The data traffic demand is exponentially growing and becoming increasingly diverse. Indeed, future 5G mobile networks shall support multiple technical and service requirements, such as high throughput, low-latency, high reliability and availability, large storage capacity, cost-efficiency and operational energy-efficient features. That context brings many new challenges to develop suitable 5G technology. In order to cope with the conflictual requirements, future wireless 5G networks have to change in two ways. First, network deployments have to get dense (cells smaller and part of base station closer to the mobile user) and tightly coordinated (severe interference avoidance). Second, networks have to dynamically adapt to meet the diverse requirements of multiple applications. For this, it has to be flexible, intelligent, distributed and programmable. That is why, it has been suggested using for 5G: network function virtualization (NFV), xRAN architecture (that includes SDN principles) and network slicing in order to cope with multiple constraints and demands. Inspired from SDN approach, the xRAN architecture decouples the control plane from the data plane and split in independent components (centralized and distributed BBU and RRH/RRU) the traditional one size RAN. xRAN relies basically on software-defined access points that may be configured remotely in a similar way than SDN switches. Such a design approach enables to dynamically instantiate, update, manage and delete XRAN slices in a very cost-efficient way. xRAN combined with NFV and slicing driven by data analytics, will definitely provide more open, agile and evolving RAN characterized by a simple architecture, an independence between data delivery infrastructure and control functionalities, and surely reduced deployment and operational costs.
The idea of this phD is to offer a resilient efficient low-cost xRAN infrastructure by relying on data analytics (error prediction thus policy prediction) and on process automation (collect/monitoring/decision and achievement). To reach this goal we propose a dynamic resilience solution that combines advantages of reactive and predictive approaches, i.e., optimal resource (physical and virtual) utilization and timely reaction in a xRAN architecture. The resilience reconfiguration procedure will be running as a background Machine Learning-based automatic process during the life-cycle of a slice. Decisions will be taken based on massive data collected inside the network. Here, we identify key challenges on Machine Learning associated with this idea, concerning the phases of life-cycle of a slice. They are
(1) The network state forecasting: using Machine Learning techniques working on time-varying covariates and landmark analysis applied to discretized network state maps.
(2) The resilience update instant: using Machine Learning in order to classify criticality of anomalies (massive and isolated) and therefore to decide if it is worthy or no to launch an update cycle of a resilience solution.
(3) The reconfiguration of resilience solutions: The problem shall be formulated as a multi-objective optimization one and solved using heuristics inspired from AI techniques.
Ideal candidate should a solid mathematical background, especially in probability and statistics. The candidate should be eager to tackle new challenges in the area of machine learning and deep learning, distributed optimizations. He/she should have a good background in networking technology and familiar with LTE RAN and 5G. It is imperative that the candidate has a perfect proficiency in Python/C/C++ or tensorflow/keras, or other scripting languages. Lastly, the candidate should demonstrate a good motivation and autonomy. Good communication skills (written and verbal) in both English and French are required.