Dr. Konstantinos Katsaros, Senior Researcher / Team Leader – Intelligent Networks & Services Team, I-SENSE Group, Institute of Communication & Computer Systems (ICCS)
Recent years have witnessed a boom in the adoption of AI/ML technology, largely fueled by data availability, as well as HW and SW advances in the field. The telecommunications sector has joined this stream actively developing solutions on various fronts, largely focusing on the support of autonomous operations, such as self-configuration, self-monitoring, self-healing and self-optimization; that is, without further human intervention. While emphasis has been placed on allowing Mobile Network Operators (MNOs) to simplify their operations and bring down their operational expenditure (e.g., intelligent energy management, server consolidation, etc.), other efforts have been targeting the interaction of the vertical services/applications with the network. In this context, Predictive QoS (PQoS) has been considered as a promising approach in enabling the graceful adaptation of vertical services/application to the anticipated conditions in the network. Applying analysis on historical and contextual data that describe network performance and conditions, PQoS enables the proactive identification of significant fluctuations and the generation of corresponding notifications (i.e., In-advance Quality of Service Notifications, IQNs) towards the vertical service/application. These notifications subsequently allow the vertical service/application to adjust its operation within the anticipated environment. A typical example comes from remote control services in the broader automotive sector, where a vehicle (passenger car, truck, crane, vessel) is teleoperated by a (human) operator in a remote control room: with PQoS, the network predicts that network latency will drastically increase (or supported data rates will drop below a threshold value) in a time horizon of 10 secs and issues a notification towards the control room application SW; as a result, control is handed back to a passenger, or to the autonomous driving SW stack of the vehicle or simply the vehicle is slowed down or stopped.
Industry and academia are actively exploring the technoeconomic solution space of PQoS. On the standardization front, 3GPP Rel-17 specifications support the provisioning of predictive QoS sustainability analytics to vertical service providers, based on historic network monitoring data, through the invocation of the network data analytics function (NWDAF); while the architecture further also foresees interfaces and components for the support of generic ML model management by the MNO. At the same time, the automotive vertical industry also seeks over-the-top (OTT) solutions i.e., solutions that are based on data observations available at the application layer. This state-of-affairs reveals the importance of data ownership and exposure also on a business/economic level: MNOs and vertical service providers have strong business interest to not share their data, seeking stand-alone solutions that can increase revenue (e.g., for PQoS provisioning by MNOs) or reduce costs (e.g., by employing competitive OTT solutions). However, research has shown that multi-featured data sets i.e., data sets combining observation from multiple locations and administrative domains, yield substantially superior performance in terms of accuracy, rendering stand-alone solutions sub-optimal. As a result, a series of new technological approaches are currently being investigated, including for instance Federated Learning, Split Learning and Transfer Learning techniques, which realize various forms of collaborative, yet privacy-preserving ML training process. The industry appears to follow up on this trend, with 3GPP Rel-17 specifications providing preliminary support for primitive Federated Learning operations.
In all, a multi-faceted and fast pacing environment is being developed for Predictive QoS, where research and innovation are driving developments and start to shape new business relationships and interactions centered around the value of data.