Machine learning in beam selection

By employing machine learning models to optimize beam selection in 5G NR systems, the problems of scheduling latency and insufficient network capacity in beam selection have been solved, achieving lower latency and higher throughput.

CN119013913APending Publication Date: 2024-11-22NOKIA TECHNOLOGIES OY
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202380033184.9
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Priority Date
2022-04-20
Filing Date
2023-04-19
Publication Date
2024-11-22

AI Technical Summary

Technical Problem

Existing 5G wireless telecommunications systems suffer from scheduling delays and insufficient network capacity in beam selection, especially in multi-user and multi-QoS level scenarios, where traditional methods struggle to efficiently utilize limited beam resources.

Method used

A machine learning-based beam selection method is adopted, which utilizes Q-learning and deep Q-network (DQN) models, combined with beam usage information and QoS requirements, to optimize beam selection to reduce latency and improve network capacity. The optimal beam is dynamically selected by implementing the machine learning model in gNB or UE.

Benefits of technology

It achieves lower latency and higher network capacity in 5G NR systems, especially improving the quality of service for latency-critical users, and increasing throughput and energy efficiency.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119013913A_ABST
    Figure CN119013913A_ABST
Patent Text Reader

Abstract

Systems, methods, apparatuses, and computer program products are provided for beam selection using data radio bearer specific machine learning. For example, a method may include providing one or more inputs regarding a plurality of beams to a machine learning model. The method may also include obtaining at least one output value for the plurality of beams from the machine learning model. The machine learning model may be a data radio bearer specific machine learning model, a data radio bearer group specific machine learning model, or a model trained to output selective data radio bearer specific values or data radio bearer group specific values.
Need to check novelty before this filing date? Find Prior Art

Description

CROSS-REFERENCE TO RELATED APPLICATIONS

[0001] This application claims the benefit of U.S. patent application No. 17 / 725,508, filed on April 20, 2022, and entitled “MACHINE LEARNING IN BEAM SELECTION,” the entire contents of which are incorporated herein by reference. Technical Field

[0002] Some example embodiments may generally relate to communications including mobile or wireless telecommunication systems such as long term evolution (LTE) or fifth generation (5G) radio access technology or new radio (NR) access technology or other communications systems. For example, certain example embodiments may generally relate to systems and / or methods for providing beam selection using data radio bearer specific machine learning. Background Art

[0003] Examples of mobile or wireless telecommunication systems may include Universal Mobile Telecommunications System (UMTS) Terrestrial Radio Access Network (UTRAN), Long Term Evolution (LTE) Evolved UTRAN (E-UTRAN), Advanced LTE (LTE-A), MulteFire, LTE-A Pro and / or fifth generation (5G) radio access technology or new radio (NR) access technology. 5G wireless systems refer to the next generation (NG) of radio systems and network architectures. Most 5G systems are built on 5G New Radio (NR), but 5G (or NG) networks can also be built on E-UTRA radios. It is estimated that NR provides a bit rate of about 10-20Gbit / s or higher, and can support at least service levels such as enhanced mobile broadband (eMBB) and ultra-reliable low latency communication (URLLC) and massive machine type communication (mMTC). NR is expected to deliver extremely high broadband and ultra-robust, low latency connectivity and large-scale networking to support the Internet of Things (IoT). As IoT and machine-to-machine (M2M) communications become more and more popular, the demand for networks that meet the requirements of lower power, low data rates and long battery life will grow. Next Generation Radio Access Network (NG-RAN) represents the RAN for 5G that can provide NR and LTE (and Advanced LTE) radio access. It should be noted that in 5G, a node that can provide radio access functions to user equipment (i.e., similar to a Node B, NB in ​​UTRAN or an evolved NB, eNB in ​​UTRAN in LTE) can be named a Next Generation NB (gNB) when built on an NR radio, and can be named a Next Generation eNB (NG-e NB) when built on an E-UTRA radio. Summary of the invention

[0004] Embodiments may be directed to an apparatus. The apparatus may include at least one processor and at least one memory including computer program code. The at least one memory and the computer program code may be configured to, together with the at least one processor, cause the apparatus to at least provide one or more inputs about a plurality of beams to a machine learning model. The at least one memory and the computer program code may also be configured to, together with the at least one processor, cause the apparatus to at least: obtain at least one output value about a plurality of beams from the machine learning model. The machine learning model may be a data radio bearer specific machine learning model, a data radio bearer group specific machine learning model, or a model trained to selectively output a data radio bearer specific value or a data radio bearer group specific value.

[0005] Embodiments may be directed to a method. The method may include providing one or more inputs about a plurality of beams to a machine learning model. The method may also include obtaining at least one output value about the plurality of beams from the machine learning model. The machine learning model may be a data radio bearer specific machine learning model, a data radio bearer group specific machine learning model, or a model trained to selectively output a data radio bearer specific value or a data radio bearer group specific value.

[0006] Embodiments may be directed to an apparatus. The apparatus may include means for providing one or more inputs about a plurality of beams to a machine learning model. The apparatus may also include means for obtaining at least one output value about the plurality of beams from the machine learning model. The machine learning model may be a data radio bearer specific machine learning model, a data radio bearer group specific machine learning model, or a model trained to selectively output a data radio bearer specific value or a data radio bearer group specific value. BRIEF DESCRIPTION OF THE DRAWINGS

[0007] For a proper understanding of the example embodiments, reference should be made to the accompanying drawings, in which:

[0008] Figure 1 shows beam service placements that appear to be inefficient or unfair;

[0009] Figure 2 illustrates beam selection according to certain embodiments;

[0010] Figure 3 An example of transmit / receive beam selection implemented using a neural network according to some embodiments is shown;

[0011] Figure 4 shows an example of a neural network according to some embodiments;

[0012] Figure 5 shows a simulation method for simulating a method according to some embodiments;

[0013] Figure 6 shows performance gains according to certain embodiments;

[0014] Figure 7 Further performance gains according to certain embodiments are shown;

[0015] Figure 8 shows throughput according to certain embodiments;

[0016] Fig. 9 An example flow chart showing a method according to one embodiment; and

[0017] Fig.10 An example block diagram of a system according to one embodiment is shown. DETAILED DESCRIPTION

[0018] It will be readily appreciated that the components of certain example embodiments as generally described and illustrated in the figures herein may be arranged and designed in a wide variety of different configurations. Therefore, the following detailed description of some example embodiments of systems, methods, apparatus, and computer program products for providing beam selection using data radio bearer specific machine learning is not intended to limit the scope of certain embodiments, but rather represents selected example embodiments.

[0019] The features, structures, or characteristics of the example embodiments described throughout this specification may be combined in any suitable manner in one or more example embodiments. For example, the use of the phrases "certain embodiments," "some embodiments," or other similar language throughout this specification refers to the fact that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment. Thus, the appearance of the phrases "certain embodiments," "in some embodiments," "in other embodiments," or other similar language throughout this specification does not necessarily refer to the same set of embodiments, and the described features, structures, or characteristics may be combined in any suitable manner in one or more example embodiments.

[0020] Certain embodiments may have various aspects and features. These aspects and features may be applied individually or in any desired combination with each other. Other features, processes and elements may also be applied in combination with some or all aspects and features disclosed herein.

[0021] In addition, if necessary, the different functions or processes discussed below can be performed in different orders and / or simultaneously with each other. In addition, if necessary, one or more of the described functions or processes can be optional or can be combined. Therefore, the following description should be considered as an explanation of the principles and teachings of some example embodiments, rather than a limitation.

[0022] Certain embodiments relate to beam measurement and beam reporting while in a fifth generation (5G) New Radio (NR) radio resource control (RRC) connected mode. In 5G NR, a user equipment (UE) may report measurement quantities that are used by a gNB to determine the beam to use to serve the UE.

[0023] Machine learning (ML) based beam selection may be aimed at meeting the quality of service (QoS) requirements of a particular QoS level. Certain approaches to ML based beam selection may focus on obtaining the highest achievable signal quality. In this approach, the best beam may refer to the beam that produces the strongest signal, the least interference, the highest signal to interference plus noise ratio (SINR), reference signal received power (RSRP), reference signal received quality (RSRQ), or some other beam quality measure.

[0024] Certain embodiments use machine learning in different ways. Reinforcement learning methods called Q-learning and deep Q-network (DQN) can be used to illustrate the principles of certain embodiments. However, other models can also be used. Q-learning, DQN, and reinforcement learning are generally just examples of machine learning or artificial intelligence methods. Reinforcement learning (RL) can refer to aspects of machine learning related to the way in which intelligent agents perform actions to maximize cumulative rewards. In some embodiments, the action can be a beam selection in a state where the state is a beam measurement result. As a cumulative reward, the throughput of the eMBB UE can be used instead of the strongest signal. Other reward functions can be used to maximize other key performance indicators.

[0025] Conventional beam selection may only consider reported and / or measured beam measurement quantities, such as SINR, RSRP, RSRQ, or received signal strength indicator (RSSI). In such a beam selection process, a beam that provides the best channel conditions based on the selected reported / measured quantities may be selected for the UE. This approach may be sufficient if the number of UEs to be served is small. However, if the load increases and the next generation Node B (gNB) can only utilize one of a few beams at a time, there may be scheduling delays. Such delays can be avoided in various ways. For example, a reduced beam grid (GoB) size can lead to improved ultra-reliable low-latency communication (URLLC) results. By suboptimally reducing the GoB size, scheduling delays can be minimized without providing excessively poor signals because more UEs can be scheduled per transmission time interval (TTI) using frequency domain (FD) scheduling. Some gNBs may perform beam and scheduling candidate selection in the time domain (TD) before scheduling the UE to a selected beam in the FD.

[0026] Figure 1 shows a seemingly inefficient or unfair beam service arrangement. Figure 1As shown, beam 1 serves three user equipments, each user equipment is identified as a URLLC user, beam 2 does not serve any user equipment, and beam 3 serves only one user equipment.

[0027] Certain embodiments may provide a way to efficiently utilize the entire GoB when, for example, the gNB only schedules one beam at a time, which may be applicable to frequency range 2 (FR2) products with a limited number of radio frequency (RF) chains.

[0028] As described above, some ML methods may attempt to find the beam that produces the best signal quality. While such a solution may be suitable for serving a single UE, this approach may not be optimal when there are multiple UEs and multiple QoS levels that should be scheduled using a limited number of beams and limited time / frequency domain resources. Therefore, some embodiments provide other ways to identify beams with ML.

[0029] Figure 2 Beam selection in accordance with certain embodiments is shown. Figure 2 The method shown can be applicable to 5G NR. Alternatively or in addition to the selection based on channel measurements, the proposed ML model-based method can be used at the gNB to select a beam other than the best reported one for the UE, or can be used at the UE side to report a beam suggested by the ML model.

[0030] like Figure 2 As shown, there may be a burst of synchronization signals (SS) from the gNB, which may be received and measured at the UE. The UE may make a beam decision based on an evaluation of the SS burst. Subsequently, the UE may receive a random access channel (RACH) resource including an allocation for the selected beam. The UE may send a RACH preamble to the gNB. The gNB may respond with a random access response (RAR). Msg3 and Msg4 of the RACH procedure may then be sent.

[0031] Subsequently, the gNB may send the SS burst and / or channel state information reference signal (CSI-RS) again. At 210a, the UE may measure the SS block and / or CSI-RS. The UE may also send a beam report via channel state information (CSI). The UE may communicate with the gNB via a physical uplink shared channel (PUSCH) and / or a physical uplink control channel (PUCCH). At 220a, the gNB may select a beam for the UE based on the CSI report.

[0032] The gNB may also send the SS burst and / or CSI-RS again. At 210b, the UE may measure the SS burst and / or CSI-RS again and may send a beam report to the gNB using the CSI. The UE may communicate with the gNB again via the PUSCH / PUCCH. At 220b, the gNB may select a beam other than the beam selected by the UE. The gNB may indicate the selected beam identifier (ID) to the UE.

[0033] There may be at least two options for ML implementation, which may be designated here as Option #1 and Option #2 for convenience only, and not by priority, preference, or in any other manner.

[0034] In option #1, the ML model may be implemented in the gNB. In this option, the gNB may select the beam intended to transmit the upcoming data based on the best beam ID and / or data radio bearer (DRB) or QoS ID reported by the UE. Other beam IDs or measurements reported by the UE may also be used for such selection purposes. The beam may be selected using a beam selection model. The beam selection ML model may be selected by the gNB. The model may be applied at both 220a and 220b.

[0035] In option #2, the ML model may be implemented in the UE. The UE may select candidate beams for ML and may report one or more beams provided by the ML for a particular DRB. If the UE uses the ML model, the PUSCH / PUCCH may include the beam ID suggested by the UE ML model. In addition, the UE may indicate the DRB for which the beam was selected. The UE may also or alternatively indicate any other parameters that indicate the intended use of the selection, for example, from which the DRB may be derived. This data may be included in PUSCH / PUCCH communications after 210a and PUSCH / PUCCH communications after 210b.

[0036] In either option, the PDDCH / PDSCH after 220b may include downlink control information (DCI) and / or media access control (MAC) control element (CE) or other equivalent downlink control information signaling, which may be used to inform the UE of the beam ID that the gNB will use for a certain DRB.

[0037] Therefore, certain embodiments may use ML to assist gNB transmit (Tx) / receive (Rx) beam selection for UEs. As described above, the ML model may exist in the gNB and / or the UE, such as Figure 2 In some embodiments, in addition to the single measurement-based beam selection covering all DRBs, there may also be DRB-specific beam selection. Machine learning may be used to obtain such DRB-specific beam selection.

[0038] If the UE is an ML proxy, as in option #2, then the UE may report DRB-specific beams for one or more DRBs instead or in addition to the best measured beam.

[0039] If the gNB is an ML proxy, as in option #1, the UE may provide candidate beams for certain DRBs, such as for the most delay critical DRBs. For example, a certain DRB may have a beam quality threshold, such as RSRP, for the candidate beams, from which the ML model selects the beam to use. In some embodiments, the best beam for a delay critical packet sent by the corresponding DRB in 5G-NR may not be the best beam defined by interference related measurements. This may be because delay critical packets may be quite small compared to the available bandwidth of 5G. Therefore, if the same beam can be used for multiple UEs with small packets, the gNB may be able to schedule multiple delay critical packets within a single transmission time interval (TTI). Certain embodiments may use an ML proxy to find such a beam in order to reduce latency and increase network capacity.

[0040] In an alternative approach, the UE may assume that the gNB uses the best beam that the UE has reported. Therefore, if the gNB is selecting another beam than the one reported by the UE, the gNB may beneficially inform the UE whether to use a beam other than the one reported by the UE for a certain DRB. This information may help the UE to adjust the UE's transmit / receive filters. For example, when transmitting / receiving data for the indicated DRB, the UE may adjust the transmit / receive filters based on the information provided by the gNB.

[0041] In addition to or instead of DRB differentiation, other QoS flow mapping methods may also be used.DRB may be a way to differentiate beam selection strategies.

[0042] The 3rd Generation Partnership Project (3GPP) Technical Specification (TS) 38.213 specifies that the UE reports 1-4 beams in a single CSI report based on the Radio Resource Control (RRC) parameter nrofReportedRS. Certain embodiments provide an ML method that can find the best performing beam when the 1-4 best beams of the UE are known. If the gNB is learning, the UE can report up to four best beams in the CSI. If the UE is learning, the UE can report only one beam given by its ML model.

[0043] Certain embodiments focus specifically on beam selection for a DRB or a set of DRBs that are to carry a certain type of data. If the gNB wants to improve throughput, then only one model is sufficient. However, separate models may also be used simultaneously, for example for delay critical DRBs. While the models may be mostly the same, the reward functions may be different. For example, best effort data may benefit from having a reward based on throughput, for example. In contrast, for another example, delay critical data may benefit from having a reward based on protocol data unit (PDU) delay as well as transmit buffer status.

[0044] Therefore, certain embodiments may address ML-based beam selection for delay-critical data. Certain embodiments may also address neural network-based implementation embodiments. Furthermore, as described below, certain embodiments may be applied to different QoS levels by selecting the reward function accordingly.

[0045] In certain embodiments, the ML agent (which may be located at the UE or at the gNB) may observe the latency of PDUs, such as Packet Data Convergence Protocol (PDCP) PDUs, during an observation window. The observation window may be based on, for example, a CSI reporting period. PDU latency may be the time between packet arrival and successful reception. The PDU may be, for example, a PDCP and / or Radio Link Control (RLC) PDU. In addition to PDU latency, the transmit buffer status may be monitored. To minimize latency, it may be beneficial to keep the transmit buffer as empty as possible.

[0046] So, for example, a reward function that includes both latency and buffer state terms could be as follows:

[0047]

[0048] where the first case is if the PDU is sent / received, where the second case is if the data is present in the transmit buffer but not sent / received, where n is the number of PDUs received / sent, and d sum is the cumulative sum of PDU delays within the observation window. This approach may be useful for learning beam selection for delay-critical DRBs or QoS classes.

[0049] Figure 3 An example of transmit / receive beam selection implemented using a neural network in accordance with certain embodiments is shown.

[0050] like Figure 3 As shown, at 310, a CSI report from a UE may be received at the gNB. The CSI report may include, for example, from one to four best beams and at least an RSRP value for the best beam. If the UE provides other RSRPs, the other RSRPs may be provided as differential values, such as relative differences from the best RSRP.

[0051] At 315, the gNB may obtain beam usage statistics. The gNB may obtain statistics about the most recent usage count for each beam. The gNB may track such statistics in any desired manner. If the UE is implementing an ML model, the gNB may provide the statistics to the UE, for example, using periodic broadcasts to provide statistics to UEs with ML capabilities.

[0052] At 320, the gNB may formulate a tuple and store the tuple to a replay memory. The tuple may include, for example, the following: a previous state of the UE; a previous action of the UE, including, for example, a beam selection ID; rewards, such as accumulated UE throughput, packet delay, and / or buffer status; and a state, such as a neural network (NN) input vector including a beam usage value and RSRP for each beam.

[0053] At 325, the gNB may train the neural network from the replay memory. For example, a plurality of epochs may be looped through, and a random tuple may be selected for training batches for each epoch.

[0054] At 330, the gNB may select a neural network based on the best beam. A separate small neural network may exist for each gNB beam ID.

[0055] At 335, the gNB may feed input into the selected neural network. The input may be, for example, a normalized number of times a certain gNB beam is scheduled within a time window (e.g., between CSI reports).

[0056] The neural network may suggest beams at 340. To obtain the suggestions, a neural network forward pass may be done using the latest state NN input vector.

[0057] At 345, the gNB may determine whether the beam suggested by the neural network is within the beam reported by the UE as the best. If so, at 350, the gNB may use the beam suggested by the ML model. Optionally, at 355, the gNB may indicate the selected beam to the UE.

[0058] The indication to the UE, while not mandatory, helps the UE to change the UE's beam / panel to a better matching one if the gNB chooses to use a beam other than the best beam reported by the UE. As an example, a DCI or MAC CE with 6 bits may be able to indicate the selected one of 64 beam IDs. The number 64 is an example taken from the 5G specification.

[0059] Otherwise, if the beam suggested by the ML model is not one of the reported best beams, the gNB may use the best beam reported based on the reported beams at 360. Additionally, at 365, the gNB may penalize the ML model for ineffective beam selection.

[0060] Penalization can allow the ML model to obtain feedback for substandard outputs. A beam can be considered substandard if no RSRP is reported for that beam, or if RSRP is reported but below a threshold.

[0061] If a neural network is used for learning, a small neural network can be used for each gNB beam. Alternatively, a single neural network can be used. Using a small neural network can be used to ensure that for each CSI input, the best possible beam selection can be made. For example, each beam can have a different set of other beams that are spatially correlated. Additionally, the back-propagation process and forward-propagation process of the neural network may become less demanding for the processor performing the ML. Furthermore, a single processor can be used to efficiently process multiple neural networks in parallel on different cores / threads.

[0062] like Figure 3 As shown, a neural network for determining the gNB beam to use may be selected based on a CSI report indicating the best beam. In other words, a neural network corresponding to the best beam in the CSI report may be selected. The neural network may output a recommendation for a beam to use. The recommended beam may be the same or different than the best beam based on the measurement. If the recommended output is within the best beam reported by the UE, then that beam may be used. Otherwise, the best beam among the reported beams may be used.

[0063] Figure 4 An example of a neural network in accordance with some embodiments is shown. Figure 4 Examples of neural networks may be able to learn and provide beam selection that is superior to conventional measurement-based methods.

[0064] Therefore, in Figure 4An example of a neural network capable of learning beam selection is shown in . The inputs to the network may include usage counts for each gNB beam at inputs 1 to n, and user reported RSRP values ​​at inputs n+1 to 2n. The beam usage information may be, for example, the number of times a certain beam was used for all or high priority data. As described above, such information may be obtained at the gNB by implementing, for example, storage of statistical data. If the UE's ML model wants to use such information, the gNB may, for example, broadcast such information to at least ML-capable UEs. For example, a machine learning radio network temporary identifier ML-RNTI may be defined for such usage. High priority data may be, for example, a certain set of QoS metrics with strict delay and / or reliability targets. The values ​​may be scaled to be between -1 and 1 to stabilize the neural network forward and backward propagation computations. The i-th (i-th) scheduling count may be derived from the scheduling count th )Beam usage of beam: count i / count max. If there is a beam that is not used or has no RSRP value, the value can be zero. If the RSRP value is known from the difference value, the RSRP value can be calculated before normalization.

[0065] In addition to the usage counts, a second set of inputs may be signal quality, such as RSRP reported by the user equipment, for example, for the 1-4 best beams. The value range may be from -140 to -44 dBm scaled to -1 to 1.

[0066] The neural network may output a Q value for each gNB beam index. The beam with the highest Q value may be expected to produce the highest performance. The highest performance may be, for example, the lowest packet delay or the highest throughput, depending on the reward function used.

[0067] Figure 5 A simulation method for simulating a method according to some embodiments is shown. Accurate system-level simulation can prove the principles of some embodiments. Figure 5 , Figure 6 , Figure 7 and Figure 8The simulation scenario described follows the 3GPP guidelines for dense urban macro scenarios. In this simulation, 21 base stations each have 14 beams. The simulation uses a 50MHz bandwidth in the 30GHz mmW FR2 band. A UE mobility speed of 3km / h is used. Only a single radio frequency (RF) chain is used that limits the number of simultaneous Rx / Tx beams to a single one. Frequency domain scheduling is performed with UEs with the same gNB Tx / Rx beam. Two types of UEs (URLLC and eMBB) are randomly dropped within the simulation area. 378 URLLC UEs send / receive 50B packets for UL / DL according to the File Transfer Protocol (FTP) 3 traffic model with a packet inter-arrival time of 2ms. 42 eMBB UEs have a fully buffered traffic model for both DL and UL. Scheduling in both the time and frequency domain is prioritized over URLLC UEs in order to minimize their latency. The CSI reporting period is always 5ms. It is assumed that each CSI report includes 1 or 4 best beams and at least the RSRP for the best beam is measured.

[0068] The scheduling method used by the simulator is as follows Figure 5 As shown. First, at 510, the time domain scheduler sorts the scheduling candidates based on the time domain scheduling priority metric of the time domain scheduler. Then, at 520, the gNB selects a beam for Tx / Rx in the upcoming TTI based on the best beam selected for the scheduling candidate with the highest time domain priority metric. At 530, UEs belonging to different gNB Tx / Rx beams are removed from the scheduling candidate pool. Finally, at 540, frequency domain resource blocks are shared based on the frequency domain scheduling metric.

[0069] If all candidates belong to the same QoS class, the beam to be used within the scheduled TTI is selected first based on the QoS priority of the scheduling candidates and secondly based on the scheduling priority metric priority selection.

[0070] In order to generate a sufficiently converged neural network, the learning algorithm was first run with a random exploration rate of 10%. Therefore, on average every 10 beam selections were random. The other 90% of beam selections were based on ML suggestions. The simulation approach already outperformed the measurement-based beam selection with such a random exploration rate. This is probably mainly because the random selections come only from a limited pool of beams reported by the UE.

[0071] Once the neural network has converged enough, random exploration can be stopped or reduced to a smaller number, such as 1%. After that, the model can still store the model's experience into replay memory and continue to continuously update the model without random exploration. An experience tuple can include two consecutive states, and the reward and action of the first state. Since there may be multiple neural networks, multi-core / multi-threaded processors can be used to parallelize model updates from replay memory for each beam.

[0072] Figure 6 The performance gain according to some embodiments is shown. More specifically, as described above, Figure 6 The performance gain is shown when the model according to certain embodiments is used for delay critical UEs with a reward function. The DQN option utilizes beam usage information as an additional input, which can provide further insight into which beams should be selected.

[0073] like Figure 6 As shown in the simulation data in , the ML approach achieves lower latency and higher reliability compared to CSI-based beam selection alone (i.e., based only on beam quality / power measurements). This improvement can allow the network to increase capacity for latency-critical UEs.

[0074] Figure 7 Further performance gains according to certain embodiments are shown. More specifically, Figure 7 The performance gain and the ability of DQN to continuously improve performance when the model of certain embodiments is used for a best effort UE with a throughput based reward function are shown. In this simulation example, beam usage information is used as an additional input to the DQN. The simulation bandwidth is only 50 MHz. For higher bandwidths available on the millimeter wave (mmW) frequency range FR2, higher throughput gains are expected.

[0075] More specifically, Figure 7 The cell throughput gain based on measured beam selection after 1-6 DQN training iterations is shown. Figure 7 The "no initialization" in the example refers to the starting simulation run without a lot of warm-up exploration. The "no initialization" example confirms that once a trained model exists, the model does not need to run heavy exploration that may temporarily slightly degrade network performance. Significantly higher throughput for best-effort UEs can be achieved with ML methods that do not rely solely on beam quality / power measurements.

[0076] Figure 8 The throughput according to some embodiments is shown. Figure 8 In , the UE throughput distribution is compared between the CSI and ML based alternatives. The worst performing best effort UEs may be those that do not acquire resources because the gNB prioritizes beams directed to delay critical UEs. If the learning is implemented using a DQN using the methods of certain embodiments, it is able to learn beam selections that perform well in terms of distributed throughput. The learning may provide high gains for the worst performing UEs and may not cause significant throughput degradation for the best performing UEs. A Q-learning alternative using certain embodiments can perform a similar job, but may not take into account as much input information as a trained DQN. Nonetheless, Figure 8Q-learning and DQN are used in to illustrate the gains of certain embodiments, which may be independent of the exact learning method.

[0077] More specifically, Figure 8 Best effort UE throughput distribution is shown. Certain embodiments of the ML method are particularly helpful in increasing the worst percentile throughput, thereby protecting the UE from unacceptable service.

[0078] In addition to throughput and latency improvements, certain embodiments can also improve energy efficiency. If more UEs can be served within the same TTI, as shown in simulations, certain embodiments can help reduce the amount of transmission in the time domain and shorten the transmission time, which can also reduce gNB / UE energy consumption.

[0079] Certain embodiments may include any of the following aspects in any combination. For example, certain embodiments may have a DRB or DRB group specific ML model that may be used for beam selection. The model may determine the service beam utilized for the DRB, for example only for delay critical DRBs or only for best effort DRBs. The beam selection model may use at least the PDU delay value for beam selection. The beam selection model may use the buffer status as an additional input for beam selection. The UE may be a model and the UE may report an ML recommended beam for a certain DRB. The UE may instruct the UE to provide a DRB for which the ML recommended beam selection is provided. This approach may be implemented by a method involving standardization.

[0080] In some embodiments, the gNB may have a DRB specific model. The beam selection model may use CSI reports as input. The model may suggest using other beams for a certain DRB in addition to the beam reported by the UE. The UE may provide candidate beams for a certain DRB. The candidate beams may be the n best beams based on measurements or all beams exceeding an RSRP threshold or another measurement quantity threshold. The values ​​of the DRB-related ML acquisition in the CSI report may be achieved by standardization.

[0081] If the gNB has a beam selection model, then after selecting another serving beam in addition to the gNB beam reported by the UE, in some embodiments the gNB may indicate to the UE the other serving beam to be used. For example, this may be signaled with 6 additional bits in the DCI or MACCE. Such DL control information may be achieved by standardization.

[0082] Certain embodiments may involve obtaining beam usage information to be used as input to a beam selection model. For example, the gNB may use beam usage statistics as input to the beam selection model of the gNB. The gNB may also broadcast ML-related information, such as beam usage data of the gNB, with an RNTI (e.g., the ML-RNTI described above). Such broadcasting of the ML-RNTI may be achieved via standardization.

[0083] If the UE has a beam selection model, the gNB may broadcast or otherwise send beam usage information to one or more connected ML-capable UEs. The UE may utilize the beam usage information as input to its beam selection model. As described above, these features may be implemented through standardization.

[0084] In some embodiments, the beam usage information may only consider beams of certain QoS indices. The UE may use the received beam usage information as input to the UE beam selection model.

[0085] Certain embodiments may have a separate neural network or model for each beam, antenna port, or antenna panel. Certain embodiments may select a model based on channel quality measurements such as RSRP, SINR, RSRQ, and / or RSSI. For example, a beam ID may be selected to produce the best measurement results, and a corresponding ML model may be selected to select the beam ID to be used for upcoming transmission and / or reception.

[0086] Fig. 9 Methods according to some embodiments are shown. Fig. 9 The method may be performed by the gNB and / or the UE, for example, Figure 2 , Figure 3 and Figure 5 shown.

[0087] like Fig. 9As shown, the method may include providing one or more inputs about multiple beams to a machine learning model at 910. The machine learning model may be a data radio bearer specific machine learning model, a data radio bearer group specific machine learning model, or a model trained to selectively output a data radio bearer specific value or a data radio bearer group specific value. For example, a single model may be trained to output a data radio bearer value and a data radio bearer group value at different times or in combination with each other. The input may include a protocol data unit delay value. The input may also include a buffer status. The input may also include usage statistics of multiple beams. The machine learning model may be in a next generation node B, as in option #1 discussed above. Additionally or alternatively, the machine learning model may be in a user equipment, as in option #2 discussed above. If the input includes usage statistics of multiple beams and option #2 is followed, the usage statistics may be received at the UE from the gNB, for example via a broadcast or other transmission. The method may also include obtaining at least one output value about multiple beams from the machine learning model at 920. For example, the machine learning model may provide an output value for each beam.

[0088] Fig. 9 The method may also include, at 930, obtaining a recommended beam from the machine learning model on which the radio access network node will send or receive radio communications with the user equipment. For example, when the model provides various output values ​​corresponding to various beams, the best beam may be selected based on the best output value. For example, in some embodiments, the model may output a recommended beam that the gNB will use to send / receive to / from the user equipment, regardless of the location of the ML model. If the model is in the UE, the recommendation may be indicated to the gNB, for example, with a CSI report. If the model is at the gNB, the gNB may only use the recommended beam. In this case, the gNB may indicate the beam to be used for the UE, but not necessarily. The ML model may be used to select the UE beam, but the UE beam may depend on the UE implementation.

[0089] The machine learning model may be configured to select the proposed beam based on the resulting latency or network capacity regardless of whether the proposed beam has a higher or lower signal quality than the alternative beams, for example, even if the proposed beam has a lower signal quality than the alternative beams. In some embodiments, the model may simply ignore the signal quality characteristics in its decision making process. Thus, the machine learning model may not know whether the proposed beam has equal, better, or worse signal quality characteristics than any other alternative beams.

[0090] Fig. 9The method may further include: at 940, the user device compares the suggested beam with the report of the user device. The comparison may involve checking whether the suggested beam is included in the report provided by the user device. For example, it may be determined whether the user device finds the suggested beam in the report. The method may also include selecting a beam for use based on the comparison or other checks at 950. When the suggested beam is included in the report, the beam may be selected for use based on the inclusion in the report. Conversely, when the suggested beam is not included in the report, the beam from the report may be selected for use. If the report includes only one item, the ML recommendation may still make sense. In this case, the suggested beam may be used even if the suggested beam is not found in the report. The method may also include: at 960, notifying the user device of the selected beam. If the suggested beam is not included in the report, the machine learning model may be penalized based on not including the suggested beam in the report.

[0091] It should be noted that Fig. 9 is provided as an example embodiment of a method or process. However, certain embodiments are not limited to this example, and further examples are possible as discussed elsewhere herein.

[0092] Fig.10 An example of a system including an apparatus 10 according to an embodiment is shown. In an embodiment, the apparatus 10 may be a node, a host, or a server in a communication network, or serve such a network. For example, the apparatus 10 may be a network node, a satellite, a base station, a Node B, an evolved Node B (eNB), a 5G Node B or access point associated with a radio access network (such as an LTE network, 5G or NR), a TRP, a HAPS, an integrated access and backhaul (IAB) node, and / or a WLAN access point. In some example embodiments, the apparatus 10 may be, for example, a gNB or other similar radio node.

[0093] It should be understood that in some example embodiments, the device 10 may include an edge cloud server as a distributed computing system, where the server and the radio nodes may be independent devices that communicate with each other via a radio path or via a wired connection, or they may be located in the same entity that communicates via a wired connection. For example, in certain example embodiments where the device 10 represents a gNB, it may be configured in a central unit (CU) and distributed unit (DU) architecture that divides the gNB functionality. In such an architecture, the CU may be a logical node that includes gNB functionality, such as transmission of user data, mobility control, radio access network sharing, positioning and / or session management, etc. Depending on the functional division option, the DU may be a logical node that includes a subset of the gNB functionality. It should be noted that a person of ordinary skill in the art will understand that the device 10 may include Fig.10 Components or features not shown.

[0094] like Fig.10 As shown in the example of , the device 10 may include a processor 12 for processing information and executing instructions or operations. The processor 12 may be any type of general-purpose or special-purpose processor. In fact, as an example, the processor 12 may include one or more of a general-purpose computer, a special-purpose computer, a microprocessor, a digital signal processor (DSP), a field programmable gate array (FPGA), an application-specific integrated circuit (ASIC), and a processor based on a multi-core processor architecture. Although Fig.10 A single processor 12 is shown in the figure, but multiple processors may be utilized according to other embodiments. For example, it should be understood that in some embodiments, the device 10 may include two or more processors that can form a multi-processor system that can support multi-processing (e.g., in this case, the processor 12 can represent a multi-processor). In some embodiments, the multi-processor system can be tightly coupled or loosely coupled (e.g., to form a computer cluster).

[0095] Processor 12 may perform functions associated with the operation of device 10, which may include, for example, precoding of antenna gain / phase parameters, encoding and decoding of individual bits forming communication messages, formatting of information, and overall control of device 10, including processes related to communications or management of communication resources.

[0096] The device 10 may further include or be coupled to a memory 14 (internal or external), which may be coupled to the processor 12 for storing information and instructions that can be executed by the processor 12. The memory 14 may be one or more memories and may be of any type suitable for the local application environment and may be implemented using any suitable volatile or non-volatile data storage technology, such as semiconductor-based memory devices, magnetic memory devices and systems, optical memory devices and systems, fixed memory, and / or removable memory. For example, the memory 14 may include any combination of random access memory (RAM), read-only memory (ROM), static storage (such as a magnetic disk or optical disk), a hard disk drive (HDD), or any other type of non-transitory machine or computer readable medium, or other suitable storage devices. The instructions stored in the memory 14 may include program instructions or computer program code that, when executed by the processor 12, enables the device 10 to perform tasks as described herein.

[0097] In an embodiment, the device 10 may also include or be coupled to a drive or port (internal or external) configured to accept and read an external computer-readable storage medium, such as an optical disk, a USB drive, a flash drive, or any other storage medium. For example, the external computer-readable storage medium may store a computer program or software for execution by the processor 12 and / or the device 10.

[0098] In some embodiments, the device 10 may also include or be coupled to one or more antennas 15 for sending and receiving signals and / or data to and from the device 10. The device 10 may also include or be coupled to a transceiver 18 configured to send and receive information. The transceiver 18 may include, for example, a plurality of radio interfaces that may be coupled to the antenna 15, or may include any other suitable transceiver device. The wireless interface may correspond to a variety of radio access technologies, including one or more of Global System for Mobile Communications (GSM), Narrowband Internet of Things (NB-Io T), LTE, 5G, WLAN, Bluetooth (BT), Bluetooth Low Energy (BT-LE), Near Field Communication (NFC), Radio Frequency Identifier (RFID), Ultra Wideband (UWB), MulteFire, etc. The radio interface may include components such as filters, converters (e.g., digital-to-analog converters, etc.), mappers, Fast Fourier Transform (FFT) modules, etc., to generate symbols for transmission via one or more downlinks and receive symbols (e.g., via uplinks).

[0099] Thus, the transceiver 18 may be configured to modulate information about a carrier waveform for transmission by the antenna 15, and to demodulate information received via the antenna 15 for further processing by other elements of the apparatus 10. In other embodiments, the transceiver 18 may be capable of directly sending and receiving signals or data. Additionally or alternatively, in some embodiments, the apparatus 10 may include input and / or output devices (I / O devices) or input / output devices.

[0100] In an embodiment, memory 14 may store software modules that provide functionality when executed by processor 12. The modules may include, for example, an operating system that provides operating system functionality for device 10. The memory may also store one or more functional modules, such as applications or programs, to provide additional functionality to device 10. The components of device 10 may be implemented in hardware or as any suitable combination of hardware and software.

[0101] According to some embodiments, the processor 12 and the memory 14 may be included in a processing circuit / device or a control circuit / device or may form a part thereof. In addition, in some embodiments, the transceiver 18 may be included in a transceiver circuit / device or may form a part of a transceiver circuit / device.

[0102] As used herein, the term "circuit" may refer to a hardware circuit implementation (e.g., analog and / or digital circuitry), a combination of hardware circuitry and software, a combination of analog and / or digital hardware circuitry and software / firmware, any portion of a hardware processor with software (including a digital signal processor) that works together to cause a device (e.g., device 10) to perform various functions, and / or hardware circuitry and / or processors, or portions thereof, that operate using software but where the software may not be present when not required for operation. As another example, as used herein, the term "circuit" may also encompass implementations of only a hardware circuit or processor (or multiple processors) or portions of a hardware circuit or processor and their accompanying software and / or firmware. The term circuitry may also cover, for example, a baseband integrated circuit in a server, cellular network node or device, or other computing or network device.

[0103] As described above, in some embodiments, the device 10 may be or may be part of a network element or RAN node, such as a base station, access point, Node B, eNB, gNB, TRP, HAPS, IAB node, relay node, WLAN access point, satellite, etc. In an example embodiment, the device 10 may be a gNB or other radio node, or may be a CU and / or DU of a gNB. According to some embodiments, the device 10 may be controlled by the memory 14 and the processor 12 to perform functions associated with any of the embodiments described herein. For example, in some embodiments, the device 10 may be configured to perform one or more of the processes depicted in any of the flowcharts or signaling diagrams described herein, such as Figures 1 to 9 In some embodiments, the apparatus 10 may be configured to perform processes related to providing beam selection using, for example, data radio bearer specific machine learning, as discussed herein.

[0104] Fig.10An example of an apparatus 20 according to an embodiment is further shown. In an embodiment, the apparatus 20 may be a node or element in a communication network or associated with such a network, such as a UE, a communication node, a mobile device (ME), a mobile station, a mobile device, a fixed device, an IoT device, or other device. As described herein, a UE may alternatively be referred to as, for example, a mobile station, a mobile device, a mobile unit, a mobile device, a user device, a subscriber station, a wireless terminal, a tablet, a smart phone, an IoT device, a sensor or a NB-IoT device, a watch or other wearable device, a head mounted display (HMD), a vehicle, a drone, a medical device and its applications (e.g., remote surgery), an industrial device and its applications (e.g., a robot and / or other wireless device operating in the context of an industrial and / or automated processing chain), a consumer electronic device, a device operating on a commercial and / or industrial wireless network, etc. As an example, the apparatus 20 may be implemented in, for example, a wireless handheld device, a wireless plug-in accessory, etc.

[0105] In some example embodiments, the apparatus 20 may include one or more processors, one or more computer-readable storage media (e.g., memory, storage, etc.), one or more radio access components (e.g., modems, transceivers, etc.), and / or a user interface. In some embodiments, the apparatus 20 may be configured to operate using one or more radio access technologies, such as GSM, LTE, LTE-A, NR, 5G, WLAN, WiFi, NB-IoT, Bluetooth, NFC, MulteFire, and / or any other radio access technology. It should be noted that those skilled in the art will appreciate that the apparatus 20 may include Fig.10 Components or features not shown.

[0106] like Fig.10 As illustrated in the example of , the device 20 may include or be coupled to a processor 22 for processing information and executing instructions or operations. The processor 22 may be any type of general or special purpose processor. In fact, as examples, the processor 22 may include one or more of a general purpose computer, a special purpose computer, a microprocessor, a digital signal processor (DSP), a field programmable gate array (FPGA), an application specific integrated circuit (ASIC), and a processor based on a multi-core processor architecture. Although Fig.10 A single processor 22 is shown in FIG. 1 , but multiple processors may be utilized according to other embodiments. For example, it should be understood that in some embodiments, the apparatus 20 may include two or more processors that may form a multiprocessor system that may support multiprocessing (e.g., in which case the processor 22 may represent a multiprocessor). In some embodiments, the multiprocessor system may be tightly coupled or loosely coupled (e.g., to form a computer cluster).

[0107] Processor 22 may perform functions associated with the operation of device 20, including, as some examples, precoding of antenna gain / phase parameters, encoding and decoding of individual bits forming communication messages, formatting of information, and overall control of device 20, including processes related to management of communication resources.

[0108] The device 20 may further include or be coupled to a memory 24 (internal or external), which may be coupled to the processor 22 for storing information and instructions that can be executed by the processor 22. The memory 24 may be one or more memories and may be of any type suitable for the local application environment and may be implemented using any suitable volatile or non-volatile data storage technology, such as semiconductor-based memory devices, magnetic memory devices and systems, optical memory devices and systems, fixed memory, and / or removable memory. For example, the memory 24 may include any combination of random access memory (RAM), read-only memory (ROM), static storage (such as a magnetic disk or optical disk), a hard disk drive (HDD), or any other type of non-transitory machine or computer readable medium. The instructions stored in the memory 24 may include program instructions or computer program codes that, when executed by the processor 22, enable the device 20 to perform tasks as described herein.

[0109] In an embodiment, the device 20 may also include or be coupled to a drive or port (internal or external) configured to accept and read an external computer-readable storage medium, such as an optical disk, a USB drive, a flash drive, or any other storage medium. For example, the external computer-readable storage medium may store a computer program or software for execution by the processor 22 and / or the device 20.

[0110] In some embodiments, the device 20 may also include or be coupled to one or more antennas 25 for receiving downlink signals and for transmitting from the device 20 via an uplink. The device 20 may also include a transceiver 28 configured to send and receive information. The transceiver 28 may also include a radio interface (e.g., a modem) coupled to the antenna 25. The wireless interface may correspond to a variety of radio access technologies, including one or more of GSM, LTE, LTE-A, 5G, NR, WLAN, NB-IoT, Bluetooth, BT-LE, NFC, RFID, UWB, etc. The radio interface may include other components, such as filters, converters (e.g., digital-to-analog converters, etc.), symbol demappers, signal shaping components, inverse fast Fourier transform (IFFT) modules, etc., to process symbols carried by the downlink or uplink, such as OFDMA symbols.

[0111] For example, the transceiver 28 may be configured to modulate information about a carrier waveform for transmission by the antenna 25, and to demodulate information received via the antenna 25 for further processing by other elements of the device 20. In other embodiments, the transceiver 28 may be capable of directly sending and receiving signals or data. Additionally or alternatively, in some embodiments, the device 20 may include input and / or output devices (I / O devices). In some embodiments, the device 20 may also include a user interface, such as a graphical user interface or a touch screen.

[0112] In an embodiment, the memory 24 stores software modules that provide functionality when executed by the processor 22. The modules may include, for example, an operating system that provides operating system functionality for the device 20. The memory may also store one or more functional modules, such as applications or programs, to provide additional functionality for the device 20. The components of the device 20 may be implemented in hardware or as any suitable combination of hardware and software. According to an example embodiment, the device 20 may optionally be configured to communicate with the device 10 via a wireless or wired communication link 70 according to any radio access technology such as NR.

[0113] According to some embodiments, the processor 22 and the memory 24 may be included in or may form part of a processing circuit or a control circuit. In addition, in some embodiments, the transceiver 28 may be included in or may form part of a transceiver circuit.

[0114] As described above, according to some embodiments, for example, the apparatus 20 may be a UE, a SL UE, a relay UE, a mobile device, a mobile station, a ME, an IoT device, and / or a NB-IoT device, etc. According to some embodiments, the apparatus 20 may be controlled by the memory 24 and the processor 22 to perform functions associated with any of the embodiments described herein, for example Figures 1 to 9 As described in or about Figures 1 to 9 One or more of the operations described, or any other method described herein. For example, in an embodiment, the apparatus 20 may be controlled to perform processes related to providing beam selection using data radio bearer specific machine learning, as described in detail elsewhere herein.

[0115] In some embodiments, an apparatus (e.g., apparatus 10 and / or apparatus 20) may include a unit for performing a method, process, or any variant discussed herein. An example of an apparatus may include one or more processors, memories, controllers, transmitters, receivers, and / or computer program codes for causing the execution of any operation discussed herein.

[0116] In view of the foregoing, certain example embodiments provide several technical improvements, enhancements and / or advantages over prior art processes and improve at least the technical field of wireless network control and / or management. Certain embodiments may have various benefits and / or advantages. For example, certain embodiments may allow beam selection to be controlled in a manner specific to data radio bearers, thereby allowing greater throughput, etc.

[0117] In some example embodiments, the functionality of any method, process, signaling diagram, algorithm, or flow chart described herein may be implemented by software and / or computer program code or code portions stored in a memory or other computer-readable or tangible medium and executed by a processor.

[0118] In some example embodiments, the device may include or be associated with at least one software application, module, unit, or entity, which is configured as an arithmetic operation, or as a program or part of a program (including added or updated software routines) that can be executed by at least one operating processor or controller. Programs, also known as program products or computer programs, include software routines, applets, and macros, which can be stored in any device-readable data storage medium and may include program instructions for performing specific tasks. A computer program product may include one or more computer executable components that are configured to perform some example embodiments when the program is running. One or more computer executable components may be at least one software code or code portion. Modifications and configurations required to implement the functionality of the example embodiments may be performed as routines, which may be implemented as added or updated software routines. In one example, the software routines may be downloaded to the device.

[0119] As an example, software or computer program code or code portions may be in source code form, object code form or in some intermediate form, and may be stored in some carrier, distribution medium or computer readable medium, which may be any entity or device capable of carrying the program. For example, such carriers may include recording media, computer memory, read-only memory, optoelectronic and / or electrical carrier signals, telecommunication signals and / or software distribution packages. Depending on the processing power required, the computer program may be executed in a single electronic digital computer, or it may be distributed among multiple computers. The computer readable medium or computer readable storage medium may be a non-transitory medium.

[0120] In other example embodiments, the functions of the example embodiments may be performed by hardware or circuitry included in a device, such as by using an application specific integrated circuit (ASIC), a programmable gate array (PGA), a field programmable gate array (FPGA), or any other combination of hardware and software. In yet another example embodiment, the functions of the example embodiments may be implemented as a signal carried by an electromagnetic signal downloaded from the Internet or other network, such as a non-tangible device.

[0121] According to example embodiments, an apparatus such as a node, a device or a corresponding component may be configured as a circuit, a computer or a microprocessor, such as a single-chip computer element or a chipset, which may include at least a memory for providing storage capacity for arithmetic operations and / or an operation processor for performing arithmetic operations.

[0122] The example embodiments described herein may apply to both singular and plural implementations, regardless of whether the singular or plural is used in conjunction with describing certain embodiments. For example, an embodiment describing the operation of a single network node may also apply to example embodiments including multiple instances of the network node, and vice versa.

[0123] Those of ordinary skill in the art will readily appreciate that the example embodiments described above may be practiced in different orders and / or with hardware elements in configurations different from those disclosed. Therefore, although some embodiments have been described based on these example embodiments, it will be apparent to those skilled in the art that certain modifications, variations, and alternative constructions will be apparent while remaining within the spirit and scope of the example embodiments. Partial glossary: CE control components CSI channel state information DCI downlink control information gNB 5G-NR base station MAC Media Access Control ML Machine Learning ML-RNTI Machine Learning Radio Network Temporary Identifier mmW NR New Radio RB Resource Block RL Reinforcement Learning RSRP reference signal received power RSRQ Reference Signal Received Quality RSSI Received Signal Strength Indicator SINR Signal to Interference Plus Noise Ratio TTI transmission time interval UE User Equipment

Claims

1. A device, include: at least one processor; as well as at least one memory including computer program code, wherein the at least one memory and the computer program code are configured to, together with the at least one processor, cause the apparatus to at least: providing one or more inputs regarding the plurality of beams to a machine learning model; obtaining at least one output value from the machine learning model regarding the plurality of beams, The machine learning model comprises a data radio bearer specific machine learning model, a data radio bearer group specific machine learning model, or a model trained to selectively output a data radio bearer specific value or a data radio bearer group specific value.

2. The apparatus of claim 1 , wherein the at least one memory and the computer program code are configured to, together with the at least one processor, cause the apparatus to at least: At least one suggested beam is obtained from the machine learning model, and the radio access network node is to send or receive radio communication with the user equipment on the at least one suggested beam.

3. The apparatus of claim 2, wherein the machine learning model is configured to select the at least one suggested beam based on the resulting delay or network capacity, regardless of whether the signal quality of the suggested beam is higher or lower than that of an alternative beam.

4. The apparatus of claim 2, wherein the at least one memory and the computer program code are configured to, with the at least one processor, cause the apparatus to at least: checking whether the at least one suggested beam is included in a report of the user equipment; selecting at least one beam for use based on the examining; as well as The user equipment is informed of the selected at least one beam. 5 . The apparatus of claim 4 , wherein if the at least one proposed beam is included in the report, at least one of the at least one proposed beam is selected for use based on the inclusion in the report. 6 . The apparatus of claim 4 , wherein if the at least one suggested beam is not included in the report, at least one beam included in the report is selected for use.

7. The apparatus of claim 6, wherein the machine learning model is penalized based on the at least one suggested beam not being included in the report.

8. An apparatus according to any one of claims 1 to 7, wherein the input comprises a protocol data unit delay value.

9. An apparatus as claimed in any one of claims 1 to 8, wherein the input comprises a buffer status.

10. The apparatus of any one of claims 1 to 9, wherein the input comprises usage statistics of a plurality of beams.

11. The apparatus of any one of claims 1 to 10, wherein the machine learning model is located in a next generation Node B.

12. The apparatus of claim 11, wherein the at least one memory and the computer program code are configured to, with the at least one processor, cause the apparatus to at least: The at least one output value is provided to a user device.

13. The apparatus of any one of claims 1 to 10, wherein the machine learning model is in a user device, and wherein the at least one memory and the computer program code are configured to, with the at least one processor, cause the apparatus to at least: The at least one output value is provided to a radio access node.

14. The apparatus of claim 13, wherein the input comprises usage statistics for a plurality of beams and wherein the usage statistics are received from the radio access node.

15. The apparatus of any one of claims 1 to 14, wherein the machine learning model comprises a separate neural network for each of the plurality of beams.

16. A method, include: providing one or more inputs regarding the plurality of beams to a machine learning model; obtaining at least one output value from the machine learning model regarding the plurality of beams, The machine learning model comprises a data radio bearer specific machine learning model, a data radio bearer group specific machine learning model, or a model trained to selectively output a data radio bearer specific value or a data radio bearer group specific value.

17. The method according to claim 16, further comprising: include: At least one suggested beam is obtained from the machine learning model, and the radio access network node is to send or receive radio communication with the user equipment on the at least one suggested beam.

18. The method of claim 17, wherein the machine learning model is configured to select the at least one suggested beam based on the resulting delay or network capacity, regardless of whether the signal quality of the suggested beam is higher or lower than an alternative beam.

19. The method according to claim 17, further comprising: include: checking whether the at least one suggested beam is included in a report of the user equipment; selecting at least one beam for use based on the examining; as well as The user equipment is informed of the selected at least one beam.

20. The method of claim 19, wherein if the at least one suggested beam is included in the report, at least one of the at least one suggested beam is selected for use based on the inclusion in the report.

21. The method of claim 19, wherein if the at least one suggested beam is not included in the report, at least one beam included in the report is selected for use.

22. The method of claim 21 , wherein the machine learning model is penalized based on the at least one suggested beam not being included in the report.

23. A method according to any one of claims 16 to 22, wherein the input comprises a protocol data unit delay value.

24. A method according to any one of claims 16 to 23, wherein the input comprises a buffer status.

25. A method according to any one of claims 16 to 24, wherein the input comprises usage statistics of a plurality of beams.

26. A method according to any one of claims 16 to 25, wherein the machine learning model is located in a next generation Node B.

27. The method according to claim 26, further comprising: include: The at least one output value is provided to a user device.

28. A method according to any one of claims 16 to 25, wherein the machine learning model is in a user equipment, and wherein the method further comprises providing the at least one output value to a radio access node.

29. The method of claim 28, wherein the input comprises usage statistics of a plurality of beams and wherein the usage statistics are received from the radio access node.

30. The method of any one of claims 16 to 29, wherein the machine learning model comprises a separate neural network for each of the plurality of beams.

31. A device, include: means for providing one or more inputs regarding a plurality of beams to a machine learning model; means for obtaining at least one output value from the machine learning model with respect to the plurality of beams, The machine learning model comprises a data radio bearer specific machine learning model, a data radio bearer group specific machine learning model, or a model trained to selectively output a data radio bearer specific value or a data radio bearer group specific value.

32. The device according to claim 31, further comprising: include: Means for obtaining from the machine learning model at least one suggested beam on which a radio access network node is to send or receive radio communications with a user equipment.

33. An apparatus according to claim 32, wherein the machine learning model is configured to select the at least one suggested beam based on the resulting delay or network capacity, regardless of whether the signal quality of the suggested beam is higher or lower than that of an alternative beam.

34. The device according to claim 32, further comprising: include: means for checking whether the at least one proposed beam is included in a report of the user equipment; means for selecting at least one beam for use based on said examining; as well as Means for notifying the user equipment of the selected at least one beam.

35. The apparatus of claim 34, wherein if the at least one proposed beam is included in the report, at least one of the at least one proposed beam is selected for use based on the inclusion in the report.

36. The apparatus of claim 34, wherein if the at least one suggested beam is not included in the report, at least one beam included in the report is selected for use.

37. An apparatus according to claim 36, wherein the machine learning model is penalized based on the at least one suggested beam not being included in the report.

38. An apparatus as claimed in any one of claims 31 to 37, wherein the input comprises a protocol data unit delay value.

39. An apparatus as claimed in any one of claims 31 to 38, wherein the input comprises a buffer status.

40. An apparatus according to any one of claims 31 to 39, wherein the input comprises usage statistics of a plurality of beams.

41. An apparatus according to any one of claims 31 to 40, wherein the machine learning model is located in a next generation Node B.

42. The device according to claim 41, further comprising: include: Means for providing the at least one output value to a user device.

43. An apparatus according to any one of claims 31 to 40, wherein the machine learning model is in a user equipment, and wherein the apparatus further comprises means for providing the at least one output value to a radio access node.

44. The apparatus of claim 43, wherein the input comprises usage statistics for a plurality of beams and wherein the usage statistics are received from the radio access node.

45. An apparatus according to any one of claims 31 to 44, wherein the machine learning model comprises a separate neural network for each of the plurality of beams.