Apparatus, method and computer program for supporting machine learning function

Abstract

An apparatus comprises: means for receiving configuration information from a base station, the configuration information comprising information related to a machine learning function; means for selecting one or more machine learning options from a plurality of different machine learning options providing the machine learning functionality; and means for causing performance information to be transmitted to the base station, said performance information relating to the performance of the selected one or more machine learning options.

Description

TECHNICAL FIELD

[0001] The present disclosure relates to apparatuses, methods, and computer programs, and in particular, but not exclusively, to apparatuses, methods, and computer programs related to apparatuses, methods, and computer programs supporting machine learning functionality. BACKGROUND

[0002] A communication system can be seen as a facility that enables communication between two or more communication devices or provides access to a data network.

[0003] The communication system can be a wireless communication system. Examples of wireless communication systems include public land mobile networks (PLMN) operating based on a radio access technology standard, such as provided by 3GPP (Third Generation Partnership Project) or ETSI (European Telecommunications Standards Institute), satellite communication systems, and different wireless local area networks, e.g., wireless local area networks (WLAN). Wireless communication systems operating based on a radio access technology are typically divided into cells, and are therefore often referred to as cellular systems.

[0004] Communication systems and related devices typically operate in accordance with one or more radio access technologies defined in a given specification of a standard, such as provided by 3GPP or ETSI, which specifies what the various entities associated with the communication system and the communication devices accessing or connected to the communication system are allowed to do and how that should be implemented. The communication protocols and / or parameters used by the communication devices for accessing or connecting to the communication system are also typically defined in the standard. An example of a standard is the so-called 5G (Fifth Generation) standard provided by 3GPP. SUMMARY

[0005] According to one aspect, there is provided an apparatus comprising: means for receiving configuration information from a base station, the configuration information comprising information related to a machine learning functionality; means for selecting one or more machine learning options from a plurality of different machine learning options providing the machine learning functionality; and means for causing performance information to be transmitted to the base station, the performance information relating to performance of the selected one or more machine learning options.

[0006] The information related to the machine learning functionality can comprise: information identifying a machine learning functionality identity and / or information identifying a machine learning model.

[0007] The machine learning options can comprise one or more of: a machine learning algorithm, or a machine learning architecture.

[0008] The configuration information can comprise data associated with use of one or more of the different machine learning options by one or more other user equipment, and the means for selecting a machine learning option can be for selecting the one or more machine learning options based on the data.

[0009] The data can comprise information about one or more cell conditions when the respective machine learning option is used by one or more other user equipment.

[0010] The apparatus can comprise means for determining one or more cell conditions, and wherein the means for selecting a machine learning option can be for selecting the one or more machine learning options based on the determined one or more cell conditions and data about one or more cell conditions when the respective machine learning option is used by one or more other user equipment.

[0011] The performance information can comprise feedback information about consensus of data associated with use of the selected one or more machine learning options by one or more other user equipment.

[0012] The performance information can be provided together with information about the selected one or more machine learning options.

[0013] The performance information can comprise information about one or more network conditions when the selected one or more machine learning options are used.

[0014] The performance information can comprise one or more of: one or more predictions made by use of the selected one or more machine learning options; accuracy of the one or more predictions; and statistical bias information.

[0015] The performance information can be provided together with timestamp information.

[0016] The configuration information can comprise information indicating performance information to be transmitted to the base station.

[0017] The apparatus can comprise means for switching between a plurality of the selected machine learning options.

[0018] The configuration information can comprise configuration information for a handover procedure for a handover from the base station to a target base station.

[0019] The selected machine learning option can be provided by a machine learning model.

[0020] The machine learning model can comprise a proprietary machine learning model.

[0021] The components of any of the examples discussed above can be provided at least in part by circuitry.

[0022] The components of any of the examples discussed above can be provided, at least in part, by circuitry capable of supporting machine learning functionality.

[0023] The components of any of the examples discussed above can be provided, at least in part, by at least one processor and at least one memory. The at least one memory can comprise computer executable code.

[0024] The apparatus can be provided in or be a user equipment.

[0025] According to a further aspect, there is provided a method comprising: receiving configuration information from a base station, the configuration information comprising information relating to a machine learning functionality; selecting one or more machine learning options from a plurality of different machine learning options providing the machine learning functionality; and causing performance information to be transmitted to the base station, the performance information relating to performance of the selected one or more machine learning options.

[0026] The information relating to the machine learning functionality can comprise: information identifying an identity of the machine learning functionality and / or information identifying a machine learning model.

[0027] The machine learning options can comprise one or more of: a machine learning algorithm, or a machine learning architecture.

[0028] The configuration information can comprise: data associated with use of one or more of the different machine learning options by one or more other user equipment, and the means for selecting the machine learning options can be for selecting the one or more machine learning options based on the data.

[0029] The data can comprise: information about one or more cell conditions when the respective machine learning options were used by the one or more other user equipment.

[0030] The method can comprise: determining one or more cell conditions, and selecting the one or more machine learning options based on the determined one or more cell conditions, and the data about one or more cell conditions when the respective machine learning options were used by the one or more other user equipment.

[0031] The performance information can comprise: feedback information about consensus of data associated with use of the selected one or more machine learning options by the one or more other user equipment.

[0032] The performance information can be provided with information about the selected one or more machine learning options.

[0033] The performance information can comprise information about one or more network conditions when the selected one or more machine learning options are used.

[0034] The performance information can comprise one or more of: one or more predictions acquired by use of the selected one or more machine learning options; accuracy of the one or more predictions; and statistical bias information.

[0035] The performance information can be provided together with timestamp information.

[0036] The configuration information can comprise information indicating the performance information to be transmitted to the base station.

[0037] The method can comprise switching between a plurality of the selected machine learning options.

[0038] The configuration information can comprise configuration information for a handover procedure for a handover from the base station to a target base station.

[0039] The selected machine learning option can be provided by a machine learning model.

[0040] The machine learning model can comprise a proprietary machine learning model.

[0041] The method can be performed by an apparatus. The apparatus can be provided in or be a user equipment.

[0042] According to a further aspect, there is provided an apparatus comprising at least one processor and at least one memory storing instructions that, when executed by the at least one processor, cause the apparatus at least to: receive, from a base station, configuration information comprising information relating to a machine learning function; select one or more machine learning options from a plurality of different machine learning options providing the machine learning function; and cause performance information to be transmitted to the base station, the performance information relating to performance of the selected one or more machine learning options.

[0043] The information relating to the machine learning function can comprise information identifying a machine learning function identity and / or information identifying a machine learning model.

[0044] The machine learning options can comprise one or more of: a machine learning algorithm, or a machine learning architecture.

[0045] The configuration information can comprise data associated with use of one or more of the different machine learning options by one or more other user equipment, and the means for selecting the machine learning option can be for selecting the one or more machine learning options based on the data.

[0046] The data can comprise information about one or more cell conditions when the respective machine learning option was used by one or more other user equipment.

[0047] The apparatus can be caused to determine one or more cell conditions, and wherein the means for selecting a machine learning option can be to select one or more machine learning options based on the determined one or more cell conditions and the data about one or more cell conditions when the respective machine learning option was used by one or more other user equipment.

[0048] The performance information can comprise feedback information about a consensus of data associated with the use of the selected one or more machine learning options by one or more other user equipment.

[0049] The performance information can be provided together with information about the selected one or more machine learning options.

[0050] The performance information can comprise information about one or more network conditions when the selected one or more machine learning options were used.

[0051] The performance information can comprise one or more of: one or more predictions made by the use of the selected one or more machine learning options; accuracy of the one or more predictions; and statistical bias information.

[0052] The performance information can be provided together with timestamp information.

[0053] The configuration information can comprise information indicating performance information to be transmitted to the base station.

[0054] The apparatus can be caused to switch between a plurality of the selected machine learning options.

[0055] The configuration information can comprise configuration information for a handover procedure for a handover from the base station to a target base station.

[0056] The selected machine learning option can be provided by a machine learning model.

[0057] The machine learning model can comprise a proprietary machine learning model.

[0058] The apparatus can be provided in or be a user equipment.

[0059] According to a further aspect, there is provided an apparatus comprising: means for providing configuration information to a first user equipment, the configuration information comprising information related to a machine learning function; and means for receiving performance information from the first user equipment, the performance information relating to performance of one or more machine learning options selected by the first user equipment to provide the machine learning function.

[0060] The information related to the machine learning function can comprise information identifying an identity of the machine learning function and / or information identifying a machine learning model.

[0061] The configuration information can comprise data associated with use of one or more different machine learning options of different machine learning options to provide the machine learning function by one or more other user equipment.

[0062] The data can comprise information about one or more cell conditions when the respective machine learning option was used by the one or more other user equipment.

[0063] The performance information can comprise feedback information about a consensus of the first user equipment’s data associated with use of the one or more machine learning options selected by the first user equipment by the one or more other user equipment.

[0064] The performance information can be provided together with information about the selected one or more machine learning options.

[0065] The performance information can comprise information about one or more network conditions when the selected one or more machine learning options was used by the first user equipment.

[0066] The performance information can comprise one or more of: one or more predictions made by use of the selected one or more machine learning options; accuracy of the one or more predictions; and statistical bias information.

[0067] The performance information can be provided together with timestamp information.

[0068] The configuration information can comprise information indicating performance information to be sent by the first user equipment.

[0069] The configuration information can comprise configuration information of a handover procedure of the first user equipment for a handover from a base station to a target base station. The apparatus can comprise means for receiving performance information for the first user equipment from the target base station.

[0070] The selected machine learning option can be provided by a machine learning model.

[0071] The machine learning model can comprise a proprietary machine learning model.

[0072] The components of any of the examples discussed above can be provided at least in part by circuitry.

[0073] The components of any of the examples discussed above can be provided at least in part by circuitry capable of supporting machine learning functionality.

[0074] The components of any of the examples discussed above can be provided at least in part by at least one processor and at least one memory. The at least one memory can include computer executable code.

[0075] The apparatus can be or be provided in a base station.

[0076] According to a further aspect, there is provided a method comprising: providing configuration information to a first user equipment, the configuration information comprising information relating to a machine learning functionality; and receiving performance information from the first user equipment, the performance information relating to performance of one or more machine learning options selected by the first user equipment to provide the machine learning functionality.

[0077] The information relating to the machine learning functionality can comprise information identifying an identity of the machine learning functionality and / or information identifying a machine learning model.

[0078] The configuration information can comprise data associated with use of one or more different machine learning options of different machine learning options to provide the machine learning functionality by one or more other user equipment.

[0079] The data can comprise information about one or more cell conditions when the respective machine learning option was used by the one or more other user equipment.

[0080] The performance information can comprise feedback information about a consensus of the first user equipment’s data associated with use of the one or more machine learning options selected by the first user equipment by the one or more other user equipment.

[0081] The performance information can be provided with information about the selected one or more machine learning options.

[0082] The performance information can comprise information about one or more network conditions when the selected one or more machine learning options were used by the first user equipment.

[0083] The performance information can comprise one or more of: one or more predictions made by use of the selected one or more machine learning options; accuracy of the one or more predictions; and statistical bias information.

[0084] The performance information can be provided with timestamp information.

[0085] The configuration information can comprise information indicating performance information to be transmitted by the first user equipment.

[0086] The configuration information can comprise configuration information of a handover procedure of the first user equipment for a handover from the base station to a target base station. The method can comprise receiving performance information for the first user equipment from the target base station.

[0087] The selected machine learning option can be provided by a machine learning model.

[0088] The machine learning model can comprise a proprietary machine learning model.

[0089] The method can be performed by an apparatus. The apparatus can be or be arranged in a base station.

[0090] According to a further aspect, there is provided an apparatus comprising at least one processor and at least one memory storing instructions that, when executed by the at least one processor, cause the apparatus at least to: provide configuration information to a first user equipment, the configuration information comprising information related to a machine learning function; and receive performance information from the first user equipment, the performance information relating to a performance of one or more machine learning options selected by the first user equipment to provide the machine learning function.

[0091] The information related to the machine learning function can comprise information identifying an identity of the machine learning function and / or information identifying a machine learning model.

[0092] The configuration information can comprise data associated with a use of one or more different machine learning options of the machine learning function by one or more other user equipment.

[0093] The data can comprise information about one or more cell conditions when the respective machine learning option is used by the one or more other user equipment.

[0094] The performance information can comprise feedback information about a consensus of the first user equipment with data associated with the use of the one or more machine learning options by the one or more other user equipment.

[0095] The performance information can be provided together with information about the selected one or more machine learning options.

[0096] The performance information can comprise information about one or more network conditions when the selected one or more machine learning options is used by the first user equipment.

[0097] The performance information can comprise one or more of: one or more predictions obtained by use of the selected one or more machine learning options; accuracy of the one or more predictions; and statistical bias information.

[0098] The performance information can be provided together with time stamp information.

[0099] The configuration information can comprise information indicating performance information to be transmitted by the first user equipment.

[0100] The configuration information can comprise configuration information of a handover procedure of the first user equipment for a handover from the base station to a target base station. The apparatus can be caused to receive performance information for the first user equipment from the target base station.

[0101] The selected machine learning option can be provided by a machine learning model.

[0102] The machine learning model can comprise a proprietary machine learning model.

[0103] The apparatus can be or be arranged in a base station.

[0104] According to a further aspect, there is provided a computer program comprising instructions which, when executed by the apparatus, cause the apparatus to carry out any of the previously stated methods.

[0105] According to a further aspect, there is provided a computer program comprising instructions which, when executed, cause any of the previously stated methods to be carried out.

[0106] According to an aspect, there is provided a computer program comprising computer executable instructions which, when executed, cause any of the previously stated methods to be carried out.

[0107] According to an aspect, there is provided a computer readable medium having stored thereon program instructions for carrying out at least one of the above methods.

[0108] According to an aspect, there is provided a non-transitory computer readable medium comprising program instructions which, when executed by the apparatus, cause the apparatus to carry out any of the previously stated methods.

[0109] According to an aspect, there is provided a non-transitory computer readable medium comprising program instructions which, when executed, cause any of the previously stated methods to be carried out.

[0110] According to an aspect, there is provided a non-transitory tangible storage medium having stored thereon program instructions for executing at least one of the above-described methods.

[0111] In the foregoing, a number of different aspects have been described. It should be appreciated that other aspects can be provided by a combination of any two or more of the aspects described above.

[0112] Various other aspects are also described in the following detailed description and the appended claims. BRIEF DESCRIPTION OF DRAWINGS

[0113] Some examples will now be described in greater detail by way of illustration only with reference to the accompanying drawings, in which:

[0114] Figure 1a A schematic illustration of a 5G system is shown;

[0115] Figure 1b A schematic illustration of an apparatus is shown;

[0116] Figure 1c A schematic illustration of a user equipment is shown;

[0117] Figure 2 An example message sequence flow for conditional handover is shown;

[0118] Figure 3 Some non-limiting example beam management (BM) use cases for AI / ML are shown;

[0119] Figure 4 An example configuration for BM case 1 is shown;

[0120] Figure 5 A signal flow of some embodiments is shown;

[0121] Figure 6 A method of some embodiments is shown; and

[0122] Figure 7 Another method of some embodiments is shown. DETAILED DESCRIPTION

[0123] The following abbreviations which can appear in the specification and / or drawings, are defined as follows:

[0124] 3GPP: Third Generation Partnership Project

[0125] 5G: Fifth Generation

[0126] 5GC: 5G Core Network

[0127] 5GS: 5G System

[0128] 6G: Sixth Generation

[0129] AF: Application Function

[0130] AI: Artificial Intelligence

[0131] AMF: Access and Mobility Management Function

[0132] BM: Beam Management

[0133] CHO: Conditional Handover

[0134] CSI: Channel State Information

[0135] CSI-RS: Channel State Information Reference Signal

[0136] DCI: Downlink Control Information

[0137] DN: Data Network

[0138] gNB (or gNodeB): Base station for 5G / NR, i.e. a node that provides NR user plane and control plane protocol terminations towards the UE and that is connected via an NG interface to a 5GC

[0139] HO: Handover

[0140] ID: Identity

[0141] KPI: Key Performance Indicator

[0142] L1: Layer

[0143] LCM: Lifecycle Management

[0144] LTE: Long Term Evolution

[0145] LSTM: Long Short-Term Memory

[0146] MAC: Medium Access Control

[0147] MAC-CE: Medium Access Control Control Element

[0148] ML: Machine Learning

[0149] MME: Mobility Management Entity

[0150] MSE: Mean Squared Error

[0151] ng or NG: New Generation

[0152] NN: Neural Network

[0153] NR: New Radio

[0154] N / W or NW: Network

[0155] PLMN: Public Land Mobile Network

[0156] PHY: Physical Layer

[0157] PRACH: Physical Random Access Channel

[0158] RA: Random Access

[0159] RACH: Random Access Channel

[0160] RAN: Radio Access Network

[0161] RRC: Radio Resource Control

[0162] RSRP: Reference Signal Received Power

[0163] RX: Receiver

[0164] SGW or S-GW: Serving Gateway

[0165] SMF: Session Management Function

[0166] SN: Sequence Number

[0167] SSB: Synchronization Signal Block

[0168] TX: Transmitter

[0169] UE: User Equipment (e.g., wireless device, typically a mobile device)

[0170] UPF: User Plane Function

[0171] In the following, certain embodiments are explained with reference to a communication device capable of communicating via a wireless cellular system and a mobile communication system serving such a communication device. Before explaining the example embodiments in detail, it is to be understood that the example described below is not limited in its application to the details explained below, but is capable of application to other examples and embodiments. Figure 1a , Figure 1b and Figure 1c Certain general principles of wireless communication systems, their access systems and communication devices are briefly explained to help understanding the technology behind the described examples.

[0172] Figure 1a A schematic diagram of a communication system operating based on the fifth generation radio access technology, often referred to as 5G system (5GS), is shown. The 5GS can comprise a (radio) access network ((R)AN), a 5G core network (5GC), one or more application functions (AFs) and one or more data networks (DN). A user equipment can access or connect to one or more DNs via the 5GS.

[0173] A 5G(R)AN can comprise one or more base stations or radio access network (RAN) nodes, such as gNodeB (gNB). A base station or RAN node can comprise one or more distributed units connected to a central unit.

[0174] A 5GC can comprise various network functions, such as an access and mobility management function (AMF), a session management function (SMF), an authentication server function (AUSF), a user data management (UDM), a user plane function (UPF), a network data repository, a network exposure function (NEF), a service communication proxy (SCP), an edge application server discovery function (EASDF), a policy control function (PCF), a network slice access control function (NSACF), a network slice specific authentication and authorization function (NSSAAF), and / or a network slice selection function (NSSF).

[0175] Figure 1b An example of an apparatus 100 is illustrated. The apparatus 100 can be provided in a communication device and / or a network entity. The apparatus 100 can have at least one processor and at least one memory storing instructions which, when executed by the at least one processor, cause one or more functions to be performed. In this example, the apparatus can comprise at least one random access memory (RAM) 111a and / or at least one read-only memory (ROM) 111b. The apparatus can comprise at least one processor 112, 113 and / or an input / output interface 114. The at least one processor 112, 113 can be coupled to the at least one memory, which in this example is the RAM 111a and the ROM 111b. The at least one processor 112, 113 can be configured to execute appropriate software code 115. For example, the software code 115 can allow the apparatus to perform one or more steps of one or more aspects of the present invention.

[0176] Figure 1c An example of a communication device 300 is illustrated. The communication device 300 can be any device capable of transmitting and receiving radio signals. Further non-limiting examples of the communication device 300 include a mobile station (MS) or mobile device such as a mobile phone or so-called “smartphone”, a computer provided with a wireless interface card or other wireless interface facility such as a USB dongle, a personal data assistant (PDA) or a tablet computer provided with wireless communication functionality, a machine-type communication (MTC) device, a cellular Internet of Things (CIoT) device, or any combination of these devices, etc.

[0177] The communication device 300 can transmit or receive, for example, radio signals carrying communications. The communications can be one or more of voice, electronic mail (email), text message, multimedia, data, machine data, etc.

[0178] The communication device 300 can receive radio signals over the air, or radio interface, 307 via the transceiver arrangement 306. The transceiver arrangement 306 can be provided, for example, by a radio and an associated antenna arrangement. The antenna arrangement can be arranged internally or externally to the mobile device, and can include a single antenna or a plurality of antennas. The antenna arrangement can be an antenna array comprising a plurality of antenna elements.

[0179] The communication device 300 can be provided with at least one processor 301 and / or at least one memory. The at least one memory can be at least one ROM 302a and / or at least one RAM 302b. Other possible components can be provided, as is deemed necessary and / or as is customary, for software and hardware assisted execution of the tasks it is designed to perform, including controlling access to and communication with access systems, such as a 5G RAN, and other communication devices. The at least one processor 301 is coupled to the RAM 302b and the ROM 302a. The at least one processor 301 can be configured to execute instructions of a software code 308. Execution of the instructions of the software code 308 can for example allow the communication device 300 to perform one or more operations. The software code 308 can be stored in the ROM 302a. It is to be understood that, in other embodiments, any other suitable memory can alternatively or additionally be used.

[0180] The at least one processor 301, the at least one ROM 302a, and / or the at least one RAM 302b can be provided on an appropriate circuit board, in an integrated circuit, and / or in a chipset. This feature is denoted by reference 304.

[0181] A machine learning module can be provided, which provides circuitry to support one or more machine learning models. The circuitry can comprise neural network circuitry, or any other suitable circuitry for supporting machine learning models.

[0182] The communication device 300 can optionally have a user interface, such as a keypad 305, a touch-sensitive screen or touchpad, combinations thereof, and the like. Optionally, the communication device can have one or more of a display, a speaker, and a microphone.

[0183] In the following examples, the term UE or user equipment is used. This term encompasses any of the examples of the communication device 300 discussed previously and / or any other communication device.

[0184] An example of a wireless communication system is an architecture standardized by the Third Generation Partnership Project (3GPP). Radio access technologies currently standardized by 3GPP2 are commonly referred to as 5G or NR. Other radio access technologies standardized by 3GPP include Long Term Evolution (LTE) or LTE-Advanced Pro of the Universal Mobile Telecommunication System (UMTS). A wireless communication system typically includes an access network, such as a radio access network operating based on a radio access technology, including base stations or radio access network nodes. A wireless communication system can also include other types of access networks, such as wireless local area networks (WLAN) and / or WiMAX (Worldwide Interoperability for Microwave Access) networks.

[0185] It should be appreciated that example embodiments can also be used with standards such as 6G and beyond future radio access technologies.

[0186] Some embodiments can generally relate to artificial intelligence (AI) / machine learning (ML). Some embodiments can relate to AI / ML for the air interface. This can be applicable to NR, 5G, or any other suitable standard.

[0187] There is a REL-18 study item (see, e.g., Artificial Intelligence (AI) / Machine Learning (ML) for NR Air Interface described in RP-213599 (https: / / www.3gpp.org / ftp / TSG_RAN / TSG_RAN / TSGR_94e / Docs / RP-213599.zip)).

[0188] The study item aims to explore the benefits of enhancing the air interface with features that enable support for AI / ML based algorithms to improve performance and / or reduce complexity and / or reduce overhead. The goal of the study item is to lay the foundation for future air interface use cases that leverage AI / ML techniques. The initial set of use cases to be covered include: Channel State Information (CSI) feedback enhancements (e.g., compression of CSI reports, reduction of overhead, increase of accuracy, prediction), beam management (e.g., temporal and / or spatial beam prediction for reduction of overhead and / or latency, increase of beam selection accuracy), and positioning accuracy enhancements. For these use cases, the benefits can be evaluated (e.g., with developed methods and defined key performance indicators (KPIs)), and potential impact on specification(s) can be evaluated, including PHY layer aspects and protocol aspects.

[0189] Some embodiments can be used with any of these use cases or any other suitable use case.

[0190] Some embodiments can relate to the following study item: “AI / ML methods for selected sub-use cases need to be diverse enough to support various requirements at gNB-UE collaboration levels.”

[0191] The level of cooperation can indicate whether the gNB or the UE or both are “running” the ML-enabled functionality, and how and which information flows are provided between the gNB and the UE to allow “running” such ML-enabled functionality. This can include aspects of life cycle management such as data collection, model training, model inference, model monitoring functionality, etc.

[0192] It should be noted that other embodiments can be related to the work item phase of “AI / ML for the air interface”. Starting from Release 18, companies can come up with a large number of use cases and applications for ML in gNB and UE. This work item aims to explore the benefits of enhancing the air interface with features that enable improved support for AI / ML-based algorithms to improve performance and / or reduce complexity and / or reduce overhead. The enhanced performance can depend on the use case considered and, for example, can improve throughput, robustness, accuracy, and / or reliability, etc. One outcome of this work item can be that sufficient use cases will be considered to enable the determination of a general AI / ML framework, including functional requirements for AI / ML architectures, which can be used for follow-on projects. The study can also determine areas in which AI / ML can improve air interface function performance. Specification impact can be evaluated to improve the overall understanding of what is needed to enable AI / ML techniques for the air interface.

[0193] Some embodiments described herein can generally relate to one or more AI / ML models. For example, the AI / ML model(s) can be located with each of the UE and the base station (e.g., split between multiple nodes, or use a separate model at each node). Alternatively, the AI / ML model can be located with one of the UE or the base station.

[0194] The AI / ML model can be implemented by a neural network. A neural network (NN) is a computational graph composed of two or more layers of computations. Each layer can be composed of one or more units, where each unit can perform a basic computation. One unit can be connected to one or more other units, and the connection can have a weight associated with it. The weight can be used to scale the signal that passes through the associated connection. The weight can be a learnable parameter, i.e., a value that can be learned from training data. There can be other learnable parameters as well, such as the parameters of a batch normalization layer.

[0195] The two most widely used architectures for neural networks are feedforward and recurrent architectures. A feedforward neural network does not include feedback loops; each layer takes input from one or more previous layers and provides output that is used as input to one or more subsequent layers. The units within a layer take input from the unit(s) in the preceding layer(s) and provide output to the unit(s) in the following layer(s).

[0196] The initial layers (i.e., the layers close to the input data) extract low-level semantic features from the received data, and the intermediate and final layers extract higher-level features. After the feature extraction layers, there can be one or more layers that perform specific tasks, such as classification, semantic segmentation, object detection, denoising, style transfer, super-resolution, etc. In recurrent neural networks, there is a feedback loop, making the network stateful, i.e., it is able to memorize or retain information or state.

[0197] As mentioned above, machine learning models can be applied in an increasing number of applications for many different types of devices, such as mobile phones. Examples of applications can include image and video analysis and processing, social media data analysis, device usage data analysis, and the examples discussed earlier.

[0198] Neural networks and other machine learning tools can be able to learn properties from input data in a supervised or unsupervised manner. This learning can be the result of a training algorithm or the result of a meta-level neural network that provides training signals.

[0199] A training algorithm can include changing some properties of the neural network so that the output of the neural network is as close as possible to the desired output. Training can include changing the properties of the neural network to minimize or reduce the error in the output, also known as loss. Examples of loss include mean squared error (MSE), cross-entropy, etc. In recent deep learning techniques, training is an iterative process, where in each iteration the algorithm modifies the weights of the neural network to gradually improve the output of the network, i.e., to gradually reduce the loss.

[0200] Training of neural networks involves an optimization process, but the ultimate goal of machine learning is different from the goal of typical optimization. In optimization, the goal is to minimize the loss. In machine learning, the goal is typically to have the model learn properties of the data distribution from a limited training dataset in addition to the optimization goal. In other words, the training process is also used to ensure that the neural network learns to use the limited training dataset in order to learn to generalize to previously unseen data, i.e., data that was not used to train the model. This additional goal is often referred to as generalization. In practice, data can be split into at least two sets, a training set and a validation set. The training set can be used to train the network, i.e., modify its learnable parameters, to minimize the loss. The validation set can be used to check the performance of the neural network using data that was not used to minimize the loss, i.e., data that is not part of the training set, where the performance of the neural network using the validation set can be indicative of the ultimate performance of the model. During the training process, errors on the training set and the validation set can be monitored to understand whether the neural network is learning, and whether the neural network is learning to generalize. In case the network is learning, the training set error should decrease. If the network is not learning, the model can be in an underfit state. In case the network is learning to generalize, the validation set error should decrease and not be much higher than the training set error. If the training set error is low, but the validation set error is much higher than the training set error, or the validation set error does not decrease or even increases, the model can be in an overfit state. Overfitting can mean that the model has memorized properties of the training set and performs well only on that set, but not on sets that were not used to tune its parameters. In other words, the model has not learned to generalize.

[0201] In this specification, the terms “model”, “AI model”, and “ML model” can be used interchangeably with each other. In case an example embodiment is described with reference to one type of model, another type of model can be substituted.

[0202] Conditional handover (CHO) aims to improve the mobility performance of a UE by reducing the number of mobility failures. Referring now to Figure 2 , an example message sequence chart of conditional handover is illustrated. Figure 2 Parts of the conditional handover can be similar to handover described in NR Rel. 15 [TS 38.300].

[0203] The configured event can trigger the UE 205 to send a measurement report 232 to the source node 210. Based on the report, the source node 210 can make a CHO decision 234 to prepare one or more target cells for handover. The source node 210 can send a CHO request 236 to the target node 215, and optionally, to one or more other potential target nodes 220. The target node 215 can perform admission control 240. Optionally, one or more other potential target nodes 220 can perform admission control 242. The target node 215 can send a CHO request acknowledgement message 244 to the source node 210. Optionally, one or more other potential target nodes 220 can send a CHO request acknowledgement message 246 to the source node 210. The source node 210 can then send an RRC reconfiguration CHO command 248 to the UE 205.

[0204] For baseline handover, the UE can immediately access the target cell to complete handover. In contrast, for CHO, the UE 205 can only access the target cell 215 when an additional CHO execution condition expires, i.e., the HO preparation and execution phases are decoupled. This condition can be configured by the source node in the HO command. Thus, the UE 205 can evaluate one or more CHO conditions 250. Optionally, the UE 205 and the source node 210 can exchange user data 252. At 254, after the CHO condition is satisfied for a cell in the target node 215, the UE 205 can stop TX / RX to / from the source node 210 and start the handover HO. At 256, the UE 205 can send a PRACH preamble to the target node 215. At 258, the target node 215 can send a RACH response to the UE 205. At 260, the UE 205 can send an RRC reconfiguration complete message to the target node 215.

[0205] After the UE 205 completes handover to the target cell 215, e.g., the UE has sent the RRC reconfiguration complete 260, the target cell 215 can send a “handover success” indication 262 to the source cell 210. Upon receiving this indication from the target cell, at 264, the source cell 210 can stop its TX / RX to / from the UE and start forwarding data to the target cell 215, e.g., SN status transfer 268 to the target node 215, data forwarding 270 to the target node 215, and S-GW / UPF 225. Further, at 272, the source node 210 can release CHO preparations in one or more other target nodes / cells (e.g., 220) that are no longer needed when the source node receives the “HO success” indication. At 274, a path switch can be performed between the source node 210, the target node 215, the S-GW / UPF 225, and the MME / AMF 230.

[0206] The advantage of CHO is that the HO command can be sent very early while the UE is still safe in the source cell without jeopardizing the stability of the access and its radio link in the target cell. That is, conditional handover can provide mobility robustness and protection against mobility failure.

[0207] Reference Figure 3 which shows some non-limiting example beam management (BM) use cases for AI / ML. In this example, a gNB 310 can be configured with a ML model 320. The gNB 310 can transmit multiple different TX beams. In a first BM use case 330, spatial beam prediction can be performed by a UE 340 configured with a ML model 350, e.g., by predicting Tx beam #1 or Tx beam #3 given Tx beam #0, Tx beam #2, and Tx beam #4. As a result of ML model inference, the spatial beam prediction provides at least a prediction of the “best beam” ID(s) (e.g., best in terms of link quality metric DL signal strength (i.e., RSRP)) from a larger set of beams (set A) predicted from a smaller set of beams (set B), thereby utilizing fewer DL reference signal resources, and can also report a confidence interval (e.g., 95%) of the prediction and RSRP for the beam ID(s). Current discussions in 3GPP indicate that up to 8 beam IDs can be predicted in a Top K predicted beam list. However, in other cases, more or less than 8 beam IDs can be used for protection.

[0208] In a second BM use case 360, time beam prediction can be performed with a UE 370 configured with a ML model 380 that can be at a first location 370 at time t, and at a second location 390 at time t+T2 by predicting the best beam that can serve the UE at other times or at other times in a future time window. This allows resources to be prepared in advance for beams that have been predicted. This can provide lower interruption time in a beam switch procedure. For example, this can improve user throughput and experience.

[0209] The legacy beam management procedures P1, P2, P3 require a time-consuming operation of sweeping through all Tx and Rx beams by configuring the UE with a large number of SSB / CSI-RS measurements.

[0210] The Rel-18 study item discussed above involves AI / ML assisted beam management to achieve overhead saving and latency reduction. Further studied sub-use cases include BM Case 1 spatial beam prediction 330 and BM Case 2 temporal beam prediction 360. For each sub-use case, optimization targets can include: DL TX beam prediction (P1 / P2 joint optimization); and DL TX-RX beam pair prediction (P1 / P2 / P3 joint optimization). Performance targets and / or KPIs can include one or more of the following: one or more beam prediction accuracy related KPIs (e.g., prediction accuracy and / or RSRP difference); and / or one or more system performance related KPIs (e.g., UE throughput, control signal overhead, and / or power consumption).

[0211] Reference is now made to Figure 4 which shows a configuration for BM Case 1 (i.e., spatial beam prediction). For the ML model 440, different ML model inputs 410 are possible. In a first alternative 420 (Alt. 1), set B can be different from set A. In other words, various different beam RSRP measurements can be input. The bitmap location of the corresponding reference signal (e.g., 422) on the grid is illustrated.

[0212] In a second alternative 430 (Alt. 2), set B can be a subset of set A. The output 450 of the ML model 440 can be that set A is the best beam ID / RSRP prediction. The bitmap location of the corresponding reference signal (e.g., 432) on the grid is illustrated.

[0213] One non-limiting example of Alt. 1 can allow the UE to perform prediction of narrow beams in set A based on wide beams in set B. The same example can allow the use of even different reference signal extensions, e.g., set B can use SSB-based beams while set A can predict CSI-RS-based beams.

[0214] A further non-limiting example of Alt. 2 can allow prediction of set A from set B based on the same type of reference signal (i.e., wide beam to wide beam or narrow beam to narrow beam, hence called a subset).

[0215] The following configuration for BM Case 1 (i.e., spatial beam prediction) has been discussed by 3GPP. For the AI / ML input 410, L1-RSRP measurements can be measurements of a subset of narrow beams and / or wide beams. For the AI / ML input 410, assistance information (i.e., beam shape information, beam ID) can be input. For the AI / ML output 450, the best narrow beam ID or the best narrow beam RSRP, or an (internal) QoS value for beam selection can be output. An (offline) 5G system level simulator can provide model training data to the ML model 440.

[0216] It can be noted that the ML model 440 can be a UE-side ML model, a network-side ML model, a one-sided ML model, or a two-sided ML model. In other words, the inference of the ML model can be performed completely at the UE side, completely at the network side, or partially at the UE side and partially at the network side, with the ML model of each side being paired together to produce joint inference.

[0217] While Figure 3 And Figure 4 relate to beam management using AI / ML models, these examples are not limiting; AI / ML models can be used for other purposes, including but not limited to CSI feedback enhancement and positioning accuracy. For the example embodiments of the present disclosure, other uses of AI / ML models can be substituted where appropriate.

[0218] Table 1 below is a list of currently used terminology.

[0219]

[0220]

[0221]

[0222] Table 1

[0223] Currently, there are two separate model format categories being considered. These are proprietary format models and open format models.

[0224] Proprietary format models are ML models in a vendor / device specific proprietary format from a standards perspective. An example is a device specific binary executable format that the 3GPP network cannot interpret as the details of the format are implementation specific or proprietary to the ML algorithm vendor.

[0225] Proprietary format models can not be mutually recognizable between vendors. Proprietary format models can hide model design information from other vendors when shared.

[0226] Open format models are ML models in a specified format that are mutually recognizable between vendors and allow interoperability from a standards perspective. An example is an ML model architecture (e.g., weights and biases corresponding to each layer in the ML algorithm, number of layers, interconnections, etc.) that is specified in a manner that allows vendors to freely interpret by referring to a standard technical specification (e.g., 3GPP TS).

[0227] Open format models can be mutually recognizable between vendors. Open format models do not hide model design information from other vendors when shared.

[0228] An open format model can support cross-vendor parameter updates and over-the-air training. One example of an open format for ML models is ONNX (Open Neural Network Exchange).

[0229] Enabling open format models can require standardization to make it interoperable between devices of different vendors (e.g., by the UE and the NW). If 3GPP specifies a format for ML models, it can be considered an open format.

[0230] ML models can be associated with functions that apply the ML model to certain decisions (inference). These can include one or more of run-time instructions, input data pre-processing, and output data post-processing algorithms.

[0231] To distinguish between AI / ML models and the functions supported by the AI / ML models, a function identification and a model identification can be used. The model identification can use a “model ID” in the identification process, and the function identification can use a “function” (with or without an explicit model ID).

[0232] In some cases, one or more ML functions (or functions) can be organized and referenced with a model ID.

[0233] In some cases, a model ID can group one or more ML functions into one or more ML algorithm implementations. For example, a given ML function (e.g., beam prediction) can be implemented by two different ML architecture options, but when tested as a black box, can exhibit the same or similar functionality. As an example only, the different ML architecture options can be a CNN (convolutional neural network) and an RNN (recurrent neural network).

[0234] It should be understood that a model ID can or can not be used. The function can instead manage a given ML algorithm for a given use case. For example, the network can indicate a required ML function. Based on that required function, the UE can select an ML algorithm and / or an implementation for that algorithm. An ML algorithm that implements a given function can support one or more combinations of functions. This can be due to different ways of configuring the function. These can be managed by a combination of capabilities that the UE supports for a given function. In 3GPP discussions, these are sometimes referred to as “applicability conditions.”

[0235] Non-limiting examples are considered below. These show examples of how the network can combine functions based on UE capabilities for a given ML use case (e.g., beam prediction).

[0236] Assuming DL Tx beam prediction, consider the following set of applicability conditions for functions associated with BM Case 1,

[0237] Support Top-K DL Tx beam prediction

[0238] o K = 1, 2, 4, [8]

[0239] - This defines support for SSB and / or CSI-RS based RSRP measurements for predicting

[0240] Support for measurement of best-K NZP (non-zero power) CSI-RS resources.

[0241] • Set B condition

[0242] o Measured DL RS (SSB, CSI-RS)

[0243] - Define support for using SSB and / or CSI-RS based RSRP measurements.

[0244] o Measured DL RS set size (4, 8, 12,

[16] )

[0245] - Indicate the minimum number of NZP-CSI-RS resources that the UE should measure and use for predicting best-K NZP CSI-RS resources

[0246] o Measured DL RS set pattern (e.g., fixed, pre-configured list, random)

[0247] - Indicate the restriction on Set B condition

[0248] • Set A condition

[0249] o Predicted DL RS (CSI-RS)

[0250] - Define support for predicting CSI-RS resources

[0251] o Predicted DL RS set size (16, 32, 64)

[0252] - Indicate the maximum number of NZP-CSI-RS resources that should be configured to predict a set of NZP-CSI-RS resources

[0253] Model identification can be considered to identify an AI / ML model for common understanding between the NW and the UE. Information about the AI / ML model can be shared during the model identification.

[0254] Function identification can be considered to identify an AI / ML function for common understanding between the NW and the UE. Information about the AI / ML function can be shared during the function identification.

[0255] In some embodiments, model identification can be used. In some embodiments, function identification can be used. In some embodiments, both model identification and function information can be used.

[0256] The NW can (de)activate or switch between different functions based on a function (ID) of different functions (each function using a proprietary model at the UE), or alternatively, based on a model ID in combination with associated metadata, if available. The function ID (i.e., a label to identify a given function / use case) can be standardized. The model ID can be standardized. The model ID can follow a global format of its symbol (e.g., International Mobile Equipment Identity, IMEI). Such global model ID can be mapped by the network to a temporary model ID. This can provide UE-NW interoperability without requiring, for example, bilateral consensus.

[0257] The model ID based lifecycle management (LCM) (model deactivation / activation, switching, and / or monitoring) can be implemented by the UE and / or handled by UE vendor specific proprietary mechanisms, the function (ID) based LCM (function deactivation / activation, switching, and / or monitoring) can be handled by the NW / NG-RAN.

[0258] The following agreements were reached in the RAN 112:

[0259] “For UE-side models, as well as the UE part of bilateral models:

[0260] - For AI / ML function identification

[0261] o Reuse legacy 3GPP feature framework as a starting point for discussion.

[0262] o UE indicates one / multiple supported functions for a given sub-use case.

[0263] o Start with UE capability reporting.

[0264] - For AI / ML model identification

[0265] o Model is identified by a model ID at the network. UE indicates supported AI / ML models.

[0266] - In function-based LCM

[0267] o Network indicates activation / deactivation / fallback / switching of AI / ML functions via 3GPP signaling (e.g., RRC, MAC-CE, DCI).

[0268] o Model can not be identifiable at the network, and UE can perform model-level LCM.

[0269] o Study whether and how much the NW should be aware / interact with model-level LCM

[0270] - In model ID based LCM, the model is identified at the network and the network / UE can activate / deactivate / select / switch individual AI / ML models via the model ID.

[0271] consensus

[0272] - AI / ML enabled feature refers to a feature in which AI / ML can be used.

[0273] consensus

[0274] - For function identification, one or more functions can be defined in an AI / ML enabled feature.

[0275] The following was agreed in RANI #110bis-e:

[0276] “Study various approaches for achieving good performance in different scenarios / configurations / sites, including:

[0277] i. Model generalization, i.e., use one model that is generalizable to different scenarios / configurations / sites ii. Model switching, i.e., switch among a set of models, where each model is for a specific scenario / configuration / site • [The models in a set of models can have different model structures, share a common model structure, or partially share a common sub-structure. The models in a set of models can have different input / output formats and / or different pre-processing / post-processing.]

[0278] iii. Model updating, i.e., use one model whose parameters are flexibly updated over time as the scenarios / configurations / sites experienced by the device change. Fine-tuning is one example.”

[0279] In the first approach (i.e., model generalization), it is suggested to use a common model to cover various scenarios, and the model is properly trained with this assumption. However, in some cases, a generalization ML model can not be found, or it can even be difficult to implement, because the number of cell-specific parameter configurations (antenna panel configurations, reference signal types, and their transmission configurations, etc.) can be large, and the training of such a model can be difficult.

[0280] A hybrid solution can be used, which can allow for a run-time check of possible model generalization. This can enable further determination of whether an ML model update is needed, or whether an ML model switching is needed.

[0281] In some embodiments, a UE can be provided with a plurality of different machine learning models. Two or more of the machine learning models can support a common function.

[0282] Some embodiments can support model switching or selection, i.e., switching or selection among a set of models, where each model is for a specific scenario / configuration / site. As previously mentioned, the models in the set of models can have different model structures, share a common model structure, or partially share a common sub-structure. The models in the set of models can have different input / output formats and / or different pre / post-processing.

[0283] Some embodiments can allow a UE to receive information from other UEs about ML model performance. This can be to allow seamless handling of ML model functionality.

[0284] Some embodiments can be used in the presence of multiple or a set of ML models for beam management. However, it should be appreciated that other embodiments can be used for ML models capable of providing one or more different functionalities, such as CSI feedback enhancement using CSI compression implementation, CSI prediction, positioning accuracy using LOS / NLOS (line of sight / non-line of sight) determination implementation, or any other different functionality.

[0285] Some embodiments can be used during different switching scenarios. Other embodiments can be used for different scenarios.

[0286] Some embodiments allow switching or selection of ML models. This can be to improve performance of the underlying ML functionality.

[0287] The ML functionality can be, for example, beam management, positioning, etc.

[0288] Some embodiments can address the issue of how to switch or select ML models to provide, for example, good performance of the underlying ML functionality.

[0289] For example, beam prediction accuracy being consistently above 90% when considering the top 1 or 2 beams reported to the network can be considered to provide good performance. This can be when the UE is experiencing mobility, for example, in a HO scenario. The HO scenario can be any suitable HO scenario, such as basic handover, CHO, or any other mobility related scenario.

[0290] Some embodiments can use previous experience from other UEs that are experiencing mobility to help similar UEs to enable them to more efficiently switch or select ML models in a HO scenario. Mobility performance can be used as a guide to improve ML model switching or selection behavior. This can provide good ML model switching or selection behavior results. Model switching or selection can involve deactivating a currently active AI / ML model, and activating a different AI / ML model for a specific functionality.

[0291] In some embodiments, a UE will collect performance data, which can then be used to assist other UEs in ML model selection.

[0292] Some embodiments may involve how to switch or select ML models to provide good performance on the underlying ML functions, such as beam management, positioning, or any other suitable functions.

[0293] Some embodiments may allow for methods for a network to orchestrate the operation of UE(s) at least from the same vendor and / or using the same proprietary format model.

[0294] To do this, one or more of the following can be standardized:

[0295] Network configuration to the UE(s) regarding one or more performance aspects to be collected and / or reported during a handover scenario;

[0296] UE-side model, or UE part of the two-sided ML model capabilities, and underlying ML model capabilities and / or model IDs to be marked during (at least) mobility scenarios;

[0297] When a UE reports a cell and the network considers it a potential target for HO, the network can assist the UE with information from the experience of other UEs (e.g., with information about which ML model to use to ensure a given performance target);

[0298] UEs from the same vendor and / or using the same proprietary format model may request the network to act as a conduit for sharing / joining the above information. This may be based on one or more conditions (e.g., for the same vendor and / or device type and / or UE category); and / or

[0299] The UE can vote for or against the network in one or more corresponding information fields to understand how a given UE perceives the shared / joint information. This can enhance useful information and / or deprioritize useless information.

[0300] In some embodiments, UEs from the same vendor and / or using the same proprietary format model may leverage experience from other UEs to decide on a given ML model for a given function and switch to that model to ensure a given performance target;

[0301] refer to Figure 5 , which shows the signal flow of some embodiments. Figure 5 In the example embodiment, data is collected from UE(s) with ML functionality enabled during HO. This will illustrate an example scenario in which some embodiments may be provided. Figure 5 The example shown in FIG shows a conditional HO scenario. However, it should be understood that other embodiments can also be used in other HO scenarios. It should be understood that other embodiments can be used in non-HO scenarios.

[0302] existFigure 5 In a middle, for a UE-side ML model or a bilateral ML model, the network can provide a configuration to the UE to track the performance of a given ML model. The given ML model can be referenced by a ML function ID and / or a ML model ID. The performance of the given ML model can be tracked in a specific procedure. In Figure 5 In the example shown in the middle, the given procedure can be a handover procedure. This allows the network to collect information about how the given procedure impacts the ML-enabled function.

[0303] As shown by reference 1, the UE is connected to a source gNB. The UE has an ML function enabled. As an example only, the ML function can be a beam prediction in time domain.

[0304] In this example, the UE can have multiple or a set of ML models that are capable of providing the beam prediction function. Different models can be associated with different performance.

[0305] Alternatively or additionally, there can be one or more ML models, and one or more of the one or more ML models can be provided by two or more different implementations of the model. Different implementations of the ML model can be associated with different performance.

[0306] As shown by reference 2, the source gNB sends a request to the AMF. The request is for retrieval of ML information.

[0307] As shown by reference 3, the AMF retrieves information stored during a previous transaction that is tagged for this UE. The previous transaction can be equivalent to step 21 and step 23 discussed later (but for the previous transaction). The AMF sends an ML information response to the source gNB, where the retrieved ML information is tagged for the UE.

[0308] As shown by reference 4, the source gNB updates a local database. The database can be updated based on one or more of the following examples. These examples show how the gNB interprets the received ML information.

[0309] In one embodiment, the ML information can contain how the UE ML model is associated with a beam prediction in time domain performed during a handover from a cell in the source gNB to a cell in the target gNB. This can provide handover performance information for a given pair of cells.

[0310] The handover performance information may, for example, include predictions performed by the underlying ML model in the UE. Cell conditions can be provided with the performance information. For example, the associated cell-specific context and / or radio configuration of the cells in the source gNB and the target gNB can be provided with the handover performance information. For example, the radio configuration can be provided by a CSI configuration.

[0311] Alternatively or additionally, the performance information can indicate the accuracy of the predictions performed by the underlying ML model. The performance information can comprise: beam prediction accuracy and / or RSRP difference.

[0312] Alternatively or additionally, the performance information can be ordered based on the ML model implementation. For example, a given ML function or model can be implemented by different neural network structures. For example, the neural network structure can be an LSTM or a transformer structure. The given ML function or model can be associated with a given ID.

[0313] Alternatively or additionally, the performance information can be ordered based on the ML model implementation, as discussed earlier, and be time-stamped. The time-stamp can be added to indicate that the UE switched ML model implementation.

[0314] Alternatively or additionally, the performance information can have a vote count. For example, this can indicate how many UEs in the population agree with the performance information when performing mobility from a cell in the source gNB to a target gNB (i.e. for a given pair of cells).

[0315] In some embodiments, statistical deviations can be recorded. The statistical deviations can comprise, for example, mean and / or variance.

[0316] Depending on the implementation, the source gNB can update its local database with the required piece(s) of information.

[0317] If the ML information contains a container that is independent of the UE vendor or vendor-independent information, then this information can be stored in the database. This can allow all UE(s) in the source gNB cell to benefit from such information.

[0318] If the ML information contains a UE vendor-specific container or vendor-specific information, and the container / information is not network vendor-specific, then this information can be stored in the database. This can allow all UE(s) in the source gNB cell from the same category or vendor to benefit from such information.

[0319] If the ML information contains a UE vendor-specific container or information, and the container / information is network vendor-specific, then this information is merged into the database so that all UE(s) in the source gNB cell from the same category can benefit from such information.

[0320] If the ML information contains only a UE vendor-specific container or information, then this information can be stored in the database. The network can only update UE(s) from a given vendor. In this case, the network can not be able to interpret such information.

[0321] In previous examples, reference was made to vendor specific containers / information. It will be appreciated that these examples can be applied to any other proprietary containers / information.

[0322] In some embodiments, the information can be updated only for UEs meeting a given criteria or grouping. For example, the information can be updated only for UEs having similar speed.

[0323] As shown by reference number 5, the UE sends a measurement report to the source gNB. The measurement report can be an RRC measurement report. The measurement report can indicate that the UE is interested in a cell in the target gNB.

[0324] As shown by reference number 6, the source gNB interprets the report and formats a container. The container can include a request to the target gNB to add / modify what the source gNB proposes. The source gNB can propose how to configure the UE to collect performance information during this particular HO. It will be appreciated that one or more of these pieces of information can be provided outside of the container.

[0325] The container can provide one or more of the following:

[0326] In some embodiments, the network (e.g., gNB) can select one or more ML functions or models for the UE to collect information. The one or more ML functions or models can be identified by a respective ID. This can be provided in the container. It will be appreciated that each ML function or model can have a respective implementation in the UE. For example, function ID X and / or ML model ID X can be implemented in the UE using two or more different ML architectures. As an example only, two different ML architectures are LSTM and transformer.

[0327] In some embodiments, the source gNB and target gNB can structure the information to require the UE to record performance information independent of a UE vendor’s ML function or model, dependent on a UE vendor’s ML function or model, dependent on a UE network vendor’s ML function or model, or a subset of one or more thereof.

[0328] In some embodiments, the UE can not be aware whether the model is dependent on a vendor or independent of a vendor.

[0329] The UE can be requested to track HO performance (example is beam prediction in time domain).

[0330] As an example, the UE can perform one or more of the following:

[0331] o Apply the ML model (and related implementation) to both the source and target cell(s) and record the beam index.

[0332] o Record the number of ML model implementation switches or selections (this can be relevant only for ML functions or models that depend on the UE vendor or the UE network vendor).

[0333] o Record ground truth and compare the accuracy for the ML function in both source and target cell(s) by recording the beam index and comparing it to the prediction.

[0334] The configuration can request the UE to apply a timestamp. This can be based on a fixed duration (or epoch) during the HO for which performance measurements are performed and / or associated with one or more execution conditions. For example, the one or more execution conditions can be that the source cell is X dB worse than the target cell, or a set of thresholds to allow the recording of the event for the entire duration of the HO. This can be refined by specific steps in the HO procedure (e.g., before initiating a RACH attempt to the target cell).

[0335] As shown by reference number 7, the source gNB sends a handover request to the target gNB. This request includes the UE configuration and the handover container as described earlier in the section of reference number 6.

[0336] As shown by reference number 8, the target gNB sends a handover request acknowledgement to the source gNB. This has a UE-specific ML-HO configuration response that can be shared with the UE. This provides a configuration guideline for taking the necessary measurements during the HO procedure.

[0337] As shown by reference number 9, the source gNB formats the ML-HO configuration container based on the response received from the target gNB (as described by reference number 8). The ML-HO configuration container includes the ML-HO configuration response received from the target gNB. The ML-HO configuration container can also include data from other UEs for which the source gNB has collected performance data at a previous occasion.

[0338] This information can be formatted under the following restrictions: it is independent of the UE vendor (i.e., applicable and visible to all UE(s)), dependent on the UE vendor (i.e., applicable and visible to all UE(s) from a given vendor), dependent on the UE network vendor pair (i.e., applicable and visible to all UE(s) from a given pair of vendors).

[0339] If there is a significant difference in the performance data and it needs to be voted for / against by the UE individually, two or more instances of performance data for the same source / target cell pair can be sent to the UE. This allows the UE to not only record the performance during the HO but also to vote on the information from the network about the experience of other UEs.

[0340] As shown by reference number 10, the UE receives an ML-HO configuration container from the source gNB. The ML-HO configuration container can be provided in a reconfiguration message. The reconfiguration message can be an RRC reconfiguration message. The reconfiguration message can also include a HO container. The UE configures or reconfigures based on the information in the reconfiguration message.

[0341] As shown by reference number 11, the UE sends a message to the source gNB to indicate that the reconfiguration has been completed. This can be sent in a reconfiguration complete message. The reconfiguration complete message can be an RRC reconfiguration complete message.

[0342] As shown by reference number 12, the UE performs measurements under the guidance of the ML-HO configuration.

[0343] The UE can determine which ML model to select and switch to using. The determination can be based on information provided by the network in the ML-HO configuration container, as shown by reference number 10. This can include data from other UEs. The data from other users can be based on performance data collected by the source gNB at a previous occasion. The data can include associated cell conditions for the performance data or information from other UEs. The data can be used to determine the best selection and switch or use of the selected model. For example, if three ML models are selected and switched, but the performance data from other UEs indicates that the second switch is below a given threshold, e.g., 90%, configured by the network, the UE can determine to only make two model selections and switches.

[0344] The UE can collect performance data and / or cell conditions related to the selected ML model(s).

[0345] As shown by reference number 13, the UE is ready to access the target cell (e.g., using a RA procedure).

[0346] As shown by reference number 14, when the UE is ready to access the target cell, the UE can format the results based on the performance measurements using the ML-HO configuration. The container can be formatted to be independent of the UE vendor (i.e., applicable and visible to all UE(s) from all vendors), dependent on the UE vendor (i.e., applicable and visible to all UE(s) from a given vendor), or dependent on the UE network vendor pair (i.e., applicable and visible to all UE(s) from a given pair of vendors). The UE can additionally vote on each entry it received in the ML-HO configuration that most closely experiences what it experiences.

[0347] As shown by reference number 15, the UE sends a random access request to the target gNB. This has an ML-HO results indication.

[0348] As shown by reference number 16, the target gNB provides a random access response with a grant for the ML-HO result.

[0349] As shown by reference number 17, the UE sends a handover complete message to the target gNB. This is the result of the ML-HO. This can include the voting information. The UE can provide the performance data to the target base station as described previously. This can include the cell conditions. As described previously, this data can be the same or similar type of data that the UE received from the source base station.

[0350] As shown by reference number 18, the UE sends an uplink message to the target gNB. This is the result of the ML-HO. This can include the voting information. Step 18 is an optional step, which is a follow-up to steps 15-17 if the message size is large and does not fit in the previous steps.

[0351] Thus, in the random access procedure, the UE can be provided with an initial or full grant to send the ML-HO result.

[0352] As shown by reference number 19, the feedback from the UE is used to update the target gNB database and / or the UE / network specific container of the UE.

[0353] As shown by reference number 20, handover feedback is provided from the target gNB to the source gNB.

[0354] As shown by reference number 21, the feedback from the target gNB is used to update the source gNB database and / or the UE / network specific container of the UE.

[0355] Thus, in steps 19-21, the feedback from the UE is appended to the target gNB database and also synchronized to the source gNB database. The updating of the database can be as described in step 4. This allows the source and target gNB to update the performance data and the results of the UE voting to mark which aspects the UE considers useful. For example, if the UE does not consider a particular ML model (implementation) handover to be beneficial from its perspective, the gNB can take this into account in its database in order to warn other UEs in advance. Thus, the UE can learn from the previous experience of other UEs that a ML model implementation handover is not beneficial and can retain the same ML model implementation in the HO procedure.

[0356] It will be appreciated that the AMF or other network entity can have a “master” database that stores the information collected by the gNBs. In steps 2 and 3, the UE needs ML information is retrieved from this database. The database of the gNBs can be considered as a “cached” version of the database.

[0357] As shown in Figure 22, due to the sharing of feedback, the UEs can benefit from each other’s experience in terms of network participation and control.

[0358] As shown by reference number 23, the target gNB can send a path switch request to the AMF. The request can include updated ML information.

[0359] As shown by reference number 24, the AMF updates its ML information for later retrieval. Thus, the target gNB can update the results for a given UE to the AMF for later retrieval, i.e., when the UE next comes back to the same cell or gNB. This can also allow for building an initial database at the gNB(s).

[0360] As shown by reference number 25, the AMF can send a path switch response to the target gNB.

[0361] In some embodiments, real-time performance of ML models for a given function and / or model ID can be collected in a manner that is independent of the UE, dependent on the UE, and / or dependent on the UE vendor relationship.

[0362] Some embodiments can allow UEs to jointly experience each other during handover. This can provide a baseline for performance of ML models for a given function and / or model ID.

[0363] Within the same UE vendor (or proprietary model), some embodiments provide a technique for sharing what a given UE experiences with respect to what the ML model implementation faces when the UE moves from cell X to cell Y. This can help to evaluate whether the deployed ML models are operating in a generic or specific manner. For example, it can be determined whether the ML models show cell-specific variations in performance, or whether they are robust enough to handle the environment of different cells.

[0364] When applied in the context of a specific UE network vendor relationship, some embodiments can allow for fine-tuning of features in a specific deployment or niche market (e.g., factory environment).

[0365] When applied in the context of a vendor-independent UE, some embodiments can help to provide a joint approach to collect and share performance data to allow the UE to know which works well and which does not. The UE can use the performance data from other UEs as a pre-warning, and the UE can agree or disagree with the UE based on its own experience.

[0366] Based on the joint experience from other UEs received by a given UE from the network, the UE can determine how to do ML model switching for a given function (for a given use case, such as beam prediction) in a more meaningful way, rather than encountering the same problems as the other UE(s) earlier.

[0367] The UE and / or network can also build ML models based on the data that provides the recommendations as described above.

[0368] Thus, in some embodiments, experiences from the UE(s) are shared by the network to other UE(s). This can provide a robust and / or efficient way to proactively implement ML model switching of a given functionality during HO.

[0369] In the above examples, the information is provided in a container. It will be appreciated that this is merely an example, and the information can be provided in other ways.

[0370] Reference is made to Figure 6 which illustrates a method of some embodiments.

[0371] The method can be performed by an apparatus. The apparatus can be in or be a communication device.

[0372] The apparatus can comprise suitable circuitry for providing the method.

[0373] The circuitry can support machine learning functionality.

[0374] Alternatively or additionally, the apparatus can comprise: at least one processor; and at least one memory storing instructions that, when executed by the at least one processor, cause the apparatus at least to provide the method.

[0375] Alternatively or additionally, the apparatus can be as shown in Figure 1b or Figure 1c .

[0376] The method can be provided by computer program code or computer executable instructions.

[0377] The method can comprise: receiving, from a base station, configuration information, the configuration information comprising information related to a machine learning functionality, as indicated by reference sign Al.

[0378] The method can comprise: selecting, from a plurality of different machine learning options providing the machine learning functionality, one or more machine learning options, as indicated by reference sign A2.

[0379] The method can comprise: causing performance information to be transmitted to the base station, the performance information relating to performance of the selected one or more machine learning options, as indicated by reference sign A3.

[0380] It will be appreciated that Figure 6 The method outlined in

[0381] Reference is made to Figure 7 which illustrates a method of some embodiments.

[0382] The method can be performed by an apparatus. The apparatus can be in or be a base station.

[0383] The apparatus can comprise suitable circuitry for providing the method.

[0384] The circuitry can support machine learning functionality.

[0385] Alternatively or additionally, the apparatus can comprise: at least one processor; and at least one memory storing instructions that, when executed by the at least one processor, cause the apparatus at least to provide the method.

[0386] Alternatively or additionally, the apparatus can be as shown in Figure 1c

[0387] The method can be provided by computer program code or computer executable instructions.

[0388] The method can comprise providing configuration information to the first user equipment, the configuration information comprising information relating to machine learning functionality, as indicated by reference sign B1.

[0389] The method can comprise receiving performance information from the first user equipment, the performance information relating to performance of one or more machine learning options selected by the first user equipment to provide the machine learning functionality, as indicated by reference sign B2.

[0390] It will be appreciated that Figure 7 The method outlined in the Summary can be modified to include any of the previously described features.

[0391] The computer program code can be downloaded and stored in one or more memories of the relevant apparatus or device.

[0392] Thus, although certain embodiments have been described above with reference to a certain example architecture of wireless networks, technologies and standards by way of example, embodiments can be applied to any other suitable form of communications system than the one(s) illustrated and described herein. In this example, some embodiments have been described in conjunction with a 5G network. It should be appreciated that other embodiments can be provided in any other suitable network.

[0393] It should also be noted that while example embodiments have been described above, various changes and modifications can be suggested to one skilled in the art. These changes and modifications can be made only within the scope of the application and within the scope of the following claims.

[0394] ​As used herein, “at least one of ” and “one or more of ” and similar phrases, where a list of two or more elements is preceded by “at least one of” or “one or more of”, means that at least any one of the listed elements is present, or that at least any two or more of the listed elements are present, or that at least all of the listed elements are present.

[0395] In general, the various embodiments can be implemented in hardware or special purpose circuits, software, logic or any combination thereof. Some aspects of the disclosure can be implemented in hardware, while other aspects can be implemented in firmware or software which can be executed by a controller, microprocessor or other computing device, although the disclosure is not limited thereto. While various aspects of the disclosure can be illustrated and described as block diagrams, flow charts, or using some other pictorial representation, it is well understood that these blocks, apparatus, systems, techniques or methods described herein can be implemented in hardware, software, firmware, special purpose circuits or logic, general purpose hardware or controler or other computing devices, or some combination thereof.

[0396] As used in this application, the term “circuitry” can refer to one or more or all of the following:

[0397] (a) hardware-only circuit implementations (such as implementations in only analog and / or digital circuitry) and

[0398] (b) combinations of hardware circuits and software, such as (as applicable):

[0399] (i) combinations of analog and / or digital hardware circuit(s) with software / firmware and (ii) portions of hardware processor(s) with software (including digital signal processors), software, and memory(ies) that work together to cause an apparatus, such as a mobile phone or server, to perform various functions and / or

[0400] (c) hardware circuit(s) and or processor(s), such as a microprocessor(s) or a portion of microprocessor(s), that requires software (e.g., firmware) for operation, but is not hardware or processor(s) that

[0401] performs a function or a portion of a function using software only.

[0402] This definition of circuitry applies to all uses of this term in this application, including in any claims. As a further example, as used in this application the term circuitry also includes an implementation comprising only hardware circuitry or only processor(s) (or only a combination of hardware, software, and / or firmware), and that the term circuitry can

[0403] Embodiments of the disclosure can be implemented by computer software executable by a data processor of the mobile device (such as in the processor entity), or by hardware, or by a combination of software and hardware. Computer software or program, also called program product, including software routines, applets and / or macros, can be stored in any apparatus-readable data storage medium and they comprise program instructions that, when executed by the data processor, implement the embodiments. The apparatus-readable data storage medium can include any data storage medium that is suitable to store program instructions adopted to implement the embodiments. The apparatus-readable data storage medium can include magnetic storage media such as magnetic diskette, magnetic tape, and hard disk drive; optical storage media such as compact disc (CD), DVD, Blu- ray Disc, and laser disc; electronic storage media such as flash memory, solid state memory, and electrically programmable memory (EPROM, EEPROM); and non-transitory storage media such as RAM with firmware. The computer program product can also include program instructions confined within a single device, to program instructions that are spread across multiple devices, and to program instructions that are stored within and implemented by a MIMO device.

[0404] Moreover, in this regard, it should be noted that any blocks of the logic flow described herein can represent program steps, or interconnected logic circuit, blocks, and functions, or a combination of program steps and logic circuit, blocks, and functions. The software can be stored on such physical media as memory chips, or memory blocks implemented within the processor, magnetic media such as hard disk or floppy disks, and optical media such as DVD and the data variants thereof CD. The physical media is non-transitory. The term "non-transitory" expressly excludes propagating signals as the term is used in the context of data storage persistency. The term "non-transitory" as used herein is to be taken as a limitation of the medium itself (i.e., tangible, not a signal) and not as a limitation of data storage persistency (e.g., RAM vs. ROM).

[0405] The memory can be of any type appropriate for the local technical environment and can be implemented using any appropriate data storage technology, such as semiconductor-based memory devices, magnetic memory devices and systems, optical memory devices and systems, fixed memory and removable memory. The data processor can be of any type appropriate for the local technical environment, and can include one or more of general purpose computers, special purpose computers, microprocessors, digital signal processors (DSPs), application specific integrated circuits (ASIC), FPGAs, gate level circuits and processors based on multi-core processor architectures, as non-limiting examples.

[0406] Embodiments of the disclosure can be implemented in various components such as integrated circuit modules. The design of integrated circuits is by nature a highly automated process. Complex and powerful software tools are available for converting a logic level design into a semiconductor circuit design ready to be etched and formed on a semiconductor substrate.

[0407] The above description provides a complete and informative description of exemplary embodiments of the present disclosure based on specific examples. However, various modifications and adaptations can become apparent to those skilled in the relevant art in view of the above description, taught with the aid of the accompanying drawings and the appended claims. Indeed, other embodiments can be used in addition or in place of the embodiments explicitly described herein. While the DETAILED DESCRIPTION has been described with reference to various embodiments, it will be understood that these are included to illustrate and provide context for various embodiments of the present disclosure. It will also be understood that the claims are not limited to the embodiments just described and / or shown. Rather, embodiments anticipated by the claims and to which the claims are entitled can include other combinations and / or permutations of one or more of the embodiments previously disclosed, as well as other embodiments that are already known to those of ordinary skill in the art and that can include other features not specifically disclosed herein.

Claims

1. A device comprising: means for receiving configuration information from a base station, the configuration information including information related to a machine learning function; means for selecting one or more machine learning options from a plurality of different machine learning options that provide the machine learning functionality; as well as Means for causing performance information to be sent to the base station, the performance information relating to performance of the selected one or more machine learning options.

2. The apparatus of claim 1 , wherein the machine learning options comprise one or more of: a machine learning algorithm, or a machine learning architecture.

3. The apparatus according to claim 1 or 2, wherein the configuration information comprises: Data associated with usage of one or more of the different machine learning options by one or more other user devices, and the component for selecting the machine learning option is used to select the one or more machine learning options based on the data.

4. The apparatus according to claim 3, wherein the data comprises: Information about one or more cell conditions when the corresponding machine learning option is used by the one or more other user equipments.

5. The apparatus according to claim 4, comprising: means for determining one or more cell conditions, and wherein the means for selecting the machine learning options is for selecting the one or more machine learning options based on the determined one or more cell conditions and data regarding the one or more cell conditions when the corresponding machine learning options are used by the one or more other user devices.

6. The apparatus according to any one of claims 3 to 5, wherein the performance information comprises: Feedback information regarding a consensus of data associated with usage of the one or more machine learning options selected by one or more other user devices.

7. An apparatus according to any one of the preceding claims, wherein the performance information is provided together with information about the one or more machine learning options selected.

8. The apparatus of any preceding claim, wherein the performance information comprises: Information about one or more network conditions when the selected one or more machine learning options are used.

9. The apparatus of any preceding claim, wherein the performance information comprises one or more of: one or more predictions obtained through use of the selected one or more machine learning options; the accuracy of the one or more predictions; and Statistical deviation information.

10. An apparatus according to any preceding claim, wherein the performance information is provided together with time stamp information.

11. The apparatus according to any one of the preceding claims, wherein the configuration information comprises: Information indicating the capability information to be sent to the base station.

12. The device according to any one of the preceding claims, comprising: Means for switching between a plurality of the selected ones of the machine learning options.

13. The apparatus according to any one of the preceding claims, wherein the configuration information comprises: Configuration information of a handover procedure for handover from the base station to a target base station.

14. An apparatus comprising: a component for providing configuration information to the first user equipment, the configuration information including information related to the machine learning function; as well as A component for receiving performance information from the first user device, the performance information relating to the performance of one or more machine learning option machines selected by the first user device to provide the machine learning function.

15. The apparatus according to claim 14, wherein the configuration information comprises one or more of the following: data associated with usage of one or more of the different machine learning options that provide the machine learning functionality by one or more other user devices; data comprising information regarding one or more cell conditions when the corresponding machine learning option is used by one or more other user equipment; information indicating the performance information to be sent by the first user equipment; or Configuration information of a handover procedure for the first user equipment used for handover from the base station to a target base station.

16. The apparatus according to claim 14 or 15, wherein the performance information comprises one or more of the following: feedback information regarding a consensus of data of the first user device associated with usage of the one or more machine learning options selected by the first user device by one or more other user devices; information about one or more network conditions when the selected one or more machine learning options are used by the first user device; one or more predictions obtained through use of the selected one or more machine learning options; the accuracy of the one or more predictions; or Statistical deviation information.

17. The device according to any one of claims 14 to 16, comprising: means for receiving, from the target base station, capability information for the first user equipment.

18. An apparatus according to any one of the preceding claims, wherein the selected machine learning option is provided by a machine learning model.

19. A method comprising: receiving configuration information from a base station, the configuration information including information related to a machine learning function; selecting one or more machine learning options from a plurality of different machine learning options that provide the machine learning functionality; and causing performance information to be sent to the base station, the performance information relating to performance of the selected one or more machine learning options.

20. A method comprising: Providing configuration information to the first user equipment, the configuration information including information related to the machine learning function; as well as Performance information is received from the first user device, the performance information relating to performance of one or more machine learning option machines selected by the first user device to provide the machine learning function.

21. A computer program comprising computer executable code which, when run, causes the method according to claim 19 or 20 to be performed.

Images