ML-based Monitoring and Management of QoE for User-centric Communication Networks (UserNet)

• Applicant: Professor Dr. Michael Seufert
• Subject Area: Security and Dependability, Operating-, Communication- and Distributed Systems
• Term: since 2022
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 500105691

Due to the best effort principle on the Internet, the desired subjectively perceived quality of a service (Quality of Experience, QoE) cannot be guaranteed at the same time for all users of different Internet applications. Thus, user-centric network management is necessary in case of bottlenecks to allocate the network resources, e.g., bandwidth, such that the largest possible number of users are satisfied and an allocation is reached, which is as fair as possible with respect to the resulting QoE (QoE fairness). However, the visibility of network operators on QoE is increasingly limited due to end-to-end encryption of data traffic.For that reason, the data-driven approach of machine learning (ML) is brought into focus to create fine-granular models, which represent the high complexity of the interplays between users, applications, and networks better than previous models. As has already been shown for video streaming, ML models can retrieve relevant QoE information despite encryption of data flows. In order to allow QoE monitoring for arbitrary Internet applications, the interplays between QoE and user interactions shall be investigated and modelled based on measurements and subjective studies. In addition, ML methods shall be adapted to the domain in order to apply them to encrypted network traffic. This allows to quantify the QoE by monitoring interactions and the resulting changes in the encrypted application traffic. Based on this, a data-driven improvement of QoE and QoE fairness shall be enabled by using reinforcement learning to find optimal network configurations by interacting with the dynamic network environment.By means of powerful, software-defined networking (SDN) technologies like P4, together with available computing resources in the network, such fine-granular models can now be implemented in the network for the first time, such that network management becomes more dynamic. Thus, the implementation of the required ML-based algorithms and components and their integration into network operation shall be researched. This applies to monitoring and flexible configuration of network components close to end users, e.g., at the Broadband Network Gateway or the Wide Area Network edge. There, the collaboration of ML-based models and new SDN technologies provides new opportunities for scalable concepts and technologies, which can monitor and proactively improve QoE and QoE fairness of the users of arbitrary Internet applications despite encryption. This allows to flexibly tailor communication networks to the requirements, such that QoE and QoE fairness of users improve, given the same resource usage. This complements the expansion of network infrastructure and allows network operators to cope with increasing requirements in communication networks.

• DFG Programme: Independent Junior Research Groups