This application concerns
Machine Learning (ML) based beam prediction applied at a UE. This involves measuring only a subset of beams and then applying a ML model to predict the best beams or beam pairs. However, it is important that the model is trained under configurations that closely resemble the actual conditions which the UE will operate in; there must be consistency between training and
inference. The UE receives 204 a data collection configuration comprising a plurality of lists of
channel state information (CSI) resource configurations. One of the lists is associated with the identity of the serving
cell, which may be represented by a
cell global identity (CGI). Each
list comprises a plurality of resource configurations (308-312, Fig. 3). The resource configurations are associated with respective network side additional condition identifiers (NW-additional-condition-IDs). The UE may
train several different ML models each associated with a respective NW-additional-condition-ID. At
inference, the UE applies 224 an ML model associated with both the serving
cell identity and a NW-additional-condition-ID communicated by a network device in a CSI reporting configuration message 222.