Method of pre-mapping based model signaling

Abstract

The present disclosure describes methods of using the pre-configured AI / ML (artificial intelligence / machine learning) based model operation modes in wireless mobile communication system including base station (e.g., gNB, TN, NTN) and mobile station (e.g., UE). In AI / ML model is applied to radio access network, signaling of model information exchange can be heavily congested. Therefore, model operation (e.g., model training / inferencing / monitoring / updating) can be set up between network and UE by configuring mapping relation look-up table.

Description

[0001] TITLE

[0002] Method of pre-mapping based model signaling

[0003] TECHNNICAL FIELD

[0004] The present disclosure relates to AI / ML based model operation with mapping relation look-up table, where techniques for pre-configuring and signaling the specific information about pre-mapping based model operation applicable to radio access network are presented.

[0005] BACKGROUND

[0006] In 3GPP (Third Generation Partnership Project), one of the selected study items as the approved Release 18 package is AI / ML (artificial intelligence / machine learning) as described in the related document (RP-213599) addressed in 3GPP TSG (Technical Specification Group) RAN (Radio Access Network) meeting #94e. The official title of AI / ML study item is “Study on AI / ML for NR Air Interface”. The goal of this study item is to identify a common AI / ML framework and areas of obtaining gains using AI / ML based techniques with use cases. According to 3GPP, the main objective of this study item is to study AI / ML framework for air-interface with target use cases by considering performance, complexity, and potential specification impact. In particular, AI / ML model, terminology and description to identify common and specific characteristics for framework are included as one of key work scopes. Regarding AI / ML framework, various aspects are under consideration for investigation and one of key items is about lifecycle management of AI / ML model where multiple stages are included as mandatory for model training, model deployment, model inference, model monitoring, model updating etc.

[0007] Also in 3GPP, two-sided (AI / ML) model is defined as a paired AI / ML model(s) over which joint inference is performed, where joint inference comprises AI / ML Inference whose inference is performed jointly across the UE and the network. Also for onesided (AI / ML) model, UE-side (AI / ML) model is defined as an AI / ML model whose inference is performed entirely at the UE and network-side (AI / ML) model is defined as an AI / ML model whose inference is performed entirely at the network. Currently, AI / ML specification work is at the stage of work item discussion for Release 19. Earlier, in 3GPP TR 37.817 for Release 17, titled as Study on enhancement for Data Collection for NR and EN-DC, UE (user equipment) mobility was also considered as one of AI / ML use cases and one of scenarios for model training / inference is that both functions are located within RAN node. Followingly, in Release 18 the new work item of “Artificial Intelligence (AI)ZMachine Learning (ML) for NG-RAN” was initiated to specify data collection enhancements and signaling support within existing NG-RAN interfaces and architecture. In RAN, signaling is crucial for communication between the UE and the network as this signaling occurs across different layers of the protocol stack, primarily L1 (Layer-1 ), L2 (Layer-2), and RRC (radio resource control). For RRC, it is a Layer-3 protocol used on the air interface between UE and the base station (e.g., gNB in 5G, eNB in LTE) where the main role is to establish, configure, maintain, and release radio resources and connections needed for communication.

[0008] For the above active standardization works, RAN-based AI / ML model is considered very significant for both network and UE to meet any desired model operations (e.g., model training, inference, selection, switching, update, monitoring, etc.). Model information can be signaled to pair both network-side and UE-side models for various lifecycle management (LCM) operations. However, signaling overhead indicating model information can be very high especially when model based LCM is processed between base station (BS / gNB) and multiple UEs. In LCM, model training is one of the most important parts for model deployment and currently there is no specification defined for signaling methods and network-UE behaviors so as to identify the required dataset when model updating / re-training as any activated model can be also impacted due to model / data drift.

[0009] When ML condition changes, the enabled AI / ML model(s) can be impacted for model performance due to data / model drift. In this case, model re-training / updating can be executed. For example, when the trained ML model is deployed in RAN, model performance for inferencing can be easily degraded if target ML condition is not well aligned with real ML condition measured for specific model operation. US 2023069342 describes how to assist determination of the model update time in consideration of cost for the update of a model.

[0010] US 2023022737 explains supporting generation of machine learning model when a certain machine learning model is changed.

[0011] US 2019012876 provides projections, predictions, and recommendations for computing system.

[0012] US 2019332895 shows that the monitored states are to decide to change a trained ML model as currently used.

[0013] EP 4075348 describes control of machine learning model, which can be based on a federated learning method collectively performed by nodes of a decentralized distributed database.

[0014] US 2021019612 provides the self-healing system that can automatically provide a diagnostic, and it can also automatically provide an action if the performance of the model predictions has changed over time.

[0015] The present disclosure solves the cited problem by the proposed embodiments and describes a method of pre-mapping based model signaling by configuring a set of ML model operation modes in a wireless communication system, comprising presetting mode-1 for mapping-based model, mode-2 for non-mapping based model and mode- 3 for hybrid model; configuring mapping relation look-up table(s) for the clustered dataset and quantized model output; activating ML model operation mode(s).

[0016] In some embodiments of the method according to the first aspect, the method is characterized by, that one or multiple mapping relation look-up tables can be predefined via offline ML model operation based on specific ML configuration information and / or ML conditions such as ML model applications and / or model properties / types and / or environmental attribute data and / or applicable device capability. In some embodiments of the method according to the first aspect, the method is characterized by, that the configured mapping relation look-up table represents the mapping relation information about the dataset (model input) and the associated model output, comprising clustering dataset based on statistical characteristics of data values; quantizing the associated model output as the quantized value selection from the pre-configured set of output values.

[0017] In some embodiments of the method according to the first aspect the method is characterized by, that with ML model operation mode-1 there is no need of running real-time model inferencing operation to generate model output as the configured mapping relation look-up table can be used to provide model output when a new model input data is given.

[0018] In some embodiments of the method according to the first aspect the method is characterized by, that with ML model operation mode-2 model output has no dependency on any pre-defined input / output mapping relation and model is then run to generate output for the given input dataset.

[0019] In some embodiments of the method according to the first aspect the method is characterized by, that with ML model operation mode-3 model operation is based on using both mode-1 and mode-2 for model output generation.

[0020] In some embodiments of the method according to the first aspect the method is characterized by, that both mode-1 and mode-2 can be operated simultaneously so that model outputs of mode-1 and mode-2 can be selectively chosen or combined together to achieve target model accuracy level with mode-3 operation.

[0021] In some embodiments of the method according to the first aspect the method is characterized by, that output of mode-1 firstly can provide a specific output that is referenced for mode-2 for faster operation of mode-2 with higher accuracy of model output. In some embodiments of the method according to the first aspect the method is characterized by, that the pre-configured ML model operation modes and the associated mapping relation look-up tables can be sent via system information or dedicated RRC signaling.

[0022] In some embodiments of the method according to the first aspect the method is characterized by, that L1 / L2 or RRC signaling can be used for indication of activating mapping relation look-up table(s) for use if applicable.

[0023] In some embodiments of the method according to the first aspect the method is characterized by, that the updated mapping relation look-up table can be signaled as unicast, multicast or broadcast information.

[0024] In some embodiments of the method according to the first aspect the method is characterized by, that model input data is categorized into a finite set of dataset ID or index as DS #i (i=1 ,2,3, ... ) and model output data is quantized into a finite set of output ID or index as Q_Out #j (j=1 ,2,3, ... ) for mapping relation look-up table.

[0025] In some embodiments of the method according to the first aspect the method is characterized by, that any combinations of model input data and output data in mapping relation look-up table can be represented as index information if applicable.

[0026] In some embodiments of the method according to the first aspect the method is characterized by, that new model input data can be identified as one of DS #i or clustered data as the quantized model output can be a value from the pre-configured set of output values that have mapping relation with model input.

[0027] In some embodiments of the method according to the first aspect the method is characterized by, that different levels of quantization and / or clustering can be applied to generate multiple versions of mapping relation look-up tables with different sizes of tables. In some embodiments of the method according to the first aspect the method is characterized by, that model performance accuracy can vary in accordance with multiple versions of mapping relation look-up tables.

[0028] In some embodiments of the method according to the first aspect the method is characterized by, that online operation to update mapping relation look-up table can be applied depending on implementation scenarios so that dataset can be reclustered with the associated quantized output after online training of ML model.

[0029] In some embodiments of the method according to the first aspect the method is characterized by, that criteria of dataset clustering and output quantization can be implementation-specific for varying scenarios or ML use cases with different model applications / functionalities.

[0030] In some embodiments of the method according to the first aspect the method is characterized by, that the number of mapping relation look-up tables and the size of each mapping relation look-up table can be pre-configured depending on implementation scenarios.

[0031] In some embodiments of the method according to the first aspect the method is characterized by, that at network side ML model operation modes are preset for specific ML applications / functionalities with target model(s) for activation.

[0032] In some embodiments of the method according to the first aspect the method is characterized by, that based on the determined ML model operation mode, target model(s) can be selected for activation.

[0033] In some embodiments of the method according to the first aspect the method is characterized by, that either network side or UE side can determine a specific ML model operation mode and mapping relation look-up table for activation based on either one-Ztwo-sided model and / or ML configuration after ML configuration information with the preset ML model operation mode and mapping relation look-up table is received by UE. In some embodiments of the method according to the first aspect the method is characterized by, that activation of a specific ML model operation mode with or without mapping relation look-up table can be triggered by indication signaling sent from network side or by UE autonomous decision or by any preset threshold.

[0034] In some embodiments of the method according to the first aspect the method is characterized by, that any specific ML model operation mode and use of the preset mapping relation look-up table can be enabled by network side using indication message sent to UE.

[0035] In some embodiments of the method according to the first aspect the method is characterized by, that ML condition is measured at UE side for reporting to network side as ML condition can be configured to include information such as on-device ML model applicability, environmental condition, etc. influencing performance of the selected ML model operation mode before sending activation indication to UE.

[0036] In some embodiments of the method according to the first aspect the method is characterized by, that any specific ML model operation mode and use of the preset mapping relation look-up table can be enabled by UE based in device autonomous decision.

[0037] In some embodiments of the method according to the first aspect the method is characterized by, that applicable mode(s) can be identified to determine activation along with ML condition measurement before autonomous decision of activating a specific ML model operation mode.

[0038] In some embodiments of the method according to the first aspect the method is characterized by, that UE sends status feedback of the activated ML model operation mode after ML model operation mode activation.

[0039] In some embodiments of the method according to the first aspect the method is characterized by, that ML model operation mode-3 (hybrid mode) combines the advantages of both mapping-based inference (mode-1 ) and full AI / ML model execution (mode-2) to optimize the trade-off between inference latency, computational load, and accuracy.

[0040] In some embodiments of the method according to the first aspect the method is characterized by, that the UE and / or network entity dynamically select or fuse outputs from the pre-configured mapping relation look-up table and the model inference path according to ML conditions, signaling instructions, or preset thresholds under mode- 3.

[0041] According to a second aspect, the present disclosure relates to an apparatus for premapping based model signaling by configuring a set of ML model operation modes in a wireless communication system in a wireless communication system, that are applied to multiple cells, the apparatus comprising a wireless transceiver, a processor coupled with a memory in which computer program instructions are stored, said instructions being configured to carry out steps of the first aspect of this application.

[0042] According to a third aspect, the present disclosure relates to an user Equipment comprising an apparatus according to the second aspect.

[0043] According to a fourth aspect, the present disclosure relates to a gNB comprising an apparatus according to the second aspect.

[0044] According to a fifth aspect, the present disclosure relates to a wireless communication system for pre-mapping based model signaling by configuring a set of ML model operation modes, wherein the wireless communication systems comprises user equipment according to the third aspect, gNB according to the fourth aspect, whereby the user equipment and the gNB each comprises a processor coupled with a memory in which computer program instructions are stored, said instructions being configured to carry out the steps according to the first aspect.

[0045] According to a sixth aspect, the present disclosure relates to a computer program product comprising instructions which, when executed by at least one processor, configure said at least one processor to carry out a method according to the first aspect of the present disclosure. The computer program product can use any programming language, and can be in the form of source code, object code, or in any intermediate form between source code and object code, such as in a partially compiled form, or in any other desirable form.

[0046] According to a seventh aspect, the present disclosure relates to a computer-readable storage medium comprising instructions which, when executed by at least one processor, configure said at least one processor to carry out a method according to any one of the embodiments of the present disclosure.

[0047] BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 is an exemplary mapping relation look-up table.

[0049] Figure 2 is an exemplary block diagram of mapping relation between clustered dataset and quantized output.

[0050] Figure 3 is an exemplary flow chart of configuring ML model operation modes and mapping relation look-up tables at network side.

[0051] Figure 4 is an exemplary flow chart of activating a specific ML model operation mode at UE side.

[0052] Figure 5 is an exemplary signaling flow of activating ML model operation mode by network side.

[0053] Figure 6 is an exemplary signaling flow of activating ML model operation mode by UE side.

[0054] DETAILED DESCRIPTION The detailed description set forth below, with reference to annexed drawings, is intended as a description of various configurations and is not intended to represent the only configurations in which the concepts described herein may be practiced. The detailed description includes specific details for the purpose of providing a thorough understanding of the various concepts. However, it will be apparent to those skilled in the art that these concepts may be practiced without these specific details. In particular, although terminology from 3GPP 5G NR may be used in this disclosure to exemplify embodiments herein, this should not be seen as limiting the scope of the invention.

[0055] Some of the embodiments contemplated herein will now be described more fully with reference to the accompanying drawings. Other embodiments, however, are contained within the scope of the subject matter disclosed herein, the disclosed subject matter should not be construed as limited to only the embodiments set forth herein; rather, these embodiments are provided by way of example to convey the scope of the subject matter to those skilled in the art.

[0056] Generally, all terms used herein are to be interpreted according to their ordinary meaning in the relevant technical field, unless a different meaning is clearly given and / or is implied from the context in which it is used. All references to a / an / the element, apparatus, component, means, step, etc. are to be interpreted openly as referring to at least one instance of the element, apparatus, component, means, step, etc., unless explicitly stated otherwise. The steps of any methods disclosed herein do not have to be performed in the exact order disclosed, unless a step is explicitly described as following or preceding another step and / or where it is implicit that a step must follow or precede another step. Any feature of any of the embodiments disclosed herein may be applied to any other embodiment, wherever appropriate. Likewise, any advantage of any of the embodiments may apply to any other embodiments, and vice versa. Other objectives, features and advantages of the enclosed embodiments will be apparent from the following description.

[0057] In some embodiments, a more general term “network node” may be used and may correspond to any type of radio network node or any network node, which communicates with a UE (directly or via another node) and / or with another network node. Examples of network nodes are NodeB, MeNB, ENB, a network node belonging to MCG or SCG, base station (BS), multi-standard radio (MSR) radio node such as MSR BS, eNodeB, gNodeB, network controller, radio network controller (RNC), base station controller (BSC), relay, donor node controlling relay, base transceiver station (BTS), access point (AP), transmission points, transmission nodes, RRU, RRH, nodes in distributed antenna system (DAS), core network node (e.g. Mobile Switching Center (MSC), Mobility Management Entity (MME), etc), Operations & Maintenance (O&M), Operations Support System (OSS), Self Optimized Network (SON), positioning node (e.g. Evolved- Serving Mobile Location Centre (E-SMLC)), Minimization of Drive Tests (MDT), test equipment (physical node or software), etc.

[0058] In some embodiments, the non-limiting term user equipment (UE) or wireless device may be used and may refer to any type of wireless device communicating with a network node and / or with another UE in a cellular or mobile communication system. Examples of UE are target device, device to device (D2D) UE, machine type UE or UE capable of machine to machine (M2M) communication, PDA, PAD, Tablet, mobile terminals, smart phone, laptop embedded equipped (LEE), laptop mounted equipment (LME), USB dongles, UE category Ml, UE category M2, ProSe UE, V2V UE, V2X UE, etc.

[0059] Additionally, terminologies such as base station / gNodeB and UE should be considered non-limiting and do in particular not imply a certain hierarchical relation between the two; in general, “gNodeB” could be considered as device 1 and “UE” could be considered as device 2 and these two devices communicate with each other over some radio channel. And in the following the transmitter or receiver could be either gNodeB (gNB), or UE.

[0060] As will be appreciated by one skilled in the art, aspects of the embodiments may be embodied as a system, apparatus, method, or program product. Accordingly, embodiments may take the form of an entirely hardware embodiment, an entirely software embodiment (including firmware, resident software, micro-code, etc.) or an embodiment combining software and hardware aspects.

[0061] For example, the disclosed embodiments may be implemented as a hardware circuit comprising custom very-large-scale integration (“VLSI”) circuits or gate arrays, off- the-shelf semiconductors such as logic chips, transistors, or other discrete components. The disclosed embodiments may also be implemented in programmable hardware devices such as field programmable gate arrays, programmable array logic, programmable logic devices, or the like. As another example, the disclosed embodiments may include one or more physical or logical blocks of executable code which may, for instance, be organized as an object, procedure, or function.

[0062] Furthermore, embodiments may take the form of a program product embodied in one or more computer readable storage devices storing machine readable code, computer readable code, and / or program code, referred hereafter as code. The storage devices may be tangible, non- transitory, and / or non-transmission. The storage devices may not embody signals. In a certain embodiment, the storage devices only employ signals for accessing code.

[0063] Any combination of one or more computer readable medium may be utilized. The computer readable medium may be a computer readable storage medium. The computer readable storage medium may be a storage device storing the code. The storage device may be, for example, but not limited to, an electronic, magnetic, optical, electromagnetic, infrared, holographic, micromechanical, or semiconductor system, apparatus, or device, or any suitable combination of the foregoing.

[0064] More specific examples (a non-exhaustive list) of the storage device would include the following: an electrical connection having one or more wires, a portable computer diskette, a hard disk, a random-access memory (“RAM”), a read-only memory (“ROM”), an erasable programmable read-only memory (“EPROM” or Flash memory), a portable compact disc readonly memory (“CD-ROM”), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing. In the context of this document, a computer readable storage medium may be any tangible medium that can contain or store a program for use by or in connection with an instruction execution system, apparatus, or device.

[0065] Code for carrying out operations for embodiments may be any number of lines and may be written in any combination of one or more programming languages including an object- oriented programming language such as Python, Ruby, Java, Smalltalk, C++, or the like, and conventional procedural programming languages, such as the “C” programming language, or the like, and / or machine languages such as assembly languages. The code may execute entirely on the user’s computer, partly on the user’s computer, as a stand-alone software package, partly on the user’s computer and partly on a remote computer or entirely on the remote computer or server. In the latter scenario, the remote computer may be connected to the user’s computer through any type of network, including a local area network (“LAN”), wireless LAN (“WLAN”), or a wide area network (“WAN”), or the connection may be made to an external computer (for example, through the Internet using an Internet Service Provider (“ISP”)).

[0066] Furthermore, the described features, structures, or characteristics of the embodiments may be combined in any suitable manner. In the following description, numerous specific details are provided, such as examples of programming, software modules, user selections, network transactions, database queries, database structures, hardware modules, hardware circuits, hardware chips, etc., to provide a thorough understanding of embodiments. One skilled in the relevant art will recognize, however, that embodiments may be practiced without one or more of the specific details, or with other methods, components, materials, and so forth. In other instances, well-known structures, materials, or operations are not shown or described in detail to avoid obscuring aspects of an embodiment. Reference throughout this specification to “one embodiment,” “an embodiment,” or similar language means that a particular feature, structure, or characteristic described in connection with the embodiment is included in at least one embodiment. Thus, appearances of the phrases “in one embodiment,” “in an embodiment,” and similar language throughout this specification may, but do not necessarily, all refer to the same embodiment, but mean “one or more but not all embodiments” unless expressly specified otherwise. The terms “including,” “comprising,” “having,” and variations thereof mean “including but not limited to,” unless expressly specified otherwise. An enumerated listing of items does not imply that any or all of the items are mutually exclusive, unless expressly specified otherwise. The terms “a,” “an,” and “the” also refer to “one or more” unless expressly specified otherwise.

[0067] Aspects of the embodiments are described below with reference to schematic flowchart diagrams and / or schematic block diagrams of methods, apparatuses, systems, and program products according to embodiments. It will be understood that each block of the schematic flowchart diagrams and / or schematic block diagrams, and combinations of blocks in the schematic flowchart diagrams and / or schematic block diagrams, can be implemented by code. This code may be provided to a processor of a general-purpose computer, special purpose computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the fimctions / acts specified in the flowchart diagrams and / or block diagrams.

[0068] The code may also be stored in a storage device that can direct a computer, other programmable data processing apparatus, or other devices to function in a particular manner, such that the instructions stored in the storage device produce an article of manufacture including instructions which implement the function / act specified in the flowchart diagrams and / or block diagrams.

[0069] The code may also be loaded onto a computer, other programmable data processing apparatus, or other devices to cause a series of operational steps to be performed on the computer, other programmable apparatus, or other devices to produce a computer implemented process such that the code which execute on the computer or other programmable apparatus provide processes for implementing the functions / acts specified in the flowchart diagrams and / or block diagrams. The flowchart diagrams and / or block diagrams in the Figures illustrate the architecture, functionality, and operation of possible implementations of apparatuses, systems, methods, and program products according to various embodiments. In this regard, each block in the flowchart diagrams and / or block diagrams may represent a module, segment, or portion of code, which includes one or more executable instructions of the code for implementing the specified logical function(s).

[0070] It should also be noted that, in some alternative implementations, the functions noted in the block may occur out of the order noted in the Figures. For example, two blocks shown in succession may, in fact, be executed substantially concurrently, or the blocks may sometimes be executed in the reverse order, depending upon the functionality involved. Other steps and methods may be conceived that are equivalent in function, logic, or effect to one or more blocks, or portions thereof, of the illustrated Figures.

[0071] Although various arrow types and line types may be employed in the flowchart and / or block diagrams, they are understood not to limit the scope of the corresponding embodiments. Indeed, some arrows or other connectors may be used to indicate only the logical flow of the depicted embodiment. For instance, an arrow may indicate a waiting or monitoring period of unspecified duration between enumerated steps of the depicted embodiment. It will also be noted that each block of the block diagrams and / or flowchart diagrams, and combinations of blocks in the block diagrams and / or flowchart diagrams, can be implemented by special purpose hardware-based systems that perform the specified functions or acts, or combinations of special purpose hardware and code.

[0072] The description of elements in each figure may refer to elements of proceeding figures. Like numbers refer to like elements in all figures, including alternate embodiments of like elements.

[0073] The detailed description set forth below, with reference to the figures, is intended as a description of various configurations and is not intended to represent the only configurations in which the concepts described herein may be practiced. The detailed description includes specific details for the purpose of providing a thorough understanding of the various concepts. However, it will be apparent to those skilled in the art that these concepts may be practiced without these specific details. For instance, although 3GPP terminology, from e.g., 5G NR, may be used in this disclosure to exemplify embodiments herein, this should not be seen as limiting the scope of the present disclosure.

[0074] The disclosure is related to wireless communication system, which may be for example a 5G NR wireless communication system. More specifically, it represents a RAN of the wireless communication system, which is used exchange data with UEs via radio signals. For example, the RAN may send data to the UEs (downlink, DL), for instance data received from a core network (CN). The RAN may also receive data from the UEs (uplink, UL), which data may be forwarded to the CN.

[0075] In the examples illustrated, the RAN comprises one base station, BS. Of course, the RAN may comprise more than one BS to increase the coverage of the wireless communication system. Each of these BSs may be referred to as NB, eNodeB (or eNB), gNodeB (or gNB, in the case of a 5G NR wireless communication system), an access point or the like, depending on the wireless communication standard(s) implemented.

[0076] The UEs are located in a coverage of the BS. The coverage of the BS corresponds for example to the area in which UEs can decode a PDCCH transmitted by the BS.

[0077] An example of a wireless device suitable for implementing any method, discussed in the present disclosure, performed at a UE corresponds to an apparatus that provides wireless connectivity with the RAN of the wireless communication system, and that can be used to exchange data with said RAN. Such a wireless device may be included in a UE. The UE may for instance be a cellular phone, a wireless modem, a wireless communication device, a handheld device, a laptop computer, or the like. The UE may also be an Internet of Things (loT) equipment, like a wireless camera, a smart sensor, a smart meter, smart glasses, a vehicle (manned or unmanned), a global positioning system device, etc., or any other equipment that may run applications that need to exchange data with remote recipients, via the wireless device.

[0078] The wireless device comprises one or more processors and one or more memories. The one or more processors may include for instance a central processing unit (CPU), a digital signal processor (DSP), a field-programmable gate array (FPGA), an application specific integrated circuit (ASIC), etc. The one or more memories may include any type of computer readable volatile and non-volatile memories (magnetic hard disk, solid-state disk, optical disk, electronic memory, etc.). The one or more memories may store a computer program product, in the form of a set of programcode instructions to be executed by the one or more processors to implement all or part of the steps of a method for exchanging data, performed at a UE’s side, according to any one of the embodiments disclosed herein.

[0079] The wireless device can comprise also a main radio, MR, unit. The MR unit corresponds to a main wireless communication unit of the wireless device, used for exchanging data with BSs of the RAN using radio signals. The MR unit may implement one or more wireless communication protocols, and may for instance be a 3G, 4G, 5G, NR, WiFi, WiMax, etc. transceiver or the like. In preferred embodiments, the MR unit corresponds to a 5G NR wireless communication unit.

[0080] AI / ML Model is a data driven algorithm that applies AI / ML techniques to generate set of outputs based on set of inputs.

[0081] AI / ML model delivery is a generic term referring to delivery of an AI / ML model from one entity to another entity in any manner. Note is An entity could mean network node / function (e.g., gNB, LMF, etc.), UE, proprietary server, etc.

[0082] AI / ML model Inference is a process of using trained AI / ML model to produce set of outputs based on set of inputs. AI / ML model testing is a subprocess of training, to evaluate the performance of final AI / ML model using dataset different from one used for model training and validation. Differently from AI / ML model validation, testing does not assume subsequent tuning of the model.

[0083] AI / ML model training is a process to train an AI / ML Model [by learning the input / output relationship] in data driven manner and obtain the trained AI / ML Model for inference.

[0084] AI / ML model transfer is a delivery of an AI / ML model over the air interface in manner that is not transparent to 3GPP signalling, either parameters of model structure known at the receiving end or new model with parameters. Delivery may contain full model or partial model.

[0085] AI / ML model validation is a subprocess of training, to evaluate the quality of an AI / ML model using dataset different from one used for model training, that helps selecting model parameters that generalize beyond the dataset used for model training.

[0086] Data collection is a process of collecting data by the network nodes, management entity, or UE for the purpose of AI / ML model training, data analytics and inference.

[0087] Federated learning I federated training is a machine learning technique that trains an AI / ML model across multiple decentralized edge nodes e.g., UEs, gNBs each performing local model training using local data samples. The technique requires multiple interactions of the model, but no exchange of local data samples.

[0088] Functionality identification is a process / method of identifying an AI / ML functionality for the common understanding between the NW and the UE. Note is Information regarding the AI / ML functionality may be shared during functionality identification. Where AI / ML functionality resides depends on the specific use cases and sub use cases.

[0089] Model activation means enable an AI / ML model for specific AI / ML-enabled feature. Model deactivation means disable an AI / ML model for specific AI / ML-enabled feature.

[0090] Model download means Model transfer from the network to UE.

[0091] Model identification is A process / method of identifying an AI / ML model for the common understanding between the NW and the UE. The process / method of model identification may or may not be applicable and regarding the AI / ML model may be shared during model identification.

[0092] Model monitoring is A procedure that monitors the inference performance of the AI / ML model.

[0093] Model parameter update is Process of updating the model parameters of model. Model selection is the process of selecting an AI / ML model for activation among multiple models for the same AI / ML enabled feature. Model selection may or may not be carried out simultaneously with model activation.

[0094] Model switching is deactivating currently active AI / ML model and activating different AI / ML model for specific AI / ML-enabled feature.

[0095] Model update is Process of updating the model parameters and / or model structure of model.

[0096] Model upload is Model transfer from UE to the network.

[0097] Network-side (AI / ML) model is an AI / ML Model whose inference is performed entirely at the network.

[0098] Offline field data is the data collected from field and used for offline training of the AI / ML model.

[0099] Offline training is an AI / ML training process where the model is trained based on collected dataset, and where the trained model is later used or delivered for inference. Note is This definition only serves as guidance. There may be cases that may not exactly conform to this definition but could still be categorized as offline training by commonly accepted conventions.

[0100] Online field data is the data collected from field and used for online training of the AI / ML model.

[0101] Online training is an AI / ML training process where the model being used for inference) is (typically continuously) trained in (near) real-time with the arrival of new training samples. Note is the notion of (near) real-time vs. non real-time is context- dependent and is relative to the inference time-scale. This definition only serves as guidance.

[0102] There may be cases that may not exactly conform to this definition but could still be categorized as online training by commonly accepted conventions. Note is Fine- tuning / re-training may be done via online or offline training. This note could be removed when we define the term fine-tuning.

[0103] Reinforcement Learning (RL) is a process of training an AI / ML model from input (a.k.a. state) and feedback signal (a.k.a. reward) resulting from the model’s output (a.k.a. action) in an environment the model is interacting with.

[0104] Semi-supervised learning is a process of training model with mix of labelled data and unlabelled data.

[0105] Supervised learning is a process of training model from input and its corresponding labels.

[0106] Two-sided (AI / ML) model is a paired AI / ML Model(s) over which joint inference is performed, where joint inference comprises AI / ML Inference whose inference is performed jointly across the UE and the network, i.e, the first part of inference is firstly performed by UE and then the remaining part is performed by gNB, or vice versa. UE-side (AI / ML) model is an AI / ML Model whose inference is performed entirely at the UE.

[0107] Unsupervised learning is a process of training model without labelled data.

[0108] Proprietary-format models is ML models of vendor-Zdevice-specific proprietary format, from 3GPP perspective. They are not mutually recognizable across vendors and hide model design information from other vendors when shared.

[0109] Open-format models is ML models of specified format that are mutually recognizable across vendors and allow interoperability, from 3GPP perspective. They are mutually recognizable between vendors and do not hide model design information from other vendors when shared.

[0110] The following explanation will provide the detailed description of the mechanism about pre-configuring and signaling the specific information about model online training by configuring a set of UE behaviors. AI / ML based techniques are currently applied to many different applications and 3GPP also started to work on its technical investigation to apply to multiple use cases based on the observed potential gains. AI / ML lifecycle can be split into several stages such as data collection / pre- processing, model training, model testing / validation, model deployment / update, model monitoring etc., where each stage is equally important to achieve target performance with any specific model(s). In applying AI / ML model for any use case or application, one of the challenging issues is to manage the lifecycle of AI / ML model. It is mainly because the data / model drift occurs during model deployment / inference and it results in performance degradation of AI / ML model. Fundamentally, the dataset statistical changes occur after model is deployed and model inference capability is also impacted with unseen data as input. In a similar aspect, the statistical property of dataset and the relationship between input and output for the trained model can be changed with drift occurrence. In this context, model training or re-training is one of key issues for model performance maintenance as model performance such as inferencing and / or training is dependent on different model execution environment with varying configuration parameters.

[0111] To handle this issue, collaboration between UE and gNB is highly important to track model performance and re-configure model corresponding to different environments. AI / ML model needs model monitoring after deployment because model performance cannot be maintained continuously due to drift and update feedback is then provided to re-train / update the model or select alternative model. When AI / ML model enabled wireless communication network is deployed, it is then important to consider how to handle AI / ML model in activation with re-configuration for wireless devices under operations such as model training, inference, updating, etc. When UE device ML resource (e.g., computing power, memory size, battery capacity) is limited, AI / ML model operation at UE cannot be guaranteed to sustain any configured target performance as model performance is vulnerable to the associated input dataset availability (e.g., dataset size / quality). In this method, a set of ML model operation modes are preset for model activation such as mode-1 for mapping-based model, mode-2 for non-mapping based model and mode-3 for hybrid model. For mode-1 , when applying a specific ML model, one or multiple mapping relation look-up tables can be pre-defined via offline ML model operation based on specific ML configuration information and / or ML conditions such as ML model applications, model properties / types, environmental attribute data and / or applicable device capability, etc. For example, the configured mapping relation look-up table represents the mapping relation information about the dataset (model input) and the associated model output, where the dataset can be clustered based on statistical characteristics of data values and the associated model output can be also the quantized value selection from the pre-configured set of output values.

[0112] Therefore, with ML model operation mode-1 , there is no need of running real-time model inferencing operation to generate model output since the configured mapping relation look-up table can be used to provide model output when a new model input data is given. For mode-2, model output has no dependency on any pre-defined input / output mapping relation and model is then run to generate output for the given input dataset. For mode-3, model operation is based on using both mode-1 and mode-2 for model output generation where processing of model output generation is implementation-specific.

[0113] However, for example, both mode-1 and mode-2 can be operated simultaneously and model outputs of mode-1 and mode-2 can be selectively chosen or combined together to achieve target model accuracy level with mode-3 operation. As another example, output of mode-1 firstly provides a specific output that is referenced for mode-2 so that faster operation of mode-2 can be executed where any quantized output with mode-1 can be further accurately generated using mode-2 based model output. The pre-configured ML model operation modes and the associated mapping relation look-up tables can be sent via system information or dedicated RRC signaling. Or L1 / L2 or RRC signaling can be used for indication of activating mapping relation look-up table(s) for use if applicable. And whenever mapping relation look-up table need to be updated, the updated table can be signaled as unicast, multicast or broadcast information. Regarding the structure of mapping relation look-up table, model input data is categorized into a finite set of dataset ID or index as DS #i (i=1 ,2,3, ... ) and model output data is quantized into a finite set of output ID or index as Q_Out #j (j=1 ,2,3, ... ) where any combinations of model input data and output data can be also represented as index information if applicable. By using mapping relation look-up table, dataset of model input can be clustered using the statistical characteristics so that new model input data can be identified as one of DS #i or clustered data and the quantized model output can be a value from the preconfigured set of output values that have mapping relation with model input. Different levels of quantization and / or clustering can be applied to generate multiple versions of mapping relation look-up tables with different sizes of tables where model performance accuracy can vary in accordance with multiple versions of mapping relation look-up tables. Based on clustering method, dataset of model input can be clustered (indicated as DS #i) using the statistical characteristics. The quantized model output (indicated as Q_Out #j) can be a value representing a number of model output data values that have mapping relation with model input data values. Mapping relation look-up table representing both DS #i and Q_Out #j can be generated by running ML model(s) via offline operation. Also online operation to update mapping relation look-up table can be applied depending on implementation scenarios. For example, dataset can be re-clustered with the associated quantized output after online training of ML model.

[0114] ML model operation mode-3 (hybrid mode) combines the advantages of both mapping-based inference (mode-1 ) and full AI / ML model execution (mode-2) to optimize the trade-off between inference latency, computational load, and accuracy, where the UE and / or network entity dynamically select or fuse outputs from the preconfigured mapping relation look-up table and the model inference path according to ML conditions, signaling instructions, or preset thresholds. For example, upon receiving mode-3 activation signaling via RRC or L1 / L2, the UE may first attempt to map an incoming data sample to a clustered dataset index and retrieve the corresponding quantized output from the ML look-up table. Concurrently or subsequently, the UE can execute the full AI / ML model on the same input and compare the fast, low-complexity look-up table result with the high-confidence model output. In another use case using mode-3, the mapping-based output for mode-1 may serve as a preliminary initiation for the model inference in mode-2, where the UE initializes the AI / ML model’s internal state or a subset of its parameters with the quantized output, reducing convergence time and computation. This two-stage approach permits rapid coarse inference followed by a lightweight refinement step, further balancing computational efficiency and accuracy.

[0115] Criteria of dataset clustering and output quantization are implementation-specific for varying scenarios or ML use cases with different model applications / functionalities. The number of mapping relation look-up tables and the size of each mapping relation look-up table also can be pre-configured depending on implementation scenarios. At network side, ML model operation modes are preset for specific ML applications / functionalities with target model(s) for activation. Based on the determined ML model operation mode, target model(s) can be selected for activation.

[0116] For mode-1 (mapping based model), mapping relation look-up table is enabled for model inferencing without running ML model. For mode-2 (non-mapping based model), ML model is run for model inferencing. For mode-3 (hybrid model), model outputs of mode-1 and mode-2 can be selectively chosen or combined together for model inferencing. After ML configuration information with the preset ML model operation mode and mapping relation look-up table is received by UE, either network side or UE side can determine a specific ML model operation mode and mapping relation look-up table for activation based on either one- / two-sided model and / or ML configuration. Activation of a specific ML model operation mode with or without mapping relation look-up table can be triggered by indication signaling sent from network side or by UE autonomous decision or by any preset threshold. Any specific ML model operation mode and use of the preset mapping relation look-up table can be enabled by network side using indication message sent to UE. Before sending activation indication to UE, ML condition is measured at UE side for reporting to network side where ML condition can be configured to include information such as on-device ML model applicability, environmental condition, etc. influencing performance of the selected ML model operation mode. Any specific ML model operation mode and use of the preset mapping relation look-up table can be enabled by UE based in device autonomous decision. Before autonomous decision of activating a specific ML model operation mode, applicable mode(s) can be identified to determine activation along with ML condition measurement. After ML model operation mode activation, UE sends status feedback of the activated ML model operation mode.

[0117] Figure 1 shows an exemplary mapping relation look-up table. In this example, model input data is categorized into a finite set of dataset ID or index as DS #i (i=1 ,2,3, ... ). Model output data is quantized into a finite set of output ID or index as Q_Out #j (j=1 ,2,3, ... ). The combinations of model input data and output data can be also represented as index information if applicable as well. Depending on use cases, dataset of model input can be clustered using the statistical characteristics so that new model input data can be identified as one of DS #i or clustered data. The quantized model output can be a value from the pre-configured set of output values that have mapping relation with model input. Different levels of quantization and / or clustering can be applied to generate multiple versions of mapping relation look-up tables with different sizes of tables where model performance accuracy can vary in accordance with multiple versions of mapping relation look-up tables. Figure 2 shows an exemplary block diagram of mapping relation between clustered dataset and quantized output. In this example, based on clustering method, dataset of model input can be clustered (indicated as DS #i) using the statistical characteristics.

[0118] The quantized model output (indicated as Q_Out #j) can be a value representing a number of model output data values that have mapping relation with model input data values. Mapping relation look-up table representing both DS #i and Q_Out #j can be generated by running ML model(s) via offline operation. Also online operation to update mapping relation look-up table can be applied depending on implementation scenarios. For example, dataset can be re-clustered with the associated quantized output after online training of ML model. Criteria of dataset clustering and output quantization are implementation-specific for varying scenarios or ML use cases with different model applications / functionalities. The number of mapping relation look-up tables and the size of each mapping relation look-up table also can be pre-configured depending on implementation scenarios. However, different levels of quantization and / or clustering can be applied to generate multiple versions of mapping relation look-up tables with different sizes of tables where model performance accuracy can vary in accordance with multiple versions of mapping relation look-up tables.

[0119] Figure 3 shows an exemplary flow chart of configuring ML model operation modes and mapping relation look-up tables at network side. In this example, at network side ML model operation modes are preset for specific ML applications / functionalities with target model(s) for activation. Based on the determined ML model operation mode, target model(s) can be selected for activation. For mode-1 (mapping based model), mapping relation look-up table is enabled for model inferencing without running ML model. For mode-2 (non-mapping based model), ML model is run for model inferencing. For mode-3 (hybrid model), model outputs of mode-1 and mode-2 can be selectively chosen or combined together for model inferencing.

[0120] Figure 4 shows an exemplary flow chart of activating a specific ML model operation mode at UE side. In this example, after ML configuration information with the preset ML model operation mode and mapping relation look-up table is received by UE, either network side or UE side can determine a specific ML model operation mode and mapping relation look-up table for activation based on either one- / two-sided model and / or ML configuration. Activation of a specific ML model operation mode with or without mapping relation look-up table can be triggered by indication signaling sent from network side or by UE autonomous decision or by any preset threshold. Figure 5 shows an exemplary signaling flow of activating ML model operation mode by network side. In this example, any specific ML model operation mode and use of the preset mapping relation look-up table can be enabled by network side using indication message sent to UE. Before sending activation indication to UE, ML condition is measured at UE side for reporting to network side where ML condition can be configured to include information such as on-device ML model applicability, environmental condition, etc. influencing performance of the selected ML model operation mode. Figure 6 shows an exemplary signaling flow of activating ML model operation mode by UE side. In this example, any specific ML model operation mode and use of the preset mapping relation look-up table can be enabled by UE based in device autonomous decision.

[0121] Before autonomous decision of activating a specific ML model operation mode, applicable mode(s) can be identified to determine activation along with ML condition measurement. After ML model operation mode activation, UE sends status feedback of the activated ML model operation mode.

Claims

CLAIMS1. A method of pre-mapping based model signaling by configuring a set of ML model operation modes at network entity in a wireless communication system, comprising:• Presetting mode-1 for mapping-based model, mode-2 for non-mapping based model and mode-3 for hybrid model;• Configuring mapping relation look-up table(s) for the clustered dataset and quantized model output; and• Activating ML model operation mode(s).

2. The method according to previous claim 1 , wherein one or multiple mapping relation look-up tables can be pre-defined via offline ML model operation based on specific ML configuration information and / or ML conditions such as ML model applications and / or model properties / types and / or environmental attribute data and / or applicable device capability.

3. The method according to any of the previous claims, wherein the configured mapping relation look-up table represents the mapping relation information about the dataset (model input) and the associated model output, comprising:• Clustering dataset based on statistical characteristics of data values; and• Quantizing the associated model output as the quantized value selection from the pre-configured set of output values.

4. The method according to any of the previous claims, wherein with ML model operation mode-1 there is no need of running real-time model inferencing operation to generate model output as the configured mapping relation look-up table can be used to provide model output when a new model input data is given.

5. The method according to any of the previous claims, wherein with ML model operation mode-2 model output has no dependency on any pre-definedinput / output mapping relation and model is then run to generate output for the given input dataset.

6. The method according to any of the previous claims, wherein with ML model operation mode-3 model operation is based on using both mode-1 and mode-2 for model output generation.

7. The method according to any of the previous claims, wherein both mode-1 and mode-2 can be operated simultaneously so that model outputs of mode-1 and mode-2 can be selectively chosen or combined together to achieve target model accuracy level with mode-3 operation.

8. The method according to any of the previous claims, wherein output of mode-1 firstly can provide a specific output that is referenced for mode-2 for faster operation of mode-2 with higher accuracy of model output.

9. The method according to any of the previous claims, wherein the pre-configured ML model operation modes and the associated mapping relation look-up tables can be sent via system information or dedicated RRC signaling.

10. The method according to any of the previous claims, wherein L1 / L2 or RRC signaling can be used for indication of activating mapping relation look-up table(s) for use if applicable.11 .The method according to any of the previous claims, wherein the updated mapping relation look-up table can be signaled as unicast, multicast or broadcast information.

12. The method according to any of the previous claims, wherein model input data is categorized into a finite set of dataset ID or index as DS #i (i=1 ,2,3, ... ) and model output data is quantized into a finite set of output ID or index as Q_Out #j (j=1 ,2,3, ... ) for mapping relation look-up table.

13. The method according to any of the previous claims, wherein any combinations of model input data and output data in mapping relation look-up table can be represented as index information if applicable.

14. The method according to any of the previous claims, wherein new model input data can be identified as one of DS #i or clustered data as the quantized model output can be a value from the pre-configured set of output values that have mapping relation with model input.

15. The method according to any of the previous claims, wherein different levels of quantization and / or clustering can be applied to generate multiple versions of mapping relation look-up tables with different sizes of tables.

16. The method according to any of the previous claims, wherein model performance accuracy can vary in accordance with multiple versions of mapping relation lookup tables.

17. The method according to any of the previous claims, wherein online operation to update mapping relation look-up table can be applied depending on implementation scenarios so that dataset can be re-clustered with the associated quantized output after online training of ML model.

18. The method according to any of the previous claims, wherein criteria of dataset clustering and output quantization can be implementation-specific for varying scenarios or ML use cases with different model applications / functionalities.

19. The method according to any of the previous claims, wherein the number of mapping relation look-up tables and the size of each mapping relation look-up table can be pre-configured depending on implementation scenarios.

20. The method according to any of the previous claims, wherein at network side ML model operation modes are preset for specific ML applications / functionalities with target model(s) for activation.21 . The method according to any of the previous claims, wherein based on the determined ML model operation mode, target model(s) can be selected for activation.

22. The method according to any of the previous claims, wherein either network side or UE side can determine a specific ML model operation mode and mapping relation look-up table for activation based on either one- / two-sided model and / or ML configuration after ML configuration information with the preset ML model operation mode and mapping relation look-up table is received by UE.

23. The method according to any of the previous claims, wherein activation of a specific ML model operation mode with or without mapping relation look-up table can be triggered by indication signaling sent from network side or by UE autonomous decision or by any preset threshold.

24. The method according to any of the previous claims, wherein any specific ML model operation mode and use of the preset mapping relation look-up table can be enabled by network side using indication message sent to UE.

25. The method according to any of the previous claims, wherein ML condition is measured at UE side for reporting to network side as ML condition can be configured to include information such as on-device ML model applicability, environmental condition, etc. influencing performance of the selected ML model operation mode before sending activation indication to UE.

26. The method according to any of the previous claims, wherein any specific ML model operation mode and use of the preset mapping relation look-up table can be enabled by UE based in device autonomous decision.

27. The method according to any of the previous claims, wherein applicable mode(s) can be identified to determine activation along with ML condition measurement before autonomous decision of activating a specific ML model operation mode.

28. The method according to any of the previous claims, wherein UE sends status feedback of the activated ML model operation mode after ML model operation mode activation.

29. Apparatus for pre-mapping based model signaling by configuring a set of ML model operation modes in a wireless communication system in a wireless communication system, that are applied to multiple cells, the apparatus comprising a wireless transceiver, a processor coupled with a memory in which computer program instructions are stored, said instructions being configured to carry out steps of the claims 1 to 28.

30. User Equipment comprising an apparatus according to claim 29.31 . gNB comprising an apparatus according to claim 29.

32. Wireless communication system for pre-mapping based model signaling by configuring a set of ML model operation modes, wherein the wireless communication systems comprises user equipment according to claim 30, gNB according to claim 31 , whereby the user equipment and the gNB each comprises a processor coupled with a memory in which computer program instructions are stored, said instructions being configured to carry out the steps of the claims 1 to 29.

Images