central cloud server. The purpose is to provide computationally intense services (e.g., smart voice assistants) to users with reduced communication latency due to reduced physical proximity to the user. FL is a recently proposed approach to performing distributed learning in MEC environments. Succinctly, FL considers a central, global machine learning model (e.g., deep arti?cial neural network for image classi?ca- tion) that is stored in the cloud and sent to all edge clouds in the MEC system. Then, each edge cloud trains and updates its local model with the data they receive to improve their local model’s performance accuracy. At certain points in time, the central cloud will select a proportion of edge clouds to send their local model parameters to the central cloud. The central cloud then essentially averages the parameters among the re- ceived parameters (known as a federated average) to update its global model. It then sends the new global model parameters to each edge cloud. The bene?t of FL is that it protects user privacy by not sending user data to the central cloud, only the model parameters. The proposal, in short, is interested in studying theoretical bounds and practical applications of FL and distributed learning techniques to the problem of resource management in MEC while considering a litany of problems. More speci?cally, this work is inter- ested in joint optimization of service placement and request scheduling for a three-tier MEC system (layers include, a central cloud, a layer of multiple edge clouds, and a set of mobile users sending requests). We propose to ?rst ?nd theoretical bounds through mathematical formulation and algorithm design (Thrust I), then ?nding more practical and adaptive sloutions using multi-agent reinforcement learning and the feder- ated learning framework (Thrust II), and ?nally comparing the results of Thrusts I and II through numerical analysis and implementation on a MEC test-bed. 2 Undergraduate Mentorship History and Outcomes In the past few years, I have had the pleasure of mentoring undergraduate students on both theoretical and application-based research projects during my postdoc position at Penn State University and later on at the University of Kentucky as an assistant professor. In the following, I brie?y describe the research areas and the outcomes of my mentorship for undergrad students. Mentorship expereince at the University of Kentucky. Throughout my short time at the University of Kentucky, I have worked with several undergraduate students on different research projects. Emory Huf- baeur is a former undergraduate student that has worked closely alongside myself and my PhD student, Nathaniel Hudson, for more than a year now. Emory ?rst worked on developing a theoretical model for generating synthetic and random network topologies that closely resemble the structure of real-world on- line social networks (OSNs). This work has led to a conference paper that has been published at the IEEE 1
Effective start/end date3/1/208/31/23


  • National Science Foundation


Explore the research topics touched on by this project. These labels are generated based on the underlying awards/grants. Together they form a unique fingerprint.