Signaling for coordinating ML model adaptation for wireless networks

Abstract

A method includes confirming, by a user equipment and a network node, a set of machine learning (ML) function adaptation parameters for the user equipment to perform adaptation of ML functions associated with at least one ML model used by the user equipment to perform radio access network (RAN) related functions. The set of ML function adaptation parameters indicates at least one adaptation period during which the user equipment is to perform ML function adaptation and a valid period for which the set of ML function adaptation parameters is valid. The method also includes performing, by the user equipment, adaptation of the ML function during at least one adaptation period.

Description

Technical Field

[0001] This manual relates to wireless communication. Background Technology

[0002] A communication system can be a facility that enables communication between two or more nodes or devices (such as fixed communication devices or mobile communication devices). Signals can be carried on wired or wireless carriers.

[0003] An example of a cellular communication system is the architecture standardized by the 3rd Generation Partnership Project (3GPP). Recent developments in this field are often referred to as Long Term Evolution (LTE) of Universal Mobile Telecommunications System (UMTS) radio access technology. E-UTRA (Evolved UMTS Terrestrial Radio Access) is the air interface for the 3GPP LTE upgrade path for mobile networks. In LTE, base stations or access points (APs), referred to as enhanced node APs (eNBs), provide radio access within a coverage area or cell. In LTE, mobile devices or mobile stations are referred to as User Equipment (UEs). LTE includes many improvements or developments. All aspects of LTE continue to improve.

[0004] The development of 5G New Radio (NR) is part of the ongoing evolution of mobile broadband to meet the requirements of 5G, similar to the early evolution of 3G and 4G wireless networks. Furthermore, in addition to mobile broadband, 5G targets new and emerging use cases. The goal of 5G is to deliver significant improvements in wireless performance, which can include new levels of data rates, latency, reliability, and security. 5G NR can also be extended to efficiently connect massive Internet of Things (IoT) networks and can provide new types of mission-critical services. For example, ultra-reliable and low-latency communication (URLLC) devices may require high reliability and very low latency. Other wireless networks, such as 6G, are also currently under development. Summary of the Invention

[0005] One method may include: a user equipment (UE) and a network node confirming a set of machine learning (ML) function adaptation parameters for the UE to perform adaptation of ML functions associated with at least one ML model used by the UE to perform radio access network (RAN) related functions, the set of ML function adaptation parameters indicating at least one adaptation period and a valid time period during which the UE performs ML function adaptation, and the set of ML function adaptation parameters being valid for the valid time period; and the UE performing the ML function adaptation during the at least one adaptation period.

[0006] An apparatus may include: at least one processor; and at least one memory, including computer program code; the at least one memory and the computer program code being configured, together with the at least one processor, to enable the apparatus to at least: have a user equipment (UE) confirm a set of machine learning (ML) function adaptation parameters with a network node for the UE to perform adaptation of ML functions associated with at least one ML model used by the UE to perform radio access network (RAN) related functions, the set of ML function adaptation parameters indicating at least one adaptation period and a valid time period during which the UE performs ML function adaptation, the set of ML function adaptation parameters being valid for the valid time period; and have the UE perform adaptation of the ML function during at least one adaptation period.

[0007] Other example embodiments are provided or described for each example method in the example methods, the other example embodiments including: components for performing any of the example methods in the example methods; a non-transitory computer-readable storage medium including instructions stored thereon, the instructions being configured, when executed by at least one processor, to cause a computing system to perform any of the example methods in the example methods; and means including at least one processor and at least one memory including computer program code, the at least one memory and the computer program code being configured, together with the at least one processor, to cause the means to perform at least any of the example methods in the example methods.

[0008] Details of one or more examples of embodiments are set forth in the accompanying drawings and the following description. Other features will be apparent from the specification, drawings, and claims. Attached Figure Description

[0009] Figure 1 This is a block diagram of a wireless network according to an example embodiment.

[0010] Figure 2 This is a flowchart illustrating the operation of a UE according to an example embodiment.

[0011] Figures 3A to 3C This is a diagram showing examples of parameters and / or configurations for different example MAPs.

[0012] Figure 4 This is a diagram illustrating the adaptation cycle and another example of one or more parameters and / or configurations for the MAP.

[0013] Figure 5 This is a diagram illustrating the adaptation cycle and another example of one or more parameters and / or configurations for the MAP.

[0014] Figure 6This is a diagram illustrating the operation of UE 614 and gNB 612 according to an example embodiment, wherein ML model adaptation is performed at the UE, and a set of model adaptation parameters is provided by the UE to the network node or gNB.

[0015] Figure 7 This is a diagram illustrating the operation of UE 614 and gNB 612 according to an example embodiment, wherein ML model adaptation is performed at the UE, and a set of model adaptation parameters is provided to the UE by a network node or gNB.

[0016] Figure 8 It is a block diagram of a wireless station or node (e.g., a network node, such as a gNB), a user node or UE, a relay node or other node. Detailed Implementation

[0017] Figure 1 This is a block diagram of a wireless network 130 according to an example embodiment. Figure 1 In the wireless network 130, user equipment 131, 132, 133, and 135, also referred to as mobile stations (MS) or user equipment (UE), can connect to (and communicate with) a base station (BS) 134, which can also be referred to as an access point (AP), enhanced node B (eNB), gNB, or network node. The terms user equipment and user equipment (UE) are used interchangeably. The BS may also include or be referred to as a RAN (Radio Access Network) node and may include a portion of the BS or a portion of the RAN node, such as (e.g., in the case of splitting the BS or splitting the gNB, such as a centralized unit (CU) and / or a distributed unit (DU)). At least a portion of the functionality of the BS (e.g., an access point (AP), base station (BS), or (e) node B (eNB), gNB, RAN node) can also be performed by any node, server, or host that can be operatively coupled to a transceiver (such as a remote radio head). BS (or AP) 134 provides wireless coverage within cell 136 to user equipments (or UEs) 131, 132, 133, and 135. Although only four user equipments (or UEs) are shown connected to or attached to BS 134, any number of user equipments can be provided. BS 134 is also connected to core network 150 via S1 interface 151. This is just a simplified example of a wireless network, and other examples can be used.

[0018] A base station (e.g., such as BS 134) is an example of a radio access network (RAN) node within a wireless network. A BS (or RAN node) can be or can include (or may alternatively be referred to as) such as an access point (AP), gNB, eNB, or a portion thereof (such as a centralized unit (CU) and / or distributed unit (DU) in the case of a separate BS or separate gNB) or other network node.

[0019] According to illustrative examples, a BS node (e.g., BS, eNB, gNB, CU / DU, etc.) or radio access network (RAN) can be part of a mobile telecommunications system. The RAN (radio access network) can include one or more BS or RAN nodes implementing radio access technologies, for example, to allow one or more UEs to access the network or core network. Thus, for example, the RAN (RAN node, such as BS or gNB) can reside between one or more user equipments or UEs and the core network. According to example embodiments, each RAN node (e.g., BS, eNB, gNB, CU / DU, ...) or BS can provide one or more wireless communication services for one or more UEs or user equipments, for example, to allow UEs to wirelessly access the network via the RAN node. Each RAN node or BS can perform or provide wireless communication services, such as allowing a UE or user equipment to establish a wireless connection to the RAN node, and sending data to one or more UEs and / or receiving data from one or more UEs. For example, after establishing a connection to a UE, the RAN node or network node (e.g., BS, eNB, gNB, CU / DU, etc.) can forward data received from the network or core network to the UE, and / or forward data received from the UE to the network or core network. RAN nodes or network nodes (e.g., BS, eNB, gNB, CU / DU, ...) can perform various other radio functions or services, such as broadcasting control information to the UE (e.g., system information or on-demand system information), paging the UE when data is available to be delivered to the UE, assisting the UE in handover between cells, scheduling uplink data transmission from the UE and downlink data transmission to the UE, sending control information to configure one or more UEs, etc. These are just a few examples of one or more functions that a RAN node or BS can perform.

[0020] User equipment or user node (user terminal, user equipment (UE), mobile terminal, handheld wireless device, etc.) can refer to portable computing devices that operate with or without a subscriber identification module (SIM), including but not limited to the following types of devices: mobile station (MS), mobile phone, cellular phone, smartphone, personal digital assistant (PDA), handheld device, device using a wireless modem (alarm or measuring device, etc.), laptop and / or touchscreen computer, tablet computer, tablet phone, game console, laptop, vehicle, sensor and multimedia device (by way of example), or any other wireless device. It should be understood that user equipment can also be (or may include) a virtually exclusive uplink-only device, an example of which is a camera or video camera that loads image or video clips onto the network. Furthermore, user node can include user equipment (UE), user device, user terminal, mobile terminal, mobile station, mobile node, subscriber equipment, subscriber node, subscriber terminal, or other user node. For example, a user node can be used for wireless communication with one or more network nodes (e.g., gNB, eNB, BS, AP, CU, DU, CU / DU) and / or with one or more other user nodes, regardless of the technology or radio access technology (RAT). In LTE (as an illustrative example), the core network 150 may be referred to as the evolved packet core (EPC), which may include a mobility management entity (MME) that can handle or assist user equipment mobility / handover between BSs, one or more gateways that can forward data and control signals between the BS and a packet data network or the Internet, and other control functions or blocks. As an example, other types of wireless networks (such as 5G (which may be referred to as New Radio (NR)) and / or 6G) may also include a core network.

[0021] Furthermore, the technologies described in this paper can be applied to various types of user equipment or data service types, or to user equipment that can have multiple applications running on it, which can be different data service types. New radio (5G) development can support several different applications or several different data service types, such as, for example: machine-type communication (MTC), enhanced machine-type communication (eMTC), Internet of Things (IoT), and / or narrowband IoT user equipment, enhanced mobile broadband (eMBB), and ultra-reliable and low-latency communication (URLLC). Many of these new 5G (NR) related applications typically require higher performance than previous wireless networks. 6G and other wireless networks may continue to require even greater performance.

[0022] The Internet of Things (IoT) can refer to a growing set of objects that may possess Internet or network connectivity, enabling them to send and receive information from other network devices. For example, many sensor-type applications or devices can monitor physical conditions or states and, for instance, send reports to servers or other network devices when events occur. Machine-type communication (MTC or machine-to-machine communication) can be characterized, for example, through fully automated data generation, exchange, processing, and actuation between intelligent machines, with or without human intervention. Enhanced Mobile Broadband (eMBB) can support data rates significantly higher than those currently available in LTE.

[0023] Ultra-Reliable and Low-Latency Communication (URLLC) is a new type of data service or a new use case that can be supported for new radio (5G) systems. This enables emerging new applications and services such as industrial automation, autonomous driving, vehicle safety, and eHealth services. 3GPP aims to provide reliable connections, corresponding to a block error rate (BLER) of 10⁻⁵ and a U-plane (user / data plane) latency of at most 1 ms, as an illustrative example. Therefore, for example, URLLC user equipment / UEs may require significantly lower block error rates and lower latency than other types of user equipment / UEs (with or without the need for high reliability simultaneously). Thus, for example, a URLLC UE (or URLLC applications on a UE) may require shorter latency compared to an eMBB UE (or an eMBB application running on a UE).

[0024] The technologies described herein can be applied to a variety of wireless technologies or wireless networks, such as 5G (New Radio (NR)), cmWave and / or mmWave band networks, IoT, MTC, eMTC, eMBB, URLLC, 6G, and any other wireless network or wireless technology. These example networks, technologies, or data service types are provided as illustrative examples only.

[0025] According to example embodiments, machine learning (ML) models can be used within a wireless network to perform (or assist in performing) one or more tasks. Typically, one or more nodes within a wireless network (e.g., BS, gNB, eNB, RAN node, user node, UE, user equipment, relay node, or other wireless node) can use or employ ML models, such as neural network models (e.g., which may be referred to as neural networks, artificial intelligence (AI) neural networks, AI neural network models, AI models, machine learning (ML) models or algorithms, models, or other terms), to perform or assist in performing one or more ML-enabled tasks. Other types of models may also be used. ML-enabled tasks can include tasks that can be performed (or assisted in performing) by an ML model, or tasks that an ML model has been trained to perform or assist in performing.

[0026] ML-based algorithms or models can be used to perform and / or assist in performing various radio and / or radio resource management (RRM) and / or RAN-related functions or tasks to improve network performance. Examples include beam prediction (e.g., predicting the optimal beam or optimal beam pair based on measured reference signals) in the UE, antenna panel or beam control, RRM (Radio Resource Measurement) measurements and feedback (Channel State Information (CSI) feedback), link monitoring, transmit power control (TPC), etc. In some cases, ML models can be used to improve the performance of a wireless network in one or more aspects, or measured by one or more performance metrics or standards.

[0027] A model (e.g., a neural network or ML model) can be or can include, for example, a computational model used in machine learning consisting of nodes organized in layers. Nodes, also called artificial neurons, or simply neurons, perform a function on given input to produce some output value. Neural network or ML models typically require training cycles to learn parameters, or weights, used to map inputs to desired outputs. This mapping can occur via a function learned from given data for the problem in question. Therefore, weights are the weights of the mapping function used in a neural network. Each neural network model or ML model can be trained for a specific task.

[0028] To provide an output for a given input, a neural network model or the ML function of an ML model should be trained. This can involve learning appropriate values ​​for a large number of parameters (e.g., weights and / or biases) used for the mapping function (or the ML function of the ML model). For example, parameters can be used to weight and / or adjust terms in the mapping function. This training can be an iterative process in which the values ​​of the weights and / or biases are fine-tuned over many (e.g., tens, hundreds, and / or thousands) training episodes or iterations until the optimal or most accurate values ​​(or weights and / or biases) are reached. In the context of a neural network (neural network model) or ML model, the parameters are typically initialized with random values, and the training optimizer iteratively updates the parameters (e.g., weights) of the neural network to minimize the error in the mapping function. In other words, during each round or step of iterative training, the network updates the values ​​of its parameters such that the values ​​eventually converge to the optimal values.

[0029] As an example, ML models can be trained in a supervised or unsupervised manner. In supervised learning, training examples are provided to the ML model or other machine learning algorithm. Training examples include inputs and the expected or previously observed output. Training examples are also called labeled data because the inputs are labeled with the expected or observed output. In the case of neural networks (which can be a specific case of ML models), the network (or ML model) learns the values ​​of the weights used in the ML model's mapping function or ML function that most frequently lead to the expected output given the training inputs. In unsupervised training, the ML model learns to identify structures or patterns in the provided inputs. In other words, the model identifies implicit relationships in the data. Unsupervised learning is used for many machine learning problems and often requires large amounts of unlabeled data.

[0030] According to the example implementation, depending on the presence of learning "signals" or "feedback" available to the model, ML models can be classified into (or may include) two broad categories: supervised and unsupervised. Thus, for example, within the field of machine learning, there can be two main types of learning or training of models: supervised and unsupervised. The main difference between these two types is that supervised learning uses known or prior knowledge of what the output values ​​of certain data samples should be. Therefore, the goal of supervised learning can be to learn a function that best approximates the relationship between observable inputs and outputs in the data, given a data sample and a desired output. On the other hand, unsupervised learning does not have labeled outputs, so its goal is to infer the natural structure existing within a set of data points.

[0031] Supervised learning: The computer is presented with example inputs and their expected outputs, and the goal can be to learn general rules that map inputs to outputs. Supervised learning can be performed, for example, in the context of classification or regression where the computer or learning algorithm attempts to map inputs to output labels, where the computer or algorithm can map inputs to continuous outputs. Common algorithms in supervised learning can include, for example, logistic regression, Naive Bayes, support vector machines, artificial neural networks, and random forests. In both regression and classification, the goal can include finding specific relationships or structures in the input data that allow us to efficiently produce correct output data. In some example cases, the input signal may be only partially available or limited to specific feedback. Semi-supervised learning: The computer may be given only incomplete training signals; a training set with some (usually many) missing target outputs. Active learning: The computer can only obtain training labels from a limited set of instances (based on a budget) and can also optimize its selection of objects for obtaining labels. These can be presented to the user for labeling when used interactively.

[0032] Unsupervised learning: No labels are given to the learning algorithm, allowing it to find structure on its own from its input. Some example tasks within unsupervised learning can include clustering, representation learning, and density estimation. In these cases, the computer or learning algorithm attempts to learn the inherent structure of the data without using explicitly provided labels. Some common algorithms include k-means clustering, principal component analysis, and autoencoders. Because no labels are provided, there may not be a specific way to compare model performance in most unsupervised learning methods.

[0033] Initially, for example, in at least some cases, the (pre)trained ML model used in the UE can be deployed "offline" during UE production. Additionally, in at least some cases, for optimized performance, the ML model used by the UE may need to be adapted (e.g., adapted, retuned, or adjusted in terms of ML model weights and / or biases) by the UE, for example, based on radio information and conditions available only during operation in NG-RAN (the UE's idle / non-active or connected modes). Therefore, ML functions (e.g., ML model biases and / or weights) can be adjusted or adapted. Adapting ML functions associated with at least one ML model can include, for example, adapting, retraining, retuning, and / or adjusting in terms of ML model weights and / or biases. Thus, ML functions associated with an ML model can be ML model weights and / or biases, or can include ML model weights and / or biases (or can be described by ML model weights and / or biases). Depending on the use case, this adaptation of the UE to ML functions associated with at least one ML model (e.g., adapting and / or adjusting the weights and / or biases of the ML model) can be optimally performed by the UE based on coordination between the UE and the network (e.g., coordination between the UE and the gNB or network nodes). For example, in some cases, UE ML model (re)training (or ML function retraining or adaptation) may require a new configuration of reference signals or other updated signals or inputs from the network to the ML model, which may require assistance from network nodes for data collection (e.g., tagging) or additional radio resources. Therefore, in some cases, proper and / or accurate adaptation of the ML model (or adaptation of ML functions associated with one or more ML models) may require the network node to be aware of the requirements and / or inputs for the UE ML model, which may require collaboration and / or communication between the UE and the network node (or gNB or RAN node). This collaboration between the network node (or gNB) and the UE can also improve the quality of the input data / signals required by the UE for ML model adaptation or (re)training.

[0034] Furthermore, computational limitations in the UE hardware and software platforms (CPU / GPU cores, L1 / L2 memory, MAC / cycle, etc.) (which are unlikely to be explicitly exposed to the 3GPP network) can typically impose restrictions on the amount, frequency, and / or extent to which the deployed ML model at the UE can be (re)trained or adapted on the device (e.g., by the UE) after its deployment and during normal operation of the UE in the radio access network. In addition, retraining or adapting ML functions or adapting the weights or biases of the ML model can consume significant UE resources (e.g., hardware and / or software) and prolong the process over extended periods. Therefore, if the UE will use the ML model to perform (or assist in performing) RAN-related functions (e.g., beam prediction, UE transmit power control, channel state information (CSI) compression, CSI prediction, mobility prediction, access protocol adaptation, or other RAN-related functions) or other RAN-related functions, it may be difficult for the UE to simultaneously use the ML model to perform RAN-related functions (e.g., beam prediction) and perform ML model (or ML function) adaptation or retraining in inference mode due to UE resource (e.g., hardware and / or software) limitations. Furthermore, in some cases, the time required for the UE to perform ML model adaptation (e.g., based on new network conditions or updated reference signals) may be significant. Therefore, in some cases, the UE may need to terminate or suspend the use of the ML model in inference mode to perform RAN-related functions while simultaneously performing ML model adaptation or retraining (adaptation of at least one ML function of the ML model). Therefore, in at least some cases, it may be impractical or difficult for the UE to interrupt or suspend the use of the ML model in inference mode to perform RAN-related functions for such a long period of time, while the UE is performing ML model adaptation, since there may be no significant degrade in UE or network performance.

[0035] Therefore, in some cases, retraining or adapting a UE to perform ML models or ML functions associated with one or more ML models may present one or more challenges or problems. For example, one or more of the following technical challenges may arise: 1) collaboration between the UE and the gNB may be required or at least expected, for example, requesting adjustments to reference signals or other inputs to the ML model for adaptation, and / or otherwise coordinating the UE's ML model adaptation, and there are currently no known techniques or protocols in 3GPP regarding how the UE and network nodes should coordinate for UE ML model adaptation; 2) the inputs or signals used for ML model adaptation should remain constant during ML model adaptation; otherwise, if such inputs to the ML model change during ML model adaptation, this may (at least in some cases) render such ML model adaptation invalid or erroneous. This results in wasted UE resources performing such ML model adaptation; and / or in some cases, UE resource constraints (e.g., UE hardware and / or software) may make it difficult for the UE to perform both ML model inference and ML model adaptation simultaneously. Therefore, as noted, the UE may need to terminate or suspend the use of the ML model while performing ML model adaptation; and / or in some cases, the UE performing ML model adaptation while suspending the use of the ML model at the UE for RAN-related functions for such a significant period of time to complete ML model adaptation / retraining may negatively impact UE performance and / or network performance.

[0036] Therefore, various technologies, solutions, and / or features are provided and / or described according to example embodiments. According to example embodiments, the UE (or user equipment) and network nodes (e.g., gNB) may agree or confirm a set of machine learning (ML) function adaptation parameters for the UE to perform adaptation of ML functions associated with at least one ML model used by the UE to perform radio access network (RAN) related functions (e.g., beam prediction, UE power control, or other RAN related functions). Adaptation of the ML function may include, for example, adapting or retraining the weights and / or biases of the ML model. This set of ML function adaptation parameters may indicate (or may include information indicating the following) at least one adaptation period and valid time period during which the UE (or user equipment) will perform ML function (or ML model) adaptation (e.g., ML model retraining), and this set of ML function adaptation parameters is valid for the valid time period.

[0037] The UE can perform ML function adaptation (or ML model adaptation or retraining) during at least one adaptation period. This allows the UE and network nodes (e.g., gNB) to have a common or agreed-upon set of adaptation parameters, such as effective time periods and at least one adaptation period (or multiple adaptation periods) during which the UE can perform ML function (or ML model) adaptation or retraining. Gaps between adaptation periods (e.g., the presence or absence of gaps, and / or the duration, length, or time period of such gaps between adaptation periods) may or may not be present, configured, or agreed upon as part of a set of ML function adaptation parameters.

[0038] For example, ML function adaptation may include at least one adaptation period (or may be performed therein), and in some cases, ML function adaptation may include multiple adaptation periods (or may be performed therein). According to example embodiments, ML function (or ML model) adaptation or retraining, or a portion thereof, may be performed by the UE during each adaptation period. As noted, in some cases, gaps may be provided between each adaptation period, for example, to allow the UE to perform other functions or tasks, such as allowing the UE to perform RAN-related functions in the inference model during one or more periods of time or intervals between each adaptation period.

[0039] Therefore, for example, UE performance of ML function adaptation may include the UE performing a portion (or adaptation iteration) of ML function adaptation during each of multiple adaptation cycles. Thus, for example, ML function adaptation or retraining may be broken down or divided into smaller chunks of ML function adaptation retraining performed in each adaptation cycle, thereby freeing up (or making available) UE resources, allowing RAN-related functions (associated with one or more ML models) to perform between each adaptation cycle or time period until ML function adaptation is complete. While this may prolong or extend the time period required for ML function adaptation or retraining, it allows UE performance of ML function adaptation with minimal negative impact on UE or network performance (and / or improved UE or network performance) because the operation of the ML model in inference mode is only paused or terminated for a short period, rather than pausing or terminating the use of the ML model in inference mode throughout the entire ML function (or ML model) adaptation or retraining process. For example, allowing the UE to use the ML model in inference mode during gaps or periods between adaptation cycles can allow the UE to perform or update beam prediction more frequently, thereby improving UE or network performance and / or mitigating performance degradation that might otherwise occur if the ML model is not used throughout the entire ML function (or ML model) adaptation period without such gaps (or if the ML model used for UE beam prediction in inference mode is paused or offline). Therefore, the same ML model (or the same ML function associated with at least one ML model) can be adapted during an adaptation cycle and then used in the inference model during gaps between adaptation cycles. Furthermore, for example, an adapted ML function associated with at least one ML model may not be fully or completely adapted or retrained until the end of model adaptation (e.g., executed across multiple adaptation cycles). After the ML function adaptation is complete, the adapted or retrained ML function associated with at least one ML model can be used in inference mode to perform RAN-related functions.

[0040] Furthermore, for example, two different ML models intended for the same RAN-related functions can be adapted and used in inference mode. Thus, an ML function associated with a first ML model (which will be used to perform the first RAN-related function) can be adapted during the adaptation period, while a second ML model (or even a non-ML algorithm) can be used in inference mode to perform the same RAN-related function. Therefore, this alternative approach does not necessarily involve the same ML model adapted during the adaptation period and also used for inference between adaptation periods. These can be two different ML models supporting / associated with the same ML function (e.g., supporting or used to perform the same RAN-related function). The UE can even use non-ML algorithms between adaptation periods to perform or assist the UE in performing RAN-related functions.

[0041] Therefore, for example, the UE can use the ML model in inference mode to perform or assist in performing RAN-related functions between adaptation cycles. Thus, for example, this can allow the UE to divide ML function (or ML model) adaptation into multiple adaptation cycles, and the UE can perform a portion (or partially perform ML function (or ML model) adaptation or retraining (e.g., using the ML model to retrain or adapt the ML model in training mode) during each adaptation cycle (and, for example, pause or terminate the use of the ML model to perform RAN-related functions during a portion of such UE ML function adaptation), while allowing the UE to resume or perform RAN-related functions using the ML model (e.g., in inference mode) during periods between each adaptation cycle, thereby avoiding long periods where the UE does not use the ML model while the UE performs ML function or ML model adaptation or retraining.

[0042] Furthermore, for example, one or more inputs to an ML function or ML model configured by a network node can remain constant (unchanged) within or during an ML function adaptation (or across multiple adaptation cycles of an ML function adaptation). This allows the UE to perform more accurate ML function adaptations because the inputs to the ML model will remain the same or constant during the ML function adaptation, as agreed upon by the UE and gNB. For example, if multiple ML function adaptations exist (e.g., each ML function adaptation includes at least one adaptation cycle or multiple adaptation cycles), the inputs to the ML function associated with at least one ML model can remain constant during or within each ML function adaptation, but the inputs to the ML function associated with at least one ML model can change between each ML function adaptation.

[0043] Note that in the example embodiment, an adaptation of an ML function associated with at least one ML model is performed within a valid time period, where the valid time period includes one or more adaptation cycles. Therefore, for example, the input to the ML function associated with the ML model should remain constant during or within the adaptation period of the ML function, which is performed within the valid time period. Thus, for example, the input to the ML function should remain constant or unchanged within the valid time period.

[0044] In an example embodiment, the set of ML function adaptation parameters may include information indicating the following: a valid time period for which the ML function adaptation parameters are valid; the number of adaptation periods within the valid time period; and the adaptation period duration for each of the at least one adaptation period. Furthermore, for example, the at least one adaptation period may include multiple adaptation periods, and the adaptation period duration for the multiple adaptation periods may include at least one of the following: the adaptation period duration for the multiple adaptation periods within the valid time period, wherein the adaptation period duration is the same for each of the adaptation periods; or the average adaptation period duration for the adaptation periods within the valid time period.

[0045] Furthermore, according to the example embodiment, at least one adaptation period includes multiple adaptation periods, wherein a set of ML function adaptation parameters may include at least one of the following: the number of ML function adaptations; the number of adaptation periods for each ML function adaptation; the duration or average duration of the adaptation period; the time period between each adaptation period in the adaptation period; or the average time period between each adaptation period in the adaptation period.

[0046] Figure 2 This is a flowchart illustrating the operation of a user equipment (or UE) according to an example embodiment. Operation 210 includes the user equipment (e.g., UE) confirming a set of machine learning (ML) function adaptation parameters with a network node (e.g., gNB) for the user equipment to adapt ML functions for performing at least one ML model used by the user equipment to perform radio access network (RAN) related functions. The set of ML function adaptation parameters indicates at least one adaptation period and a valid time period during which the user equipment performs ML function adaptation. The set of ML function adaptation parameters is valid for the valid time period. Operation 220 includes the user equipment performing ML function adaptation during at least one adaptation period.

[0047] about Figure 2 The method includes at least one adaptation cycle comprising multiple adaptation cycles, wherein performing adaptation includes the adaptation of ML functions by the user equipment during the multiple adaptation cycles; the method further includes: the user equipment using at least one ML model in inference mode to perform or assist in the performance of RAN-related functions between adaptation cycles.

[0048] about Figure 2 The method of confirmation may include: the user equipment transmitting a set of ML function adaptation parameters to the network node for performing ML function adaptation; and the user equipment receiving confirmation from the network node that the set of ML function adaptation parameters is acceptable.

[0049] about Figure 2 The method of confirmation may include: the user equipment receiving a set of ML function adaptation parameters from the network node for performing ML function adaptation; and the user equipment transmitting confirmation to the network node that the set of ML function adaptation parameters is acceptable.

[0050] about Figure 2 The method may include a set of ML function adaptation parameters that may include information indicating the following: a valid time period for which the ML function adaptation parameters are valid; the number of adaptation periods within the valid time period; and the duration of the adaptation period for each adaptation period in at least one adaptation period.

[0051] about Figure 2 The method may include at least one adaptation period, which may include multiple adaptation periods, wherein the adaptation period duration for the multiple adaptation periods includes at least one of the following: the adaptation period duration for the multiple adaptation periods within a valid time period, wherein the adaptation period duration is the same for each adaptation period; or the average adaptation period duration for the adaptation periods within a valid time period.

[0052] about Figure 2 The method, wherein at least one adaptation period may include multiple adaptation periods, wherein a set of ML function adaptation parameters includes at least one of the following: the number of ML function adaptations; the number of adaptation periods for each ML function adaptation; the duration or average duration of the adaptation period; the time period between each adaptation period in the adaptation period; or the average time period between each adaptation period in the adaptation period.

[0053] about Figure 2 The method involves ensuring that one or more inputs to the ML function, configured by the network node, remain unchanged during the effective time period.

[0054] about Figure 2 The method for performing ML function adaptation may include: performing multiple ML function adaptations, wherein each ML function adaptation includes multiple adaptation periods; wherein one or more inputs of the ML function, configured by the network node, remain unchanged within each ML function adaptation; and wherein one or more inputs of the ML function, configured by the network node, are changed between two ML function adaptations during a valid time period.

[0055] about Figure 2The method may further include: transmitting a capability response from the user equipment to the network node, the capability response indicating that the user equipment has the capability to perform at least one of the following: ML function adaptation; receiving (or receiving) a set of ML function adaptation parameters by the user equipment; or sending or providing (or sending or providing) a set of ML function adaptation parameters or a set of ML function adaptation parameters, or a set of ML function adaptation parameters, suggested or requested by the user equipment to the network node.

[0056] about Figure 2 The method by which the user equipment performs the adaptation of the ML function during at least one of a plurality of adaptation cycles can be based on at least one of the following: the user equipment receiving a request from a network node to perform the adaptation of the ML function; or the user equipment detecting the need to perform the adaptation of the ML function based on the performance of the relevant RAN function being less than a threshold.

[0057] about Figure 2 The method may further include the user equipment transmitting a request to a network node for resources to be used by the user equipment to perform ML functions during multiple adaptation periods within a valid time period.

[0058] about Figure 2 The method allows the adaptation of ML functions to be performed in part (e.g., each adaptation cycle performs a portion of the adaptation of the ML function or ML model) and / or iteratively during each of multiple adaptation cycles.

[0059] about Figure 2 The method of adapting an ML function performed by a user device may include performing at least one of the following: adapting one or more weights or biases of at least one ML model; adapting at least one ML model; adapting multiple ML models (or at least one ML pattern) that may be associated with an ML function (e.g., where the weights and / or biases of the ML function can be used on or for the ML model); or adapting the architecture and / or model structure of at least one ML model.

[0060] According to the example implementation, the UE does not need to retrain the entire model, but can adjust only some layers (e.g., only some weights and / or biases) or a portion of the model, and may not require, for example, a very large set of new input data, but only some new data. The model state can be retained between update cycles and is not discarded, and adjusting / updating the model based on a portion of the updated data does not change the model architecture. The weights and biases are adjusted. Therefore, for example, some layers, weights, and / or biases can be adjusted or updated during each adaptation cycle of at least one adaptation cycle (or multiple adaptation cycles).

[0061] ML feature adaptation or ML model adaptation: This refers to the process by which an ML feature (e.g., a portion thereof) associated with one or more ML models can be adapted (e.g., tuned, retrained) locally or on another node side in response to a trigger. For example, the trigger can be based on or in response to: periodic adaptation, adaptation in response to a trigger, performance degradation, a request to perform ML model adaptation, or other triggers. Adaptation of an ML feature (or ML model adaptation) associated with one or more ML models can include any one (or a combination thereof): changes due to model retraining (training model parameters with new data and / or changes to model structure), changes in model pruning, changes in model quantization, and / or changes in model structure or architecture. Adaptation of an ML feature (or ML model adaptation) associated with one or more ML models can include adaptation of one or more weights and / or biases and / or other changes to the ML model.

[0062] An ML model adaptation cycle (also referred to as an ML model adaptation period) can be a finite-time iterative step (or period) during the model adaptation process that produces a target ML model update (or ML function update) or intermediate update state as part of the adaptation cycle sequence, ending (or completing) with the updated ML model. The length of an adaptation cycle can be estimated using or can be assumed to use some information about how quickly the required input data becomes available in the network (e.g., periodicity of the SSB / synchronization block reference signal or the CSI-RS / channel state information reference signal configuration). For example, a model adaptation may include one or more adaptation cycles, and each adaptation cycle may be based on some new input data. A complete ML model adaptation is achieved after all ML model adaptations are performed in a sequence of one or more adaptation cycles.

[0063] (ML) Model Adaptation Parameters: ML Function Adaptation Parameters (ML MAP or MAP): A MAP (or ML MAP) can be or includes a set of ML function (or ML model) adaptation parameters that define or indicate various parameters or details of ML function (or ML model) adaptation, such as information indicating one or more of the following: that the UE can or will perform at least one adaptation period of ML function (or ML model) adaptation, that the MAP (a set of ML model adaptation parameters) is a valid valid time period, and other possible parameters. For example, a MAP (which may be referred to as an ML mapping) may indicate or define parameters such as when and for how long the UE can / will update / adjust its ML model. For example, an ML mapping may indicate various update scheduling parameters so that the UE and gNB can coordinate ML model updates and what resources or signal / input updates the UE may need to perform ML model adaptation, and / or allow the UE and gNB to coordinate the time or time period (e.g., gaps between adaptation periods) when the UE can resume or use ML model execution in inference mode to perform RAN-related functions.

[0064] Prior to ML model adaptation, a UE can use or be configured to use one or more ML models (or ML functions associated with one or more ML models) to perform RAN-related functions or Radio Access Network (RAN) functions. After the UE has performed adaptation of the ML functions associated with one or more ML models, the UE can continue to use the adjusted, adapted, or retrained ML functions for the ML models in inference mode to perform or assist in performing RAN-related functions or RAN functions. Prior to ML model adaptation, RAN-related functions may have already been configured and activated, and the underlying ML model (or ML functions associated with one or more ML models) may have been fully trained and deployed, and can now be (further) adapted or retrained at the UE based on the ML MAP (the set of ML functions or ML model adaptation parameters). For example, cooperation agreed upon or confirmed between UEs on the ML MAP can enable or allow coordinated ML model adaptation or retraining at the UE, for example, without the UE explicitly exposing or indicating its vendor-specific ML computing capabilities / capacity to the serving 3GPP network to the gNB.

[0065] Several additional terms or abbreviations are used in this text and / or figures: -MAP or ML MAP: A set of ML model (or a set of ML functions) adaptation parameters. ML MAP may include one or more parameters, such as one or more of the following parameters, or other parameters indicated in this document: - Valid Time Period. The time period indicated by time Tstart to time Tend in the diagram may include one or more model adaptations (e.g., one or more ML model adaptation cycles), where each model adaptation may include one or more adaptation cycles. The valid time period is the period during which the ML MAP (or a set of adaptation parameters) will remain valid. Moreover, the inputs used by the ML function will remain constant or unchanged during the valid time period.

[0066] - Information indicating one or more adaptation periods. This information can be indicated in different ways and / or using different parameters. For example, if the adaptation periods are periodic or have the same duration, the number of adaptation periods (Nac) can be used to indicate one or more adaptation periods within a valid time period. Other parameters and / or techniques can also be used to indicate one or more adaptation periods. For example, alternatively, Nac and Tma or Tma_avg can be provided in the ML Map to indicate the adaptation period.

[0067] - ML Model Adaptation Period. The period for adapting an ML function (associated with one or more ML models) (e.g., it may be a valid period) may include one or more adaptation cycles. Furthermore, the gNB may typically maintain one or more ML model (or ML function) inputs or signals constant or invariant during the ML model adaptation period (e.g., during a valid period) to allow the UE to perform ML function adaptation using a consistent or invariant set of inputs (such as a reference signal configuration). The UE may perform ML function (or ML model) adaptation within or during each adaptation cycle. Gaps or periods may be provided between each adaptation cycle, and during each of these gaps, the UE may, for example, terminate or interrupt ML function (or ML model) adaptation and resume in inference mode using an ML model (or any other algorithm or model, ML or non-ML) to perform or assist in performing RAN-related functions, such as RAN functions. Therefore, the ML model adaptation period may include an interleaved structure that may include alternating or interleaved adaptation cycles (for ML model or ML function adaptation) and gaps (during which the ML model or algorithm may be used in inference mode to perform RAN-related functions, such as RAN-related functions). Therefore, during each adaptation period, the UE can, for example, pause or terminate the use of the ML model in inference mode to allow the UE to use its resources to perform ML model adaptation or training / retraining, and during each gap between adaptation periods, ML model adaptation or training is paused or terminated, and the UE can use the ML model (or other ML models or algorithms) to use or resume performing RAN-related functions. Alternatively, during an adaptation period, the UE can continue to use the ML model in inference mode to perform RAN-related functions while simultaneously adjusting or performing ML model adaptation, such as adjusting the weights and / or biases of the ML model (or adjusting or performing ML function adaptation).

[0068] - Multiple ML model fits within the valid time period. This indicates the number of ML model fits within the valid time period.

[0069] - The number of adaptation cycles within the adaptation period of the Nac-ML model.

[0070] -Tma- The duration or period (or length) of the adaptation period (which can be a periodic period or the same period, or it can be aperiodic). Therefore, the duration of the adaptation period can be periodic (for multiple adaptation periods, the same period or duration, and the gaps between adaptation periods can be the same duration or length) or aperiodic (for example, for multiple adaptation periods, the periods or durations can be different, and the gaps between adaptation periods are not necessarily the same, but can be the same duration or length, or the gaps can be different durations or lengths).

[0071] -Tma_avg - The average duration of the adaptation period, for example, if it is non-periodic.

[0072] Other parameters can be included in the ML MAP.

[0073] Instead of providing the gNB with its sensitive vendor (chipset) specific ML computing (or UE resource) capabilities / capacity, the UE can indicate to the serving NG-RAN node a set of ML adaptation parameters (MAP or ML MAP) for each RAN-related function, including, for example: i) Nac - number or expected adaptation period; ii) estimated average (Tma_avg) or precise duration (Tma) - the duration or period of the model adaptation period for each adaptation period (but not necessarily periodic); and iii) valid time period - for example, the ML MAP is valid for this valid time period, which may be indicated as Tend-Tstart, and may be indicated via time offset and duration, or Tstart and Tend or other information. The ML MAP is (or should remain) consistent or the same (or unchanged) during the valid time period because the UE may need to perform multiple ML model adaptations during this time window, and each ML model adaptation may include multiple adaptation periods, and the UE may need to use the same input / signal or signal configuration during this valid time period, and maintaining the same MAP for this valid time period can allow ML model adaptations (e.g., between multiple ML models) to be consistent.

[0074] According to the illustrative example embodiment, the average (Tma_avg) or exact (Tma) duration of each adaptation cycle can, for example, indicate the total estimated time for: collecting the required data (measurements), preprocessing the input data (if needed), performing ML model retraining / adaptation, and post-processing the ML output (if needed). Therefore, this parameter covers not only ML training / inference latency but also the time spent adapting the entire RAN-related functions.

[0075] We note that when a UE (chipset) vendor's ML-specific server exists (to store ML model IDs, datasets, ML firmware, etc.) and has an interface to a 3GPP network node or core network, this information can also be used in conjunction with the techniques described herein (e.g., an exchanged set of ML model adaptation parameters or MAP) to further refine / adapt the triggering signaling from NG-RAN to the UE.

[0076] For simplicity, as an illustrative example, we describe a scenario where a UE is performing RAN-related functions and the underlying ML model needs to be adapted to achieve a target performance level. For this, we assume the UE needs to collaborate with the serving gNB, for example, to configure and schedule radio resources used by the UE to measure KPIs (Key Performance Indicators) as input data to its ML model. The techniques described in this paper can also be used for some use cases where the roles of the UE and gNB are exchanged. The principles described below can be extended to situations where RAN-related functions, for example, use two-sided ML models in both the UE and gNB. Furthermore, the proposed techniques can also be used in UE-to-UE communication scenarios (sidelink, ProSe, V2V, UE-to-UE communication, also known as device-to-device communication), where the involved UEs need to establish a certain level of collaboration.

[0077] Finally, feedback from the UE can include the scheduling of ML function adaptation, the history of model adaptation, or the instantaneous status of model adaptation. This information can be used to avoid changes to the network configuration during the adaptation phase, or to optimize the MAP, and / or, if changes are necessary during the adaptation phase, to instruct the UE to use the default AI / ML or a fallback model. Using this signaling, the UE can also indicate that immediate ML model adaptation is required outside of the predefined scheduling defined by the MAP.

[0078] exist Figures 3A to 3C and Figure 4 The diagram illustrates examples of the adaptation period (also known as the model adaptation period) and some sample ML MAP configuration parameters to be agreed upon via signal transmission (e.g., from the UE to the gNB) and / or between the UE and the gNB when ML adaptation is performed in the UE. For example, these parameters can be estimated by ML control firmware specific to hardware and / or software implementations that cooperate with 3GPP UE functions. Figures 3A to 3C and Figure 4 The timelines depicted are for visualization purposes only and illustrate possible outcomes of the adaptation cycle sequence. During the "white" intervals or gaps between adaptation cycles in the UE, it is assumed that the same ML model, a different ML model, or no ML model are used (or can be used by the UE) in inference operation mode to perform RAN-related functions (e.g., RAN-related functions). The reason for extending the adaptation cycle in time is, for example, to allow new or old models to adapt to time-varying radio conditions, and / or to utilize computational resources to adapt the ML model during "inference idle" times.

[0079] Figures 3A to 3C This is a diagram illustrating the parameters and configuration examples for different sample ML MAP configurations. Figures 3A to 3CFor example, the UE's internal ML implementation can trigger the start of each adaptation (e.g., based on detected changes in radio conditions, detected beam changes, etc.). In these cases, the expected number of adaptations (Nac) and its average duration (Tma_avg), or the exact duration (Tma) for an adaptation cycle, can be estimated by the UE up to the time interval Tstart to Tend (which can be an effective time period).

[0080] exist Figures 3A to 3C In this context, the effective time period is defined as the period from Tstart to Tend. Nac (the number of adaptation cycles within the effective time period) equals 5, meaning that there are 5 adaptation cycles within the effective time period.

[0081] exist Figure 3A The diagram illustrates non-periodic adaptation cycles 1, 2, 3, 4, and 5 (each adaptation cycle having a different time interval or length between its start), and also shows that some of these five adaptation cycles have different durations (and therefore, some adaptation cycles may have different gaps or time intervals between them). For example, the duration of adaptation cycle 4 is longer than the durations of adaptation cycles 1, 2, 3, and 5. Adaptation cycle 3 has... Figure 3A The shortest duration among the 5 adaptation cycles. While the 5 adaptation cycles have different durations, they can all have an average duration of Tma_avg. Note that a gap is provided between each pair of adjacent adaptation cycles. For example, in... Figure 3AA gap 312 is provided between adaptation period 3 and adaptation period 4. Similarly, gaps (or time periods) are provided between adaptation period 1 and adaptation period 2, between adaptation period 2 and adaptation period 3, and between adaptation period 4 and adaptation period 5. Therefore, ML adaptation (or a portion of ML model adaptation or retraining) can be performed during each adaptation period from adaptation period 1 to adaptation period 5 (e.g., the ML model operates in training mode during these adaptation periods). During gaps (such as gap 312), ML model adaptation can be terminated or paused to allow the UE to apply its resources and use the ML model to perform RAN-related functions previously trained on the ML model (e.g., using the ML model in inference mode during these gaps between adaptation periods). Therefore, an interleaved structure or arrangement of adaptation cycles alternating with gaps can be provided, wherein the UE can perform (or partially perform) a portion of the ML model adaptation during each or one or more adaptation cycles (e.g., when the ML model is used to train the ML model in training mode), and then the ML model can be used in inference mode during each gap (e.g., such as gap 312) to perform RAN-related functions or RAN-related functions, such as beam prediction (as an example). In this illustrative example, for example, ML model adaptation or retraining can be completed after adaptation cycle 5 is completed.

[0082] exist Figure 3B In this context, Nac (the number of adaptation periods within the effective time frame) is also 5, and the adaptation periods are aperiodic but have the same duration (the same Tma). Figure 3C In addition, there are five adaptation cycles (Nac=5), which are periodic (the time interval between the beginning of each consecutive adaptation cycle is the same), and the five adaptation cycles have the same duration (Tma for...). Figure 3C (All five adaptation cycles shown are the same). For example, in Figure 3B and Figure 3C as well as Figure 4 Similarly, gaps are provided between the adaptation cycles shown.

[0083] Therefore, for example, Figure 3A i) shows an adapted period with a non-periodicity of average period duration; Figure 3B ii) shows an adaptation period with a fixed adaptation period duration that is non-periodic, and Figure 3C iii) shows a periodic adaptation cycle with a fixed adaptation cycle duration. For example, it can be assumed that one model adaptation is performed per adaptation cycle (Nma = 1). During the time interval (or gap) between adaptation cycles, the ML model is used in inference operation mode as part of the normal operation of RAN-related functions (e.g., to perform beam prediction or other RAN-related functions).

[0084] exist Figures 3A to 3C and Figure 4 The example illustrates how a UEML implementation can leverage available model adaptation cycles. And... Figures 3A to 3C In this process, each adaptation cycle generates a target update for the model. Figure 4 In this model, each model adaptation takes two cycles. Therefore, an additional parameter that can be included in the MAP is the number of cycles per model adaptation (Ncpma), or alternatively, the number of model adaptations (Nma), not shown. In this case, by utilizing the multiple cycles required for model updates, another ML model (or a version of the model) can be used to provide the desired inference results for the same RAN-related functions during the "white" intervals between adaptation Ncpma cycles.

[0085] Figure 4 This is a diagram illustrating the adaptation cycle and another example of one or more parameters and / or configurations for the ML MAP. Figure 4 In the example shown, there are 3 ML model adaptations and 6 (total) adaptation periods (Nac = 6). Furthermore, Ncpma (the number of adaptation periods per model adaptation) = 2, meaning that two adaptation periods are used to perform each of the three ML model adaptations. Adaptation periods 412 and 414 are provided for the first ML model adaptation; adaptation periods 416 and 418 are provided for the second ML model adaptation; and adaptation periods 420 and 422 are provided for the third ML model adaptation. For example, these ML model adaptations could be performed on three different ML models, or multiple ML adaptations of the same ML model could be performed.

[0086] For example, a complete ML model adaptation or retraining may require several adaptation cycles within these cycles until the model inference and / or KPIs (Key Performance Indicators, or the performance of ML-enabled features using the ML model) are sufficiently accurate. The UE can indicate to the gNB when model adaptation or retraining is complete and can / will be used in the UE.

[0087] Figure 5 This is another example of a diagram illustrating the adaptation cycle and one or more parameters and / or configurations for the ML MAP. Figure 5 Another example of an ML model adaptation cycle configuration is shown, where the ML model is divided into model segments (1-3) and each segment is adapted individually. Each segment of the ML model can have a different update rate (segment #1 every cycle, segment #2 every two cycles, and segment #3 every four cycles).

[0088] According to an example embodiment, the UE and gNB may confirm and / or agree on ML function adaptation parameters (e.g., which may include one or more adaptation cycles during which ML model adaptation can be performed) and a set of ML function (or ML model) adaptation parameters that constitute a valid effective time period. The effective time period may be shown in some diagrams as a period from Tstart to Tend. Furthermore, for example, as part of the confirmed (negotiated or agreed) set of ML function adaptation (or ML model adaptation) parameters, the gNB and / or UE may agree that: 1) the gNB will not change the network configuration (e.g., reference signal configuration) of any input parameters for the ML model during the adaptation time period and / or during the effective time period; 2) the gNB will generally not require or request the UE to use the ML model to perform functions during the effective time period (or at least during the adaptation cycles within the effective time period), for example, when the ML model is undergoing training and the ML model is invalid, but in some cases, the UE may have the ability to perform both training and inference in parallel. The gNB does not schedule or request the UE to perform any operations for the UE to perform using the ML model. The gNB may use UL / DL (uplink and / or downlink) traffic to schedule the UE, but the radio configuration used by the ML function (as input data for ML model adaptation) (such as reference signal configuration) will not be changed by the gNB during the effective period, and / or will not be changed during ML function adaptation (model adaptation period) which may include one or more adaptation cycles.

[0089] Figure 6 This is a diagram illustrating the operation of UE 614 and gNB 612 according to an example embodiment, wherein ML model adaptation is performed at the UE, and a set of model adaptation parameters is provided by the UE to the network node or gNB. For example, Figure 6 The operations shown can correspond to or be associated with Figures 3A to 3C .exist Figure 6 At step 1, the UE / gNB performs capability exchange, and UE 614 indicates its ML model capabilities, such as inference and training / adaptation. For example, UE 614 may indicate that it performs one or more of the following capabilities: ML function adaptation; receiving a set of ML function adaptation parameters by UE 614; or sending or providing a set of ML function adaptation parameters or a suggested or requested set of ML function adaptation parameters to the gNB / network node.

[0090] exist Figure 6 At step 2, gNB 612 can acknowledge receiving these capability indications and enable UE 614 to perform ML model adaptation and / or enable UE to send or receive a set of ML model adaptation parameters.

[0091] exist Figure 6At step 3, UE 614 and / or gNB 612 can monitor the performance of RAN-related functions, such as those for specific RAN-related functions (e.g., beam prediction or other functions). UE 614 or gNB 612 can detect that the performance of the RAN-related function is below a threshold. In this example, UE 614 detects that the performance of the RAN-related function is below the threshold or not performing to the required level, therefore, for example, the ML model (e.g., the bias and / or weights of at least one ML model or ML function) needs to be adapted or retrained based on updated data, conditions, or updated signals. For example, at step 3, gNB or the network can monitor the performance of the RAN-related function; gNB 612 can notify UE 614 to take corrective action if the performance degrades below a level or outside a boundary. For example, gNB can determine that the beam prediction from the UE no longer matches the actual beam, therefore the beam prediction ML model should be updated, and gNB 612 notifies UE 614 to update the ML model. The gNB 612 performs performance monitoring when the UE uses the ML model in inference mode, but not when the ML model is in the training / adaptation cycle. Therefore, at step 4, the UE can determine that the ML model needs adaptation.

[0092] exist Figure 6 In step 5, the UE determines or estimates a set of ML function (or ML model) adaptation parameters (MAP), such as Nac (the number of adaptation cycles in the effective time period is 10 adaptation cycles), default Ncpma=1 (which may or may not be transmitted via signaling), the effective time period can be indicated as the time period between Tend and Tstart=30 s, and Tma_avg (the average duration of the adaptation cycle) is 100 ms.

[0093] In step 6, UE 614 sends (e.g., a suggested) MAP (or a set of ML model adaptation parameters) to gNB 612 via an RRC message. Therefore, in steps 5 and 6: UE 614 estimates the required adaptation period, determines the MAP, and transmits the MAP to gNB 612 via a signaling message.

[0094] exist Figure 6 At step 7, UE 614 can also request resource allocation for ML model adaptation, such as requesting specific reference signal configuration. The UE requests resources for model adaptation. If the UE needs certain reference signals, such as CSI-RS signals on a specific beam or any signals specific to the ML model, the UE can make this request to the gNB. Therefore, UE 614 will receive the required input signals during the valid time period, and thus the UE can perform ML model adaptation.

[0095] exist Figure 6At step 8, gNB 612 and / or UE 614 can acknowledge or approve the (recommended) MAP. Therefore, at step 8, if the MAP configuration requested by the UE is acceptable, gNB 612 responds with an ACK (acknowledgment), or if the MAP configuration is unacceptable, gNB 612 responds with a NACK. In cases where there is no guarantee that the UE's configuration will remain unchanged during Tstart-Tend, a NACK is generated by the gNB; for example, this can occur under conditions such as high traffic load or MAP signaling received from too many UEs (step 5).

[0096] exist Figure 6 At step 9, if the UE receives an ACK in step 8, then when the start time instance (e.g., the number of SFNs (system frames) or UTC time) is reached, the UE begins the planned adaptation period (with the parameters determined in step 5) and performs ML model adaptation. If a NACK is received in step 8, the UE may begin its model adaptation at its own risk; that is, the gNB cannot guarantee that the configuration will not change during the Tstart-Tend period if it decides to begin such a process. Therefore, at step 9, the UE 614 performs ML model adaptation or retraining, or partial adaptation or retraining, at one or more adaptation periods during the effective time period, where the ML model is in training mode, depending on its specific implementation-specific algorithm and ML platform capabilities, based on the acknowledged or agreed MAP.

[0097] exist Figure 6 At step 10, the gNB will use (or provide) the agreed configuration and / or radio resources, and / or provide the required inputs from Tstart to Tend (e.g., transmit the required reference signals) to allow the UE to perform ML model adaptation based on (e.g., according to) the MAP. These resources, configurations, and / or input signals provided by gNB 612 should remain constant during the valid time period or at least during the ML model adaptation period (and multiple ML model adaptations may exist in each valid time period). Furthermore, for example, the gNB may not request the UE 614 to perform RAN-related functions during the valid time period and / or at least during the adaptation cycle within the valid time period.

[0098] exist Figure 6 At step 11, UE 614 can use the ML model in inference mode to perform RAN-related functions during the gaps between consecutive adaptation cycles (e.g., depending on the MAP configuration).

[0099] exist Figure 6At step 12, when the Tend time instance (e.g., the number of system frames (SFN) or UTC time) is reached, the UE stops, terminates or deactivates the planned ML model adaptation during the adaptation period.

[0100] exist Figure 6 At step 13, the UE can resume using the adapted or retrained (updated) ML model in inference mode to perform RAN-related functions (similar to step 3). Then the UE and / or gNB can monitor the execution of RAN-related functions to determine whether the UE needs another ML model adapted or retrained.

[0101] Figure 7 This is a diagram illustrating the operation of UE 614 and gNB 612 according to an example embodiment, wherein ML model adaptation is performed at the UE, and the set of model adaptation parameters is provided to the UE by a network node or gNB, and the gNB can detect the UE's need to perform ML model adaptation. Figure 7 The operation shown is very similar to Figure 6 The operations shown are identical to those noted here. Figure 7 In step 4, gNB 612 determines that the ML model needs to be adapted or retrained (e.g., gNB can detect beam prediction performance below a threshold level).

[0102] exist Figure 7 In step 5, gNB determines or estimates a set of ML function (or ML model) adaptation parameters (MAP), such as, for example, Nac (the number of adaptation cycles in the effective time period is 10 adaptation cycles), the default Ncpma = 1 (which may or may not be transmitted via signaling), the effective time period can be indicated as the period between Tend and Tstart = 30 s, and Tma_avg (the average duration of the adaptation cycle) is 100 ms. Figure 7 At step 6, gNB 612 sends (e.g., a suggested) MAP (or a set of ML model adaptation parameters) to UE 614 via an RRC message. Therefore, in Figure 7 In steps 5 and 6: gNB 612 estimates the required adaptation period, determines the MAP, and sends the MAP to UE 614 via signaling message.

[0103] Figure 6 and Figure 7The operation at step 4, “Determine if the UE ML model needs adaptation,” can be used as a “trigger” to activate the ML adaptation process and signaling. Examples of step 4 could include mechanisms such as those for data and / or model drift detection. Furthermore, the implementation of step 4 can depend, for example, on the details of the ML model used in the UE (including any proprietary hardware and / or software solutions) and also on how the ML model was trained and how much it can “generalize” (performing with acceptable performance under various input conditions). In this example, radio conditions also consider the number of beams that the UE can detect during its movement and the signal strength of the FR2 radio beams. Radio beams in FR2 can be easily blocked by relatively small physical obstacles such as people, cars, and trees. Therefore, in this given example, we assume that the ML model was initially trained under stationary radio channel conditions. To maintain beam prediction performance, the model needs to be retrained / adapted when the radio conditions around the UE change due to the UE's slow movement.

[0104] Figure 6 and Figure 7 The operation at step 5, “Estimate MAP,” can depend on the type of model adaptation determined in step 4 and the specific use case when using the ML model. Similar to step 4, step 5 can also depend on an implementation-specific solution. In a given example for UE beam prediction, based on the frequency band used (FR2 carrier, 25 GHz) and the approximate UE movement speed (or, alternatively, the number of radio beam changes detected per time unit), the radio environment around the UE is expected to change significantly after approximately 200 wavelengths at 25 GHz. Because the long-term (time-range) movement direction and speed of the UE is impossible to estimate for typical pedestrian scenarios, the UE algorithm in this example can set the maximum time window for adaptation to 30 seconds, after which both steps 4 and 5 may need to be re-evaluated.

[0105] Various additional illustrative examples and variations will now be briefly described.

[0106] Example use case (numerical values ​​are within realistic ranges, but not the only possible ones): Assume an ML-based function is used for UE beam management (e.g., beam prediction, a contemporary use case in 3GPP Release 18). The UE first estimates its mobility state, for example, based on the number of previous beam changes and the corresponding recorded RSRP level. As an example, for a pedestrian UE moving in a dense city at a speed of 3 km / h and in an FR2 radio environment, the mobility state estimated by the UE is “medium,” and therefore the UE determines that the ML model may require Nac = 10 adaptations during the next 30 seconds (Tend-Tstart), corresponding to one adaptation after every 2.5 m distance (approximately 200 wavelengths at 25 GHz). Furthermore, it is a reasonable assumption that these adaptations can only run when there is no / low user plane traffic due to UE power consumption, and are therefore non-periodic. The UE estimates the duration of a model adaptation cycle, averaging Tma_avg = 100 ms (internal algorithm, based on implementation-specific hardware / software), as the total time required for data collection, data preprocessing, retraining of some NN layers, potential test inference, and output post-processing. Depending on the UE hardware / software platform, the retraining and potential inference steps can take much less time than Tma_avg; however, the data collection and data preparation steps depend on the type of radio measurements and signals used, and therefore the cycle duration can easily be 100 times longer than typical ML model inference time. According to this illustrative example, using the timing values ​​from the example above, the UE notifies the service gNB that during the next Tend-Tstart = 30 seconds for Nac = 10 cycles, where each adaptation cycle has a length of Tma_avg = 100 ms, the UE cannot send / receive data, or it can receive / send data but cannot perform any RRC reconfiguration.

[0107] Additional Implementation: The MAP can be defined by the network (e.g., via the gNB's ML orchestrator or ML training function) and transmitted to the UE as a strategy to allow for the mapping of detected radio condition changes and the corresponding required ML model adaptations for a given RAN-related functions. Alternatively, when the MAP is defined and provided by the network or gNB (e.g., see...), Figure 7 The network can also trigger UEML adaptation cycles individually or in whole within the Tstart to Tend time window. An adaptation cycle begins when predefined conditions are validated (e.g., performance degradation), and the optimal ML model can be selected for adaptation based on performance feedback or relevant output observations. Aperiodic MAP cycles can be defined using a specified minimum time (Tmaintv_min) between consecutive adaptations and a maximum duration (Tma_max) for the adaptation cycle. Combined with any other implementation (aperiodic or periodic adaptation cycles), Figures 3A to 3Cand Figure 4 (For example), ML model adaptation can be decomposed / divided into different model parts (e.g., groups of neural network (NN) layers), and such parts will be adapted / trained separately at different rates (see [example]). Figure 5 ).

[0108] In conjunction with any other implementation (aperiodic or periodic adaptation cycle), Figures 3A to 3C and Figure 4 (For example, the ML model adaptation cycle can be paused and resumed during the negotiated cycle / time window or alternatively upon request from the gNB before triggering a new Tstart-Tend adaptation window.)

[0109] Example use cases and possible implementations: For example, a deep neural network (DNN or ML model) with L layers can have segment 1: layer 1->K and segment 2: layer K+1->L. Segment 1 can be rate-adapted R1 (while keeping segment 2 constant), and vice versa. Figure 3 shows an example with periodic adaptation cycles and 3 model segments, adapted at different rates: segment #1 every cycle, segment #2 every two cycles, and segment #3 every four cycles.

[0110] Figure 5 The arrangement of adaptation cycles can also be used to train different ML models (for the same RAN-related functions) instead of different segments of the same ML model (i.e., segment #1 -> model #1, segment #2 -> model #2, segment #3 -> model #3). In this case, if the model update of one of the models is completed before Tend (model #3), the UE can also switch to using it for inference during the "white" time interval (or gap).

[0111] Additional embodiments for processing MAP: MAP configuration can be (partially or completely) part of the initial UE ML capability exchange.

[0112] In D2D (device-to-device, UE-to-UE, or sidelink communication) scenarios, the UE receives MAP from neighboring UEs via sidelink (SL), ProSe (proximity service), D2D, and / or UE-to-UE communication protocols to avoid the UE searching for optimal parameters.

[0113] The gNB can configure the UE to provide event-triggered or periodic reports on the status / result of the adaptation period between Tstart and Tend. The UE performing model adaptation is configured by the serving gNB to, for example, interrupt the signaling transmission ('interruption of adaptation period') using the initially configured MAP in the event of failure (e.g., failure to meet the target accuracy after adaptation – prior art); this signaling can use RRC or MAC signaling channels. For example, the need for 'interruption of adaptation period' can be based on changing radio conditions such as the UE moving from indoors to outdoors or from low speed to high speed. Alternatively or additionally, a simple status indication (pass / fail) can be transmitted via signaling after the configured number of periods or when Tend expires; this signaling can use RRC or MAC signaling channels. Additionally, the UE can transmit an extended status indication including context information via signaling; this signaling can use RRC or MAC signaling channels. Context information may include: conditions that trigger an 'interruption of the adaptation cycle' at the UE; SFN (system frame number or other timing indication) when an adaptation failure has been detected; additional ML model accuracy / performance metrics (if / when available); statistics of the input data used for adaptation; and / or statistics of the output after or during adaptation (if / when available).

[0114] The 'Interruption of Adaptation Cycle' signal may include additional information such as the reason for the interruption (see above) and / or what ML model was used after the signal was sent (another / older ML model or a non-ML model, based on fallback functionality). Alternatively or additionally, the 'Interruption of Adaptation Cycle' signal is sent by the serving gNB to the UE performing model adaptation. For example, the need for 'Interruption of Adaptation Cycle' may be based on a change in the required NW configuration that is incompatible with the current configuration the UE uses to update the model and cannot be delayed until Tend.

[0115] Example use cases: Security Attack Detection: The ML model runs at the UE to detect / predict the occurrence of security attacks (such as malicious base stations). In practice, the UE uses its measurement reports (typically for other processes such as handover) to detect security attacks before harmful consequences are observed. However, primarily due to environmental variations, the ML model needs to be updated within a limited time window (time range) to ensure optimal performance in terms of detection accuracy. The UE determines that ML model adaptation requires Nac = 10 non-periodic adaptation cycles (e.g., when resources are available) over the next 30 seconds (Tend-Tstart), where the average cycle duration Tma_avg = 100 ms.

[0116] Beam Prediction: This occurs when an ML model is employed at the UE to predict the next optimal beam and trigger a beam handover / transfer towards that beam without explicitly measuring it. For example, the UE might first estimate its mobility state based, for instance, on the number of previous beam / cell changes and the corresponding recorded RSRP level. For a pedestrian UE moving in a dense urban, FR2 radio environment, the UE might estimate its mobility state as "moderate," and therefore determine that the ML model requires Nac = 10 periodic adaptation cycles over the next 30 seconds (Tend-Tstart), with each cycle lasting Tma = 100 ms.

[0117] CSI Prediction / Compression: This refers to the use of an ML model at the UE to predict CSI for a given time range without explicitly measuring it. For example, the UE initially estimates its mobility state based, for instance, on the number of previous beam / cell changes. If a significant change in mobility, for example, from low to high, is detected, the ML model used for CSI prediction needs to be adapted, and thus the UE determines that the ML model requires Nac = 10 periodic adaptation periods over the next 1 second (Tend-Tstart), with each periodic adaptation period having a period duration Tma = 10 ms, to retune the model for the new channel conditions.

[0118] Location: When using the ML model at the UE or gNB to predict the LOS / NLOS state of the UE's channel conditions. For example, the UE ML algorithm estimates (predicts) the LOS / NLOS state based on CSI measurements and channel impulse response (CIR) estimation. Due to the detection of changing radio conditions, based on S(I)NR and fast fading conditions (mobility), the UE determines that the ML model requires Nac = 10 aperiodic adaptations during the next 10 seconds (Tend-Tstart), each aperiodic adaptation having a periodic duration Tma = 200 ms.

[0119] The numerical values ​​given for the examples and illustrative use cases described in this article are provided purely for illustrative or explanatory purposes, and other numerical values ​​may be used.

[0120] Now some examples will be described.

[0121] Example 1. A method includes: a user equipment (UE) and a network node confirming a set of machine learning (ML) function adaptation parameters for the UE to perform adaptation of ML functions associated with at least one ML model used by the UE to perform radio access network (RAN) related functions, the set of ML function adaptation parameters indicating at least one adaptation period and a valid time period during which the UE performs ML function adaptation, and the set of ML function adaptation parameters being valid for the valid time period; and the UE performing the adaptation of the ML function during the at least one adaptation period.

[0122] Example 2. According to the method of Example 1, wherein at least one adaptation period includes multiple adaptation periods, wherein performing adaptation includes the user equipment performing adaptation of ML functions during the multiple adaptation periods; the method further includes the user equipment using at least one ML model in inference mode to perform or assist in the performance of RAN-related functions between adaptation periods.

[0123] Example 3. The method according to any one of Examples 1 to 2, wherein the confirmation includes: the user equipment transmitting a set of ML function adaptation parameters to the network node for performing ML function adaptation; and the user equipment receiving confirmation from the network node that the set of ML function adaptation parameters is acceptable.

[0124] Example 4. The method according to any one of Examples 1 to 2, wherein the confirmation includes: receiving a set of ML function adaptation parameters from a network node by the user equipment for performing ML function adaptation; and transmitting confirmation to the network node that the set of ML function adaptation parameters is acceptable.

[0125] Example 5. The method according to any one of Examples 1 to 4, wherein a set of ML function adaptation parameters includes information indicating the following: a valid time period for which the ML function adaptation parameters are valid; the number of adaptation periods within the valid time period; and the adaptation period duration for each adaptation period in at least one adaptation period.

[0126] Example 6. The method according to Example 5, wherein at least one adaptation period includes multiple adaptation periods, wherein the adaptation period duration for the multiple adaptation periods includes at least one of the following: the adaptation period duration for the multiple adaptation periods within a valid time period, wherein the adaptation period duration is the same for each adaptation period; or the average adaptation period duration for the adaptation periods within a valid time period.

[0127] Example 7. The method according to any one of Examples 1 to 4, wherein at least one adaptation period includes multiple adaptation periods, wherein a set of ML function adaptation parameters includes at least one of the following: the number of ML function adaptations; the number of adaptation periods for each ML function adaptation; the duration or average duration of the adaptation period; the time period between each adaptation period in the adaptation period; or the average time period between each adaptation period in the adaptation period.

[0128] Example 8. The method according to any one of Examples 1 to 7, wherein one or more inputs of the ML function, configured by the network node, remain unchanged during the effective time period.

[0129] Example 9. A method according to any one of Examples 1 to 7: wherein performing an adaptation of an ML function includes: performing multiple ML function adaptations, wherein each ML function adaptation includes multiple adaptation periods; wherein one or more inputs of the ML function, configured by the network node, remain unchanged within each ML function adaptation; and wherein one or more inputs of the ML function, configured by the network node, are changed between two ML function adaptations during a valid time period.

[0130] Example 10. The method according to any one of Examples 1 to 9 further includes: the user equipment transmitting a capability response to the network node indicating that the user equipment has the capability to perform at least one of the following: ML function or ML model adaptation; the user equipment receiving a set of ML function adaptation parameters; or the user equipment sending or providing a set of ML function adaptation parameters or a suggested or requested set of ML function adaptation parameters to the network node.

[0131] Example 11. The method according to any one of Examples 1 to 10, wherein the adaptation of the ML function by the user equipment during at least one of a plurality of adaptation cycles is performed based on at least one of the following: the user equipment receiving a request from a network node to perform the adaptation of the ML function; or the user equipment detecting the need to perform the adaptation of the ML function based on the performance of the RAN-related function being less than a threshold.

[0132] Example 12. The method according to any one of Examples 1 to 11 further includes: transmitting a request from the user equipment to a network node for resources to be used by the user equipment during multiple adaptation periods within a valid time period to perform adaptation of the ML function.

[0133] Example 13. The method according to any one of Examples 1 to 12, wherein the adaptation of the ML function is performed in part iteratively during each of a plurality of adaptation cycles.

[0134] Example 14. The method according to any one of Examples 1 to 13, wherein the adaptation of ML functions performed by the user equipment includes performing at least one of the following: adapting one or more weights or biases of at least one ML model; adapting at least one ML model; adapting multiple ML models; or adapting the architecture and / or model structure of at least one ML model.

[0135] Example 15. An apparatus comprising: at least one processor; and at least one memory including computer program code; the at least one memory and the computer program code being configured, together with the at least one processor, to cause the apparatus to perform at least the method of any one of Examples 1 to 14.

[0136] Example 16. A non-transitory computer-readable storage medium including instructions stored thereon, which, when executed by at least one processor, are configured to cause a computing system to perform any of the methods of Examples 1 to 14.

[0137] Example 17. An apparatus comprising components for performing the methods of any of Examples 1 to 14.

[0138] Example 18. An apparatus includes: at least one processor; and at least one memory, including computer program code; the at least one memory and the computer program code are configured, together with the at least one processor, to enable the apparatus to at least: have a user equipment (UE) confirm a set of machine learning (ML) function adaptation parameters with a network node for the UE to perform adaptation of ML functions associated with at least one ML model used by the UE to perform radio access network (RAN) related functions, the set of ML function adaptation parameters indicating at least one adaptation period and a valid time period, during which the UE performs ML function adaptation, the set of ML function adaptation parameters being valid for the valid time period; and have the UE perform adaptation of the ML function during at least one adaptation period.

[0139] Example 19. The apparatus according to Example 18, wherein at least one adaptation cycle comprises a plurality of adaptation cycles, wherein performing adaptation includes adaptation of ML functions performed by the user equipment during the plurality of adaptation cycles; at least one processor and computer program code are configured to further enable the apparatus to perform or assist in the performance of RAN-related functions between adaptation cycles by the user equipment using at least one ML model in inference mode.

[0140] Example 20. An apparatus according to any one of Examples 18 to 19, wherein at least one processor and computer program code are configured to cause the apparatus to confirm that: at least one processor and computer program code are configured to cause the apparatus to: transmit a set of ML function adaptation parameters from the user equipment to the network node for performing ML function adaptation; and receive from the network node an acknowledgment that the set of ML function adaptation parameters is acceptable.

[0141] Example 21. An apparatus according to any one of Examples 18 to 19, wherein at least one processor and computer program code are configured to cause the apparatus to confirm that: at least one processor and computer program code are configured to cause the apparatus to: receive a set of ML function adaptation parameters from a network node by a user equipment for performing ML function adaptation; and transmit confirmation to the network node that the set of ML function adaptation parameters is acceptable.

[0142] Example 22. An apparatus according to any one of Examples 18 to 21, wherein a set of ML function adaptation parameters includes information indicating the following: a valid time period for which the ML function adaptation parameters are valid; the number of adaptation cycles within the valid time period; and the duration of the adaptation cycle for each of at least one adaptation cycle.

[0143] Example 23. The apparatus according to Example 22, wherein at least one adaptation period comprises a plurality of adaptation periods, wherein the adaptation period duration for the plurality of adaptation periods comprises at least one of the following: an adaptation period duration for the plurality of adaptation periods within a valid time period, wherein the adaptation period duration is the same for each adaptation period; or an average adaptation period duration for the adaptation periods within a valid time period.

[0144] Example 24. The apparatus according to any one of Examples 18 to 21, wherein at least one adaptation cycle comprises a plurality of adaptation cycles, wherein a set of ML function adaptation parameters comprises at least one of the following: the number of ML function adaptations; the number of adaptation cycles for each ML function adaptation; the duration or average duration of the adaptation cycle; the time period between each adaptation cycle in the adaptation cycle; or the average time period between each adaptation cycle in the adaptation cycle.

[0145] Example 25. An apparatus according to any one of Examples 18 to 24, wherein one or more inputs of the ML function, configured by the network node, remain unchanged during the effective time period.

[0146] Example 26. An apparatus according to any one of Examples 18 to 24, wherein at least one processor and computer program code are configured to cause the apparatus to confirm the adaptation of performing an ML function, comprising: at least one processor and computer program code being configured to cause the apparatus to: perform a plurality of ML function adaptations, wherein each ML function adaptation includes a plurality of adaptation cycles; wherein one or more inputs of the ML function, configured by a network node, remain unchanged within each ML function adaptation; and wherein one or more inputs of the ML function, configured by a network node, are changed between two ML function adaptations during a valid time period.

[0147] Example 27. An apparatus according to any one of Examples 18 to 26, wherein at least one processor and computer program code are configured to further enable the apparatus to: transmit a capability response from the user equipment to a network node indicating that the user equipment has the capability to perform at least one of the following; receive a set of ML function adaptation parameters from the user equipment; or send or provide a set of ML function adaptation parameters or a suggested or requested set of ML function adaptation parameters to the network node.

[0148] Example 28. The apparatus according to any one of Examples 18 to 27, wherein at least one processor and computer program code are configured to cause the apparatus to perform adaptation of ML functions by a user equipment during at least one of a plurality of adaptation cycles, based on at least one processor and computer program code being configured to cause the apparatus to perform at least one of the following: receiving a request from a network node by the user equipment to perform adaptation for ML functions; or detecting a need for adaptation to perform ML functions by the user equipment based on the performance of the RAN-related functions being less than a threshold.

[0149] Example 29. An apparatus according to any one of Examples 18 to 28, wherein at least one processor and computer program code are configured to further enable the apparatus to: transmit requests from the user equipment to a network node for resources to be used by the user equipment during multiple adaptation cycles within a valid time period to perform adaptation of ML functions.

[0150] Example 30. The apparatus according to any one of Examples 18 to 29, wherein the adaptation of the ML function is performed in part iteratively during each of a plurality of adaptation cycles.

[0151] Example 31. An apparatus according to any one of Examples 18 to 30, wherein adaptation of at least one processor and computer program code to enable the apparatus to perform ML functions by a user equipment comprises: at least one processor and computer program code being configured as an apparatus to: adapt one or more weights or biases of at least one ML model; adapt at least one ML model; adapt multiple ML models; or adapt the architecture and / or model structure of at least one ML model.

[0152] Figure 8 This is a block diagram of a wireless station or node (e.g., UE, user equipment, AP, BS, eNB, gNB, RAN node, network node, TRP, or other node) 1300 according to an example embodiment. The wireless station 1300 may include, for example, one or more (e.g., such as...) Figure 8 The two RF (radio frequency) or wireless transceivers 1302A and 1302B shown include a transmitter for transmitting signals and a receiver for receiving signals. The wireless station also includes: a processor or control unit / entity (controller) 1304 for executing instructions or software and controlling the transmission and reception of signals; and a memory 1306 for storing data and / or instructions.

[0153] Processor 1304 may also make decisions or determinations, generate frames, packets, or messages for transmission, decode received frames or messages for further processing, and perform other tasks or functions described herein. For example, processor 1304, which may be a baseband processor, may generate messages, packets, frames, or other signals for transmission via wireless transceiver 1302 (1302A or 1302B). Processor 1304 may control the transmission of signals or messages via a wireless network and may control the reception of signals or messages via a wireless network (e.g., after down-conversion by wireless transceiver 1302). Processor 1304 may be programmable and capable of executing software or other instructions stored in memory or other computer media to perform the various tasks and functions described above, such as one or more of the tasks or methods described above. Processor 1304 may be (or may include) hardware, programmable logic, a programmable processor executing software or firmware, and / or any combination of these. For example, using other terms, processor 1304 and transceiver 1302 together may be considered a wireless transmitter / receiver system.

[0154] Additionally, refer to Figure 8 The controller (or processor) 1308 can execute software and instructions, and can provide overall control for station 1300, and can provide... Figure 8Other systems, not shown, provide control, such as controlling input / output devices (e.g., a display, a keypad), and / or can execute software for one or more applications that can be provided on the wireless station 1300, such as, for example, an email program, an audio / video application, a word processor, a VoIP application, or other applications or software.

[0155] In addition, a storage medium may be provided that includes stored instructions, which, when executed by a controller or processor, may cause processor 1304 or other controllers or processors to perform one or more of the functions or tasks described above.

[0156] According to another example embodiment, the RF or wireless transceiver 1302A / 1302B can receive signals or data and / or transmit or send signals or data. The processor 1304 (and possibly the transceiver 1302A / 1302B) can control the RF or wireless transceiver 1302A or 1302B to receive, transmit, broadcast, or transmit signals or data.

[0157] Embodiments of the various technologies described herein can be implemented in digital electronic circuit systems or in computer hardware, firmware, software, or combinations thereof. Embodiments can be implemented as computer program products, i.e., computer programs tangibly embodied in an information carrier (e.g., in a machine-readable storage device or in a propagating signal) for execution by or control of a data processing apparatus (e.g., a programmable processor, a computer, or multiple computers). Embodiments can also be provided on a computer-readable medium or a computer-readable storage medium, which may be a non-transitory medium. Embodiments of the various technologies may also include embodiments provided via transient signals or media, and / or program and / or software embodiments downloadable via the Internet or other networks (wired and / or wireless networks). Furthermore, embodiments can be provided via machine-type communication (MTC) and also via the Internet of Things (IoT).

[0158] Computer programs can be in the form of source code, object code, or some intermediate form, and they can be stored on some kind of carrier, distribution medium, or computer-readable medium, which can be any entity or device capable of carrying the program. Such carriers include, for example, recording media, computer memory, read-only memory, optoelectronic and / or electrical carrier signals, telecommunication signals, and software distribution packages. Depending on the required processing power, a computer program can be executed in a single electronic digital computer or distributed across multiple computers.

[0159] Furthermore, embodiments of the various technologies described herein can utilize network-physical systems (CPS) (systems that control collaborative computing elements of physical entities). CPS embodiments can be implemented and utilize a large number of interconnected ICT devices (sensors, actuators, processors, microcontrollers, etc.) embedded in physical objects at different locations. Mobile network-physical systems, which are inherently mobile physical systems, are a subcategory of network-physical systems. Examples of mobile physical systems include mobile robots and electronic devices transported by humans or animals. The growing popularity of smartphones has increased interest in the field of mobile network-physical systems. Therefore, various embodiments of the technologies described herein can be provided via one or more of these technologies.

[0160] Computer programs (such as those described above) can be written in any programming language (including compiled or interpreted languages) and can be deployed in any form, including as standalone programs or as modules, components, subroutines, or other units or parts thereof adapted to a computing environment. Computer programs can be deployed to execute on one or more computers at a single site, or distributed across multiple sites and interconnected via a communication network.

[0161] The method steps can be executed by one or more programmable processors that execute a computer program or a portion thereof to perform a function by manipulating input data and generating output. The method steps can also be executed by special-purpose logic circuitry, and the apparatus can be implemented as special-purpose logic circuitry, such as an FPGA (Field-Programmable Gate Array) or an ASIC (Application-Specific Integrated Circuit).

[0162] For example, processors suitable for executing computer programs include general-purpose and special-purpose microprocessors, as well as any type of digital computer, chip, or chipset and any one or more processors. Typically, a processor receives instructions and data from read-only memory or random access memory, or both. The components of a computer may include at least one processor for executing instructions and one or more memory devices for storing instructions and data. Typically, a computer may also include, or be operatively coupled to, receiving data from or transferring data to, or both to, one or more mass storage devices (e.g., magnetic disks, magneto-optical disks, or optical disks) for storing data. Information carriers suitable for embodying computer program instructions and data include all forms of non-volatile memory, including, for example, semiconductor memory devices such as EPROM, EEPROM, and flash memory devices; magnetic disks, such as internal hard disks or removable disks; magneto-optical disks; and CD-ROM and DVD-ROM disks. The processor and memory may be supplemented by or incorporated into a dedicated logic circuit system.

[0163] To provide interaction with the user, the embodiments can be implemented on a computer having a display device for displaying information to the user, such as a cathode ray tube (CRT) or liquid crystal display (LCD) monitor, and a user interface through which the user can provide input to the computer, such as a keyboard and a pointing device, such as a mouse or trackball. Other types of devices can also be used to provide interaction with the user; for example, the feedback provided to the user can be any form of sensory feedback, such as visual feedback, auditory feedback, or tactile feedback; and input from the user can be received in any form, including acoustic, voice, or tactile input.

[0164] The embodiments can be implemented in a computing system that includes backend components (e.g., as a data server), middleware components (e.g., an application server), or frontend components (e.g., a client computer having a graphical user interface or web browser through which a user can interact with the embodiments), or any combination of such backend, middleware, or frontend components. The components can be interconnected via digital data communication (e.g., a communication network) of any form or medium. Examples of communication networks include local area networks (LANs) and wide area networks (WANs), such as the Internet.

[0165] While certain features of the embodiments have been shown as described herein, many modifications, substitutions, alterations, and equivalents will now occur to those skilled in the art. Therefore, it should be understood that the appended claims are intended to cover all such modifications and variations falling within the true spirit of the various embodiments.

Claims

1. A method comprising: A set of machine learning (ML) function adaptation parameters is confirmed by the user equipment and the network node for the user equipment to perform adaptation of ML functions associated with at least one ML model used by the user equipment to perform radio access network (RAN) related functions. The set of ML function adaptation parameters indicates at least one adaptation period and a valid time period during which the user equipment performs the ML function adaptation. The set of ML function adaptation parameters is valid for the valid time period. as well as The user equipment performs the adaptation of the ML function during the at least one adaptation period.

2. The method according to claim 1, wherein the at least one adaptation period comprises a plurality of adaptation periods, wherein performing adaptation comprises the user equipment performing adaptation of the ML function during the plurality of adaptation periods; The method further includes: The user equipment uses the at least one ML model in inference mode to perform or assist in the performance of the RAN-related functions between the adaptation cycles.

3. The method of claim 1, wherein the confirmation includes: The user equipment transmits the set of ML function adaptation parameters to the network node for performing the adaptation of the ML function. as well as The user equipment receives confirmation from the network node that the set of ML function adaptation parameters is acceptable.

4. The method of claim 1, wherein the confirmation includes: The user equipment receives the set of ML function adaptation parameters from the network node to perform the adaptation of the ML function. as well as The user equipment transmits a confirmation to the network node that the set of ML function adaptation parameters is acceptable.

5. The method of claim 1, wherein the set of ML function adaptation parameters includes information indicating the following: The effective time period, and the ML function adaptation parameters are valid for the effective time period; The number of adaptation periods within the effective time period; and The duration of the adaptation period for each of the at least one adaptation period.

6. The method of claim 5, wherein the at least one adaptation period comprises a plurality of adaptation periods, wherein the duration of the adaptation period for the plurality of adaptation periods comprises at least one of the following: The adaptation period duration for the plurality of adaptation periods within the effective time period, wherein the adaptation period duration is the same for each of the adaptation periods; or The average adaptation cycle duration within the effective time period.

7. The method according to claim 1, wherein the at least one adaptation period comprises multiple adaptation periods, and wherein the set of ML function adaptation parameters comprises at least one of the following: The number of ML features adapted; The number of adaptation cycles for each ML function adaptation; The duration or average duration of the adaptation period; The time period between each adaptation cycle in the adaptation cycle; or The average time period between each adaptation cycle in the adaptation cycle.

8. The method of claim 1, wherein one or more inputs of the ML function configured by the network node remain unchanged during the effective time period.

9. The method according to claim 1: The adaptation for performing the ML function includes: Perform multiple ML function adaptations, where each ML function adaptation includes multiple adaptation cycles; The one or more inputs of the ML function, configured by the network node, remain unchanged in each ML function adapter. and The one or more inputs of the ML function, configured by the network node, are changed between two ML function adaptations during the effective time period.

10. The method according to any one of claims 1, further comprising: The user equipment transmits a capability response to the network node, indicating that the user equipment has the ability to perform at least one of the following: ML functionality or ML model adaptation; The user equipment receives the set of ML function adaptation parameters; or The user equipment sends or provides the set of ML function adaptation parameters or the suggested or requested set of ML function adaptation parameters to the network node.

11. The method of claim 1, wherein the adaptation performed by the user equipment during at least one of the plurality of adaptation periods is based on at least one of the following: The user equipment receives an adaptation request from the network node to perform the ML function; or The user equipment detects the need for adaptation to perform the ML function based on the performance of the RAN-related functions being less than a threshold.

12. The method according to claim 1, further comprising: The user equipment transmits a request to the network node for resources to be used by the user equipment during the plurality of adaptation periods within the effective time period in order to perform the adaptation of the ML function.

13. The method of claim 1, wherein the adaptation of the ML function is performed in part iteratively during each of the plurality of adaptation cycles.

14. The method of claim 1, wherein the adaptation by the user equipment to perform the ML function comprises performing at least one of the following: Adapt one or more weights or biases to the at least one ML model; Adapt to at least one of the ML models; Adapt to multiple ML models; or Adapt to the architecture and / or model structure of at least one ML model.

15. An apparatus comprising: At least one processor; as well as At least one memory, including computer program code; The at least one memory and the computer program code are configured, together with the at least one processor, to make the device at least: A set of machine learning (ML) function adaptation parameters is confirmed by the user equipment and the network node for the user equipment to perform adaptation of ML functions associated with at least one ML model used by the user equipment to perform radio access network (RAN) related functions. The set of ML function adaptation parameters indicates at least one adaptation period and a valid time period during which the user equipment performs the ML function adaptation. The set of ML function adaptation parameters is valid for the valid time period. as well as The user equipment performs the adaptation of the ML function during the at least one adaptation period.

16. The apparatus of claim 15, wherein the at least one adaptation period comprises a plurality of adaptation periods, wherein performing the adaptation comprises the user equipment performing the adaptation of the ML function during the plurality of adaptation periods; The at least one processor and the computer program code are configured to further enable the device to: The user equipment uses the at least one ML model in inference mode to perform or assist in the performance of the RAN-related functions between the adaptation cycles.

17. The apparatus of claim 15, wherein the at least one processor and the computer program code configured to enable the apparatus to acknowledge comprise: The at least one processor and the computer program code configured to cause the device to perform the following: The user equipment transmits the set of ML function adaptation parameters to the network node for performing the adaptation of the ML function; and The user equipment receives confirmation from the network node that the set of ML function adaptation parameters is acceptable.

18. The apparatus of claim 15, wherein the at least one processor and the computer program code configured to enable the apparatus to acknowledge comprise: The at least one processor and the computer program code configured to cause the device to perform the following: The user equipment receives the set of ML function adaptation parameters from the network node to perform the adaptation of the ML function; and The user equipment transmits a confirmation to the network node that the set of ML function adaptation parameters is acceptable.

19. The apparatus of claim 15, wherein the set of ML function adaptation parameters includes information indicating the following: The effective time period, and the ML function adaptation parameters are valid for the effective time period; The number of adaptation periods within the effective time period; and The adaptation cycle for each adaptation cycle.

20. The apparatus of claim 19, wherein the at least one adaptation period comprises a plurality of adaptation periods, wherein the duration of the adaptation period for the plurality of adaptation periods comprises at least one of the following: The adaptation period duration for the plurality of adaptation periods within the effective time period, wherein the adaptation period duration is the same for each of the adaptation periods; or The average adaptation cycle duration within the effective time period.

Images