Conclusions
In this contribution, we present our views on specification support for AI/ML-based beam management. Based on the discussions in the previous sections we propose the following:
Proposal 1: For Rel-19 AI/ML-based BM, for BM-Case1 and BM-Case2 with a UE-sided AI/ML model, for Option 2 (UE-assisted performance monitoring), there is no need to support other alternatives in addition to the already agreed Alt 1: Top 1 or Top K beam prediction accuracy (with or without margin).
The full set of Set A is configured as the resource set/resources for monitoring, preferably with a much longer period than Set B to reduce overhead.
Proposal 2: For Rel-19 AI/ML-based BM, at least for BM-Case1 with a UE-sided AI/ML model, for Option 2 (UE-assisted performance monitoring), introduce a new quantity (i.e., beam accuracy indicator, BAI) for beam prediction accuracy in the CSI report for monitoring, where BAI is Np (0 ≤ Np≤ N), N is implicitly or explicitly configured by NW as the number of reported inference result(s) linked with performance monitoring instance(s), and Np is the number of reported inference result(s) linked with performance monitoring instance(s), for which at least one of the Top M beam(s) of the resource set(s) for monitoring is among Top-K predicted beam(s),
Where K is the number of predicted beam(s) in the corresponding inference report
Where Top M beam(s) is the beam(s) with largest measured value(s) of L1-RSRP(s) of the resource set(s) for monitoring
M is configured by NW in CSI report configuration for monitoring
M=1
FFS: other values
Proposal 3: For Rel-19 AI/ML-based BM, at least for inference for network-sided AI/ML model of BM-Case1, for the content in a beam report in L1 signaling,
Support L1-RSRPs and corresponding beam information of up to M beams within X dB gap to the largest measured value of L1-RSRP, X and M are configured by gNB, and the number of reported beams is indicated in the beam report.
The beam information is CRI/SSBRI.
Increase the maximum value M from the existing “4” to “8” as a starting point.
Proposal 4: For Rel-19 AI/ML-based BM, for UE-sided model, at least for BM-Case1, for content in the report of inference results, do NOT support Opt 3 and Opt 4.
Opt 3: Beam information on predicted Top K beam(s) among a set of beams and probability information of predicted Top K beam(s) among a set of beams
Opt 4: Beam information on predicted Top K beam(s) among a set of beams, RSRP of predicted Top K beam(s) among a set of beams, and confidence information of the RSRP
Proposal 5: For Rel-19 AI/ML-based BM, for data collection for UE-sided AI/ML model of BM-Case1 and BM-Case2, support that NW provides/signals multiple possible configurations of DL RS transmission to the UE and the UE reports its supported/preferred one(s) out of the multiple configurations.
Proposal 6: For Rel-19 AI/ML-based BM, for data collection for AI/ML model of BM-Case1 and BM-Case2, support using RS ID as implicit indication of beam ID and reusing legacy L1-RSRP reporting as much as possible.
Proposal 7: Support to separate CPU Counting for Legacy and AI/ML-based CSI Reporting, i.e., legacy CPU and AI/ML-based CPU are from different resource pools.
Proposal 8: Support the sharing of AI/ML-based computing resources among different AI/ML features and functionalities.
Proposal 9: On AI/ML-based beam management, support at least the following RRC parameters:
Configuration of Set A
Configuration of Set B
For NW-sided model, for inference report, at least for BM-Case 1, RRC parameters to configure M for the content in a beam report in L1 signaling, where L1-RSRPs and corresponding beam information of Top M beam(s) with largest M measured value(s) of L1-RSRP(s) of a measurement resource set are reported
For UE-sided model for inference, when Set A and Set B are configured within CSI report configuration, two CSI-ResourceConfigId s are configured for Set A and Set B separately
For UE-sided model, in CSI-ReportConfig for inference, one or two associated IDs can be configured in CSI-ReportConfig. When Set B is equal or a subset of set A, one associated ID is configured. Otherwise, one associated ID is configured for Set A and another one associated ID is configured for Set B.
For UE-sided model for BM-Case 2, for inference results report, configure UE with N future time instance(s) for inference, N = [1, 2, 4, 8]
For UE-sided model for BM-Case 2, for inference, the time gap between two consecutive future time instances is configured by RRC, time gap is [10ms, 20ms, 40ms, 80ms, 160ms].
Configuration of one set or multiple sets of inference related parameters for applicability report only (not for inference), e.g.
The associated ID
Set A related information
Set B related information
For BM-Case 2,
Time instances related information for measurements
Time instances related information for prediction
At least for monitoring Type 1 Option 2 of UE-side model monitoring (when applicable), for the configuration for monitoring result report in L1 signaling, dedicated resource set(s) for monitoring and report configuration for monitoring are configured in a dedicated CSI report configuration used for monitoring
The ID of an inference report configuration is configured in the configuration for monitoring to link the inference report configuration and monitoring report configuration
For UE-sided model, CSI-ReportConfig can be used for configuring the resources for data collection purpose without CSI report.
One CSI-ResourceConfigId is configured for Set A.
One CSI-ResourceConfigId is configured for Set B.
One or two associated IDs can be configured in CSI-ReportConfig. When Set B is equal or a subset of set A, one associated ID is configured. Otherwise, one associated ID is configured for Set A and another one associated ID is configured for Set B.
|
Conclusions
In the previous sections we made the following observations:
Observation 1 In an Urban macro deployment with 64 beams, having a prediction performance metric based on the Top-4 beams in a monitoring resource set, will in 20% of the cases also label a beam with more than 4dB difference to the strongest as a correct prediction.
Observation 2 The monitoring resource set should be selected that best represent the UE inference channel conditions and its predictions. RAN1 should further discuss an exact definition.
Observation 3 The resource mapping of an NZP-CSI-RS-resources may not be the same in the monitoring and training data collection step. The NW does not need to configure UE with training data collection resources if UE has a trained model, and can optimize the time-domain allocation of the resource for monitoring (with possible subsets of set A beams).
Observation 4 It is up to UE implementation on whether to use dedicated AI hardware or reusing the hardware resources for legacy non-AI based CSI reporting to generate an AI based CSI report.
Based on the discussion in the previous sections we propose the following:
Proposal 1 For UE-sided model, in CSI-ReportConfig, the resourceConfig for set A can only include a single resource set.
Proposal 2 For UE-sided model, for aPeriodic CSI inference report configuration, AssociatedReportConfigInfo comprises the UE assumptions regarding set B.
Proposal 3 For UE-sided model, for aPeriodic CSI inference report configuration, when set B is NOT a subset of set A, the resourceConfig for set B can only include a single resourceSet.
Proposal 4 For UE-sided model, for configuring the data collection resources, do not support aPeriodic CSI-RS
Proposal 5 For the UE-sided inference result report, if not supporting UE report of probabilities, support UE indicating in UCI if the inference result report is inaccurate/invalid (e.g. setting the CRI or SSB-RI to a certain “dummy” value)
Proposal 6 For the UE-sided BM-Case 2 inference report configuration, support NW configuring the number K of UE-reported beams for each time instance.
Proposal 7 For the UE-sided model performance metric, support, at least one of the Top M beam(s) of the resource set(s) for monitoring is among Top-K predicted beam(s). Where K is the number of predicted beam(s) in the corresponding inference report,
Where Top-M beams are the M beams that are within x dB of the best beam with the largest measured value(s) of L1-RSRP(s) of the resource set(s)
X is NW configured on CSI report configuration for monitoring
X= 0 dB,1 dB, 3 dB
Proposal 8 For UE-sided performance monitoring, for the reportConfigType of the UE-sided model performance metric,
Support periodic and semi-persistent reportConfigType if the linked inference report is also periodic or semi-persistent
Support aPeriodic monitoring report only for the scenario where the UE has first been configured with a periodic or semi-persistent monitoring/inference resource config type
Proposal 9 For UE-sided performance monitoring, support joint activation of the monitoring and linked inference report configuration
Proposal 10 For UE-sided model, to identify the connection between RSs in the resource set(s) for monitoring and Set A beams:
UE can assume the same NW transmission parameters for the resourceIDs (all or a subset) that are present both in the monitoring and set A resourceSet,
Note: The resource mapping for monitoring resourceIDs may not be the same as during the UE training data collection
Proposal 11 For UE-assisted performance metric calculation, at least support an event where the UE can early indicate if a functionality is not working properly, for example prediction accuracy is below 50% after collecting N/2 samples
Proposal 12 For enabling a P2 sweep based on predicted Top-K beams, RAN1 considers the following two solutions:
Option 1: NW indicates that the corresponding beam IDs/TCI states for the P2 ResourceSet follows the UE predicted Top-K beams, FFS on details including,
i. linkage of the inference report and the P2 (prediction result measurements) CSI-Report configuration, including BM-Case 2 when UE predicts N time instances
ii. transmission order of the beams in the subsequent P2 (prediction result measurements)
Option 2: For both UE/NW sided model, NW configures dynamically as part of the MAC CE the beam IDs or TCI states that are part of the Top-K beam sweep.
i. FFS on details including the signaling overhead
FFS other options, including
i. Option 3: NW configures a number of aPeriodicTriggerStates, and the UE provides Top-K predictions restricted to the valid TCI states matching one of the aPeriodicTrigger states,
ii. Option 4: The NW uses legacy signaling, and in a best-effort manner configure a set of aPeriodicTriggerStates
iv. Note 1: Option 3 and 4 would imply reduced performance since P2 cannot support all Top-K combinations
Note 2: NW indicating beam IDs could be used during monitoring
Proposal 13 For NW-sided model, regarding max number of reported beam related information in one report, support reporting all measured beams in a resourceSet.
Proposal 14 For NW-sided model inference, support NW configuration for UEs to pre-process set B beams to reduce reporting overhead, via:
Support configuring reporting of only beams within X dB of the strongest beam,
Support configuring reporting of at most N strongest set B beams.
Proposal 15 For NW-sided model inference, support methods for UEs to compress the set B temporal domain measurement results to reduce the reporting overhead.
Proposal 16 For NW-sided data collection, RAN1 studies possible “omission/selection of collected data” by the following aspects as a starting point,
Possibility for UE to avoid signalling “duplicated” samples,
Possibility for UE to avoid signalling data based on certain events, one event can comprise that the UE experienced large channel variation during set A measurements.
Note: RAN2 can use such study when designing data collection procedures
Proposal 17 RAN1 discuss and align on the assumptions of UE implementation for AI based CSI reporting first, before discussing the detailed solutions for the CSI processing criteria and timeline issues.
Proposal 18 If dedicated AI hardware will be used at least for some AI-based CSI reporting, then, support separate processing unit pools (i.e., legacy CPU pool and AI-CPU pool) between CSI reporting generated using legacy CPU and CSI reporting generated using AI-CPU.
Proposal 19 If a UE can use different types of process unit pools for different AI based CSI reporting features, then, support UE indicating the type of process unit pool (i.e., legacy CPU pool and AI-CPU pool) for CSI processing for an AI-based CSI reporting feature.
Proposal 20 Regarding active CSI-RS resources for AI/ML
Predicted CSI-Resources that are part of set A for inference is not considered as active,
Measured CSI-Resources that are part of set A for training/monitoring is considered as active,
The number of simultaneously active associated IDs needs to be restricted and indicated via UE capability.
Proposal 21 For UE-sided model CSI processing timeline, existing values of Z3 and Z3’ for aPeriodic CSI reporting may be sufficient. RAN1 to discuss and align on the assumptions of UE implementation before concluding if new values are needed.
|