META-learning for beam prediction

Abstract

Method comprising requesting to provide, for each first prediction beam among a set A of prediction beams, respective at least K measurement samples, wherein each of the measurement samples indicates a measurement result of a received reference signal strength on a respective transmit beam from among a set B of plural transmit beams and an index of a best beam among the set A; receiving, for each of the prediction beams, the respective at least K measurement samples; adapting a meta machine learning model using the received measurement samples to obtain a variant of the meta machine learning model, predicting Top-N predicted best beams among the prediction beams for a data communication between a terminal and a base by inputting actual measurement results of the received reference signal strength on the first transmit beams into the variant of the meta machine learning model; performing the data communication between the terminal and the base station using one of the Top-N predicted best beams.

Description

[0001] META-LEARNING FOR BEAM PREDICTION Technical Field The present disclosure relates to beam prediction. Abbreviations3GPP 3rd Generation Partnership Project4G / 5G / 6G 4th / 5th / 6th generationAI Artificial IntelligenceBM Beam ManagementBP Beam PredictionCRI CSI Resource IndicatorCSI Channel State InformationCSI-RS CSI Reference SignalDCI Downlink Control InformationDL RS Downlink Reference SignalDMRS Demodulation Reference SignalDL DownlinkFG Feature GroupgNB 5G / NR base stationID IdentifierKPI Key Performance IndicatorL1 Layer 1 (Physical layer)LCM Lifecycle ManagementMeta ML Meta learning model MIMO Multiple-Input and Multiple-Output ML Machine Learning NN Neural Network NR New RadioNW NetworkNZP Non-zero PowerQCL Quasi collocatedPDSCH Physical Downlink Shared Channel RAN Radio Access NetworkRRC Radio Resource ControlRS Reference Signal RSS Reference Signal Strength Rx Receive(r) SGD Stochastic Gradient Descent SSB Synchronization Signal BlockTR Technical ReportTx Transmit(ter)UE User EquipmentUL Uplink Background Beam management includes beamforming transmission, including beam sweeping, beam measurements and reporting, beam maintenance and recovery. To enable a UE to find the best beam, 3GPP TR 38.802 defined procedures P1, P2, and P3 as follows: 1. P1 provides beam sweeping implemented for the gNB. The gNB scans the coveragearea periodically transmitting SSBs with wide angular beams. Conversely, the UE scans different SSBs to identify the best beam and corresponding time / frequency resources for requesting access. In the cell, each SSB is unambiguously related to onebeam. 2. During P2, the gNB performs beam refinement transmitting CSI-RSs with narrowbeams to identify a more precise direction towards the UE after establishing the wide beam in P1. 3. During P3, beam refinement is implemented at the UE side to scan a set of Rx narrowbeams while the gNB transmits CSI-RSs using the best beam identified in P2. The procedures P1, P2 and P3, are executed sequentially to establish the data transmission between gNB and UE. Summary It is an object to improve the prior art. According to a first aspect, there is provided an apparatus comprising means for requesting, from a sample provider, to provide, for each first prediction beamamong a first set A of m first prediction beams, respective at least K measurement samples,wherein each of the measurement samples indicates a measurement result of a receivedreference signal strength on a respective Tx beam from among a first set B of plural first TXbeams and an index of a best beam among the first set A of the m first prediction beams, and m is larger than 1; means for receiving, for each of the first prediction beams, the respective at least K measurement samples from the sample provider in response to the requesting; means for adapting a meta ML model using the received respective at least K measurement samples for each of the first prediction beams to obtain a first variant of the meta ML model, means for predicting Top-N predicted best beams among the first m prediction beamsfor a data communication between a terminal and a base by inputting actual measurementresults of the received reference signal strength on the first TX beams into the first variant ofthe meta ML model, wherein N is 1 or larger than 1;means for performing the data communication between the terminal and the base station using one of the Top-N predicted best beams. The first variant of the meta ML model may be adapted to a first task, wherein the first task comprises predicting the top-N beams among the m prediction beams for the data communication between the terminal and the base station in a first source environment if, foreach beam of the first set B of the plural first TX beams, a measurement result of the receivedreference signal strength on the respective beam is input into the first variant of the meta ML model. The meta ML model may be adapted for solving plural second tasks, wherein each of the second tasks comprises predicting Top-N’ predicted best beams for the data communication between the terminal and the base station in a respective second source environment among beams of a respective second set A of second prediction beams if measurement results of the received reference signal strength on the beams of a respective second set B of second TX beams are input into a respective second variant of the meta ML model, and N’ is equal to 1 or larger than 1. Each of the first and second tasks may differ from each of the respective other first and second tasks by at least one of a configuration or a scenario, wherein the configuration defines at least one of the set A of prediction beams or the set B of measurement beams, and the scenario defines the source environment. The apparatus may further comprise means for informing the sample provider on the Top-N best predicted beams. The apparatus may further comprise means for deciding whether a quality of the data communication is sufficient; means for inhibiting the means for adapting the meta ML model from adapting the meta ML model if the quality of the data communication is sufficient. The apparatus may be installed in the terminal, wherein the base station may comprise the sample provider. The apparatus may further comprise means for informing the base station that the terminal is capable of adapting the meta ML model; means for receiving, from the base station, an activation command; means for inhibiting the means for requesting from requesting, from the sample provider, to provide at least K measurement samples per prediction beam. The apparatus may further comprise means for informing, independently from the requesting to provide the K measurement samples per prediction beam, the base station on the value of K. The apparatus may further comprise measurement means configured to measure the actual measurement results of the received reference signal strength on the first TX beams. The apparatus may be installed in the base station, wherein the terminal may comprise the sample provider. The apparatus may further comprise means for transmitting the reference signals of the first TX beams; means for instructing the terminal to measure the received reference signal strength on the first TX beams and to provide the measurement results to the base station. According to a second aspect, there is provided an apparatus, comprising means for receiving, from a meta ML trainer, a request requesting to provide, for each prediction beam among a set A of m prediction beams, respective at least K measurement samples, wherein each of the measurement samples indicates a measurement result of a received reference signal strength on a respective TX beam from among a set B of plural TX beams and an index of a best beam among the set A of the m prediction beams, and m is larger than 1; means for providing, to the meta ML trainer for each of the prediction beams, the respective at least K measurement samples in response to the request. The apparatus may be installed in a base station, wherein a terminal may comprise the meta ML trainer. The apparatus may further comprise means for receiving a capability information informing that the terminal is capable of adapting a meta ML model; means for activating a meta ML training of the meta ML model in the terminal if the capability information is received; means for inhibiting the means for activating from activating the meta ML training if the capability information is not received. A value of K may be received independently from the request requesting to provide the respective at least K measurement samples for each of the prediction beams.The apparatus may be installed in a terminal, wherein a base station may comprise the metaML trainer. The apparatus may further comprise means for receiving, from the meta ML trainer, an indication of a beam to be used; means for performing a data communication between the apparatus and the meta ML trainer using the beam to be used. According to a third aspect, there is provided a method comprising requesting, from a sample provider, to provide, for each first prediction beam among afirst set A of m first prediction beams, respective at least K measurement samples, whereineach of the measurement samples indicates a measurement result of a received referencesignal strength on a respective transmit beam from among a first set B of plural first transmitbeams and an index of a best beam among the first set A of the m first prediction beams, and m is larger than 1; receiving, for each of the first prediction beams, the respective at least K measurement samples from the sample provider in response to the requesting; adapting a meta machine learning model using the received respective at least K measurement samples for each of the first prediction beams to obtain a first variant of the meta machine learning model, predicting Top-N predicted best beams among the first m prediction beams for a datacommunication between a terminal and a base by inputting actual measurement results of thereceived reference signal strength on the first transmit beams into the first variant of the metamachine learning model, wherein N is 1 or larger than 1;performing the data communication between the terminal and the base station using one of the Top-N predicted best beams. The first variant of the meta machine learning model may be adapted to a first task, wherein the first task comprises predicting the top-N beams among the m prediction beams for the data communication between the terminal and the base station in a first source environment if, foreach beam of the first set B of the plural first transmit beams, a measurement result of thereceived reference signal strength on the respective beam is input into the first variant of the meta machine learning model. The meta machine learning model may be adapted for solving plural second tasks, wherein each of the second tasks comprises predicting Top-N’ predicted best beams for the data communication between the terminal and the base station in a respective second source environment among beams of a respective second set A of second prediction beams if measurement results of the received reference signal strength on the beams of a respective second set B of second transmit beams are input into a respective second variant of the meta machine learning model, and N’ is equal to 1 or larger than 1. Each of the first and second tasks may differ from each of the respective other first and second tasks by at least one of a configuration or a scenario, wherein the configuration defines at least one of the set A of prediction beams or the set B of measurement beams, and the scenario defines the source environment. The method may further comprise informing the sample provider on the Top-N best predicted beams. The method may further comprise deciding whether a quality of the data communication is sufficient; inhibiting the adapting the meta machine learning model if the quality of the data communication is sufficient. The method may be performed by the terminal, wherein the base station may comprise the sample provider. The method may further comprise informing the base station that the terminal is capable of adapting the meta machine learning model; receiving, from the base station, an activation command; inhibiting the requesting, from the sample provider, to provide at least K measurement samples per prediction beam. The method may further comprise informing, independently from the requesting to provide the K measurement samples per prediction beam, the base station on the value of K. The method may further comprise measuring the actual measurement results of the received reference signal strength on the first transmit beams. The method may be performed by the base station, wherein the terminal may comprise the sample provider. The method may further comprise transmitting the reference signals of the first transmit beams; instructing the terminal to measure the received reference signal strength on the first transmit beams and to provide the measurement results to the base station. According to a fourth aspect, there is provided method, comprising receiving, from a meta machine learning trainer, a request requesting to provide, for each prediction beam among a set A of m prediction beams, respective at least K measurement samples, wherein each of the measurement samples indicates a measurement result of a received reference signal strength on a respective transmit beam from among a set B of plural transmit beams and an index of a best beam among the set A of the m prediction beams, and m is larger than 1; providing, to the meta machine learning trainer for each of the prediction beams, the respective at least K measurement samples in response to the request. The method may be performed by a base station, wherein a terminal may comprise the meta machine learning trainer. The method may further comprise receiving a capability information informing that the terminal is capable of adapting a meta machine learning model; activating a meta machine learning training of the meta machine learning model in the terminal if the capability information is received; inhibiting the activating the meta machine learning training if the capability information is not received. A value of K may be received independently from the request requesting to provide the respective at least K measurement samples for each of the prediction beams. The method may be performed by a terminal, wherein a base station may comprise the meta machine learning trainer. The method may further comprise receiving, from the meta machine learning trainer, an indication of a beam to be used; performing a data communication between an apparatus performing the method and the meta machine learning trainer using the beam to be used. Each of the methods of the third and fourth aspects may be a method of beam prediction. According to a fifth aspect, there is provided a computer program product comprising a set of instructions which, when executed on an apparatus, is configured to cause the apparatus tocarry out the method according to any of the third and fourth aspects. The computer programproduct may be embodied as a computer-readable medium or directly loadable into a computer. According to some example embodiments, at least one of the following advantages may be achieved: ^Faster selection of a best beam;^ latency may be reduced;^ Less consumption of radio resources for reference signals;^ Small data volume for transmitting K samples between UE and gNB.Brief description of the drawings Further details, features, objects, and advantages are apparent from the following detailed description of the preferred example embodiments which is to be taken in conjunction with the appended drawings, wherein: Fig.1 shows a method of DL Tx-Rx beam pair prediction using ML model at gNB; Fig.2 shows a method of DL Tx-Rx beam pair prediction using ML model at UE;Fig. 3 shows an example of meta-training task definitions for beam prediction;Fig. 4 shows a method for beam prediction according to some example embodiments;Fig. 5 shows a block diagram for the method of Fig. 2;Fig. 6 shows a method for beam prediction according to some example embodiments;Fig. 7 shows a method for beam prediction according to some example embodiments;Fig. 8 illustrates a use case according to some example embodiments;Fig.9 shows an apparatus according to an example embodiment; Fig.10 shows a method according to an example embodiment; Fig.11 shows an apparatus according to an example embodiment; Fig.12 shows a method according to an example embodiment; and Fig.13 shows an apparatus according to an example embodiment. Detailed description of certain example embodiments Herein below, certain example embodiments are described in detail with reference to the accompanying drawings, wherein the features of the example embodiments can be freely combined with each other unless otherwise described. However, it is to be expressly understood that the description of certain example embodiments is given by way of example only, and that it is by no way intended to be understood as limiting the disclosure to the disclosed details. Moreover, it is to be understood that the apparatus is configured to perform the corresponding method, although in some cases only the apparatus or only the method are described. In case of beam failure and recovery, all the procedures P1 to P3 are repeated. In addition, P2 and P3 are also periodically repeated for beam maintenance. A fundamental problem of the beam management procedures is that with a larger number of beams supported by high-dimensional MIMO arrays, the CSI-RS measurements and feedback overhead for enabling beam selection radically increase. In addition, the time required for gNBand UE to complete the beam sweeping and establish the best beam increases accordingly,mainly due to the frequency of SSB / CSI-RS transmission during procedures P1, P2, P3. Thus, the support of low latency communication is limited. 3GPP aims at supporting beam management by AI / ML. A motivation for supporting AI / ML- based beam management is overhead savings and latency reduction. It has been shown that ML algorithms enable predicting the serving beam for different UE locations and time instances, thus avoiding measuring the actual beam quality and saving those resources, whichthen may be employed for data transmission. On the other hand, beam scanning operationslike those performed in P1, P2 and P3 are time inefficient and not scalable when the size ofantenna arrays increases. Therefore, ML algorithms can replace sequential beam scanning byrecommending a reduced set of beams likely to contain the best beam index of the full scan. In a use case, AI / ML may be used for spatial-domain DL beam prediction for a Set A of beams(also called “prediction beams”) based on measurement results of a Set B of beams (alsocalled “TX beams”) (“BM-case 1”), and in another usecase AI / ML may be used for temporal-domain DL beam prediction for a Set A of beams (also called “prediction beams”) based on measurement results of a Set B of beams (also called “TX beams”) (“BM-case 2), where Set A and Set B may comprise the same or different beams. That is, spatial domain ML approachesinfer the best beam(s) in different spatial locations and temporal domain ML approaches inferthe best beam(s) based on previous beam measurements in time domain . Thus, beamselection accuracy may be improved, wherein system performance (such as reliability andoutage) is considered.options of spatial-domain and temporal-domain beam prediction can be implemented asfollows:: 1. DL Tx beam prediction at the gNB2. DL Tx-Rx (DL Tx) beam prediction at the UEThe effect of applying ML (e.g. a neural network) to these predictions are discussed at greater detail.For DL Tx beam prediction at gNB, a 2D planar array at gNB considering of multiple analog Txbeams and fixed or optimal Rx beam at the UE where only a subset of gNB beams is measuredat the UE and the ML model at the gNB predicts the Top-N beam indices explicitly or implicitlyfrom these subsets of measurements leading to latency reduction for optimal beam selectionon analog beams at the gNB. DL Tx-Rx (DL Tx) beam prediction at the UE a 2D planar at theUE with multiple analog Rx beams and fixed or optimal Tx beams at the gNB can be considered. ML model at the UE is provided by measurements for subset of the UE side beamsand the Top-N beams indices can be predicted explicitly or implicitly from a subset ofmeasurements leading to reducing the latency of beam acquisition or tracking at the UE. In other words, a sub-set of all the potential UE side narrow beam measurements may be usedto predict the best narrow beam(s) at the UE side. Thus, measuring and reporting all narrowbeams may be avoided. Different beams differ in at least one of azimuth, elevation, or beamwidth. For example, gNB may configure the UE to get the measurement report only from a subset of beams while excluding the other beams from the set of measurementsAs an output, the ML model at the UE may provide a list of CRIs associated with both measuredbeams and predicted beams, which are ranked, for instance, in descending order from the most likely best beam to the less likely best beam. Temporal domain beam prediction can take each of 3 prediction variants in which best beamsat the gNB or UE can be predicted for future time instants (prediction window) givenmeasurement observation of beams from a window of precoding time instants (observation window). Time series data can be collected from UEs moving across an environment.Both options 1 and 2 beam predictions can be implemented in two possible alternatives forboth BM-Case1 and BM-Case2: 1) Set B is a subset of Set A in which both Set A and Set Brefer to narrow beams 2) Set A and Set B are different (e.g., Set A consists of narrow beams and Set B consists of wide beams)Fig.1 depicts an example of option 1 beam prediction when the ML model is implemented atthe gNB (if Set B is subset of SetA):.Action 1: gNB sweeps SSB beams in Set B, where Set B beam pattern is cell specific.Action 2: UE reports configured L1-RSRP of measured CSI-RSsAction 3: gNB obtains CSI-RS reports as inputs to the ML model at gNB. The output of themodel are the index and / or L1-RSRP of Top-N best beams of Set A.Action 4: gNB transmits CSI-RSs over predicted Top-N beams in Action 3. UE measures L1-RSRP of Top-N best beams.Action 5: UE reports the index and / or L1-RSRP of practical Top-1 best beam or part of Top-Nbest beams. Legacy beam reporting mechanism is reused.Action 6: gNB indicates the beam for DL data transmission and transmits source RSscorresponding to the activated Transmission Configuration Indicator (TCI) states over a DCImessage which includes configurations such as Quasi-Co location (QCL)relationships between the DL RSs in one CSI-RS set and in the PDSCH DMRS ports.Fig. 2 illustrates an example of option 2 beam prediction when the ML model is implementedat the UE (if Set B is subset of Set A).Action 1: gNB sweeps SSB beams in Set B, where Set B beam pattern might be fixed orvariable for different UEs. The detailed configuration of beams in Set B is up to gNB or reportedby UE. Once SSB beam sweeping is transmitted, subset of SSB / CSI-RSs are measured at UEbased on UE’s Rx beam selection mechanism.Action 2: UE inputs L1-RSRP of measured CSI-RS beams into ML model and outputs indexand / or L1-RSRP of Top-N best beams among all beams.Action 3: UE reports the index and / or L1-RSRP of practical Top-1 predicted CSI beam or partof Top-N beams based on the legacy reporting mechanism.Action 4: gNB transmits CSI-RSs over the received Top-N beams in Action 3. UE measuresL1-RSRP of Top-N best beams.Action 5: UE reports the index and / or L1-RSRP of practical Top-1 best beam or part of Top-Nbest beams. Legacy beam reporting mechanism is reused. Action 6: gNB indicates the beam for DL data transmission and transmits source RSs corresponding to the activated Transmission Configuration Indicator (TCI) states over in a DCI message which includes configurations such as Quasi-Co location (QCL)relationships between the DL RSs in one CSI-RS set and in the PDSCH DMRS ports.Between UE and gNB, a AI / ML model may be identified by a respective model identification.Correspondingly, a AI / ML functionality may be identified by a respective functionalityidentification. Functionality refers to an AI / ML-enabled Feature / FG enabled byconfiguration(s), where configuration(s) is(are) supported based on conditions indicated by UEcapability. Correspondingly, functionality-based LCM operates based on, at least, oneconfiguration of AI / ML-enabled Feature / FG or specific configurations of an AI / ML-enabledFeature / FG. For example, functionality refers to different CSI reports with respect to indicatedUE capabilities to an AI / ML-enabled Feature / FG (e.g. predict narrow beams from wide beam measurements, or predict narrow beams from narrow beam measurement. An example of conditions for functionalities associated with the BM-Case1 assuming DL Tx beam prediction is as follows: ^Support Top-N DL Tx beam predictiono N = 1, 2, 4, [8]- This defines the support of predicting best-N NZP CSI-RS resourcesbased on SSB and / or CSI-RS-based RSRP measurements. ^Set B conditionso Measured DL RS (SSB, CSI-RS)- Defines support of using SSB and / or CSI-RS-based RSRPmeasurements. oMeasured DL RS set dimension (4, 8, 12,

[0016] )- Indicates the minimum number of NZP-CSI-RS resources that shall bemeasured and used by the UE for predicting best-K NZP CSI-RS resources oMeasured DL RS set pattern (e.g., fixed, pre-configured list, random)- Indicates the limitations on Set B conditions^ Set A conditionso Predicted DL RS (CSI-RS)- Defines support of predicting CSI-RS resourceso Predicted DL RS set dimension (16, 32, 64)- Indicates the maximum number of NZP-CSI-RS resources that shall beconfigured as the prediction NZP-CSI-RS resource set ^NW-side performance monitoring conditionso Support measurements of Predicted DL RS set (full Set A, partial Set A)- Defines the support of measuring the NZP-CSI-RS resources thatcorrespond to Set A. oMeasurement periodicity (100 ms, 200 ms)- Indicates the minimum periodicity when supporting NZP-CSI-RSresources that correspond to Set A. ^Conditions on supporting ML functionalitieso Max number of supported functionalities (1, 2, 4, 8,)- Indicates the maximum number of functionalities (e.g., number ofparameter combinations that enable ML-enabled feature) that can be configured toward the UEo Delay in activating a functionality (2 ms, 4 ms, .)- Indicates the delay required when activating or switching afunctionality oGeneralization condition of functionalities (yes, no)- Indicates that the UE supports any functionality configuredconsidering the parameter combinations of 1-4 and can be used towards the UE without any validation of whether the functionality is applicable or not. In some example embodiments, only a subset of these conditions and / or other conditions may be set. Meta-learning Meta-learning, as a subset of metacognition, learns how to learn. Learning to learn has emerged as one of the prominent methods for few-shot learning where a learning framework learns adapting to novel tasks and generalize under few-data regime. Few-shot learning isoften called k-shot learning when k number of support examples is available for each task. Inother words, the learning objective in meta-learning is not to solve a specific problem like supervised learning, but it is to find a solver (a ML model) that can solve a type of problems. Meta model learns through classifying given samples with respect to each entity, wherein the entity is called “class” in the context of meta-learning. A “task” intends to solve this classification problem (associate a number of samples to different classes). Meta learning includes two main phases: a) Meta-training, and b) Meta-testing. In the meta- training phase, a solver can be trained to solve a set of tasks, where the classes of different tasks can be different. For example, task 1 is classification of a number of images into the 5different classes of bird, tank, dog, singer, and piano (for example), while task 2 isclassification of a number of images of into 6 different classes of gymnast, amusement park,rock coast, barrel, and mushroom (for example).The goal of Meta-training phase is to find a solver to solve a set of tasks. Preferably, the tasksare similar, such that the solver includes model parameters that can be easily adapted to anew task. Meta-training phase uses two sets of data: a) Support set ^, and b) the query set^. The support set ^ includes training samples for each meta-training task, which can be usedto obtain a ML model (a “solver”) that is able to perform all the meta-training tasks. The queryset ^ may be used to evaluate how well the trained solver can classify inputs to the differentmeta-training tasks. Each ML model included in the meta ML model is defined by its hyperparameters. The hyperparameters define the number of layers and the number of nodes in each layer of the ML model. Each of the ML models solving one of the respective tasks is a variant (version, branch) of the meta ML model defined by the same hyperparameters as those of the meta ML model but by different parameters (e.g. weights).Meta-testing phase is the process of adapting the solver (meta ML model) trained at the meta-training phase to a new task (also called “meta-testing task). Meta-testing phase also includestwo sets a) Support set ^^, and b) the query set ^^. The support set ^^includes training samples for the new task, which can be used to adapt the solver (an ML model) to the classesor properties of the new task. The query set ^^ can be used to evaluate how well the adaptedsolver (the adapted ML model) can perform the new task.Meta-learning for beam prediction Meta-learning may be used to derive from a ML model for several differentscenarios / configurations (tasks) another variant of the ML model for another task. . In thecontext of beam management, a “task” is defined by a combination of scenario and configuration, shortly denoted as scenario / configuration. “Scenario” means the environment of the gNB and the UE (e.g. urban, rural, indoor, outdoor, or a combination thereof), and configuration means the configuration of the antenna array of the gNB or the UE (e.g. an array of 2*4 antennas, or an array of 4*4 antennas).ML model featured with meta-learning (^^^^ ^^) may comprise the following properties:Meta-training dataset: for each task ^, a dataset in Meta-training (^^^^^^^^^^^^^^^ ) comprisesthe support set ^. It comprises ^^(^)training samples per task ^. A training data set for task ^ refers to a possible combination of arbitrary scenarios and configuration. The training data set comprises measurement samples, wherein each of the measurement samples comprisesmeasurement values (e.g. L1-RSRP values) dependent on a beam index of set B and arespective best beam of set A.K-Shot Task Adaptation: In this application, ^ indicates the minimum number of requiredsamples per class from the corresponding ^ ^ ^^^^^^^^^^^^(e.g., a new scenario / environment or configuration) during each meta-testing. For example, if Set A of TX codebook includes 64 beams (which is equivalent to 64 classes), K-shot task adaptation requires 64×K samples. Please note that, the number of required samples for a new task, i.e. ^, is rather small. Forexample, K=2, or K=3, or K=4, etc. K is a (typically static) design parameter of the meta MLmodel. K-Shot Learning: as mentioned in previous section where k number of support samples ineach task is used for k-shot learning and , to make ^^^^ ^^ model ready for adaptation to atask with only K samples per class, similar conditions (adaptation with K samples) can beconsidered / simulated during training phase. So, the parameters of ^^^^ ^^ model can betrained in an iterative approach. At each iteration and for each Meta-training task ^, K samplesare randomly chosen out of ^^(^) training samples for the task i. Then, the adapted ^^^^ ^^model to Task ^ (using the selected K samples) is evaluated and used to update parametersof ^^^^ ^^ model. An example of an update of the Meta ML model is shown below in Table 1(Algorithm 2, action 10).Fig. 3 illustrates an example depicting some example definitions of meta-training taskscompiled by the ^^^^ ^^ model. As shown in Fig. 3, the meta-training dataset comprises foreach task i (i=0…M-1), a set of training data for the respective task. The tasks are, for example (task 0), determining the best beams for an indoor scenario with a 4*2 antenna array. Thebeam patterns of Set B beams may be defined as follows (see 3GPP RP-233133, Section6.3.1), for example: -Option 1: Set B is fixed across training and inference- Option 2: Set B is variable (e.g., different beams (pairs) patterns in each timeinstance / report / measurement during training and / or inference) -Opt 2A: Set B is changed following a set of pre-configured patterns- Opt 2B: Set B is randomly changed among pre-configured patterns- Opt 2C: Set B is randomly changed among Set A beams (pairs)- Opt 2D: Set B is a subset of measured beams (pairs) Set C (including Set B = Set C),e.g. Top-N beams(pairs) of Set CSome example embodiments provide a framework for enabling Top-N spatial domain beamprediction with meta-learning ML model. It enables configuration and determination of training samples required for using meta ML model during inference as well as signalling for meta- learning ML model adaptation to a new environment employing Set B measurements as the main input. The meta-learning-based beam selection framework according to some example embodiments: 1. enables adaptation to a new scenario / configuration with minimum required^ training samples including measurements with Set B (for example: wide beamsor narrow beams) as inputs and Set A predictions (for example: all narrow beams) as output. The output may be labelled (indicating the predicted best beam(s) (such as the beam(s) with highest RSRP) and / or the sequence of the predicted quality of the beams). 2. Enables the UE to indicate its capability to apply meta ML learning using meta MLmodel parameters (e.g., ^^, K). UE may use functionality based LCM or model-IDbased LCM. Functionality refers to an AI / ML-enabled Feature / FG enabled byconfiguration(s), where configuration(s) is(are) supported based on conditions indicated by UE capability. Correspondingly, functionality-based LCM operatesbased on, at least, one configuration of AI / ML enables feature / FG or specific configurations of an AI / ML-enabled Feature / FG. 3. configuration and determination of number of available training samples based onthe reported meta ML model parameters in Item 2 (applicable to UE sided meta-learning model sub usecase).Fig. 4 illustrates a method according to some example embodiments.In action 1, an entity 1 collects a support set S. The entity 1 may be a UE or the network (e.g. gNB). The support set S includes, for different combinations of scenarios (e.g. rural, urban, suburban, indoor, outdoor or a combination thereof) and configurations (i.e., antenna configurations) training datasets that may be used for spatial beam prediction. The training dataset comprises for each of the beams of set B a measurement result indicating the receivedsignal strength, such as an RSRP value (in particular: a L1-RSRP value)) and an index of thebest beam among the beams of set A.In action 2, the entity 1 trains a meta ML model with a number of meta-training tasks. Eachcombination of a scenario and a configuration of the training dataset is considered as a meta- training task. In action 3, an adaptation of the meta ML model to a new task (beam prediction for anothercombination of scenario and configuration) is requested (for example, because UE performeda handover to a new cell with a different scenario and / or configuration). For this purpose, the entity 1 requests, from an entity 2, (at least) K samples per class. If the entity 1 is the UE, the entity 2 is the network, and vice versa. A class is a set of beams (“Set A”) which may be output by the ML model. In action 4, the entity 2 shares a training data set comprising at least K samples (e.g., via CSI Report reconfiguration) per class with the entity 1. The training data set comprises a support set comprising at least K samples per class, and may comprise a query set comprising additional samples per class. In action 5, the entity 1 tunes (adapts) the trained meta ML model of action 2 to the combination of scenario and configuration requested in action 3. For this purpose, it inputs the samples ofthe support set received in action 4 into the trained ML model and obtains an ML model, wherethe hyperparameters are tuned for the new task. In action 6, the entity 1 uses the tuned ML model in inference for the new task (combination ofscenario and configuration). In particular, by the query set of meta-testing tasks, theparameters (but not the hyperparameters) of the tuned ML model may be adapted for inference.Fig. 5 shows a block diagram for the meta learning based beam prediction procedure of Fig.4. The blocks in Fig.5 correspond to the actions of Fig.4.Hereinafter, the framework according to some example embodiments is described at greater detail. If the meta-learning model is implemented at the UE (=entity 1 of Fig.4), the actions are as follows:1. Data collection from source scenario / configuration: UE collects training datasamples (e.g., L1-RSRP measurements with Set B and Set A) and constructs a support Set S including ^ training datasets of all meta-trainingtasks. The measurements are obtained based on Set B selected measurements (e.g. wide beams or narrow beams as the inputs to the ML model) and Set A (e.g. narrow beams as the output of the ML model). 2. UE indicates ^ and ^ corresponding to Meta ML model conditionsassociated with ML feature group in the UE capability framework. 3. NW configures ML-enabled CSI report (as a functionality) allowing the UEto share the K requested samples during inference.4. Meta-learning model to a new combination of scenario and configuration:UE sided performance (based on current ^^^^ ^^ ^^^^^) degrades. Theperformance may be measured by the configured (e.g., RRC configured)KPI. Consequently, UE makes request to NW for sharing (at least) K training data samples per class corresponding to a new scenario / configuration (e.g.,due to handover to a new cell). 5. NW shares (at least) K training data samples per class corresponding thenew scenario / configuration 6. ^^^^ ^^ adapts to the new (target) scenario / configuration stated in item 4using the shared ^ samples per class for the new (target) task. Once theML model is adapted to the new scenario / configuration, it is used in inference. The adapted model may be used for test samples at inference phase. If the meta-learning model is implemented at the network (e.g. gNB) (=entity 1 of Fig.4), the actions are as follows: 1. Data collection from source scenario / configuration: NW collects trainingdata samples (e.g., CSI-RS measurements vs. beam indices) andconstructs a support Set S including ^ training datasets of all meta-training tasks with respect to all the combinations of scenario and configuration. 2. Meta-learning model to a new scenario / configuration: NW sided meta-learning model detects the need of switching to a newscenario / configuration via monitoring the Meta ML performance degradation (e.g, from UE CSI-RS measurement reporting). Switching to anew scenario / configuration may result in using different codebook atthe NW). 3. UE provides K training data samples per class corresponding to the newscenario / configuration. 4. Meta ML adapts the ML model to the new (target) scenario / configurationstated in item 2 using the shared ^ new training samples per class. Oncethe ML model is adapted to the new scenario / configuration, it is used in inference. The adapted model may be used for test samples at inference phase.As described at greater detail hereinafter, in an example embodiment, a UE meta learning MLmodel is assumed to be capable of performing DL Tx Rx beam prediction (Option 2) inspatial and / or and temporal domain which processes RSRP / RSS measurements with beamsfrom Set B beams.Action 0:UE may apply one of the following approaches identifying meta-ML capability: 1. The UE may report meta-learning capability as a condition within UE capability (e.g.,ML feature group) applying functionality-based LCM framework including the following parameters: ^UE may indicate number of ^ training samples per class (with respectto all meta-trainings) employed by ^^^^ ^^ model. ^ indication can becommon to all ^ configured functionalities, ^^ = 1, … , ^ or specific perfunctionality. and / or ^UE may indicate number of task training samples (training samples isequal to all the existing tasks in meta-training) in meta-trainings (^ )employed with respect to ^^^^ ^^ model. ^ indication can becommon to all the meta ML models (^ = 1, … . , ^) or specific to any metaML model. and / or^ UE may indicate number of M-1 tasks employed for meta-training withrespect to ^^^^ ^^ model corresponding . M-1 indicated tasks can becommon to all the meta ML models (^ = 1, … . , ^) or specific to any metaML model. Or 2. The UE may report meta-learning capability as a condition within UE capability (e.g.,ML feature group) applying UE assisted functionality-based LCM framework (introduced as model-ID based LCM) As such, UE is able to run ^ ^^^^ ^^ models,^^^^ ^^^ (^ = 1, … . , ^) in which each model training aspects including the followingparameters: ^UE may indicate number of ^ training samples per class (with respectto all meta-trainings) employed by ^^^^ ^^^ (^^^ meta ML model)corresponding to Model IDs. ^ indication can be common to all the ^meta ML models within a functionality. Additional condition framework (e.g., via UE Assistant Information (UAI)) can be applied for all the training aspects with respect to meta-learning model(s). and / or ^UE may indicate number of task training samples in meta-trainings (^^) employed with respect to ^^^^ ^^^ (^^^ meta ML model)corresponding to Model IDs. ^ indication can be common to all the metaML models (^ = 1, … . , ^) or specific to any ^^^^ ^^. For ^ trainingassumptions (meta-trainings) in meta-learning, the UE and NW may identify the conditions and additional conditions (scenario / site / UE-distribution, etc.) associated with the meta-learning model(s). The number of task training samples may or may not be included in the UE capability report. Model identification may include Model ID + metainformation = Model ID + conditions + additional conditions + some other info (may be model delivery related, or other) + meta-learning training assumption info. Additional condition framework (e.g., via UE Assistant Information (UAI)) can be applied for all the training aspects with respect to meta-learning model(s). 3. After receiving UE capability (applying either 1. or 2.) gNB configures / indicates numberof available training samples used for inference (^) (e.g., via RRC.configuration or higher-layer signalling) during functionality configuration where: ^configured ^^ corresponding to a new task ^ can be common for anyrequested new task during inference. ^In a common configuration, the gNB may configure the same^ for each new task (scenarios / environment or configurations ^) and have to rely on the common indication to determine thefuture Top-N beam prediction.Or ^^^ can be configured corresponding to each new task ^,^ In this configuration, the gNB may configure ^^ for each foreach new task (scenarios / environment or configurations ^). The following actions 1 to 9 are shown in Fig.6. ^Action 1: UE collects datasets from all the combinations of scenarios andconfigurations for training (training tasks). Each dataset includes samples, wherethe inputs to the ML model and their ground truth labels are stored. depicts first input sample, depicts first ground truth sample of the first dataset sample pair out of ^^samples. ^Action 2: UE trains a meta model using the obtained datasets from the combinationsof scenarios and configurations for training (training tasks).. ^Action 3: UE asks if gNB can provide K support samples obtained in the targetenvironment (combination of scenario and configuration) of the UE. K can beconsidered as part of the functionality / model id.^ Action 4: gNB communicates K support samples, where each sample includes theinput (obtained with Set B) and its label (best beam from Set A). ^Action 5: UE adapts the ^^^^ ^^ using the shared support samples from Action (4).^ Action 6: RSRP / CSI-RS measurements of Set B beams is performed at the UE.^ Action 7: Infer the adapted meta model for new scenario.^ Action 8: The UE selects the beam pair(s) with the highest likelihood of being theoptimal beam pair. ^Action 9: The UE shares the selected beam pair(s) (in the form of beam indices) withthe gNB. ^Action 10: gNB and UE use the results of action (9) for data transmission.If the meta-learning model is implemented at the network (e.g. gNB =entity 1 of Fig. ), the actions are as follows: 1. Data collection from source scenario / configuration: NW collects trainingdata samples (e.g., CSI-RS measurements vs. beam indices) andconstructs a support Set S including ^ training datasets of all meta-training tasks with respect to all the source environments. 2. Meta-learning model to a new scenario / configuration: NW sided meta-learning model () detects switching to a new scenario / configuration (e.g.,due to using different codebook). 3. UE provides K training data samples corresponding to the newconfiguration. 4. Meta ML adapts to the new (target) scenario / configuration stated in action2 using the shared ^ new training samples. Then, the adapted ML modelmay be used in inference for predicting the best beam.At greater detail: A gNB meta learning ML model is assumed to be capable of performing DLTx beam prediction (Option 1) in the spatial domain which processes RSRP / RSSmeasurements with beams from Set B beams. The actions are shown in Fig.7: ^Action 1: The provider of the meta training (e.g. the gNB vendor) collects datasets fromall the source environments (combinations of scenarios and configurations for training). Each dataset includes^samples, where the inputs to the ML model and their ground truth labels are stored. ^ depicts first input sample, depicts first ground truth labels of the first datasetsample pair out of ^^samples. ^Action 2: The provider (e.g. gNB vendor) trains a meta model using the obtaineddatasets from various source environments. ^Action 3: In the target environment, the gNB asks UE to collect and share K supportsamples per class.^ Action 4: UE communicates K support samples per class, where each sample includesthe input (measurement value (e.g. RSRP) obtained with Set B) and its label (obtained from Set A, indicating e.g. the strongest beam). ^Action 5: The gNB adapts the ^^^^ ^^ using the shared support samples from Action(4). ^Action 6: RSRP / CSI-RS measurements of Set B beams is performed at the UE.^ Action 7: The UE shares the measured RSRP / CSI-RS in Set B with the gNB^ Action 8: the gNB Infers the adapted meta model using the shared measurements inSet B. ^Action 9: The gNB selects the beam pair(s) predicted to have the highest likelihood ofbeing the optimal beam pair ^Action 10: The gNB shares the selected beam pair(s) with the UE.^ Action 11: The gNB and UE use the results of action (9) for data transmission.Note the following implementation detail for some example embodiments: Model agnostic machine learning model (MAML) encodes prior knowledge into a learnable initialization that serves as a good initial set of values for weights of a base learner network across tasks. This formulation, in which meta-learning initialization for a base learner, leads to bi-level optimization: inner-loop optimization and outer-loop optimization. For the inner-loop optimization, a base learner is fine-tuned with support examples from the learnable initialization θ to each task for a fixed number of weight updates via gradient descent. Thus, after initializing θi,0 = θ, the task adaptation objective (Equation (1)) is minimized via gradient-descent. As anexample, if we consider meta-ML model as function ^^, where ^ represents the parameters ofthe ^^^^ ^^,at each iteration of the training phase, the meta-ML model is adapted to a trainingtask ^^, which results in updating the model’s parameters ^ to ^^^, i.e., where ^ denotes the step size. ^^^ represents the loss function of task ^^. In addition, stochasticgradient descent (SGD) algorithm can be used for the meta-optimization phase, the meta-model parameters ^ can be updated using the corresponding loss function of the training tasksas Note that beam prediction is a classification task, and categorical cross entropy can beconsidered as loss function ℒ^^ (^^) of task ^^. The detailed steps of few-shot learning algorithmfor training a ^^^^ ^^ is explained in Table 1: Table 1: K-shot learning algorithm (taken from C. Finn, P. Abbeel, and S. Levine, “Model- agnostic meta-learning for fast adaptation of deep networks,” in Proc.34th Int. Conf. Mach. Learn., 2017, pp.1126–1135).Fig.8 shows a use case example. In detail, Fig.8 shows an example setup on how the processof training Meta ML model using different tasks datasets (with respect to different combinations of configurations and scenarios), Meta ML model capability indication using UE capability reporting (ML related features) and UE Meta ML model adaptation to a new scenario / configuration may be defined for a UE-sided model BM-Case1 beam prediction. The actions are as follows: ^Action 1: UE collects data (best beam ID and / or strongest beam L1-RSRP) fromdifferent scenarios and build up the corresponding meta-training tasks datasets. ^Action 2: UE trains the Meta ML model using meta-training datasets based on thecollected training samples. ^Action 3: UE performance (e.g., measured by a RRC configured KPI) degradesalthough UE applies the beam pair predicted by the ML model (e.g., due to moving toa new cell). ^Action 4: UE indicates capability of applying Meta ML model learning and conditionsusing UE capability reporting framework. ^Action 5: gNB shares the requested K samples (per beam in Set A) and activates thefunctionality of Meta ML learning at the UE. ^Action 6 &7: Meta ML at the UE adapts to new environment using the samples sharedin action 5. UE uses the obtained ML model in inference.Fig.9 shows an apparatus according to an example embodiment. The apparatus may be a meta ML trainer (which may be implemented in a terminal (e.g. UE) or a base station (e.g. eNB or gNB)) or an element thereof. Fig.10 shows a method according to an example embodiment. The apparatus according to Fig.9 may perform the method of Fig.10 but is not limited to this method. The method of Fig.10 may be performed by the apparatus of Fig.9 but is not limited to being performed by this apparatus. The apparatus comprises means for requesting 110, means for receiving 120, means for adapting 130, means for predicting 140, and means for performing 150. The means for requesting 110, means for receiving 120, means for adapting 130, means for predicting 140,and means for performing 150 may be a requesting means, receiving means, adapting means,predicting means, and performing means, respectively. The means for requesting 110, means for receiving 120, means for adapting 130, means for predicting 140, and means for performing 150 may be a requester, receiver, adapter, predictor, and performer, respectively. The means for requesting 110, means for receiving 120, means for adapting 130, means for predicting 140, and means for performing 150 may be a requesting processor, receiving processor, adapting processor, predicting processor, and performing processor, respectively.The means for requesting 110 requests, from a sample provider, to provide for each predictionbeam of a first set A of plural prediction beams, respective at least K measurement samples (S110). Each of the measurement samples indicates a measurement result of a receivedreference signal strength on a respective TX beam. The TX beams belong to a set B of pluralTX beams. The request may indicate that the measurements have to be performed in a sourceenvironment. The means for receiving 120 receives the respective at least K measurement samples per prediction beam from the sample provider in response to the requesting of S110 (S120).The means for adapting 130 adapts a meta ML model using the received respective at leastK measurement samples per first prediction beam received in S120 to obtain a first variant ofthe meta ML model (S130).The means for predicting 140 predicts one or more (“Top-N”) best beams among the pluralprediction beams for the data communication between a terminal and a base station byinputting actual measurement results of the received reference signal strength on the TXbeams into the first variant of the meta ML model (S140).The means for performing 150 performs the data communication between the terminal and thebase station using one of the one or more (“Top-N”) predicted best beams (S150).The first variant of the meta ML model may be considered to be adapted to a task. The taskcomprises predicting the one or more best beams among the plural prediction beams for adata communication between the terminal and the base station in the source environment.Fig. 11 shows an apparatus according to an example embodiment. The apparatus may besample provider (which may be implemented in a terminal (e.g. UE) or a base station (e.g.eNB or gNB)) or an element thereof. Fig. 12 shows a method according to an exampleembodiment. The apparatus according to Fig. 11 may perform the method of Fig.12 but is not limited to this method. The method of Fig.12 may be performed by the apparatus of Fig.11 but is not limited to being performed by this apparatus.The apparatus comprises means for receiving 210 and means for providing 220. The meansfor receiving 210 and means for providing 220 may be a receiving means and providing means,respectively. The means for receiving 210 and means for providing 220 may be a receiver andprovider, respectively. The means for receiving 210 and means for providing 220 may be areceiving processor and providing processor, respectively.The means for receiving 210 receives, from a meta ML trainer (e.g. terminal or base station),a request requesting to provide, for each beam of a set A of plural prediction beams, respective at least K measurement samples (S210). Each of the measurement samples indicates ameasurement result of a received reference signal strength of the respective prediction beam.The request may indicate a source environment, and the received reference signal strengthsmay be measured in the source environment.The means for providing 220 provides, to the meta ML trainer, the respective at least K measurement samples per prediction beam in response to the request of S210 (S220). Fig.13 shows an apparatus according to an example embodiment. The apparatus comprises at least one processor 810, at least one memory 820 storing instructions that, when executed by the at least one processor 810, cause the apparatus at least to perform the method according to at least one of the following figures and related description: Fig.10 or Fig.12. In some example embodiments, meta ML learning may always be activated in the UE. In such example embodiments, gNB need not to activate meta ML learning in the UE. Some example embodiments are explained with respect to 6G. However, other example embodiments may be employed in other 3GPP generations, such as 4G, 5G, 7G, etc., or inother environments where data blocks as a whole are transmitted in respective transportblocks. One piece of information may be transmitted in one or plural messages from one entity to another entity. Each of these messages may comprise further (different) pieces of information. Names of network elements, network functions, protocols, and methods are based on current standards, or are current proposals. These names are not limiting. For example, in other versions or other technologies, the names of these network elements and / or network functions and / or protocols and / or methods may be different, as long as they provide a corresponding functionality. The same applies correspondingly to the terminal. If not otherwise stated or otherwise made clear from the context, the statement that two entities are different means that they perform different functions. It does not necessarily mean that they are based on different hardware. That is, each of the entities described in the present description may be based on a different hardware, or some or all of the entities may be based on the same hardware. It does not necessarily mean that they are based on different software. That is, each of the entities described in the present description may be based on different software, or some or all of the entities may be based on the same software. Each of the entities described in the present description may be deployed in the cloud. According to the above description, it should thus be apparent that example embodiments provide, for example, a device for meta-learning (such as a UE or a network element) or an element thereof (which may or may not be actually integrated in the device for meta-learning), an apparatus embodying the same, a method for controlling and / or operating the same, and computer program(s) controlling and / or operating the same as well as mediums carrying such computer program(s) and forming computer program product(s). Implementations of any of the above described blocks, apparatuses, systems, techniques or methods include, as non-limiting examples, implementations as hardware, software, firmware, special purpose circuits or logic, general purpose hardware or controller or other computing devices, or some combination thereof. Each of the entities described in the present description may be embodied in the cloud. It is to be understood that what is described above is what is presently considered the preferred example embodiments. However, it should be noted that the description of the preferred example embodiments is given by way of example only and that various modifications may be made without departing from the scope of the disclosure as defined by the appended claims. The terms “first X” and “second X” include the options that “first X” is the same as “second X” and that “first X” is different from “second X”, unless otherwise specified. As used herein, “at least one of the following: ” and “at least one of ” and similar wording, where the list of two or more elements are joined by “and” or “or”, mean at least any one of the elements, or at least any two or more of theelements, or at least all the elements. The term "or" refers to a non-exclusive “or” unlessotherwise indicated (e.g., use of “or else” or “or in the alternative”).

Claims

Claims 1. Apparatus comprising means for requesting, from a sample provider, to provide, for each first prediction beamamong a first set A of m first prediction beams, respective at least K measurement samples,wherein each of the measurement samples indicates a measurement result of a receivedreference signal strength on a respective transmit beam from among a first set B of plural firsttransmit beams and an index of a best beam among the first set A of the m first prediction beams, and m is larger than 1; means for receiving, for each of the first prediction beams, the respective at least K measurement samples from the sample provider in response to the requesting; means for adapting a meta machine learning model using the received respective at least K measurement samples for each of the first prediction beams to obtain a first variant of the meta machine learning model, means for predicting Top-N predicted best beams among the first m prediction beamsfor a data communication between a terminal and a base by inputting actual measurementresults of the received reference signal strength on the first transmit beams into the first variantof the meta machine learning model, wherein N is 1 or larger than 1;means for performing the data communication between the terminal and the base station using one of the Top-N predicted best beams.

2. The apparatus according to claim 1, wherein the first variant of the meta machine learning model is adapted to a first task, wherein the first task comprises predicting the top-N beams among the m prediction beams for the data communication between the terminal and the basestation in a first source environment if, for each beam of the first set B of the plural first transmitbeams, a measurement result of the received reference signal strength on the respective beam is input into the first variant of the meta machine learning model.

3. The apparatus according to any of claims 1 and 2, wherein the meta machine learning model is adapted for solving plural second tasks, wherein each of the second tasks comprises predicting Top-N’ predicted best beams for the data communication between the terminal and the base station in a respective second source environment among beams of a respectivesecond set A of second prediction beams if measurement results of the received referencesignal strength on the beams of a respective second set B of second transmit beams are inputinto a respective second variant of the meta machine learning model, and N’ is equal to 1 or larger than 1.

4. The apparatus according to any of claims 2 to 3, wherein each of the first and second tasks differs from each of the respective other first and second tasks by at least one of a configuration or a scenario, wherein the configuration defines at least one of the set A of prediction beams or the set B of measurement beams, and the scenario defines the source environment.

5. The apparatus according to any of claims 1 to 4, further comprising means for informing the sample provider on the Top-N best predicted beams.

6. The apparatus according to any of claims 1 to 5, further comprising means for deciding whether a quality of the data communication is sufficient; means for inhibiting the means for adapting the meta machine learning model from adapting the meta machine learning model if the quality of the data communication is sufficient.

7. The apparatus according to any of claims 1 to 6 installed in the terminal, wherein the base station comprises the sample provider.

8. The apparatus according to claim 7, further comprising means for informing the base station that the terminal is capable of adapting the meta machine learning model; means for receiving, from the base station, an activation command; means for inhibiting the means for requesting from requesting, from the sample provider, to provide at least K measurement samples per prediction beam.

9. The apparatus according to any of claims 7 and 8, further comprising means for informing, independently from the requesting to provide the K measurement samples per prediction beam, the base station on the value of K.

10. The apparatus according to any of claims 7 to 9, further comprising means for measuring the actual measurement results of the received reference signal strength on the first transmit beams.

11. The apparatus according to any of claims 1 to 6 installed in the base station, wherein the terminal comprises the sample provider.

12. The apparatus according to claim 11, further comprising means for transmitting the reference signals of the first transmit beams; means for instructing the terminal to measure the received reference signal strength on the first transmit beams and to provide the measurement results to the base station.

13. Apparatus, comprising means for receiving, from a meta machine learning trainer, a request requesting to provide, for each prediction beam among a set A of m prediction beams, respective at least K measurement samples, wherein each of the measurement samples indicates a measurement result of a received reference signal strength on a respective transmit beam from among a set B of plural transmit beams and an index of a best beam among the set A of the m prediction beams, and m is larger than 1; means for providing, to the meta machine learning trainer for each of the prediction beams, the respective at least K measurement samples in response to the request.

14. The apparatus according to claim 13 installed in a base station, wherein a terminal comprises the meta machine learning trainer.

15. The apparatus according to claim 14, further comprising means for receiving a capability information informing that the terminal is capable of adapting a meta machine learning model; means for activating a meta machine learning training of the meta machine learning model in the terminal if the capability information is received; means for inhibiting the means for activating from activating the meta machine learning training if the capability information is not received.

16. The apparatus according to any of claims 14 and 15, wherein a value of K is received independently from the request requesting to provide the respective at least K measurement samples for each of the prediction beams.

17. The apparatus according to claim 13 installed in a terminal, wherein a base station comprises the meta machine learning trainer.

18. The apparatus according to any of claims 13 to 17, further comprising means for receiving, from the meta machine learning trainer, an indication of a beam to be used; means for performing a data communication between the apparatus and the meta machine learning trainer using the beam to be used.

19. Method comprising requesting, from a sample provider, to provide, for each first prediction beam among afirst set A of m first prediction beams, respective at least K measurement samples, whereineach of the measurement samples indicates a measurement result of a received referencesignal strength on a respective transmit beam from among a first set B of plural first transmitbeams and an index of a best beam among the first set A of the m first prediction beams, and m is larger than 1; receiving, for each of the first prediction beams, the respective at least K measurement samples from the sample provider in response to the requesting; adapting a meta machine learning model using the received respective at least K measurement samples for each of the first prediction beams to obtain a first variant of the meta machine learning model, predicting Top-N predicted best beams among the first m prediction beams for a datacommunication between a terminal and a base by inputting actual measurement results of thereceived reference signal strength on the first transmit beams into the first variant of the metamachine learning model, wherein N is 1 or larger than 1;performing the data communication between the terminal and the base station using one of the Top-N predicted best beams.

20. Method, comprising receiving, from a meta machine learning trainer, a request requesting to provide, for each prediction beam among a set A of m prediction beams, respective at least K measurement samples, wherein each of the measurement samples indicates a measurement result of a received reference signal strength on a respective transmit beam from among a set B of plural transmit beams and an index of a best beam among the set A of the m prediction beams, and m is larger than 1; providing, to the meta machine learning trainer for each of the prediction beams, the respective at least K measurement samples in response to the request.

Images