Machine learning (ML) pipeline node: Each functionality in the ML pipeline is defined as a node (e.g., source, collector, pre-processor, model, policy, distributor and sink).
- src (source): This node is the source of data that can be used as input for the ML function. Examples of an src are: user equipment (UE), session management function (SMF), access and mobility management function (AMF) or any other entity in the network, including an application function (AF).
- C (collector): This node is responsible for collecting data from the src. It can use specific control protocols to configure the src. Example: it may use the 3rd Generation Partnership Project (3GPP) radio resource control (RRC) protocol to configure user equipment (UE) acting as an src. It may use vendor specific operations, administration and maintenance (OAM) protocols to configure an AMF acting as an src.
- PP (pre-processor): This node is responsible for cleaning data, aggregating data or performing any other pre-processing needed for the data so that are is in a suitable form for the ML model to consume it.
- M (model): This is an ML model. Example could be a prediction function.
- P (policy): This node provides a control for an operator to put a mechanism to minimize impacts into place on a live network, so that operation is not impacted. Specific rules can be put in place by an operator to safeguard the sanity of the network, e.g., major upgrades may be done only at night time or when traffic is low.
- D (distributor): This node is responsible for identifying the sinks and distributing the ML output to the corresponding sinks. it may use 3GPP RRC protocol to configure a UE acting as a sink.
- Sink: This node is the target of the ML output, on which it takes action, e.g., a UE adjusting the measurement periodicity based on ML output.
NOTE – The nodes are logical entities that are proposed to be managed in a standard manner (by a machine learning function orchestrator (MLFO) and hosted in a variety of network functions (NFs).
The realization of such an ML pipeline (in, say, 3GPP release 16 (R16) or R17 networks) will result in a standard method of introducing and managing ML functionality in a 5G network.
Auto214
References
Focus group on Machine Learning for Future Networks including 5G
Unified architecture for ML in 5G and future networks
A Flexible Machine Learning-Aware Architecture for Future WLANs