Working paper Open Access
Kaloxylos, Alexandros; Gavras, Anastasius; Camps Mur, Daniel; Ghoraishi, Mir; Hrasnica, Halid
This white paper on AI/ML as enablers of 5G and B5G networks is based on contributions from 5G PPP projects that research, implement and validate 5G and B5G network systems.
The white paper introduces the main relevant mechanisms in Artificial Intelligence and Machine Learning currently investigated and exploited for 5G and beyond 5G networks. A family of neural networks is presented, which are generally speaking, non-linear statistical data modelling and decision making tools. They are typically used to model complex relationships between input and output parameters of a system or to find patterns in data. Feed-forward neural networks, deep neural networks, recurrent neural networks, and convolutional neural networks belong to this family. Reinforcement learning is concerned about how intelligent agents must take actions in order to maximize a collective reward, e.g. to improve a property of the system. Deep reinforcement learning combines deep neural networks and has the benefit that is can operate on non-structured data. Hybrid solutions are presented such as combined analytical and machine learning modelling as well as expert knowledge aided machine learning. Finally other specific methods are presented, such as generative adversarial networks and unsupervised learning and clustering.
In the sequel the white paper elaborates on use case and optimisation problems that are being tackled with AI/ML, partitioned in three major areas, namely: network planning, network diagnostics/insights, and network optimisation and control. In network planning, attention is given to the network element placement problem and to dimensioning considerations for C-RAN clusters. In network diagnostics, attention is given to forecasting network conditions, characteristics and undesired events, such as security incidents. Estimating user location is part of network insights. Finally, in network optimisation and control attention is given to the different network segments, including RAN, transport networks, fronthaul and backhaul, virtualisation infrastructure, end-to-end network slicing, security and application functions.
The white paper discusses the application of AI/ML in the 5G network architecture. In this context is identifies solutions pertaining to AI-based autonomous slice management, control and orchestration, AI/ML-based scaling operations in network service orchestration, AI/ML as a Service in network management and orchestration, enablement of ML for the verticals' domain, cross-layer optimization, management analytics in general, 3rd party ML analytics for network operation optimization in particular, anomaly detection using AI/ML. In the context of architecture it discusses the requirements for ML model lifecycle and interface management. Furthermore it investigates the global efforts for the enablement of AI/ML in networks, including the network data analytics function, the lack of availability of data-sets for training the AI/ML models and the associated privacy concerns. Finally, it identifies the challenges in view of trust in AI/ML-based networks and potential solutions such as the zero-trust management approach. The section concludes with a brief overview of AI/ML-based KPI validation and system troubleshooting.
In summary the findings of this white paper conclude that for enhancing future network return on investment the following areas need further attention (research and development work):
(a) building standardized interfaces to access relevant and actionable data,
(b) exploring ways of using AI to optimize customer experience,
(c) running early trials with new customer segments to identify AI opportunities,
(d) examining use of AI and automation for network operations, including planning and optimization,
(e) ensuring early adoption of new solutions for AI and automation to facilitate introduction of new use cases, and
(f) establish/launch an open repository for network data sets that can be used for training and benchmarking algorithms by all
|All versions||This version|
|Data volume||6.2 GB||6.2 GB|