Adaptive model signaling for RAN with associated datasets
By configuring the mapping relationship between the model and the measurement range of the input dataset in the radio access network, the degradation of model performance caused by drift is solved, adaptive model signaling is realized, signaling overhead is reduced and model performance is maintained.
Patent Information
- Application Number
- CN202580029442.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Priority Date
- 2024-05-06
- Filing Date
- 2025-04-30
- Publication Date
- 2026-11-17
AI Technical Summary
In radio access networks, existing technologies have failed to effectively manage the lifecycle of AI/ML models, resulting in degraded model performance due to data/model drift, especially with excessive signaling overhead during model training and updates between base stations and multiple user equipment.
By pre-determining the mapping relationship between the model and the measurement range of the input dataset, the model set is configured to adapt to different dataset quality and size. The mapping relationship information is sent using system information or dedicated RRC signaling, allowing user devices or the network side to decide on model selection and complexity adjustment, thereby achieving adaptive model signaling.
It reduces signaling overhead, improves the model's performance maintenance capability in different environments, and ensures that the model can continuously maintain the target performance on user devices.
Smart Images

Figure CN123125841A_ABST
Abstract
Description
Technical Field
[0001] This disclosure relates to AI / ML-based model adaptation, which presents techniques for pre-configuring and signaling specific information about adapting ML models for radio access networks. Background Technology
[0002] Within 3GPP (3rd Generation Partnership Project), one of the selected research projects as part of the approved Release 18 package is AI / ML (Artificial Intelligence / Machine Learning), as described in the relevant document (RP-213599) submitted at 3GPP TSG (Technical Specification Group) RAN (Radio Access Networks) meeting #94e. The formal title of the AI / ML research project is "Study on AI / ML for NR Air Interface." The objective of this research project is to identify general AI / ML frameworks and areas where benefits can be derived from using AI / ML-based technologies and use cases.
[0003] According to 3GPP, the main objective of this research project is to study an AI / ML framework for the air interface by considering performance, complexity, and potential specification impacts, leveraging target use cases. Specifically, AI / ML models, terminology, and descriptions for identifying general and specific characteristics of the framework are included as one of the key work areas. Regarding the AI / ML framework, various aspects are being considered for investigation, and one key project concerns the lifecycle management of AI / ML models, which mandates multiple phases for model training, deployment, inference, monitoring, and updates. Furthermore, in 3GPP, a two-sided (AI / ML) model is defined as a paired AI / ML model on which joint inference is performed, where joint inference includes AI / ML inference performed jointly across the UE and network. Additionally, for one-sided (AI / ML) models, a UE-side (AI / ML) model is defined as an AI / ML model that performs inference entirely at the UE, and a network-side (AI / ML) model is defined as an AI / ML model that performs inference entirely at the network. Currently, the AI / ML specification work is in the work project discussion phase for version 19. Earlier, in 3GPP TR 37.817, version 17, entitled "Study on enhancement for Data Collection for NR and EN-DC," UE (User Equipment) mobility was also considered as one of the AI / ML use cases, and one of the scenarios for model training / inference was that both functions resided within the RAN node. Subsequently, in version 18, a new work project, "Artificial Intelligence (AI) / Machine Learning (ML) for NG-RAN," was initiated to specify data collection enhancements and signaling support within the existing NG-RAN interface and architecture. For the standardization work of the aforementioned activities, RAN-based AI / ML models were considered crucial for both the network and UE to meet any desired model operations (e.g., model training, inference, selection, handover, updates, monitoring, etc.). Model information could be signaled to pair the network-side model and the UE-side model for various lifecycle management (LCM) operations. In the RAN, signaling is crucial for communication between the UE and the network because it occurs across different layers of the protocol stack—primarily L1 (Layer 1), L2 (Layer 2), and RRC (Radio Resource Control). RRC, a Layer 3 protocol used on the air interface between the UE and the base station (e.g., gNB in 5G, eNB in LTE), is primarily responsible for establishing, configuring, maintaining, and releasing the radio resources and connections required for communication.
[0004] However, the signaling overhead for indicating model information can be very high, especially when processing model-based LCM between the base station (BS / gNB) and multiple UEs. Model training is one of the most critical parts of model deployment in LCM, and currently there are no defined specifications for signaling methods and network-UE behavior to identify the datasets required for model updates / retraining, as any activated model can be affected by model / data drift. When ML conditions change, the enabled AI / ML model may suffer performance degradation due to data / model drift. In such cases, model retraining / update may be necessary. For example, when a trained ML model is deployed in the RAN, model inference performance may easily degrade if the target ML conditions do not perfectly match the true ML conditions measured for a specific model operation.
[0005] US2016092786A1 describes a feature superset that includes features used by a machine learning model and additional features generated therefrom for various feature groups used by the machine learning model.
[0006] US2019130247A1 describes feature ranking within a deep neural network in a multi-task and / or multi-label setting, where task-specific features as well as features shared across multiple tasks are identified.
[0007] US2021035021A1 describes a method for monitoring the performance of a machine learning model, wherein monitoring data elements are fed into a machine learning model trained on a training dataset of historical training data elements, wherein each data element is associated with a corresponding time after the time associated with the training dataset.
[0008] US2021374562A1 describes a method for determining the resource overhead of a baseline version of a machine learning model that uses a feature set to generate entity rankings, as well as the number of features to be removed to reduce the resource overhead to a target resource overhead. Summary of the Invention
[0009] This disclosure addresses the cited problems through proposed embodiments and describes a method for adaptive model signaling for a RAN with associated datasets by configuring a model set to be associated with multiple input dataset measurement (e.g., quality / size) ranges in a wireless communication system. The method includes: pre-determining a mapping relationship between the model and the dataset measurement ranges; defining the dataset measurement ranges; setting a threshold to configure the dataset measurement ranges; and providing the mapping relationship information to the UE.
[0010] In some embodiments of the method according to the first aspect, the method is characterized in that one or more models can be mapped to an associated range of input dataset measurements.
[0011] In some embodiments of the method according to the first aspect, the method is characterized in that mapping relationship information can be sent via system information or dedicated RRC signaling.
[0012] In some embodiments of the method according to the first aspect, the method is characterized in that any additional updates regarding the mapping relationship can be provided to the UE via L1 / L2 or RRC signaling.
[0013] In some embodiments of the method according to the first aspect, the method is characterized in that, if applicable, the model input dataset measurement range can be defined, such as the dataset quality and / or dataset size for mapping relationships with the model set.
[0014] In some embodiments of the method according to the first aspect, the method is characterized in that the measurement range of the model input dataset with a configured threshold can be configured by applying multiple models with different complexity levels to the target model performance based on offline model training.
[0015] In some embodiments of the method according to the first aspect, the method is characterized in that multiple mapping tables may exist for different ML model functionalities, model applications, and / or model input dataset attributes.
[0016] In some embodiments of the method according to the first aspect, the method is characterized in that the UE can autonomously select the relevant model to use.
[0017] In some embodiments of the method according to the first aspect, the method is characterized in that the network side can determine any specific model for the UE based on a dataset measurement report.
[0018] In some embodiments of the method according to the first aspect, the method is characterized in that the UE can decide to adjust the applicable model to the measured input dataset attributes by changing the model complexity level for the configured target model performance.
[0019] In some embodiments of the method according to the first aspect, the method is characterized in that, if applicable, the mapping table can indicate that a lower-complexity model can be used with an associated higher-quality and / or larger-size dataset.
[0020] In some embodiments of the method according to the first aspect, the method is characterized in that the quality / size range information of the model input dataset can be represented as a range indicator to be associated with each model.
[0021] In some embodiments of the method according to the first aspect, the method is characterized in that, if applicable, one or more mapping tables can be shared by multiple UEs or groups of UEs.
[0022] In some embodiments of the method according to the first aspect, the method is characterized in that the dataset measurement can be performed in a periodic or non-periodic manner or in combination.
[0023] In some embodiments of the method according to the first aspect, the method is characterized in that the quality / size range of a specific input dataset used for mapping relationships can be measured via offline model training between the network side and the UE side.
[0024] According to a second aspect, this disclosure relates to an apparatus for adaptive model signaling for a RAN with associated datasets by configuring a model set to be associated with a range of multiple input dataset measurements (e.g., quality / size) applied to multiple cells in a wireless communication system. The apparatus includes a wireless transceiver and a processor coupled to a memory storing computer program instructions configured to implement the steps of claims 1 to 15.
[0025] According to a third aspect, this disclosure relates to a user equipment that includes the means according to the second aspect.
[0026] According to the fourth aspect, this disclosure relates to a gNB that includes the means according to the second aspect.
[0027] According to a fifth aspect, this disclosure relates to a wireless communication system for adaptive model signaling for a RAN with associated datasets by configuring a model set to be associated with a range of measurements (e.g., quality / size) of a plurality of input datasets, wherein the wireless communication system includes a user equipment according to a third aspect and a gNB according to a fourth aspect, wherein the user equipment and the gNB each include a processor coupled to a memory storing computer program instructions configured to perform the steps according to a first aspect of this application.
[0028] According to a sixth aspect, this disclosure relates to a computer program product comprising instructions that, when executed by at least one processor, configure the at least one processor to perform the method according to the first aspect, and to perform a method for exchanging data according to any embodiment of the embodiments of this disclosure. The computer program product may be used in any programming language and may be in the form of source code, object code, or any intermediate form between source code and object code, such as a partially compiled form, or any other desired form.
[0029] According to a seventh aspect, this disclosure relates to a computer-readable storage medium including instructions that, when executed by at least one processor, configure the at least one processor to perform a method according to any embodiment of the present disclosure. Attached Figure Description
[0030] Figure 1 This is an example table showing the mapping between the model and the measurement range of the input dataset.
[0031] Figure 2 This is an exemplary flowchart for configuring mapping relationship information on the network side.
[0032] Figure 3 This is an exemplary flowchart of switching the adaptive model on the UE side.
[0033] Figure 4 This is an exemplary signaling flow for the UE's autonomous decision-making regarding the adaptive model.
[0034] Figure 5 This is an exemplary signaling flow for network decisions regarding adaptive models. Detailed Implementation
[0035] The detailed description set forth below with reference to the accompanying drawings is intended as a description of various configurations and is not intended to represent only the 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. Specifically, although terms from 3GPP 5G NR may be used in this disclosure to exemplify embodiments herein, this should not be construed as limiting the scope of the invention.
[0036] Some embodiments of the ideas contemplated herein will now be described more fully with reference to the accompanying drawings. However, other embodiments are also included within the scope of the subject matter disclosed herein, and the disclosed subject matter should not be construed as being limited to 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.
[0037] Generally, all terms used herein should be interpreted according to their ordinary meaning in the relevant art, unless a different meaning is expressly given and / or implied in the context of their use. Unless otherwise expressly stated, all references to a / an / said element, device, component, part, step, etc., should be interpreted openly as referring to at least one instance of that element, device, component, part, step, etc. The steps of any method disclosed herein need not be performed in the exact order disclosed, unless a step is explicitly described as following or preceding another step, and / or it implies that a step must follow or precede another step. Where appropriate, any feature of any embodiment in the embodiments disclosed herein may be applied to any other embodiment. Similarly, any advantage of any embodiment in the embodiments may be applied to any other embodiment, and vice versa. Other objects, features, and advantages of the appended embodiments will become apparent from the following description.
[0038] In some embodiments, the more general term "network node" may be used, and this term may correspond to any type of radio network node or any network node that communicates with the UE (directly or via another node) and / or with another network node. Examples of network nodes are NodeB, MeNB, ENB, network nodes belonging to MCG or SCG, base station (BS), multi-standard radio (MSR) radio nodes (such as MSR BS), eNodeB, gNodeB, network controller, radio network controller (RNC), base station controller (BSC), repeater, donor node controlling repeater, base transceiver (BTS), access point (AP), transport point, transport node, RRU, RRH, nodes in distributed antenna system (DAS), core network nodes (e.g., Mobile Switching Center (MSC), Mobility Management Entity (MME), etc.), Operations and Maintenance (O&M), Operations Support System (OSS), Self-Optimizing Network (SON), location nodes (e.g., Evolved Services Mobile Location Center (E-SMLC)), Minimized Drive Test (MDT), test equipment (physical node or software), etc.
[0039] In some embodiments, the non-limiting terms User Equipment (UE) or Wireless Device may be used, and the term may refer to any type of wireless device that communicates with a network node and / or with another UE in a cellular or mobile communication system. Examples of UEs are target devices, device-to-device (D2D) UEs, machine-type UEs or UEs capable of machine-to-machine (M2M) communication, PDAs, PADs, tablet computers, mobile terminals, smartphones, laptop embedded devices (LEEs), laptop mounted devices (LMEs), USB dongles, UE class M1, UE class M2, ProSe UEs, V2V UEs, V2X UEs, etc.
[0040] Furthermore, terms such as base station / gNodeB and UE should be considered non-restrictive and, in particular, do not imply any hierarchical relationship between the two; generally, "gNodeB" can be considered as device 1 and "UE" can be considered as device 2, and the two devices communicate with each other via a radio channel. Also, in the following text, a transmitter or receiver can be a gNodeB (gNB) or a UE.
[0041] As those skilled in the art will understand, aspects of the embodiments can be embodied as a system, apparatus, method, or program product. Therefore, embodiments can take the form of purely hardware embodiments, purely software embodiments (including firmware, resident software, microcode, etc.), or embodiments combining software and hardware aspects.
[0042] For example, the disclosed embodiments may be implemented as hardware circuitry including custom-designed very large-scale integration (“VLSI”) circuitry 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, etc.). As another example, the disclosed embodiments may include one or more physical or logical blocks of executable code, which may, for example, be organized as objects, procedures, or functions.
[0043] Furthermore, embodiments may take the form of a program product embodied in one or more computer-readable storage devices that store machine-readable code, computer-readable code, and / or program code (hereinafter referred to as code). The storage device may be tangible, non-transitory, and / or non-transferable. The storage device may not contain signals. In one embodiment, the storage device uses only signals to access the code.
[0044] Any combination of one or more computer-readable media may be used. A computer-readable medium may be a computer-readable storage medium. A computer-readable storage medium may be a storage device for storing code. A storage device may be, for example, but not limited to, electronic, magnetic, optical, electromagnetic, infrared, holographic, micromechanical, or semiconductor systems, apparatuses, or devices, or any suitable combination of the foregoing.
[0045] More specific examples of storage devices (a non-exhaustive list) will include the following: electrical connections having one or more wires, portable computer floppy disks, hard disks, random access memory (“RAM”), read-only memory (“ROM”), erasable programmable read-only memory (“EPROM” or flash memory), portable optical disc read-only memory (“CD-ROM”), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing. In the context of this document, a computer-readable storage medium can be any tangible medium that can contain or store programs for use by or in conjunction with an instruction execution system, apparatus, or device.
[0046] The code used to perform the operations of the embodiments can be any number of lines and can be written in any combination of one or more programming languages, including object-oriented programming languages such as Python, Ruby, Java, Smalltalk, C++, etc., as well as conventional procedural programming languages such as the "C" programming language, and / or machine languages such as assembly language. The code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In the latter scenario, the remote computer can be connected to the user's computer via any type of network, including a local area network ("LAN"), a wireless LAN ("WLAN"), or a wide area network ("WAN"), or can be connected to an external computer (e.g., via the Internet through an Internet Service Provider ("ISP").
[0047] Furthermore, the features, structures, or characteristics described in the embodiments can be combined in any suitable manner. Numerous specific details, such as examples of programming, software modules, user selection, network transactions, database queries, database structures, hardware modules, hardware circuits, hardware chips, etc., are provided in the following description to provide a thorough understanding of the embodiments. However, those skilled in the art will recognize that the embodiments can be practiced without one or more of these specific details or using other methods, components, materials, etc. In other instances, well-known structures, materials, or operations have not been shown or described in detail to avoid obscuring aspects of the embodiments. Throughout the specification, references to “an embodiment,” “embodiment,” or similar language mean that a particular feature, structure, or characteristic described in connection with the embodiment is included in at least one embodiment. Therefore, unless expressly stated otherwise, the phrases “in one embodiment,” “in an embodiment,” and similar language appearing throughout the specification may, but not necessarily all, refer to the same embodiment, but rather mean “one or more, but not all, embodiments.” Unless expressly stated otherwise, the terms “including,” “comprising,” “having,” and variations thereof mean “including, but not limited to.” Unless expressly stated otherwise, the enumerated list of items does not imply that any or all of these items are mutually exclusive. Unless otherwise expressly stated, the terms “a,” “an,” and “the / described” also mean “one or more.”
[0048] The following description of aspects of embodiments is based on schematic flowcharts and / or schematic block diagrams of methods, apparatus, systems, and program products according to embodiments. It should be understood that each block of the schematic flowcharts and / or schematic block diagrams, and combinations of blocks of the schematic flowcharts and / or schematic block diagrams, can be implemented by code. This code can 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 instructions executable via the processor of the computer or other programmable data processing apparatus create components for implementing the functions / actions specified in the flowcharts and / or block diagrams.
[0049] The code may also be stored in a storage device that can instruct a computer, other programmable data processing apparatus or other device to operate in a particular manner, such that the instructions stored in the storage device produce an article of art including instructions that implement the functions / actions specified in the flowchart and / or block diagram.
[0050] The code may also be loaded onto a computer, other programmable data processing apparatus or other device to cause a series of operational steps to be performed on the computer, other programmable apparatus or other device, thereby producing a computer-implemented process, such that the code executing on the computer or other programmable apparatus provides a process for implementing the functions / actions specified in the flowchart and / or block diagram.
[0051] The flowcharts and / or block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of apparatus, systems, methods, and program products according to various embodiments. In this regard, each block in the flowcharts and / or block diagrams may represent a module, segment, or portion of code, which includes one or more executable instructions for implementing a specified logical function.
[0052] It should also be noted that in some alternative implementations, the functions indicated within the boxes may not appear in the order shown in the figures. For example, depending on the functionality involved, two boxes shown consecutively may actually be executed substantially simultaneously, or these boxes may sometimes be executed in reverse order. Other steps and methods that are functionally, logically, or effectively equivalent to one or more boxes or portions thereof in the illustrated figures are conceivable.
[0053] While various arrow and line types may be used in flowcharts and / or block diagrams, they should be understood not to limit the scope of the corresponding embodiments. In fact, some arrows or other connectors may be used only to indicate the logical flow of the depicted embodiment. For example, an arrow may indicate a wait or monitoring period of unspecified duration between enumeration steps in a depicted embodiment. It should also be noted that each block in a block diagram and / or flowchart, and combinations of blocks in block diagrams and / or flowcharts, may be implemented by a system based on dedicated hardware or a combination of dedicated hardware and code that performs the specified function or action.
[0054] The description of the elements in each figure can be found in the preceding figures. Throughout all figures, the same numbers refer to the same elements, including alternative embodiments of the same elements.
[0055] The detailed description set forth below with reference to the accompanying drawings is intended to describe various configurations and is not intended to represent only the 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 example, although 3GPP terminology from, for example, 5G NR may be used in this disclosure to illustrate embodiments herein, this should not be considered as limiting the scope of this disclosure.
[0056] This disclosure relates to a wireless communication system, which may be, for example, a 5G NR wireless communication system. More specifically, it refers to a RAN (Radio Access Network) within the wireless communication system for exchanging data with a UE via radio signals. For example, the RAN may send data (downlink DL) to the UE, such as data received from the core network (CN). The RAN may also receive data from the UE (uplink UL), which may be forwarded to the CN.
[0057] In the example shown, the RAN includes a base station (BS). Of course, the RAN can include more than one BS to increase the coverage of the wireless communication system. Depending on the implemented wireless communication standard, each of these BSs may be referred to as an NB, eNodeB (or eNB), gNodeB (or gNB in the case of a 5G NR wireless communication system), access point, etc.
[0058] The UE is located within the coverage area of the BS. The coverage area of the BS corresponds, for example, to the region where the UE can decode the PDCCH sent by the BS.
[0059] Examples of wireless devices suitable for implementing any of the methods discussed in this disclosure at the UE correspond to means that provide wireless connectivity to a RAN of a wireless communication system and can be used to exchange data with said RAN. Such wireless devices may be included in the UE. The UE may be, for example, a cellular phone, a wireless modem, a wireless communication device, a handheld device, a laptop computer, etc. The UE may also be an Internet of Things (IoT) device, such as a wireless camera, a smart sensor, a smart meter, smart glasses, a (manned or unmanned) vehicle, a GPS device, etc., or any other device capable of running applications that require the exchange of data with a remote receiver via a wireless device.
[0060] The wireless device includes one or more processors and one or more memories. The one or more processors may include, for example, 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 memory (hard disk, solid-state drive, optical disk, electronic storage, etc.). The one or more memories may store a computer program product in the form of a set of program code instructions that will be executed by the one or more processors to implement all or part of the steps of a method for exchanging data performed on the UE side according to any embodiment of the embodiments disclosed herein.
[0061] The wireless device may also include a main radio (MR) unit. The MR unit corresponds to the main wireless communication unit of the wireless device, which is used to exchange data with the BS of the RAN using radio signals. The MR unit can implement one or more wireless communication protocols and can be, for example, a 3G, 4G, 5G, NR, WiFi, WiMax, or other transceivers. In a preferred embodiment, the MR unit corresponds to a 5G NR wireless communication unit.
[0062] AI / ML models are data-driven algorithms that apply AI / ML techniques to generate output sets based on input sets.
[0063] AI / ML model delivery is a general term that refers to delivering an AI / ML model from one entity to another in any way. It should be noted that "entity" can mean network node / function (e.g., gNB, LMF, etc.), UE, proprietary server, etc.
[0064] AI / ML model inference is the process of using a trained AI / ML model to produce an output set based on an input set.
[0065] AI / ML model testing is a sub-process of training that evaluates the performance of the final AI / ML model using a different dataset than that used for model training and validation. Unlike AI / ML model validation, testing does not assume subsequent tuning of the model.
[0066] AI / ML model training is the process of training an AI / ML model in a data-driven manner [by learning input / output relationships] and obtaining a trained AI / ML model for inference.
[0067] AI / ML model transfer is the delivery of AI / ML models via the air interface in a manner opaque to 3GPP signaling. The delivery consists of parameters of a model structure known at the receiving end, or a new model with parameters. The delivery may include a complete model or a partial model.
[0068] AI / ML model validation is a subprocess of training that evaluates the quality of an AI / ML model using a different dataset than the one used for model training. This helps in selecting model parameters that generalize to datasets other than those used for model training.
[0069] Data collection is the process by which network nodes, management entities, or user-defined users (UEs) collect data for the purposes of AI / ML model training, data analysis, and inference.
[0070] Federated learning / federated training is a machine learning technique that trains AI / ML models across multiple decentralized edge nodes (e.g., UE, gNB), with each node using local data samples to perform local model training. This technique requires multiple interactions between models but does not require exchanging local data samples.
[0071] Functionality identification is the process / method of identifying AI / ML functionality to achieve a shared understanding between the New Work (NW) and the User Experience (UE). It should be noted that information regarding AI / ML functionality can be shared during functionality identification. The location where AI / ML functionality resides depends on the specific use case and sub-use case.
[0072] Model activation means enabling an AI / ML model for a specific AI / ML-enabled feature.
[0073] Model deactivation means deactivating the AI / ML model for a specific AI / ML enabled feature.
[0074] Model download refers to the transfer of a model from the network to the UE.
[0075] Model identification is the process / method of identifying AI / ML models to achieve a common understanding between the New Work (NW) and the User Experience (UE). The process / method of model identification may or may not be applicable, and information about the AI / ML model can be shared during model identification.
[0076] Model monitoring is the process of monitoring the inference performance of AI / ML models.
[0077] Model parameter update is the process of updating the model parameters.
[0078] Model selection is the process of choosing an AI / ML model from among multiple models used for the same AI / ML enabling feature to be activated. Model selection may or may not be performed concurrently with model activation.
[0079] Model switching deactivates the currently active AI / ML model and activates a different AI / ML model for a specific AI / ML enabling feature.
[0080] Model update is the process of updating the model parameters and / or model structure.
[0081] Model upload is the transfer of the model from the UE to the network.
[0082] Network-side (AI / ML) models are AI / ML models that perform inference entirely at the network level.
[0083] Offline field data is data collected from the field and used for offline training of AI / ML models.
[0084] Offline training is an AI / ML training process in which a model is trained on a collected dataset, and the trained model is later used for inference or delivered for inference. It should be noted that this definition is for guidance only. There may be cases that do not fully meet this definition but can still be classified as offline training according to accepted conventions.
[0085] Online field data refers to data collected from the field and used for online training of AI / ML models.
[0086] Online training is an AI / ML training process in which the model used for inference is trained (typically continuously) as new training samples arrive (nearly) in real time. It should be noted that the concepts of (near) real-time and non-real-time depend on the context and are related to the inference timescale. This definition is for guidance only.
[0087] There may be cases that do not fully meet this definition but can still be classified as online training according to accepted conventions. It should be noted that fine-tuning / retraining can be done via online or offline training. This note can be removed when defining the term "fine-tuning".
[0088] Reinforcement learning (RL) is the process of training an AI / ML model based on inputs (also known as states) and feedback signals (also known as rewards) generated by the model's outputs (also known as actions) in the environment in which the model interacts with it.
[0089] Semi-supervised learning is the process of training a model using a mixture of labeled and unlabeled data.
[0090] Supervised learning is the process of training a model based on the input and its corresponding labels.
[0091] A two-sided (AI / ML) model is a paired AI / ML model on which joint inference is performed, where joint inference includes AI / ML inference performed jointly across the UE and the network, i.e., the first part of the inference is performed by the UE first, and then the remaining part is performed by the gNB, or vice versa.
[0092] The UE-side (AI / ML) model is an AI / ML model that performs inference entirely at the UE.
[0093] Unsupervised learning is the process of training a model without labeled data.
[0094] From a 3GPP perspective, proprietary format models are vendor / equipment-specific, proprietary ML models. They are mutually unrecognizable across vendors and hide model design information from other vendors when shared.
[0095] From the 3GPP perspective, open format models are ML models with a specified format that are mutually identifiable across vendors and allow interoperability. They are mutually identifiable between vendors and do not hide model design information from other vendors when shared.
[0096] The following explanation provides a detailed description of the mechanism for pre-configuring and signaling specific information about online model training by configuring UE behavior sets. AI / ML-based technologies are currently applied in many different applications, and based on observed potential benefits, 3GPP has also begun its technology research to apply them to multiple use cases. The AI / ML lifecycle can be broken down into several phases, such as data collection / preprocessing, model training, model testing / validation, model deployment / update, and model monitoring, each of which is equally important for achieving the target performance of any particular model. One of the challenging issues when applying AI / ML models to any use case or application is managing the AI / ML model lifecycle. This is primarily because data / model drift occurs during model deployment / inference, and it leads to performance degradation of the AI / ML model.
[0097] Fundamentally, dataset statistics change after model deployment, and model inference capabilities are also affected by unseen data as input. Similarly, the statistical properties of the dataset and the relationship between the input and output of the trained model can change with drift. In this context, model training or retraining is one of the key issues in maintaining model performance, as model performance, such as inference and / or training, depends on different model execution environments with varying configuration parameters. To address this issue, collaboration between the UE and gNB is crucial for tracking model performance and reconfiguring models for different environments. Because model performance cannot be consistently maintained due to drift, AI / ML models require post-deployment model monitoring, followed by updated feedback for retraining / updating the model or selecting an alternative model.
[0098] When deploying AI / ML model-enabled wireless communication networks, it is important to consider how to leverage reconfiguration to handle active AI / ML models for wireless devices during operations such as model training, inference, and updates. When UE device ML resources (e.g., computing power, memory size, battery capacity) are limited, there is no guarantee that AI / ML model operations at the UE will maintain any configured target performance, as model performance is susceptible to the availability of associated input datasets (e.g., dataset size / quality). In this approach, a set of models can be configured to be associated with multiple ranges of input dataset measurements (e.g., quality / size), allowing the mapping between models and these ranges to be predetermined, where one or more models can be mapped to associated input dataset measurement ranges. As an example, lower-complexity models can be used with high-quality input datasets, if applicable.
[0099] For signaling, mapping information can be sent via system information or dedicated RRC signaling. Any additional updates to the mapping can be provided to the UE via L1 / L2 or RRC signaling. Where applicable, the model input dataset measurement range can be defined, such as dataset quality and / or dataset size for mapping to the model set. For dataset quality, it refers to the suitability of data for its intended use in training, validating, and deploying AI / ML models for a task, indicating characteristics affecting the reliability and usefulness of the input dataset used for AI / ML model inference or training. For dataset size, it refers to a measure of the amount of data available for model input, characterized by parameters such as the number of data points collected during a measurement period, sample size, number of features, or data volume. Larger dataset sizes typically provide richer information for model inference, allowing for the application of more complex models to enhance performance. The model input dataset measurement range with configured thresholds can be configured based on offline model training by applying multiple models with different levels of complexity to the target model performance. Multiple mapping tables can exist for different ML model functionalities, model applications, and / or model input dataset attributes. Based on dataset measurements on the UE side, the UE can autonomously select the relevant model to use. Alternatively, the network side can determine any specific model to use for the UE based on dataset measurement reports. In the absence of configured mapping information, the UE can adjust the applicable model to the measured input dataset attributes by changing the model complexity level for the configured target model performance. Generally, if applicable, the mapping table can indicate the use of lower complexity models and associated higher quality and / or larger size datasets. Each model can be represented as an index or ID. The quality / size range information of the model input dataset can be represented as a range indicator to be associated with each model. The specific input dataset quality / size range used for mapping can be measured via offline model training between the network and UE sides.
[0100] Figure 1 An exemplary table illustrating the mapping between models and input dataset measurement ranges is shown. In this example, the mapping between models and input dataset measurement ranges can be predetermined, where one or more models can be mapped to an associated input dataset measurement range. By applying multiple models with different complexity levels to the target model performance, the model input dataset measurement range with configured thresholds can be configured based on offline model training. Multiple mapping tables can exist for different ML model functionalities, model applications, and / or model input dataset attributes. If applicable, one or more mapping tables can be shared with multiple UEs or groups of UEs.
[0101] Figure 2An exemplary flowchart for configuring mapping relationship information on the network side is shown. In this example, a set of models to be associated with multiple input dataset measurement (e.g., quality / size) ranges can be configured on the network side. The mapping relationship information can be sent via system information or dedicated RRC signaling. Any additional updates regarding the mapping relationship can be provided to the UE via L1 / L2 or RRC signaling.
[0102] Figure 3 An exemplary flowchart illustrating the handover of an adaptive model on the UE side is shown. In this example, after dataset measurement, the UE can identify any relevant model for handover based on mapping information. Dataset measurement can be performed periodically, aperiodically, or in combination. For example, a lower-complexity model can be used in conjunction with an associated higher-quality and / or larger-size dataset.
[0103] Figure 4 An exemplary signaling flow is shown for a UE's autonomous decision-making regarding an adaptive model. In this example, based on dataset measurements on the UE side, the UE can autonomously select the relevant model to use. As another example, where no configured mapping information is available, the UE can decide to adjust the applicable model to the measured input dataset attributes by changing the model complexity level for the configured target model performance.
[0104] Figure 5 An exemplary signaling flow for network decisions regarding adaptive models is illustrated. In this example, the network side can determine any specific model for the UE based on dataset measurement reports. The quality / size range information of the model input dataset can be represented as a range indicator to be associated with each model, and the specific input dataset quality / size range used for mapping relationships can be measured via offline model training between the network side and the UE side.
Claims
1. A method for adaptive model signaling for a RAN with associated datasets by configuring a model set at a network entity to be associated with a range of multiple input dataset measurements (e.g., quality / size) in a wireless communication system, comprising: • Predetermine the mapping relationship between the model and the measurement range of the dataset; • Define the measurement range of the dataset; • Set a threshold to configure the measurement range of the dataset; • Provide mapping relationship information to the UE.
2. The method according to claim 1, wherein one or more models are capable of mapping to an associated input dataset measurement range.
3. The method according to any one of the preceding claims, wherein mapping relationship information can be transmitted via system information or dedicated RRC signaling.
4. The method according to any one of the preceding claims, wherein any additional updates regarding the mapping relationship can be provided to the UE via L1 / L2 or RRC signaling.
5. The method according to any one of the preceding claims, wherein, if applicable, the measurement range of the model input dataset can be defined, such as the dataset quality and / or dataset size for mapping to the model set.
6. The method according to any one of the preceding claims, wherein the measurement range of the model input dataset with a configured threshold can be configured based on offline model training by applying multiple models with different complexity levels to the target model performance.
7. The method according to any one of the preceding claims, wherein multiple mapping tables can exist for different ML model functionalities, model applications and / or model input dataset attributes.
8. The method according to any one of the preceding claims, wherein the UE is able to autonomously select the relevant model to be used.
9. The method according to any one of the preceding claims, wherein the network side is able to determine any specific model for the UE based on the dataset measurement report.
10. The method according to any one of the preceding claims, wherein the UE is able to determine to adjust the applicable model to the measured input dataset attributes by changing the model complexity level for the configured target model performance.
11. The method according to any one of the preceding claims, wherein, if applicable, the mapping table is capable of indicating that a lower complexity model is used with an associated higher quality and / or larger size dataset.
12. The method according to any one of the preceding claims, wherein the quality / size range information of the model input dataset can be represented as a range indicator to be associated with each model.
13. The method according to any one of the preceding claims, wherein, if applicable, one or more mapping tables can be shared by multiple UEs or groups of UEs.
14. The method according to any one of the preceding claims, wherein dataset measurement can be performed in a periodic or non-periodic manner or in combination.
15. The method according to any one of the preceding claims, wherein the quality / size range of a specific input dataset used for mapping relationships can be measured via offline model training between the network side and the UE side.
16. An apparatus for performing adaptive model signaling for a RAN with associated datasets by configuring a model set to correlate with a range of multiple input dataset measurements (e.g., quality / size) applied to multiple cells in a wireless communication system, the apparatus comprising a wireless transceiver and a processor coupled to a memory storing computer program instructions configured to implement the steps of claims 1 to 15.
17. A user equipment comprising the means according to claim 16.
18. A gNB comprising the apparatus according to claim 16.
19. A wireless communication system for adaptive model signaling for a RAN with associated datasets by configuring a model set to be associated with a range of measurements (e.g., quality / size) of a plurality of input datasets, wherein the wireless communication system includes a user equipment according to claim 17 and a gNB according to claim 18, wherein the user equipment and the gNB each include a processor coupled to a memory storing computer program instructions configured to implement the steps of claims 1 to 16.
Citation Information
Patent Citations
Selection and modification of features used by one or more machine learned models used by an online system
US20160092786A1
Multi-task feature selection neural networks
US20190130247A1
Systems and methods for monitoring of a machine learning model
US20210035021A1
Feature removal framework to streamline machine learning
US20210374562A1
