AI/ML in the RAN — Learning inside the radio, and at what timescale
Procedures over time
Where it sits
What it is
What it is. Learned models making decisions the specification leaves open. This is the folder's layer 4, and the honest position is that most of it is not in a frozen specification yet — but one piece now is.
In the specification, today. csi-InferencePrediction-r19 (TS 38.214 clause 5.2.1.4.2), with
the predicted-PMI codebooks of clauses 5.2.2.2.10–.11. A device reports precoders for slots that
have not happened yet, and the standard does not say how it should arrive at them.
PDSCH §13.2 quotes it.
The pattern is the same as link adaptation. 3GPP fixes the vocabulary — what may be reported, in what format, when — and leaves the algorithm to the implementer. What is different in Release 19 is that the thing left open is a prediction of a time series, which is a shape a learned model fits and a rule of thumb does not.
The timescale question is the one that matters, and it is what separates this from the RIC. A
predicted precoder is useful over a few slots — hundreds of microseconds. The O-RAN RIC's xApps
operate at 10 ms to 1 s and rApps above 1 s. They cannot be the same mechanism, and working
out which decisions live where is the subject of ran-ric-and-ai.md.
The study behind it is 3GPP TR 38.843, whose three use cases are CSI feedback compression, beam management and positioning. Not in the archive; cited as a pointer only, as PDSCH §15.2 says.
Read on
This concept was first written up in ref-system, which reads the whole group as one argument.
Before this concept, the hierarchy says to learn the following — the full chain, in order: