Artificial Intelligence and Dynamic Modeling tools for Future Flexible Networks – AIDY-F2N

Davidson Paris and Telecom SudParis wish to strengthen their links by setting up the ANR AIDY-F2N LabCom, a joint research laboratory dedicated to Dynamic Modeling for Next Generation networks and services. This initiative reflects the aspiration of both parties to develop and strengthen their know-how in these areas.

AIDY-F2N will naturally enhance the synergy between public research, often theoretical, and industrial R&D activities inspired and guided by short and medium terms market needs. In particular, the joint laboratory AIDY-F2N aims to increase the knowledge of both parties in the field of analysis and optimization of telecommunication networks and services. The objective is to propose artificial intelligence tools and methodologies in order to model and understand, in real time, the behavior of these systems and predict their evolution. This will allow finding new means and tools to ease autonomous adaptation and reconfiguration of complex systems on the fly.

Next-generation networks and services (5G) are expected to be modular, programmable and shared. Access and Transport networks’ architectures are moving towards more virtualization, pooling and cooperation. They will use different technologies but also serve several actors and trades. Consequently, they will carry and offer heterogeneous services in terms of throughput and QoS requirements. 5G solutions are expected to be more complex to manage and traditional management means are no longer adapted to their complexity and richness. Therefore, there is a need for new tools to analyze and predict the evolution of these dynamic and complex systems and to allow them to self-adapt and reconfigure without or with minimum human intervention.

In this project, we will explore several optimization techniques including decomposition and aggregation techniques, to solve in a simpler and faster way complex problems, combinatorial optimization, stochastic bounds and learning techniques (Machine Learning and Deep Learning).
A specific effort will be done to ensure that the proposed approaches can be used in real time to allow quick decisions and reconfiguration tailored to the operational needs of 5G. These modeling and optimization solutions can then be part of the real systems itself and fed by the performance indicators and the measurements collected at low and high levels, from radio and access networks to transport and service layers.
An example to deal with in these new networks is the real-time prediction of machine-to-machine and more generally of Internet of Things traffic. This type of traffic can be very sporadic but sometimes massive (e.g. in case of critical situations. These dynamics may lead to the degradation of the QoS, not only for the considered service, but also for all other services that may share the same physical and software resources. Adaptation and reconfiguration actions may be required at the access and transport networks but also at the service level. Therefore, the system must be able to predict load evolution and get ready for it by self-re configuring (at different levels).

Another example is the dynamic orchestration of virtual networks. The new standards have proposed for several service types, a new architecture based on what called network slices which consists of virtual networks that are created and managed using Network Function Virtualization (NFV) and Software Defined Network (SDN). There is a big challenge on placement of service components on the physical infrastructure in order to create functional isolated slices satisfying services with different requirements such as (sensor networks, autonomous driving, interactive applications, broadcast...). The set up and the orchestration of these slices has to be done dynamically taking into account predictions on the network load and service types demand using machine learning techniques and real time self-reconfigurable in case of unpredictable events that may affect the network.