Support of network slicing in NG-RAN is defined in TS 38.300. The NG-RAN plays a key role in taking mobility, load balancing and Radio Resources Management decisions for the purpose of meeting target requirements derived from the SLA of each supported network slice. AI/ML function can analyze metrics related to network and UE level performance related to perform optimal resource management and mobility decisions for network slicing to meet the requirements.

The objective of NR Coverage and Capacity Optimization (CCO) function is to detect and resolve or mitigate CCO issues. An NG-RAN node may autonomously adjust within and switch among coverage configurations. When a change is executed, a NG-RAN node may notify its neighbour NG-RAN nodes with the list of cells and SSBs with modified coverage included. In the legacy CCO solution, a reactive approach is used: when the gNB (gNB-CU in case of CU-DU split architecture) detects a CCO issue which negatively impacts network and UE performance after it has already occurred, the gNB (gNB-DU in case of CU-DU split architecture) attempts to resolve or mitigate it. With an AI/ML based CCO, a more proactive approach is used to prevent (or limiting at an early stage) the rise of a CCO issue with the consequent degradation of network (and UE) performance.

Mobility in NR-DC can be optimized by means of AI/ML. Mobility Optimization for NR-DC is studied by assuming inference at the MN only. The main use case is limited to Dual Connectivity only and Conditional Dual Connectivity procedures are out of scope.

The Dual Connectivity procedures (e.g., SN Addition, MN-initiated SN Change) are enhanced to trigger the collection of measured UE performance.

The split architecture should be enhanced to support the Rel-18 use cases, e.g, Load Balancing, Energy Saving, and Mobility Optimization. In case of CU-DU architecture, the following solutions are possible:

• AI/ML Model Training is located in the OAM and AI/ML Model Inference is located in the gNB-CU(-CP);
• AI/ML Model Training and Model Inference are both located in the gNB-CU(-CP).

The following standard impacts are listed for subsequent Rel-19 normative work compared with what was specified during Rel-18:

• The details of signaling measured UE performance metrics from gNB-DU to gNB-CU-CP and/or from gNB-CU-UP to gNB-CU-CP need further discussion during normative phase.
• Measured Energy Cost from gNB-DU to gNB-CU.

A description of the AI/ML-based Network Energy Saving use case as such is presented in TR 37.817. Based on the conclusions in TR 37.817, a metric called Energy Cost was introduced during Rel-18 normative work; this metric is defined as an index representing the energy consumption at the NG-RAN node and can be exchanged between NG-RAN nodes over the Xn interface upon request with a per NG-RAN node granularity. However, a node-level Energy Cost metric might be too coarse to provide a good understanding of the energy impact of AI/ML-based Network Energy Saving actions in NG-RAN nodes handling multiple cells, so possible solutions to improve Energy Cost granularity could be beneficial.

During Rel-18 normative work, the Energy Cost (EC) metric was defined on an NG-RAN node-level granularity even though the AI/ML Energy Saving actions that we assumed in Rel-18 were on a cell-level granularity, e.g., cell switch-off actions. Enhancements to the node-level EC to a finer granularity could be beneficial, assuming that such enhancements enable network-level energy saving. The following approaches to improve the EC granularity were discussed:

1. EC per group of Cells based on energy consumption measurements associated to the hardware serving the group of cells
2. EC per Cell based on energy consumption estimations
3. EC per one or more HO event based on energy consumption estimations
4. EC per gNB-DU based on energy consumption measurements associated to the hardware serving the gNB-DU

The problem of continuous data collection for management-based MDT can be described as follows: a UE in the NG-RAN can be configured with management-based Logged MDT when in RRC_Idle and RRC_Inactive states and with management-based Immediate MDT when in RRC_Connected state. Differently from signalling-based MDT, in management-based MDT, a UE is not uniquely identified in the MDT activation. Therefore, when a UE transits to RRC_Connected state from RRC_Idle/RRC_Inactive (during which Logged MDT data have been collected) or when a UE is handed over between gNBs, the network does not have standardized means to select again the same UE for continuous MDT for subsequent MDT data collection. The Data Collection continuity in this scenario can be split into two tasks as below:

• Problem A (measurement continuity): how to ensure that the same UE collecting MDT measurements during the same RRC state and across different RRC states.
• Problem B (trace correlation): how to ensure that the TCE which eventually receives the MDT reports can associate the received logged and immediate MDT measurements to a continuous data collection period from the same UE.

Potential solutions addressing the problems above can be discussed during normative phase.

In Rel-18, the cell-based UE trajectory prediction is limited to the first-hop target NG-RAN node. Multi-hop predicted UE trajectory across gNBs consists of a list of cells belonging to one or more gNBs where the UE is expected to connect and these cells are listed in chronological order.