X. S T Y L I A N O U Vasilako

This project is designed for students interested in combining the Machine Learning (ML) and Multi-access Edge Computing (MEC) topics. Students who undertake this project will need study, select, train This project is designed for students interested in combining the Machine Learning (ML) and Multi-access Edge Computing (MEC) topics. Students who undertake this project will need to study, select, train, assess/compare appropriate ML models which will be able to predict user mobility based on time/location history data. The resulted ML models will be personalised for users (or groups of users). The final goal of this project is to use these ML models to proactively place Docker containers (or upon students missing skill, then VM(s)) that serve users to a MEC location. In any case, students are expected to be able to deliver an evaluation based on either real setup hands-on or simulation-based evaluation results, depending on lab access availability.  Communication Networks  Wireless Communications

Cognitive Network Management is a key concept for flexible service-oriented 5G communication networks [1], guaranteeing the Quality-of-Service (QoS) and even the Quality-of-Experience requirements defined in Service Level Agreements (SLAs). Nevertheless, managing (on a higher level) and controlling (on a lower level) virtualized resources at both the Core Network (CN) and Radio Access Network (RAN) is a complicated task [5][8]. The scale of parameters and alternative algorithmic solutions is beyond human management capacity. Therefore, collaborative Machine Learning (ML) solutions [5] have emerged (e.g., [6][7]) to tackle both the scale and the high network dynamics in 5G, while allowing humans to have a more efficient and effective role in 5G management. The focus of this project is to investigate a Methodology approach to Cognitive Network Management motivated by the needs of standardizing, orchestrating and automating all the necessary steps and actions for building and deploying efficient ML models as collaborative components of an integrated Cognitive Network Management system.

Virtualized resource allocation is a key flexibility enabler towards future service-oriented 5G communication networks [1]. The virtualization of both the Core Network (CN) and Radio Access Network (RAN) must be such as to provide tenants (a.k.a. the slice owners) with the necessary resources that satisfy their Quality-of-Service (QoS) or even Quality-of-Experience requirements from physical Network Operators (NOs). The focus of this project is to investigate the allocation of virtualized network resources from an End-to-End (E2E) perspective, spanning both the CN and RAN where users reside and consume 5G services. The problem is highly practical (even though based on an interesting analytical background [2][3]) of major importance for future 5G services. The ultimate goal is to improve performance by moving or replicating applications, services or/and content with respect to the network edge points where user demand comes. Within this context, different types of resources like memory, storage, link bandwidth and/or latency, CPU cores and/or cycles, etc., have to be allocated together to form a new service node or for migrating an existing service node to another network point. Specific investigation goals This project is aimed at students interested in Machine Learning (ML) solutions for 5Gcognitive resource management. Within the context of this, the students will have the opportunity to  Experience the design of resource allocation hybrid models inspired by reinforcement learning methods and, specifically, Q-learning [4] as well as  congestion pricing/cost-based models along the lines of the work in [5][6].  Investigate, elaborate and evaluate the design of the hybrid models tailored for tackling multi-type resource allocation.  Reinforcement Leaning (RL), either with Q-learning or-upon student's skills-Deep RL (DRL)  Get in know important concepts that dominate the current and future 5G industry such as Multi-access Edge Computing (MEC) and resource Slicing in 5G.

Cognitive Network Management is a key concept for flexible service-oriented 5G communication networks [1], guaranteeing the Quality-of-Service (QoS) and even the Quality-of-Experience requirements defined in Service Level Agreements (SLAs). Nevertheless, managing (on a higher level) and controlling (on a lower level) virtualized resources at both the Core Network (CN) and Radio Access Network (RAN) is a complicated task [5][8]. The scale of parameters and alternative algorithmic solutions is beyond human management capacity. Therefore, collaborative Machine Learning (ML) solutions [5] have emerged (e.g., [6][7]) to tackle both the scale and the high network dynamics in 5G, while allowing humans to have a more efficient and effective role in 5G management. The focus of this project is to investigate a Methodology approach to Cognitive Network Management motivated by the needs of standardizing, orchestrating and automating all the necessary steps and actions for building and deploying efficient ML models as collaborative components of an integrated Cognitive Network Management system. Minimum investigation goals This project expects the interest of students who wish to work on Machine Learning (ML) solutions for 5G cognitive management. Within the context of this, the students will have the opportunity to: • Get experience with Neural Networks and Reinforcement Learning. • Validate a methodology approach from designing up to deploying a resource allocation model after the popular Deep Reinforcement Learning [4] paradigm. • Provided that lab access conditions allow it, get hands-on experience with performing an experimental evaluation of the proposed model; or/and get experience with network simulator environments. • Investigate, elaborate and evaluate the design of 5G cognitive models tailored for tackling more than just single-type resource allocation (multi-objective optimization). • Get involved with important 5G concepts such as Multi-access Edge Computing (MEC) and Network Slicing.

Cognitive Network Management is a key concept for flexible service-oriented 5G communication networks [1], guaranteeing the Quality-of-Service (QoS) and even the Quality-of-Experience requirements defined in Service Level Agreements (SLAs). Nevertheless, managing (on a higher level) and controlling (on a lower level) virtualized resources at both the Core Network (CN) and Radio Access Network (RAN) is a complicated task [5][8]. The scale of parameters and alternative algorithmic solutions is beyond human management capacity. Therefore, collaborative Machine Learning (ML) solutions [5] have emerged (e.g., [6][7]) to tackle both the scale and the high network dynamics in 5G, while allowing humans to have a more efficient and effective role in 5G management. The focus of this project is to investigate a Methodology approach to Cognitive Network Management motivated by the needs of standardizing, orchestrating and automating all the necessary steps and actions for building and deploying efficient ML models as collaborative components of an integrated Cognitive Network Management system.

Grid systems tend to grow in size, but currently deployed state-of-the-art schedulers have inherent scalability limits due to centralization and high messaging cost. In this paper, we explore the feasibility of scalable grid scheduling using a peer-to-peer overlay. We propose DGSASAP, a decentralized scheduling algorithm that schedules compute-intensive jobs such that their execution starts as soon as possible. Simulations of