Machine learning based analysis at network layer

Abstract

Embodiments of the present disclosure provide machine learning based analysis at a network layer. The method includes receiving, from a first network layer, a first potential representation of first information available at the first network layer, extracting the first potential representation by a trained first machine learning model executed at the first network layer, the first machine learning model configured to generate a first predicted output of a first analysis task based on the first information; generating, at a second network layer and using a trained second machine learning model, a second predicted output of a second analysis task based on the first potential representation and second information available at the second network layer, the second network layer being different from the first network layer; and determining, at the second network layer, a target output for the second analysis task based on the second predicted output.

Description

Technical Field

[0001] The embodiments disclosed herein generally relate to the telecommunications field, and particularly to machine learning (ML) based analytics at the network layer. Background Technology

[0002] Wireless communication systems are widely deployed to provide a variety of telecommunications services, such as telephone, video, data, messaging, and broadcasting. In the telecommunications industry, artificial intelligence / machine learning (AI / ML) models have been used in communication systems to improve performance. In typical communication architectures, various analytical tasks are defined at different network protocol layers, such as serving cell selection / scheduling, radio resource allocation / scheduling, packet generation / processing, and so on. As a specific example, the serving cell selection function aims to select serving cells (e.g., primary and / or secondary cells) to add to the set of serving cell lists of a terminal device. As used herein, a primary cell (pCell) refers to the cell that operates on the primary frequency band of the terminal device and handles the terminal device's radio resource control (RRC) connections, while a secondary cell (sCell) refers to a cell that aggregates with a pCell and provides additional resources to a terminal device configured with carrier aggregation (CA). All cells to be aggregated for the terminal device are referred to as the serving cells for the terminal device.

[0003] The proposal suggests employing machine learning models to learn more about the serving cell (e.g., cell coverage) to facilitate the selection of the optimal serving cell for an end device. In such a solution, predictions of cell coverage are provided for serving cell selection by analyzing some measurements and applying machine learning techniques. Machine learning models typically rely on relevant information to output accurate predictions. In the serving cell selection example, a straightforward solution is to collect all relevant information across network protocol layers at the serving cell selection function where the machine learning model is deployed. However, this can introduce costly message exchanges between layers used for information exchange, a load on the system from such data transfers, and latency in analysis. Other AI-based analytics tasks may face similar challenges when the model deployed at a network layer requires relevant information available at another network layer. Summary of the Invention

[0004] The embodiments of this disclosure provide an improved solution for machine learning-based analytics to optimize analytics by leveraging simplified message exchange, reduced load from cross-layer information exchange, and lower decision latency.

[0005] Specifically, embodiments of this disclosure propose extracting knowledge from information available at a first network layer and transferring that knowledge to different second network layers for use, thereby improving the analysis performed at the second network layer. Knowledge transfer is achieved through a first trained machine learning model at the first network layer, a second trained machine learning model at the second network layer, and a knowledge transfer module in between. The first trained machine learning model at the first network layer extracts a latent representation of the information available at that layer, which is transferred to the second network layer for combination with information available at the second network layer by the second machine learning model. In this way, knowledge from another network layer can be used to enhance the output of the analysis task performed at the second network layer. Furthermore, because the extracted latent representations are exchanged across layers instead of the original data, the amount of information exchanged is reduced or minimized.

[0006] In a first aspect of this disclosure, a method implemented by a network device is proposed. The method includes: receiving a first latent representation of first information available at the first network layer; extracting the first latent representation by a trained first machine learning model executed at the first network layer, the first machine learning model being configured to generate a first predictive output for a first analysis task based on the first information; generating a second predictive output for a second analysis task at a second network layer and using a trained second machine learning model, based on the first latent representation and second information available at the second network layer, the second network layer being different from the first network layer; and determining a target output for the second analysis task at the second network layer based on the second predictive output, wherein the first predictive output indicates whether to schedule at least one serving cell for a terminal device, the second predictive output indicates whether to select at least one serving cell as a potential serving cell for the terminal device, and the target output indicates a subset of serving cells selected from a set of available serving cells for the terminal device, or wherein the first predictive output indicates resource allocation for at least one terminal device located in a network slice, the second predictive output indicates potential resource allocation for a traffic flow in the network slice, and the target output indicates target resource allocation for a traffic flow in the network slice.

[0007] In some embodiments of the first aspect, the first latent representation is fed into a second machine learning model along with the second information. In some embodiments of the first aspect, the first latent representation and the second latent representation of the second information are processed by the second machine learning model to generate a second prediction output.

[0008] In some embodiments of the first aspect, the first analysis task is based on the output of the second analysis task, and the method further includes: at a second network layer, generating a third predicted output of the second analysis task based on third information available at the second network layer. In some embodiments of the first aspect, a first machine learning model is configured to generate a first predicted output of the first analysis task based on the first information and the third predicted output.

[0009] In some embodiments of the first aspect, the first network layer includes a lower protocol layer, and the second network layer includes a higher protocol layer. In some embodiments of the first aspect, the first information includes first measurement information relating to at least one serving cell at the lower protocol layer, and the second information includes second measurement information relating to at least one serving cell at the higher protocol layer.

[0010] In some embodiments of the first aspect, a first machine learning model is trained for a first serving cell, and a first potential representation is specific to the first serving cell. The method further includes: mapping a first potential representation of the first measurement information related to the first serving cell to a third potential representation of the second serving cell at a higher protocol layer, based on a determination that measurement information related to a second serving cell is unavailable; or receiving a third potential representation of the second serving cell from a lower protocol layer at a higher protocol layer.

[0011] In some embodiments of the first aspect, a second prediction output is further generated based on a third potential representation to further indicate whether a second serving cell should be selected as a potential serving cell for the terminal device.

[0012] In some embodiments of the first aspect, the first network layer includes a real-time layer in an Open Radio Access Network (O-RAN), and the second network layer includes a non-real-time layer or a near-real-time layer in the O-RAN. In some embodiments of the first aspect, the first network layer includes a near-real-time layer in the O-RAN, and the second network layer includes a non-real-time layer in the O-RAN.

[0013] In some embodiments of the first aspect, the second predicted output is provided as a suggestion for determining the target output, or the second predicted output is determined as the target output.

[0014] In a second aspect of this disclosure, a method for training a machine learning model to be executed by a network device is proposed. The method includes: training a first machine learning model using first information available at a first network layer as a first ground truth output of the first information in a first analysis task and a model input; using the trained first machine learning model to extract a first latent representation of the first information; training a second machine learning model using second information available at a second network layer, the first latent representation, and a second ground truth output of the second information in the second analysis task; and providing the trained first machine learning model and the trained second machine learning model for execution by the network device, wherein the first ground truth output indicates whether to schedule at least one serving cell for a terminal device, and the second ground truth output indicates whether to select at least one serving cell as a potential serving cell for the terminal device, or wherein the first ground truth output indicates resource allocation for at least one terminal device located in a network slice, and the second ground truth output indicates potential resource allocation for traffic flows in the network slice.

[0015] In some embodiments of the second aspect, the first latent representation is fed into a second machine learning model along with the second information. In some embodiments of the second aspect, the first latent representation and the second latent representation of the second information are processed by the second machine learning model to generate a model output.

[0016] In some embodiments of the second aspect, the method further includes: performing initial training of the second machine learning model using at least third information available at the second network layer before training the first machine learning model. In some embodiments of the second aspect, the second machine learning model is retrained using the second information, the first latent representation, and the second ground truth output.

[0017] In some embodiments of the second aspect, the first network layer includes a lower protocol layer, and the second network layer includes a higher protocol layer. In some embodiments of the second aspect, the first information includes first measurement information relating to at least one serving cell at the lower protocol layer, and the second information includes second measurement information relating to at least one serving cell at the higher protocol layer.

[0018] In some embodiments of the second aspect, a first truth value output is retrieved from a serving cell scheduler at a lower protocol layer, and the first truth value output indicates whether at least one serving cell should be scheduled for the terminal device. In some embodiments of the second aspect, a second truth value output is retrieved from a serving cell selection function at a higher protocol layer, and the second truth value output indicates whether at least one serving cell should be selected as a potential serving cell for the terminal device.

[0019] In some embodiments of the second aspect, the first true value output indicates whether to schedule at least one serving cell for the terminal device within a corresponding time instance during a time period. The method further includes: aggregating the first true value output and the second true value output to obtain an aggregated true value output, the aggregated true value output indicating whether to select at least one serving cell as a potential serving cell for the terminal device within a time period. In some embodiments of the second aspect, the aggregated true value output is used to train a second machine learning model.

[0020] In some embodiments of the second aspect, a first machine learning model is trained for a first serving cell, and the first latent representation is specific to the first serving cell. The method further includes: mapping the first latent representation of the first measurement information related to the first serving cell to a third latent representation of the second serving cell based on a determination that measurement information related to the second serving cell is unavailable. In some embodiments of the second aspect, the third latent representation is used to further train the second machine learning model.

[0021] In some embodiments of the second aspect, the first network layer includes a real-time layer in an Open Radio Access Network (O-RAN), and the second network layer includes a non-real-time layer or a near-real-time layer in the O-RAN. In some embodiments of the second aspect, the first network layer includes a near-real-time layer in the O-RAN, and the second network layer includes a non-real-time layer in the O-RAN.

[0022] In the third aspect, a network device is proposed. A network device includes: at least one processor; and at least one memory coupled to the at least one processor, the at least one memory including instructions, the known instructions which, when executed by the at least one processor, implement a method comprising: receiving a first latent representation of first information available at the first network layer from a first network layer; extracting the first latent representation by a first machine learning model trained at the first network layer, the first machine learning model being configured to generate a first predictive output for a first analysis task based on the first information; generating a second predictive output for a second analysis task at a second network layer and using a second machine learning model trained at the second network layer, the second network layer being different from the first network layer; and determining a target output for the second analysis task at the second network layer based on the second predictive output, wherein the first predictive output indicates whether to schedule at least one serving cell for a terminal device, the second predictive output indicates whether to select at least one serving cell as a potential serving cell for the terminal device, and the target output indicates a subset of serving cells selected from a set of available serving cells for the terminal device, or wherein the first predictive output indicates resource allocation for at least one terminal device located in a network slice, the second predictive output indicates potential resource allocation for a traffic flow in the network slice, and the target output indicates target resource allocation for a traffic flow in the network slice.

[0023] In a fourth aspect, a computing system is proposed. The computing system includes: at least one processor; and at least one memory coupled to the at least one processor, the at least one memory including instructions that, when executed by the at least one processor, implement a method for training a machine learning model to be executed by a network device. The method includes: training a first machine learning model using first information available at a first network layer as a first ground truth output and model input for the first information in a first analysis task; extracting a first latent representation of the first information using the trained first machine learning model; training a second machine learning model using second information available at a second network layer, the first latent representation, and a second ground truth output of the second information in the second analysis task; and providing the trained first machine learning model and the trained second machine learning model for execution by the network device, wherein the first ground truth output indicates whether to schedule at least one serving cell for a terminal device, and the second ground truth output indicates whether to select at least one serving cell as a potential serving cell for the terminal device, or wherein the first ground truth output indicates resource allocation for at least one terminal device located in a network slice, and the second ground truth output indicates potential resource allocation for traffic flows in the network slice.

[0024] In a fifth aspect, an apparatus is proposed. The apparatus includes: components for receiving a first latent representation of first information available at the first network layer from a first network layer, extracting the first latent representation by a trained first machine learning model executed at the first network layer, the first machine learning model being configured to generate a first predictive output for a first analysis task based on the first information; components for generating a second predictive output for a second analysis task at a second network layer and using a trained second machine learning model, based on the first latent representation and second information available at the second network layer, the second network layer being different from the first network layer; and components for determining a target output for the second analysis task at the second network layer based on the second predictive output, wherein the first predictive output indicates whether to schedule at least one serving cell for a terminal device, the second predictive output indicates whether to select at least one serving cell as a potential serving cell for the terminal device, and the target output indicates a subset of serving cells selected from a set of available serving cells for the terminal device, or wherein the first predictive output indicates resource allocation for at least one terminal device located in a network slice, the second predictive output indicates potential resource allocation for a traffic flow in the network slice, and the target output indicates target resource allocation for a traffic flow in the network slice.

[0025] In a sixth aspect, an apparatus is proposed. The apparatus includes components for training a first machine learning model using first information available at a first network layer as a first ground truth output and model input for the first information in a first analysis task; components for extracting a first latent representation of the first information using the trained first machine learning model; components for training a second machine learning model using second information available at a second network layer, the first latent representation, and a second ground truth output for the second information in the second analysis task; and components for providing the trained first machine learning model and the trained second machine learning model for execution by a network device, wherein the first ground truth output indicates whether to schedule at least one serving cell for a terminal device, and the second ground truth output indicates whether to select at least one serving cell as a potential serving cell for the terminal device, or wherein the first ground truth output indicates resource allocation for at least one terminal device located in a network slice, and the second ground truth output indicates potential resource allocation for traffic flows in the network slice.

[0026] In a seventh aspect, a computer-readable medium is proposed. The computer-readable medium stores instructions thereon that, when executed by at least one processor, cause at least one processor to perform a method according to any embodiment of the first aspect.

[0027] In an eighth aspect, a computer-readable medium is proposed. The computer-readable medium stores instructions thereon that, when executed by at least one processor, cause at least one processor to perform a method according to any embodiment of the second aspect.

[0028] In a ninth aspect, a computer program is proposed. The computer program includes instructions that, when executed on at least one processor, cause at least one processor to perform a method according to any embodiment of the first aspect.

[0029] In a tenth aspect, a computer program is proposed. The computer program includes instructions that, when executed on at least one processor, cause at least one processor to perform a method according to any embodiment of the second aspect.

[0030] Embodiments of this disclosure enable machine learning-based network analytics using cross-layer knowledge transfer. In this solution, knowledge is extracted from available information at one network layer and transferred to another network layer for use, thereby improving the analytics implemented at that other network layer. Knowledge from the other network layer can be used to enhance the output of the analytics task implemented at one network layer. Furthermore, the amount of information exchanged is reduced or minimized because the extracted latent representations are exchanged across layers instead of the original data. Therefore, machine learning-based analytics is optimized by utilizing simplified message exchange, reduced cross-layer information exchange load, and lower decision latency.

[0031] It should be understood that the summary section is not intended to identify key or essential features of embodiments of this disclosure, nor is it intended to limit the scope of this disclosure. Other features of this disclosure will become readily apparent from the following description. Attached Figure Description

[0032] The above and other objects, features, and advantages of this disclosure will become more apparent from the more detailed description of some exemplary embodiments of the present disclosure in the accompanying drawings, in which: Figure 1 A schematic diagram illustrating a simplified architecture for analysis at the network layer in which embodiments of the present disclosure may be applied; Figure 2 A schematic diagram illustrating the radio access network (RAN) that enables serving cell selection is provided. Figure 3A A schematic diagram illustrating a radio protocol stack according to some embodiments of the present disclosure is provided. Figure 3B A schematic diagram illustrating the action space of a serving cell at different layers according to some embodiments of the present disclosure is provided. Figure 4A schematic diagram illustrating the Open Radio Access Network (O-RAN) architecture that enables resource allocation is provided. Figure 5 A schematic diagram illustrating an architecture for machine learning-based analytics at a network layer that leverages cross-layer knowledge transfer, according to some embodiments of the present disclosure; Figure 6 A schematic diagram illustrating a training process for a machine learning model utilizing knowledge transfer according to some embodiments of the present disclosure is provided. Figure 7 The illustration shows a schematic diagram illustrating an example integration of a potential representation from a machine learning model with another machine learning model, according to some embodiments of the present disclosure; Figures 8A to 8D The diagram illustrates different training stages of a machine learning model utilizing knowledge transfer according to some embodiments of the present disclosure; Figure 9 This illustration shows a schematic diagram illustrating domain adaptation between pre-selected cells and unselected cells according to some embodiments of the present disclosure; Figure 10 A signaling diagram for training a machine learning model using knowledge transfer, according to some other embodiments of this disclosure, is illustrated. Figure 11A and Figure 11B A schematic diagram illustrating reasoning using a machine learning model that utilizes knowledge transfer according to some other embodiments of the present disclosure is provided. Figure 12 A flowchart illustrating a method implemented at a network device according to some embodiments of the present disclosure is provided. Figure 13 A flowchart illustrating a method for model training according to some embodiments of the present disclosure is provided. Figure 14 A simplified block diagram illustrating a network apparatus suitable for implementing embodiments of the present disclosure is provided; and Figure 15 A simplified block diagram illustrating a virtual environment suitable for implementing model training according to some embodiments of the present disclosure is provided.

[0033] Throughout the accompanying drawings, the same or similar reference numerals represent the same or similar elements. Detailed Implementation

[0034] Some embodiments of the ideas contemplated herein will now be described more fully with reference to the accompanying drawings. The embodiments set forth below represent information enabling those skilled in the art to implement the embodiments and illustrate the best mode of implementation. Those skilled in the art will understand the concepts of this disclosure by reading the following description in conjunction with the accompanying drawings, and will recognize that the application of these concepts is not specifically proposed herein. It should be understood that these concepts and applications are within the scope of this disclosure.

[0035] Generally, unless a different meaning is explicitly given and / or implied in the context of the use of the term, all terms used herein shall be interpreted in accordance with their ordinary meaning in the relevant art. Unless otherwise expressly stated, all references to "a" or "an" element, device, component, part, step, etc., shall be openly interpreted as referring to at least one instance of that element, device, component, part, step, etc. Unless a step is explicitly described as occurring after or before another step and / or implied that a step must occur after or before another step, the steps of any method disclosed herein are not necessarily performed in the exact order disclosed. Where appropriate, any feature of any embodiment of the embodiments disclosed herein may be applied to any other embodiment. Similarly, any advantage of any embodiment of the embodiments may be applied to any other embodiment, and vice versa. Other objects, features, and advantages of the appended embodiments will become apparent from the following description.

[0036] As used herein, the term “comprising” and its variations are to be interpreted as open-ended terms meaning “including, but not limited to”. The term “based on” is to be interpreted as “at least partially based on”. The terms “one embodiment” and “embodiment” are to be interpreted as “at least one embodiment”. The term “another embodiment” is to be interpreted as “at least one other embodiment”. The terms “first,” “second,” etc., may refer to different or the same objects. Other definitions (explicit and implicit) may be included below.

[0037] As used herein, the term "network node" can also be referred to as a network function (NF), network entity, or network device, and refers to a physical, virtual, or hybrid function or entity deployed on the network side to provide one or more services to clients / consumers. For example, an NF can be deployed at a device in a radio access network (RAN) or core network (CN) of a communication system. A network node can be implemented using hardware, software, firmware, or some combination thereof. Examples of network nodes in the RAN include, but are not limited to, Node B (NodeB or NB), evolved Node B (eNodeB or eNB), next-generation Node B (gNB), Transmitter Receive Point (TRP), Remote Radio Unit (RRU), Radio Header Terminal (RH), Remote Radio Header Terminal (RRH), IAB node, low-power nodes such as femtonodes, piconodes, reconfigurable smart surfaces (RIS), and so on. Examples of network nodes in the CN include, but are not limited to, Mobility Management Entity (MME), Packet Data Network Gateway (P-GW), Service Capability Open Function (SCEF), Home Subscriber Server (HSS), and so on. Other examples of core network nodes include nodes that implement Access and Mobility Management (AMF), User Plane Function (UPF), Session Management (SMF), Authentication Server Function (AUSF), Network Slice Selection Function (NSSF), Network Open Function (NEF), Network Function (NF) Repository Function (NRF), Policy Control Function (PCF), Unified Data Management (UDM), and so on.

[0038] Communication in the communication environment can conform to any suitable standard, including but not limited to Global System for Mobile Communications (GSM), Long Term Evolution (LTE), LTE-Evolution, LTE-A Advanced, New Radio (NR), Wideband Code Division Multiple Access (WCDMA), Code Division Multiple Access (CDMA), GSM EDGE Radio Access Network (GERAN), Machine Type Communication (MTC), etc. Embodiments of this disclosure can be performed according to any generation of communication protocols, whether currently known or to be developed in the future. Examples of communication protocols include, but are not limited to, fourth-generation (4G), 4.5G, fifth-generation (5G) communication protocols, 5.5G, 5G-Advanced networks, or sixth-generation (6G) networks and above, wireless LAN communication protocols such as IEEE 802.11, and / or any other protocols currently known or to be developed in the future. Furthermore, communication may utilize any suitable wireless communication technology, including but not limited to: Code Division Multiple Access (CDMA), Frequency Division Multiple Access (FDMA), Time Division Multiple Access (TDMA), Frequency Division Duplex (FDD), Time Division Duplex (TDD), Multiple Input Multiple Output (MIMO), Orthogonal Frequency Division Multiple Access (OFDM), Discrete Fourier Transform Spread Spectrum OFDM (DFT-s-OFDM) and / or any other technology currently known or to be developed in the future.

[0039] As used herein, the term "model" refers to the association between an input and an output learned from training data, and thus, after training, can generate a corresponding output for a given input. Model generation can be based on machine learning (ML) techniques. Machine learning techniques can also be referred to as artificial intelligence (AI) techniques. Typically, a machine learning model can be constructed that receives input information and makes predictions based on that input information. For example, a classification model can predict the category of input information in a predetermined set of categories. As used herein, "model" can also be referred to as a "machine learning model," "learning model," "machine learning network," or "learning network," which are used interchangeably throughout this document.

[0040] Machine learning typically involves three phases: training, validation, and application (also known as inference). In the training phase, a given machine learning model is iteratively trained (or optimized) using a large amount of training data until the model can make consistent inferences similar to those that human intelligence can make from the training data. During training, the set of parameter values ​​for the model is iteratively updated until the training objective is achieved. Through the training process, the machine learning model can be considered to have learned the relationship between inputs and outputs (also known as input-output mapping) from the training data. In the validation phase, validation inputs are applied to the trained machine learning model to test whether the model can provide correct outputs, thus determining the model's performance. The validation phase can often be considered a step in the training process, or sometimes it can be omitted. In the inference phase, the trained machine learning model is used to process real-world model inputs based on the set of parameter values ​​obtained from the training process and determine the corresponding model output.

[0041] Example Environment In a typical communication architecture, various analysis tasks are defined at different network protocol layers, such as serving cell selection / scheduling, radio resource allocation / scheduling, packet generation / processing, and so on. Figure 1 A simplified architecture 100 for analysis at the network layer, in which embodiments of the present disclosure can be applied, is illustrated. As shown, at network layer 110 (which is sometimes referred to herein as the first network layer), analysis task 112 (which is sometimes referred to herein as the first analysis task) is performed using information 114 available at network layer 110. At network layer 120 (which is sometimes referred to herein as the second network layer), analysis task 122 (which is sometimes referred to herein as the second analysis task) is performed using information 124 available at that layer. Network layers 110 and 120 are different layers defined for one or more network devices in a communication network. Information can be exchanged between network layers 110 and 120 via specified interfaces and according to specified message formats. The analysis task can be any prediction / analysis task to be performed within the communication environment. In some embodiments, the output of the analysis task at a network layer can be used as input or partial input to an analysis task at another network layer.

[0042] Different network layers can be defined in different communication networks. In a Radio Access Network (RAN), three network layers (Layer 1, Layer 2, and Layer 3) are used to define the radio protocol stack. The analysis task for serving cell selection for terminal devices can be implemented at Layer 3 of the network device within the RAN.

[0043] The Open RAN (O-RAN) architecture aims for intelligent RAN automation by applying artificial intelligence, machine learning, and advanced analytics to manage complex and evolving network, device, and end-user requirements for optimal efficiency. In O-RAN, there can be two or more network layers, including a non-real-time (non-RT) layer (also known as a Service Management and Orchestration (SMO) layer), a near-real-time (near-RT) layer, and / or a real-time (RT) layer that implements control and optimization of RAN elements and resources. Example analytical tasks in O-RAN may include resource allocation at the network layers.

[0044] To better understand the analysis scenarios in communication networks, we will refer to... Figure 2 , Figures 3A-3B and Figure 4 This will illustrate the serving cell selection and resource allocation tasks in O-RAN.

[0045] It should also be understood that the elements shown in the figures are intended to represent the main functions provided within the environment. In this way, the boxes shown in the following figures indicate specific elements in the communication network that provide these main functions. However, some or all of the main functions represented can be implemented using other network elements. In this way, it should be understood that not all functions of the communication environment are depicted in the figures. Instead, functions are represented to facilitate explanation of illustrative embodiments.

[0046] When carrier aggregation (CA) is introduced into a communication system, serving cell selection is supported. Using carrier aggregation, multiple component carriers (CCs) can be aggregated, and the multiple component carriers (CCs) can be used jointly for transmission to or from a single wireless device. Figure 2 A schematic diagram of RAN 200 is provided, illustrating the serving cell selection process. In this example, four carriers / bands exist: F1, F2, F3, and F4. Cells 210, 212, 214, and 216 operate under F1, cell 220 under F2, cell 230 under F3, and cell 240 under F4. For a specific terminal device 202, cell 210 acts as the primary cell (pCell) for terminal device 202, handling Radio Resource Control (RRC) connections. Other cells to be aggregated with the pCell used for terminal device 202 are referred to as secondary cells (sCells). All cells to be aggregated for terminal device 202 are referred to as the serving cells for terminal device 202.

[0047] In the example shown, it is assumed that four carriers / bands F1, F2, F3, and F4 are available at the location of terminal device 202, one of which is the pCell (also referred to as the serving carrier) and the other three are adjacent carriers / bands. Furthermore, there are three adjacent cells available on the serving carrier of terminal device 202 and three sCells available on the adjacent carriers / bands.

[0048] Using CA (Cellular Aspect) allows secondary cells to be added to primary cells to increase bandwidth and thus bit rate. As described below, introducing CA into communication networks has introduced a new set of functionalities.

[0049] sCells can be added, released, or reconfigured for terminal devices. Adding, releasing, and reconfiguring sCells is the responsibility of the RRC layer. The RRC can configure sCells for terminal devices with CA capabilities. The initial SCell configuration can be blind, based on network knowledge of the configuration. Measurements are then configured on the sCell carrier frequency and other frequencies to understand all frequencies covered by the terminal device. Once an sCell is added to the set of serving cells (which may include pCells and one or more sCells), the terminal device can perform measurements on the sCell (and on all serving cells) without measurement gaps. Network devices (e.g., network device 204 in pCell 210) always attempt to configure terminal device 202 with the optimal cell (radio conditions) as the sCell to improve spectral efficiency.

[0050] After configuring an sCell as the serving cell for terminal device 202, the sCell must be activated in order for terminal device 202 to receive data on the sCell. This is the next step after configuring the sCell. This activation is accomplished via a Media Access Control (MAC) element. To enable reasonable battery consumption by terminal device 202 during CA configuration, an sCell activation / deactivation mechanism is supported. If terminal device 202 is configured with one or more sCells, network device 204 can activate and deactivate the configured sCells.

[0051] In some embodiments, network device 204 may deactivate sCell when no more data needs to be transmitted to terminal device 202 or when the channel quality of sCell becomes poor. In some embodiments, sCell is removed from pCell or from the set of serving cells via an RRC connection reconfiguration process.

[0052] Network device 204 may include a serving cell selection function 250, which is designed to select the best serving cell from a set of serving cell lists to be added to by optimizing the configuration of terminal device 202 based on a combination of frequency bands supported by the UE to match the carrier aggregation configuration supported in RAN 200. As shown, serving cell selection function 250 may include two functions (configuration discoverer 252 and capability checker 254) for performing capability checks during the serving cell selection process.

[0053] Configuration discoverer 252 can collect network information indicating supported carrier aggregation configurations and resources in RAN 200. Configuration discoverer 252 can suggest configurations to be checked by capability checker 254. The suggested configurations can indicate one or more serving cells to be configured for terminal device 202. Capability checker 254 also receives UE capabilities of terminal device 202 and checks whether terminal device 202 supports the suggested configurations based on its capabilities. The output of capability checker 254 indicates whether terminal device 202 supports the suggested configurations and is provided to configuration discoverer 252 as feedback. This process continues until all configurations have been checked or a timeout occurs. Configuration discoverer 252 can output the selected configurations supported by terminal device 202. If the serving cell selection process cannot find a configuration supported by the capabilities of terminal device 202, the UE capability is declared to have failed, and a capability failure trigger is executed.

[0054] Figure 3A A schematic diagram of a radio protocol stack 300 according to some embodiments of the present disclosure is illustrated. The radio protocol stack for terminal devices and network devices is shown to have three layers: Layer 1, Layer 2, and Layer 3. Layer 1 (L1 layer) is the lowest layer. The physical (PHY) layer 310 at the L1 layer provides information transfer services over the physical channel. Sometimes the L1 layer is referred to as the physical layer. Layer 2 (L2 layer) is above the physical layer 310 and is responsible for the link between the terminal device and the network device at the physical layer 310.

[0055] Layer 2 includes the Media Access Control (MAC) layer 320, the Radio Link Control (RLC) layer 330, and the Packet Data Convergence Protocol (PDCP) layer 340. Layer 3 (L3 layer) includes the Radio Resource Control (RRC) layer 350, the Non-Access Stratum (NAS) layer 360, and the Internet Protocol (IP) layer 370.

[0056] PHY layer 310 provides information transfer services to its higher layers over physical channels. As shown, PHY layer 310 is connected to MAC layer 320 via transport channels, and data is packaged into MAC Packet Data Units (PDUs), and data is transferred between MAC layer 320 and PHY layer 310 over the transport channels. MAC layer 320 is connected to RLC layer 330 via logical channels, and data is packaged into RLC PDUs, and data is transferred between RLC layer 330 and MAC layer 320 over the logical channels. RLC layer 330 is connected to PDCP layer 340, which provides multiplexing between different radio bearers and logical channels. Data is transferred between PDCP layer 340 and RLC layer 330 as PDCP PDUs. User services are transferred between IP layer 370 and PDCP layer 340.

[0057] RRC layer 350 is responsible for acquiring radio resources (e.g., radio bearers) and configuring lower layers using RRC signaling. RRC layer 350 provides PDCP control signaling to PDCP layer 340 and receives control traffic as RRCPDUs from PDCP layer 340. RRC layer 350 also provides RLC control signaling to RLC layer 330, MAC control signaling to MAC layer 320, and L1 configuration and measurement to PHY layer 310.

[0058] Figure 3B This illustration shows the action space of a serving cell at different protocol layers according to some embodiments of the present disclosure. For a particular terminal device, for example... Figure 2 Terminal device 202 will have a set of available serving cells 380, which may include one pCell and one or more sCells. Serving cell selection 382 is performed at the L3 layer to select a subset of serving cells from the set of available serving cells, referred to as L3-filtered serving cells 384. Then, serving cell scheduling 386 is performed at the L1 / 2 layer to schedule one or more serving cells (pCell and one or more sCells) for terminal device 202. The serving cell(s) to be scheduled are the active serving cells 390 for terminal device 202.

[0059] Use cases for cross-layer information exchange For serving cell selection, some solutions aim to maximize achievable downlink (DL) and uplink (UL) data rates based on L3 measurement information and static cell information from L3 layer (e.g., network and UE configuration, cell bandwidth (BW), TDD mode used, number of MIMO layers, etc.). Figure 3AAs shown, the serving cell selection function 302 receives measurement information and static cell information from the L3 layer to determine the configuration for selecting the serving cell for the terminal device.

[0060] However, relying solely on L3 measurement information may miss much important information from the serving cell selection process. For example, the coverage and load of the sCell at the terminal device's location are not considered, which can depend on information from lower layers (i.e., L1 and / or L2). Therefore, the initial sCell selection is blind; that is, the network device is unaware of sCell coverage information before sCell configuration. Only after the selected configuration, based on information from the terminal device (available through measurement reports), can the network device change the sCell configuration to a potentially better one, especially if an A6-based reselection is triggered. This blind selection method leads to situations where data transfer only occurs on the pCell because one or more sCells are not ready to accept traffic. The main reason behind this approach is the need for low-latency decision-making. In other words, most data communication sessions are short in duration, and this duration is always decreasing due to increased system capacity. Therefore, if you try to find the optimal sCell, there is a risk of losing the time that the sCell is needed (for some systems, it is expected that 80% of the business will be handled by pCell and 20% by sCell).

[0061] In this scenario, there is a need for interfaces and message exchange between the L1 / 2 and L3 layers, which introduces additional overhead, complexity, and requires additional storage on the hardware.

[0062] Recently, ML-driven serving cell selection methods have been proposed to improve the carrier aggregation configuration process by employing machine learning models in the serving cell selection process. For example, the machine learning model can be configured to predict sCell coverage within a pCell. In such solutions, predictions of sCell coverage based on estimates of the terminal device's radio location within the pCell are provided by analyzing some measurements and applying machine learning techniques. A straightforward solution for this is to obtain all relevant information across layers at the serving cell selection function, where a machine learning model is deployed to learn from the data and make decisions accordingly. A drawback of this solution is the expensive introduction of message exchange between layers used for information exchange, as well as the overhead of such data transfers on the system. Compared to the potential benefits of performance improvements, such solutions lack techno-economic feasibility.

[0063] The solution in patent application US2019357057A1 proposes building a machine learning model at the L3 layer based on background L3 measurements. Based on this solution, sCell coverage knowledge can be built using periodic reporting of the strongest cell (RSC) measurements triggered by a network device on an inactive terminal device (a terminal device connected to the network device but not transmitting any data) configured to measure the sCell frequency configured on the network device. However, this solution requires auxiliary terminal devices (i.e., inactive terminal devices), which can lead to energy consumption and side effects for the auxiliary terminal devices. Sometimes neighbors may be missing, for example, when the serving cell has a much stronger signal channel quality than other cells. Additional configuration and measurements of the system are also required. In other words, the model does not utilize the large amount of available data in the RAN but instead requests additional measurements. On the other hand, the period during which the learned model has effective output depends on the quality of the training, which itself depends on having neighbors for data collection. The collected data and model are specific to pCell-sCell relationships, i.e., they are lost when cell relationships or cell locks are removed or changed, resulting in limited reusability of the machine learning model and the collected data.

[0064] Another solution in patent application US2020106536A1 proposes using UE-related measurements performed on PCells in a wireless communication system, such as a neural network outside the RAN network device, in a machine learning model. Using this machine learning algorithm and pCell measurements, predictions of achievable channel quality for each sCell are derived. These predicted achievable channel qualities are then sent back to the sCell selection function for decision-making. That is, a machine learning model is constructed to convert pCell channel measurements into sCell coverage predictions. However, pCell frequencies are typically below GHz, and lossy coupling between pCell channel measurements and overlapping sCells is expected, especially at higher values ​​for the sCell carrier frequency. Particularly at higher frequencies, even device orientation and obstruction movement can significantly affect the channel state. A primary use case for this solution could be indoor dense radio access networks.

[0065] Another solution proposed in patent application US9467918B1 involves a machine learning agent outside the RAN for load prediction of cells. The output of this prediction can be fed into mobility management functions as an additional and dynamic input. Some similar work proposes enhancing this solution by applying reinforcement learning algorithms for implementations outside the RAN, aiming to balance the load across different RATs by predicting the load across different RATs. However, while providing load information for the serving cell selection function can be valuable for key performance indicators (KPIs) of interest, particularly in terms of throughput, channel state information for the sCell remains a major challenge to address. This is due to the fact that the lack of coverage by the configured sCell at the UE location can lead to over-measurement and thus energy waste and communication delays.

[0066] Several other solutions in patent applications US2021007023A1, WO2021107831A1, and WO2022084469A1 have proposed applying UE-side information and cooperation to mobility management. Based on these solutions, the terminal device uses information from its sensors to roughly estimate its location within the cell. This location indication can be shared with the RAN, or the input can be further processed on the UE side, and the output sent back to the RAN for use in making decisions regarding mobility management and serving cell selection. However, exposing information from the UE side to the network side opens the door to a range of interesting functionalities, among which the optimization of communication parameters for configuring the set of serving cells, such as Discontinuous Reception (DRX), is of particular interest. However, such solutions require addressing privacy issues (e.g., UE location) when devices share information, and require UE-side memory, processing, and power consumption, which may be unacceptable for some devices or users.

[0067] The above has discussed machine learning-based serving cell selection in the RAN and some traditional solutions proposed for this function. In summary, optimizing serving cell selection results requires more information, which may lead to an increased load on data transfer across network layers, or if information from outside the RAN or UE-side information is introduced, it may lead to latency, complexity, and / or security issues.

[0068] In addition to serving cell selection in RAN, machine learning-based analytics are also applicable to other communication systems. Figure 4 A schematic diagram of the O-RAN 400 architecture is provided, illustrating the analytical tasks that can be performed on resource allocation. O-RAN aims for intelligent RAN automation by applying artificial intelligence, machine learning, and advanced analytics to manage the complex and ever-changing network, device, and end-user requirements for optimal efficiency.

[0069] The O-RAN architecture 400 may include two or more network layers, an RT layer 410, a near-RT layer 420, and a non-RT layer 430. In some implementations, the O-RAN architecture 400 may include two layers, for example, a near-RT layer 420 and a non-RT layer 430, a near-RT layer 420 and an RT layer 410, or a non-RT layer 430 and an RT layer 410. As shown, layers 410, 420, and 430 include orchestrators 412, 422, and 432, respectively, to control radio resource management (RRM) in the RAN. The orchestrator may also be referred to as a RAN intelligent controller (RIC).

[0070] The orchestrator 432 at the non-RT layer 430 enables non-real-time control and optimization of RAN elements and network resources. The orchestrator 422 at the near-RT layer 420 enables near real-time control and optimization of RAN elements and network resources. The orchestrator 412 at the RT layer 410 enables real-time control and optimization of RAN elements and network resources. For example, in the case of network slicing, resource allocation for service flows on the network slice can be implemented in near real-time at the near-RT layer 420 or even non-real-time at the non-RT layer 430 by using a machine learning model. Resource scheduling for terminal devices in the network slice can be implemented in real-time at the RT layer 410 or near real-time at the near-RT layer 420 by using a machine learning model.

[0071] Traditionally, orchestrators at each layer utilize the information available at that layer to perform analytical tasks (e.g., resource allocation for network slices or resource scheduling for end devices). Information from other layers (one or more) can certainly optimize analysis at higher network layers, for example, to predict a more rational resource allocation for a network slice based on the existing end devices within that slice. Information exchange between network layers can lead to costly introduction of message exchanges, a load on the system from such data transfers, and latency in the analysis, which is undesirable.

[0072] Working principle and overall architecture The exemplary embodiments of this disclosure provide an improved solution for machine learning-based network analysis with cross-layer knowledge transfer. In this solution, knowledge is extracted from available information at a first network layer and transferred to different second network layers for use, thereby improving the analysis performed at the second network layer. Knowledge transfer is achieved through a first trained machine learning model at the first network layer, a second trained machine learning model at the second network layer, and a knowledge transfer module in between. The first trained machine learning model at the first network layer extracts a latent representation of the information available at that layer, which is transferred to the second network layer for combination by the second machine learning model with the information available at the second network layer. In this way, knowledge from another network layer can be used to enhance the output of the analysis task performed at the second network layer. Furthermore, since the extracted latent representations are exchanged across layers instead of the original data, the amount of information exchanged is reduced or minimized. Therefore, machine learning-based analysis at the second network layer can be optimized by utilizing simplified message exchange, reduced cross-layer information exchange load, and reduced decision latency.

[0073] Figure 5 A schematic diagram illustrates an architecture 500 with cross-layer knowledge transfer at a network layer according to some embodiments of the present disclosure. Architecture 500 is based on... Figure 1 The simplified architecture 100 described in the diagram illustrates the analysis at the network layer. Besides... Figure 1 In addition to the elements shown in architecture 100, architecture 500 further includes a machine learning model 510 (sometimes referred to herein as the first machine learning model) and a knowledge transfer (KT) function 512 at network layer 110, and a machine learning model 520 (sometimes referred to herein as the second machine learning model) at network layer 120.

[0074] Machine learning model 510 is used to extract and compress knowledge from information available at network layer 110, and machine learning model 510 is associated with analysis task 122 at network layer 120. KT function 512 is used to transfer knowledge from network layer 110 to network layer 120. Using the transferred knowledge, machine learning model 520 from network layer 120 knows the information at network layer 110 and can optimize decisions for analysis task 122 at network layer 120. Machine learning model 510 can be implemented as an agent deployed at network layer 110, and KT function 512 can also be implemented as a KT agent deployed at network layer 110. Machine learning model 520 can be implemented as an agent deployed at network layer 120.

[0075] Specifically, a machine learning model 510 is executed at network layer 110 to perform analysis task 112. The model input of machine learning model 510 includes information 114 for analysis task 112, which is available at network layer 110. The model output generated by machine learning model 510 is the predicted output for analysis task 112. During task execution, machine learning model 510 extracts latent representations from input information 114 and transfers these latent representations to network layer 120 via KT function 512. The latent representation can be considered a compressed version of information 114, and may also be referred to as a feature representation, embedding, or feature of information 114.

[0076] A machine learning model 520 is executed at network layer 120 to perform analysis task 112. The machine learning model 520 generates a predicted output for analysis task 112 based on the latent representation from network layer 110 and information 124 at network layer 120. With the introduction of the latent representation, the feature space at network layer 120 is enriched, and more accurate predictions can be achieved by the machine learning model 520. The predicted output from the machine learning model 520 is used to determine the target output for analysis task 122. In some embodiments, the predicted output of the machine learning model 520 may be provided as a suggestion for determining the target output, or the predicted output of the machine learning model 520 may be directly determined as the target output.

[0077] In architecture 500, machine learning model 510 acts as a knowledge synthesizer for network layer 110, designed to mimic the behavior of network layer 110 in analysis task 112 and allow information from this network layer to be summarized into a compressed form. Machine learning model 520 is used to leverage the knowledge extracted from machine learning model 510 for high-level decision-making. KT function 512 can perform low-overhead cross-layer signaling between the two machine learning models at the two network layers. Machine learning models 510 and 520 can be constructed as any machine learning or deep learning model (e.g., neural network) suitable for performing analysis tasks 112 and 122, respectively, and the specific model architecture is not limited to the scope of this disclosure.

[0078] In some embodiments, in the RAN 200 scenario, analysis task 122 is a serving cell selection task, and analysis task 112 is a serving cell scheduling task, both performed at a network device (e.g., network device 204). The prediction output from machine learning model 520 indicates whether to select at least one serving cell as a potential serving cell for a terminal device (e.g., terminal device 202). The prediction output from machine learning model 510 indicates whether to schedule and activate at least one serving cell for terminal device 202. That is, machine learning models 510 and 520 are used to make decisions regarding at least one serving cell for a particular terminal device. The target output indicates a subset of serving cells selected from the set of available serving cells for terminal device 202, wherein the set of available serving cells includes at least one serving cell measured by the two models 510 and 520. In some embodiments, in RAN 200, network layer 120 may include higher protocol layers, such as L3, and network layer 110 may include lower protocol layers, such as L1 and / or L2.

[0079] In an embodiment of serving cell selection, a machine learning model 510, executing at a lower protocol layer (e.g., L1 and / or L2), may receive measurement information related to at least one serving cell, such as the reference signal received power (RSRP), reference signal received quality (RSRQ), load, etc., of at least one serving cell. The machine learning model 510 infers a potential representation from the available information for transfer to a machine learning model 520. The machine learning model 520 may combine the potential representation with measurement information related to at least one serving cell that is available at a higher protocol layer (L3). The measurement information may include, for example, network configuration, UE capabilities, etc., to improve the serving cell selection mechanism at L3.

[0080] In some embodiments, the predicted output from the machine learning model 520 may be provided as a suggestion to the serving cell selection function 250 in network device 204 to assist the function in making a decision about the serving cell to be selected for terminal device 202. In this case, the serving cell selection function 250 may receive the predicted output from the machine learning model 520 and determine the target output based on the received predicted output. Alternatively, in some embodiments, the machine learning model 520 may be deployed as the serving cell selection function in network device 204, and therefore the predicted output from the machine learning model 520 may be directly provided as a decision about the serving cell to be selected for terminal device 202.

[0081] Through knowledge transfer, it can leverage the advantages of traditional serving cell selection mechanisms at Layer 3 and traditional L1 / 2 schedulers by utilizing machine learning techniques. Transferring information and knowledge learned from L1 / 2 machine learning models to L3 machine learning models can improve serving cell selection and enable Layer 3 to better understand the measurements and decisions made at lower layers.

[0082] Figure 5 The proposed functionality can also be applied to other cross-layer functions besides serving cell selection. In some embodiments, in the O-RAN 400 scenario, analysis task 122 is a slice resource allocation task to allocate network resources to traffic flows within a network slice. Analysis task 112 is a device resource allocation task to allocate network resources to one or more terminal devices within the network slice. The prediction output from machine learning model 520 indicates potential resource allocations for traffic flows within the network slice. The prediction output from machine learning model 510 indicates resource allocations for at least one terminal device located within the network slice. The target output indicates the target resource allocation for traffic flows within the network slice. In some embodiments, in O-RAN 400, network layer 120 may include a near-RT layer 420 or a non-RT layer 430, and network layer 110 may include an RT layer 410. In some embodiments, in O-RAN 400, network layer 120 may include a non-RT layer 430, and network layer 110 may include a near-RT layer 420. This depends on the actual network layer structure in the O-RAN.

[0083] In an embodiment of resource allocation, machine learning model 510 may receive information relating to terminal devices in a network slice available at RT layer 410 (or near RT layer 420). Machine learning model 510 infers a latent representation from the information for transfer to machine learning model 520. Machine learning model 520 may combine the latent representation with information relating to network slices and traffic flows available near RT layer 420 (or non-RT layer 430) to derive a more accurate decision regarding resource allocation for traffic flows in the network slice.

[0084] In some embodiments, the predicted output from the machine learning model 520 may be provided as a suggestion to the orchestrator 422 near the RT layer 420 (or to the orchestrator 432 outside the RT layer 430) to assist the orchestrator in resource allocation. Alternatively, in some embodiments, the machine learning model 520 may be deployed as a resource allocation function in the orchestrator 422 near the RT layer 420 (or to the orchestrator 432 outside the RT layer 430), and thus the predicted output from the machine learning model 520 may be provided directly as a decision regarding resource allocation.

[0085] According to embodiments of this disclosure, cross-layer knowledge transfer from another network layer can be leveraged to optimize and enhance machine learning-based analytics tasks (e.g., serving cell selection or slice resource allocation) at network layer 120, thereby reducing latency in decision-making. These benefits are achieved by learning analytics tasks performed at network layer 110 (e.g., L1 / L2 serving cell scheduling, or device resource allocation at RT or near RT). That is, analytics at network layer 120 is adapted to analytics at network layer 110. Furthermore, the amount of information exchanged is reduced or minimized due to the learned knowledge, or more specifically, the transfer of latent representations rather than original information. In some embodiments, machine learning models can be trained and executed in parallel without altering existing functionality in the network layers used for analytics tasks, and thus the machine learning models can be readily compatible with existing functionality.

[0086] To utilize machine learning models 510 and 520 for inference, a model training phase is first required. The trained machine learning models 510 and 520 are then provided for the model inference phase. The model training phase will be discussed first, followed by the model inference phase.

[0087] Figure 6 A schematic diagram illustrating a training process 600 for a machine learning model with knowledge transfer according to some embodiments of the present disclosure is provided. In some embodiments, the training process 600 may be implemented at a separate computing system that has access to training data for both models. This computing system may be a physical or virtualized system located within or outside a communication network in which the trained machine learning model is deployed. In some embodiments, the training process 600 may be implemented at a network device that executes the machine learning model.

[0088] The training data for machine learning model 510 may include sample model inputs 650 and corresponding ground truth outputs 652 for the labels used as sample inputs. The sample model inputs may be information available at network layer 110, and the ground truth outputs may be the ground truth outputs for information in analysis task 112. The training data for machine learning model 520 may include sample model inputs 640 and corresponding ground truth outputs 642 for the labels used as sample inputs. The sample model inputs may be information available at network layer 120, and the ground truth outputs may be the ground truth outputs for information in analysis task 122. Sample inputs and ground truth outputs for machine learning model 510 can be collected at network layer 110, and sample inputs and ground truth outputs for machine learning model 520 can be collected at network layer 120.

[0089] In the context of serving cell selection, truth value output 652 indicates whether to schedule at least one serving cell for the terminal device, and truth value output 642 indicates whether to select at least one serving cell as a potential serving cell for the terminal device. In the context of resource allocation, truth value output 652 indicates resource allocation for at least one terminal device located in a network slice, and truth value output 642 indicates potential resource allocation for traffic flows in the network slice.

[0090] In some embodiments, such as analysis task 122, analysis task 112 is related to analysis task 122 based on its output. In this case, an initialization step is required during the model training phase before training the machine learning model 510. Figure 6 As shown, an initial training phase 610 for the machine learning model 520 can be performed before the training of the machine learning model 510. During the initial training phase 610, the machine learning model 520 can be trained using supervised, semi-supervised, or unsupervised learning algorithms. When using an unsupervised learning algorithm, the model input (i.e., information 640) is used, but the true value output 642 may not be required. Through unsupervised learning, the machine learning model 520 can be trained to generate output clusters, and some domain knowledge from an expert (e.g., prior knowledge) can be applied to assign each category or label to one of the clusters. When using a semi-supervised learning algorithm, some, but not all, of the true value outputs (labels) are required. When using a supervised learning algorithm, both the model input 640 and the true value output 642 are required.

[0091] In training phase 620, the machine learning model 510 is trained using information available at network layer 110 as model input 650 and the corresponding true value output 652 of model input 650. During training phase 620, supervised learning algorithms can be used to train the machine learning model 510. Model input 650 is fed into the trained machine learning model 510 to provide a predicted output. The error between the predicted output of this model input 650 and the true value output 652 is determined. The training objective 622 of the machine learning model 510 is to reduce or minimize the error between the predicted output and the true value output by iteratively updating the model parameters of the machine learning model 510. Various training algorithms can be applied to achieve the training objective.

[0092] In some embodiments, if analysis task 112 is related to analysis task 122 based on the output of analysis task 122, the initially trained machine learning model 510 can be applied to first execute analysis task 122 in order to provide output to network layer 110 to implement analysis task 112. At this time, the implementation of analysis task 112 can be accomplished through conventional functions at network layer 110 to generate the true value output of machine learning model 510. In the context of serving cell selection, the true value output 652 can be retrieved from the serving cell scheduler at L1 or L2 layer to indicate whether at least one serving cell should be scheduled for the terminal device. In the context of resource allocation, the true value output 652 can be retrieved from the orchestrator 412 in RT layer 410 (if the device resource allocation task is implemented at this layer) or the orchestrator 422 in near RT layer 420 (if the device resource allocation task is implemented at this layer) to indicate resource allocation for at least one terminal device located in the network slice.

[0093] In training phase 630, the machine learning model 520 can be trained using at least the information available at network layer 120 as model input 640 and the corresponding true value output 642 of model input 640. If the machine learning model 520 is initially trained in phase 610, it can be further retrained in phase 630. During training phase 630, supervised learning algorithms can be used to train the machine learning model 520. Model input 640 is fed into the trained machine learning model 520 to provide a predicted output. The error between the predicted output of this model input 640 and the true value output 642 is determined. The training objective 632 of the machine learning model 520 is to reduce or minimize the error between the predicted output and the true value output by iteratively updating the model parameters of the machine learning model 520. Various training algorithms can be applied to achieve the training objective.

[0094] The expected true value output 642 is intended to indicate the actual optimized serving cell to be selected or the actual optimized resource allocation for the network slice, which may sometimes be unknown beforehand. In some embodiments, in the context of serving cell selection, the true value output 642 can be retrieved from the serving cell selection function at layer L3 to indicate whether at least one serving cell should be selected as a potential serving cell for the terminal device. In the context of resource allocation, the true value output 642 can be retrieved from the orchestrator 432 in layer 430 (if the slice resource allocation task is implemented at this layer) or the orchestrator 422 in near layer 420 (if the slice resource allocation task is implemented at this layer) to indicate the resource allocation for the terminal device located in the network slice. The serving cell selection function can be initially constructed based on expert knowledge according to a policy.

[0095] However, from a machine learning perspective, retraining the model using more data with suboptimal true values ​​can be a penalty for machine learning model algorithms that predict optimal outputs. Therefore, in some embodiments, unsupervised or semi-supervised algorithms can be applied to train machine learning model 520 to relabel the true value output 642. In some embodiments, true value output 642 can be aggregated with true value output 652 to enhance output labeling. True value aggregation will be described in detail below.

[0096] As mentioned above, the latent representation 624 of the model input (information) of machine learning model 510 will be transferred to machine learning model 520 at network layer 120. In some embodiments, the latent representation 624 can be extracted from the last layer before the output of machine learning model 510.

[0097] Specifically, for the same analytical objective in analytical tasks 112 and 122 (e.g., at least one serving cell in the context of serving cell selection, or a network slice in the context of resource allocation), a trained machine learning model 510 is used to extract a latent representation 624 of the input information related to the same analytical objective, and this latent representation 624 is provided for training the machine learning model 520. During the training phase 630, the machine learning model 520 can be trained using the input 640 (i.e., the information available at network layer 120), the latent representation 624, and the ground truth output 642. Note that the latent representation 624 associated with the same analytical objective is used in conjunction with the corresponding input 640.

[0098] Embodiments of this disclosure aim to transfer representations of information available at network layer 110 (e.g., L1 / 2 layer) for analysis tasks at network layer 120. The idea here is to train a machine learning model 510 that mimics the analysis mechanisms at network layer 110, and then transfer one or more layers before outputting from this trained machine learning model 510 to network layer 120. The transferred knowledge is represented as a latent representation. This allows machine learning model 520 to better understand the decisions made at network layer 110, which knows a synthesized version of the information at that layer.

[0099] The latent representation 624 can be integrated into the machine learning model 520 in any suitable manner. Figure 7 A schematic diagram illustrating some example integrations of the potential representation 624 from machine learning model 510 with machine learning model 520 is provided.

[0100] In example ensemble 701, latent representation 624 can be considered as an additional input to machine learning model 520. Therefore, latent representation 624 can be fed into machine learning model 520 along with model input 710 (i.e., information available at network layer 120) to generate a predicted output. For example, latent representation 624 can be concatenated with model input 710 and then fed together into machine learning model 520.

[0101] In some embodiments, latent representation 624 may be fed into one or more hidden layers of machine learning model 520. In this way, latent representation 624 may be processed together with the input of one or more hidden layers, and it may also be referred to as a latent representation extracted by machine learning model 520 from the input information available at network layer 120. In this way, latent representation 624 and the original latent representation may be further processed by subsequent layers in machine learning model 520.

[0102] In embodiments where input is fed into a hidden layer, in example ensemble 701, latent representation 624 can be fed into the last layer before the output layer of machine learning model 520 to influence the final prediction of machine learning model 510. Latent representations extracted from the input of machine learning model 510 by previous layers can be processed together with latent representation 624 in the last layer to generate a predicted output. In example ensemble 703, latent representation 624 can be fed into an earlier layer before the last layer, for example, the last layer before a dense layer in machine learning model 520. Latent representation 624 can therefore be processed together with latent representation 720 extracted from the input of machine learning model 520 by one or more previous layers.

[0103] After machine learning models 510 and 520 have been trained, two models can be provided for execution by a network device (e.g., network device 204) (if both models are trained in the context of serving cell selection) and for execution by an orchestrator 432 at a non-RT layer 430 (or an orchestrator 422 at a near-RT layer 420) in the context of resource allocation. Note that the example ensemble of latent representation 624 is applied to both the model training and model inference phases of machine learning model 520.

[0104] Example model training for cell selection For the purpose of better illustration, the training of machine learning models 510 and 520 will be described in the context of serving cell selection. A similar model training process can be applied in the context of resource allocation in O-RAN.

[0105] Figures 8A to 8DSchematic diagrams illustrating different training stages of machine learning models 510 and 520 with knowledge transfer according to some embodiments of the present disclosure are provided. In those embodiments, as mentioned, machine learning model 520 is configured to perform a serving cell selection task at layer L3, and machine learning model 510 is configured to perform a serving cell scheduling task at layers L1 / L2.

[0106] Figure 8A Example 810 illustrates the initial training phase (or warm-up phase) of machine learning model 520. Machine learning model 520 can be initially trained. To derive training data for machine learning model 520, conventional serving cell selection functions, such as serving cell selection function 250 at L3 layer, can be utilized. The input to machine learning model 520 includes measurement information available at L3 layer, which is first fed into serving cell selection function 250. Serving cell selection function 250 can be designed in any suitable manner (e.g., based on expert knowledge) to perform serving cell selection tasks based on L3 measurement information and possibly on static cell information. L3 measurement information may include, for example, network configuration (UL / DL cells, bandwidth, spectrum sharing, etc.), UE capabilities, etc. Serving cell selection function 250 outputs its serving cell selection decision to a serving cell scheduler at L1 / L2 layer 820 for serving cell scheduling. The input and output of serving cell selection function 250 can be received as training data for machine learning model 520, where the output is considered a ground truth output.

[0107] exist Figure 8A In this embodiment, a machine learning model 520 is trained at the L3 layer to learn a serving cell selection mechanism by taking it as input to the serving cell selection function 250 and outputting it for the selection of one or more potential serving cells for a given terminal device. Note that in addition to static cell information and L3 measurement information, other parameters can be added as input to the machine learning model 520. In some embodiments, the output of the machine learning model 520 can be a hard decision regarding the selected serving cell(s). In some embodiments, the output of the machine learning model 520 can be a soft vector of values ​​between 0 and 1, which can be interpreted as the probability of each cell being selected. For example, if the output vector length is 4, a hard decision output of [0, 1, 0, 0] can indicate the selection of cell number 2. If the output vector length is 4, a soft decision output of [0.1, 0.5, 0.3, 0.2] can indicate the selection of cell number 2, which has the highest probability.

[0108] In some embodiments, the machine learning model 520 may be initially trained using a supervised learning algorithm (e.g., a neural network (NN)) that requires the output of the serving cell selection function 250 as the ground truth output. Note that in this case, the machine learning model may reside in a multi-label classifier, where each category represents a serving cell, and several categories may be selected simultaneously. In some embodiments, a semi-supervised ranking algorithm may be used to initially train the machine learning model 520, which uses some labels and ranks from the output of the serving cell selection function 250 to help adapt to the ground truth labels.

[0109] In some embodiments, an unsupervised learning algorithm (e.g., a clustering algorithm) can be used to initially train the machine learning model 520. This unsupervised learning algorithm can be trained based on historical data from L3 measurement information (including network configuration and UE capabilities) from different terminal devices to define clusters, each representing a serving cell. The labels of the serving cells can be mapped to clusters from the output of the serving cell selection function 250 and used to help adapt to real values. Then, for new L3 measurement information from the terminal device, the algorithm can select the cluster(s) closest in distance (and their corresponding serving cells) and pass them to the L1 / L2 scheduler. The unsupervised algorithm can combine information from radio parameters (DMRS, modulation, etc.) and information from the current implementation in the product, and it consists of rule-based methods, particularly for defining the distance between a measurement and a cluster. The choice of unsupervised algorithm can depend on the distribution of the data and on time constraints. For example, if the clusters are well separated and balanced (each cluster has approximately the same number of examples), a K-means or density-based spatial clustering with noise application (DBSCAN) method can be used. Alternatively, more advanced techniques can be used, such as neural networks or spectral clustering designed to both classify and learn rules. Of course, any other suitable unsupervised learning algorithm can also be applied.

[0110] Figure 8BExample 811 illustrates the training phase of machine learning model 510. Machine learning model 510 is trained at L1 / L2 layer 820 to learn serving cell scheduling mechanisms. The inputs to machine learning model 510 include L1 and / or L2 measurement information at L1 / L2 layer 820, which may include, for example, RSRP, RSRQ, signal-to-interference-plus-noise ratio (SINR), CA / class, buffer state record, channel rank, utilized MIMO, and / or the like. Furthermore, the inputs to machine learning model 510 also include serving cell selection decisions from L3 layer. The serving cell scheduling task is to determine which(s) of the selected serving cells should be scheduled and activated for the terminal device. In some embodiments, the input serving cell decision can be provided from the initially trained machine learning model 520. In some embodiments, the input serving cell decision can be provided from the serving cell selection function 250.

[0111] The output of machine learning model 510 can thus indicate at least one serving cell to be scheduled for the terminal device, for example, the pCell to be scheduled and (one or more) possible sCells. In some embodiments, this output can be determined as a vector of values ​​between 0 and 1, which can be interpreted as the probability that each cell should be scheduled. If machine learning model 510 replaces a conventional serving cell scheduler, this output can be used for scheduling. Alternatively, this output can be provided as a suggestion to the serving cell scheduler. Furthermore, the latent representation of the input information extracted by machine learning model 510 can be transferred to an L3 layer for use in machine learning model 520. In some embodiments, the latent representation can be compressed in time, for example, as a size of ( L , n A two-dimensional (2D) array, in which L It is the size of the latent representation, and n It represents the number of time instances within the time period in the input.

[0112] In some embodiments, a supervised learning algorithm (e.g., a neural network (NN)) can be used to train the machine learning model 510, which requires the output of a conventional serving cell scheduler at the L1 / L2 layers as the true value output. In this case, the latent representation could be the last layer before classifying the selected serving cells. The advantage of such a representation is that it takes the three-dimensional (3D) input size of the machine learning model 510 ( d , t , n ) decrease to a size of ( L , n ) 2D array, where d It is the number of input features of the model. t It is the size of the time frame, andn The number of time instances used for training or decision-making needs to be considered, where L This is the output size of the last hidden layer. This results in a reduction of important dimensions because... L Typically much smaller than the original input size, i.e. L << d * t In some embodiments, the machine learning model 510 may apply dimensionality reduction techniques, such as projection onto a manifold. However, using a supervised neural network should be more efficient, as it proposes potential representations suitable for the task of scheduling and selecting serving cells.

[0113] The primary purpose of training the machine learning model 510 is to transfer information from the L1 / L2 layer to the L3 layer. This will be useful for enhancing the decision-making process for L3 serving cell selection, while reducing cross-layer signaling overhead, memory consumption, and time consumption added from this information transfer. Note that there is no need to transfer the input to the machine learning model 510 here, as the information is summarized by its latent representation.

[0114] In some embodiments, the machine learning model 510 is trained using data from several terminal devices across a network of several serving cells, and thus information from unselected cells can still be obtained from these other terminal devices, and the ML algorithm can adapt to unseen conditions. The training process from different terminal devices and different cells can lead to good stability and optimality in cell selection / scheduling, because information from different but similar terminal devices can enhance the decision for a particular terminal device.

[0115] Figure 8C Example 812 illustrates a trained machine learning model 510 and a KT function 512 preparing inference based on L1 / L2 measurement information for the L3 layer. Figure 8D Example 813 illustrates the (re)training of a machine learning model utilizing latent representations transferred from L1 / L2 layers. At this stage, a supervised learning algorithm (e.g., a neural network) can be used to train the machine learning model 520 to enrich the feature space by integrating latent representations from L1 / L2 layers. Integrating latent representations into the machine learning model 520 can be referenced in… Figure 7 Regarding the description above, in some embodiments, ensemble is simply concatenating the latent representation from machine learning model 510 with the expectation layer of machine learning model 520. The values ​​of the latent representation are frozen, and during training, only backpropagation is needed to train the weights to and from these values ​​of the latent representation. If the latent representation is concatenated with the input of machine learning model 520 (as in example ensemble 701), then only the weights from the latent representation need to be trained.

[0116] In some embodiments, the ground truth output at Layer 3 can be modified to improve the ground truth of the machine learning model 520 by leveraging the ensemble of latent representations, since the original ground truth at Layer 3 can be derived from conventional serving cell selection functionality. In this case, for a given terminal device, the ground truth output at Layer 3 (indicating which serving cell to select) and the ground truth output at Layers 1 / L2 (indicating which serving cell to schedule) can be aggregated to generate a new aggregated ground truth output for the terminal device (sometimes referred to as the third ground truth output). The aggregated ground truth output indicates whether at least one serving cell should be selected as a potential serving cell for the terminal device within a time period. The aggregated ground truth output is then used to train the machine learning model 520. Figure 8D As shown, ground truth outputs are received from the L1 / L2 layers to improve the ground truth output at the L3 layer for training the machine learning model 530. Typically, if a serving cell is actually scheduled for the terminal device, this cell can be considered the best cell for the terminal device. Using the ground truth outputs from the L1 / L2 layers, better ground truth in serving cell selection can guide the learning of the model at the L3 layer.

[0117] In some embodiments, aggregation may take into account the differences in temporal granularity between L1 / L2 and L3 layers, making the result consistent with the temporal granularity at L3 layer. The truth value output at L1 / L2 layers may indicate whether to schedule at least one serving cell for a terminal device across multiple time instances within a time period, while the truth value output at L3 layer may indicate a serving cell decision within that time period. Truth value aggregation may include averaging the initial L3 truth value output using a weighted average of the L1 / L2 truth value outputs. In some embodiments, the weight of each L1 / L2 truth value output is uniform and accounts for 50% of the final truth value output. In some embodiments, majority voting, consisting of the most active cells during the considered time period, may be used to summarize the L1 truth values. Note that truth value aggregation can be similarly applied to resource allocation in O-RAN.

[0118] In some cases, the scheduling of one or more “best” serving cells in Layer 1 / 2 is conditioned on pre-selected cells from an initially trained machine learning model 520 at Layer 3, which is performed to avoid latency due to computation time and / or resource requirements. However, this may result in the selection of cells that are not the most suitable for the terminal device compared to cells that might be selected if L1 and L3 input information were computed on all available serving cells in the network. In some embodiments, multi-source multi-target transfer is proposed to mimic the L1 / L2 scheduling mechanism if it can measure all available serving cells rather than just selected cells.

[0119] To complete the overall information about different cells in L3 (or L1 / L2 respectively), it is proposed to use multi-source multi-objective domain adaptation between one or more pre-selected serving cells (with available measurement information) and one or more unselected serving cells (without measurement information for transition to L3). Multi-source multi-objective domain adaptation can be performed either at L1 / L2 or L3 to estimate the potential representation of the serving cells that will be unmeasured.

[0120] The idea here is to train a small machine learning model 510 for each cell in the L1 / L2 layers. Each machine learning model 510 can be considered specific to a corresponding serving cell. Each machine learning model 510 has information about a specific serving cell as its input and outputs whether to select the corresponding serving cell. In some embodiments, the machine learning models 510 for different serving cells can have the same model architecture to simplify learning, although different model architectures may also be applicable. Then, one or more layers are guided from these trained machine learning models 510 to the L3 layer to enable the L3 layer to better understand the decisions and measurements at the L1 / L2 layers.

[0121] Figure 9 This illustration depicts domain adaptation between pre-selected and unselected cells according to some embodiments of the present disclosure. As shown, a machine learning model 510-1 in the L1 / 2 layer is specifically configured for cell 1, and a machine learning model 510-m in the L1 / 2 layer is specifically configured for cell m. Cell 1 and cell m are serving cells pre-selected for a terminal device by an initially trained machine learning model 520, and therefore can use available measurement information (to compare with...) Figure 8BThe machine learning models 510-1, ..., 510-m are trained in a manner similar to that shown in the diagram. Cells m+1 and N are serving cells that have not yet been selected, and therefore measurement information related to those cells is unavailable for transfer to the L3 layer. Note that each of the machine learning models 510-1, ..., 510-m, 510-(m+1), ..., 510-N can operate in a manner similar to that of the machine learning model 510 discussed above, which is executed at the L3 layer. In some embodiments, the machine learning models 510-1, ..., 510-m, 510-(m+1), ..., 510-N can be configured as smaller models or neural networks, as each of them corresponds to a single serving cell. Naturally, the output layer of those models would therefore consist of a single neuron that outputs a soft value between 0 and 1 to select a serving cell based on its measurement information. The associated latent representation is therefore specific to the serving cell. The size of the potential representation L' can be smaller than that discussed above for machine learning model 510, but the information to be transferred to the L3 layer can therefore be L' times the number of pre-selected cells.

[0122] During the entire model training process, after one or more machine learning models specific to certain service cells have been trained (e.g. Figure 8B (in Chinese) and before training the machine learning model 520 (such as...) Figure 8C In this context, multi-source, multi-target domain adaptation can be applied between pre-selected and unselected cells. A mapping or projection function 910 is proposed for domain adaptation from a source domain (e.g., where cells 1, ..., m are located) to a target domain (e.g., where cells m+1, ..., N are located). The latent representations of unselected serving cells (cells m+1, ..., N) are unavailable because measurement information associated with those cells is unavailable. The mapping or projection function 910 is configured to estimate the latent representations of unselected serving cells (e.g., cells m+1, ..., N) based on the latent representations of pre-selected serving cells (e.g., cells 1, ..., m). Figure 9 As described herein, the mapping or projection function 910 takes data from pre-selected cells (cell 1, ..., cell m) as the source domain and aims to map or project them onto unselected cells (cell m+1, ..., cell N), which are considered to be different targets.

[0123] More specifically, the latent representation of the available measurement information (which can be extracted by trained machine learning models 510-1, ..., 510-m) can be mapped or projected to generate a latent representation of the unselected serving cells. Then, an estimated latent representation of the unselected serving cells can be provided to train machine learning model 520 at layer L3. In this way, machine learning model 520 can also learn about the unselected serving cells.

[0124] In some embodiments, multi-source multi-target domain adaptation can be performed at the L1 / L2 layers, and an estimated potential representation can be provided to the L3 layer. Both the potential representation of the pre-selected serving cell and the estimated potential representation of the unselected serving cell are transferred to the L3 layer. Alternatively, in some embodiments, multi-source multi-target domain adaptation can be performed at the L3 layer after receiving the potential representation of the pre-selected serving cell from the L1 / L2 layers. In the latter case, some additional information from the network configuration and UE capabilities at the L3 layer can be used to further implement domain adaptation. Note that the choice between the two options for multi-source multi-target domain adaptation depends on the computational capabilities at the L1 / L2 layers and the overhead required to send all potential representations from the L1 / L2 layers to the L3 layer. The latter option reduces overhead.

[0125] In some embodiments, if the machine learning model 520 used for the L3 layer is initially trained, the same machine learning model 520 as the one initially trained can be retrained using the latent representation, L3 measurement information, and the corresponding ground truth output. In some embodiments, the machine learning model 520 used at the L3 layer may be different from the initially trained machine learning model, and the machine learning model 520 used at the L3 layer can be trained from scratch, or the machine learning model 520 used at the L3 layer can be warm-started from the initially trained machine learning model.

[0126] Example signaling for model training for cell selection Figure 10 Signaling diagram 1000 illustrates a method for training a machine learning model using knowledge transfer, according to some other embodiments of this disclosure. In the illustrated embodiment, it is assumed that the training of the machine learning model is implemented at a network device having a centralized unit (CU) 902 and a distributed unit (DU) 904. CU 902 corresponds to Layer 3. DU 904 includes a DU converter connected to CU 902 via an interface and an L1 / L2 scheduler for serving cell scheduling at Layer 1 / L2. It will be appreciated that signaling diagram 1000 provides an example of training the machine learning model proposed herein. As mentioned, the training of the machine learning model can be implemented in other ways.

[0127] In signaling diagram 1000, CU 902 performs (905) pre-training of machine learning model 520 and transmits (910) the output of the initially trained machine learning model 520 to L1 / L2 scheduler 908. The output of the initially trained machine learning model 520 indicates whether one or more serving cells are pre-selected for terminal device 909. By receiving (915) the output of the initially trained machine learning model 520, L1 / L2 scheduler 908 performs (920) cell scheduling for terminal device 909, for example, scheduling one or more pre-selected serving cells for terminal device 909.

[0128] Scheduling results can be collected as ground truth outputs to train (925) machine learning model 510 at L1 / L2 scheduler 908. Training data for machine learning model 510 may further include L1 / L2 measurement information available at the L1 / L2 layers. After training, the trained machine learning model 510 is used to extract latent representations of the L1 / L2 measurement information at L1 / L2 scheduler 908. L1 / L2 scheduler 908 transfers (930) the latent representations to DU converter 906. In some embodiments, L1 / L2 ground truth outputs may also be transferred to DU converter 906. By receiving (935) the required information, DU converter 906 performs (940) retraining of machine learning model 520. In some embodiments, L1 / L2 ground truth outputs may be aggregated with L3 ground truth outputs to generate aggregated ground truth outputs for training. The retrained machine learning model 520 is provided (945) to CU 902. By receiving (950) a trained machine learning model 520, CU 920 can use this model to implement serving cell selection.

[0129] The training of the machine learning model for resource allocation in the O-RAN use case can be similar to the training of the machine learning model for resource allocation discussed above in the serving cell selection use case.

[0130] Example Model Inference After the machine learning models are trained and ready to use, model inference can be similar, and the steps are the same, except that it does not require the output of true values. During the model inference phase, the inputs and outputs of machine learning models 510 and 520 are similar to those discussed above. The latent representation is transferred from the trained machine learning model 510 to machine learning model 520.

[0131] In some cases, such as those mentioned, analysis task 112 is related to analysis task 122 based on its output. In this case, an initialization step is required during the model inference phase to allow machine learning model 510 to first extract latent representations for use by a second machine learning model. The initialization step can be performed either by conventional network functions of analysis task 122 in network layer 120 or by machine learning model 520. The output of analysis task 122 is provided from network layer 120 to network layer 110. The trained machine learning model 510 can then generate the predicted output of analysis task 112 based on the input information and the predicted output of analysis task 122.

[0132] For example, in RAN 200, a trained machine learning model 520 is used at time steps. t -1 is used to perform the serving cell preselection process. After that, the preselection in L3 is based on the output at time t-1, and the preselection in L3 is updated at time t using L1 / L2 knowledge transfer.

[0133] More specifically, as an initialization step, the trained machine learning model 520 can first be used at network layer 120 to generate the predicted output for analysis task 122 based on the information available at network layer 120. The predicted output is passed to network layer 110 as part of the model input of the trained machine learning model 510. In this way, the trained machine learning model 510 can have complete input to extract latent representations for transfer to network layer 120.

[0134] In some embodiments, considering multi-source, multi-objective domain adaptation in the use case of serving cell selection, if the trained machine learning model 510 is specific to a first serving cell and the first latent representation is specific to the first serving cell, and if measurement information related to a second serving cell is unavailable, the first latent representation of the first measurement information related to the first serving cell can be mapped to a third latent representation of the second serving cell at layer L3. Alternatively, the third latent representation of the second serving cell can be estimated at layers L1 / L2 based on the first latent representation of the first serving cell, and the third latent representation of the second serving cell can be transferred to layer L3. Using the third latent representation of the second serving cell as part of its model input, the machine learning model 520 can generate a predictive output indicating whether the second serving cell should be selected as a potential serving cell for the terminal device.

[0135] In some embodiments, as mentioned, the predicted output from the machine learning model 520 may be provided as a suggestion for determining the target output, or the predicted output from the machine learning model 520 may be directly determined as the target output. Figure 11Aand Figure 11B This diagram illustrates the reasoning process of a machine learning model with knowledge transfer in the use case of serving cell selection.

[0136] exist Figure 11A In Example 1101, a serving cell selection function 250 is applied at Layer 3 using a decision provider 1112 to provide serving cell selection decisions for cell scheduling at Layers 1 / 2 820. A trained machine learning model 510 at Layers 1 / 2 extracts latent representations of L1 / L2 measurement information and transfers these latent representations to the machine learning model at Layer 3 via a KT function 512. The machine learning model 520 can generate a predictive output regarding whether to select one or more serving cells for the terminal device and provides this predictive output as a suggestion to the serving cell selection function 250. The serving cell selection function 250 can generate serving cell selection decisions based on conventional inputs such as static cell information, L3 measurement information, and suggestions from the machine learning model 520. In this example, suggestions from the machine learning model 520 can be easily excluded or included in the decision regarding serving cell selection.

[0137] exist Figure 11B In Example 1102, machine learning model 520 is used in place of serving cell selection function 250 at L3 layer. The input to machine learning model 520 may include L3 measurement information, potential representations from L1 / L2 layers, and possible static cell information. Decision provider 1114 is configured to provide the predicted output of machine learning model 520 as a serving cell selection decision for cell scheduling at L1 / L2 layers 820.

[0138] The deployment and use of machine learning models 510 and 520 can be determined according to actual requirements, and are not limited to the scope of this disclosure.

[0139] Example Method Figure 12 A flowchart illustrating a method 1200 implemented at a network device according to some embodiments of the present disclosure is provided. Figure 2 Network device 204 and / or Figure 4 Method 1200 is implemented at position 432 or 422 of the arranger.

[0140] At box 1210, a first latent representation is received at the second network layer and from the first network layer. The first information is available at the first network layer. The first latent representation is extracted by a first machine learning model trained at the first network layer, and the first machine learning model is configured to generate a first prediction output for a first analysis task based on the first information. The second network layer is different from the first network layer.

[0141] At box 1220, at the second network layer and using a trained second machine learning model, a second prediction output for the second analysis task is generated based on the first latent representation and the second information available at the second network layer.

[0142] In some embodiments, the first latent representation is fed into a second machine learning model along with the second information. In some embodiments, the first latent representation and the second latent representation of the second information are processed by the second machine learning model to generate a second prediction output.

[0143] In some embodiments, the first analysis task is based on the output of the second analysis task. In some embodiments, a third predicted output of the second analysis task is generated at the second network layer based on third information available at the second network layer. In some embodiments, the first machine learning model is configured to generate a first predicted output of the first analysis task based on the first information and the third predicted output.

[0144] At box 1230, in the second network layer, the target output of the second analysis task is determined based on the second predicted output. In some embodiments, the second predicted output is provided as a suggestion for determining the target output, or the second predicted output is determined as the target output.

[0145] In some use cases, a first prediction output indicates whether to schedule at least one serving cell for a terminal device, a second prediction output indicates whether to select at least one serving cell as a potential serving cell for the terminal device, and a target output indicates a subset of serving cells selected from the set of available serving cells for the terminal device. In those use cases, in some embodiments, the first network layer includes a lower protocol layer, and the second network layer includes a higher protocol layer. In some embodiments, the first information includes first measurement information relating to at least one serving cell at the lower protocol layer, and the second information includes second measurement information relating to at least one serving cell at the higher protocol layer.

[0146] In some embodiments, a first machine learning model is trained for a first serving cell, and a first latent representation is specific to the first serving cell. In some embodiments, based on a determination that measurement information related to a second serving cell is unavailable, a first latent representation of the first measurement information related to the first serving cell is mapped to a third latent representation of the second serving cell at a higher protocol layer. In some embodiments, the third latent representation of the second serving cell is received from a lower protocol layer at a higher protocol layer.

[0147] In some embodiments, a second prediction output is further generated based on a third potential representation to further indicate whether a second serving cell should be selected as a potential serving cell for the terminal device.

[0148] In some use cases, a first prediction output indicates resource allocation for at least one terminal device located in a network slice, a second prediction output indicates potential resource allocation for traffic flows in the network slice, and a target output indicates a target resource allocation for traffic flows in the network slice. In those use cases, in some embodiments, a first network layer includes a real-time layer in an Open Radio Access Network (O-RAN), and a second network layer includes a non-real-time layer or a near-real-time layer in the O-RAN. In those use cases, in some embodiments, a first network layer includes a near-real-time layer in the O-RAN, and a second network layer includes a non-real-time layer in the O-RAN.

[0149] Figure 13 A flowchart illustrating a method for model training according to some embodiments of this disclosure is provided. Figure 2 Network device 204 in Figure 4 The programmable unit 432 or 422, or an external computing system / device (such as...) Figure 15 The virtualization system 1500 implements method 1300.

[0150] At box 1310, a first machine learning model is trained using first information available at a first network layer as the first ground truth output and model input of the first information in a first analysis task. At box 1320, the trained first machine learning model is used to extract a first latent representation of the first information. At box 1330, a second machine learning model is trained using second information available at a second network layer, the first latent representation, and the second ground truth output of the second information in the second analysis task. At box 1340, the trained first machine learning model and the trained second machine learning model are provided for execution by a network device.

[0151] In some embodiments, the first latent representation is fed into a second machine learning model along with the second information. In some embodiments, the first latent representation and the second latent representation of the second information are processed by the second machine learning model to generate a model output.

[0152] In some embodiments, the initial training of the second machine learning model is performed using at least third information available at the second network layer before training the first machine learning model. In those embodiments, the second machine learning model is retrained using the second information, the first latent representation, and the second ground truth output.

[0153] In some use cases, a first truth value output indicates whether to schedule at least one serving cell for a terminal device, and a second truth value output indicates whether to select at least one serving cell as a potential serving cell for the terminal device. In some embodiments, the first information includes first measurement information relating to at least one serving cell at a lower protocol layer, and the second information includes second measurement information relating to at least one serving cell at a higher protocol layer.

[0154] In some embodiments, a first truth value output is retrieved from a serving cell scheduler at a lower protocol layer, and the first truth value output indicates whether at least one serving cell should be scheduled for the terminal device. In some embodiments, a second truth value output is retrieved from a serving cell selection function at a higher protocol layer, and the second truth value output indicates whether at least one serving cell should be selected as a potential serving cell for the terminal device.

[0155] In some embodiments, the first true value output indicates whether to schedule at least one serving cell for the terminal device within a corresponding time instance during a time period. In some embodiments, the first true value output and the second true value output are aggregated to obtain an aggregated true value output, which indicates whether to select at least one serving cell as a potential serving cell for the terminal device within a time period. In some embodiments, the aggregated true value output is used to train a second machine learning model.

[0156] In some embodiments, a first machine learning model is trained for a first serving cell, and a first latent representation is specific to the first serving cell. In those embodiments, based on the determination that measurement information related to a second serving cell is unavailable, a first latent representation of the first measurement information related to the first serving cell is mapped to a third latent representation of the second serving cell. In some embodiments, the third latent representation is used to further train a second machine learning model.

[0157] In some use cases, a first truth value output indicates resource allocation for at least one terminal device located in a network slice, and a second truth value output indicates potential resource allocation for traffic flows in the network slice. In those use cases, the first network layer includes a lower protocol layer, and the second network layer includes a higher protocol layer. In those use cases, in some embodiments, the first network layer includes a real-time layer in an Open Radio Access Network (O-RAN), and the second network layer includes a non-real-time layer or a near-real-time layer in the O-RAN. In some embodiments, the first network layer includes a near-real-time layer in the O-RAN, and the second network layer includes a non-real-time layer in the O-RAN.

[0158] Example device / system Figure 14A simplified block diagram of a network device 1400 suitable for implementing embodiments of the present disclosure is illustrated. As used herein, a network device refers to a device capable of, configured to, arranged to, and / or operable to communicate directly or indirectly with terminal devices and / or with other network devices or equipment in a telecommunications network. Examples of network devices include, but are not limited to, access points (APs) (e.g., radio access points) and base stations (BSs) (e.g., radio base stations, Node Bs, evolved Node Bs (eNBs), and NRNode Bs (gNBs)). In some embodiments of the present disclosure, the FSF 210 and / or model provider 110, as discussed above, may be implemented as or included in network device 1400.

[0159] Base stations can be classified based on the coverage they provide (or, in other words, their transmit power levels), and therefore, depending on the coverage provided, a base station can be referred to as a femtobase, picobase, microbase, or macrobase. A base station can be a relay node or a relay donor node controlling a relay. The network apparatus can also include one or more (or all) portions of a distributed radio base station such as a centralized digital unit and / or a remote radio unit (RRU), sometimes referred to as a remote radio headend (RRH). Such a remote radio unit may or may not be integrated with an antenna as an antenna-integrated radio apparatus. A portion of a distributed radio base station can also be referred to as a node in a distributed antenna system (DAS).

[0160] Other examples of network devices include multi-transmission point (multi-TRP) 5G access nodes, multi-standard radio (MSR) devices such as MSR BS, network controllers such as radio network controllers (RNC) or base station controllers (BSC), base transceiver stations (BTS), transmission points, transmission nodes, multi-cell / multicast coordination entities (MCE), operation and maintenance (O&M) nodes, operation support system (OSS) nodes, self-organizing network (SON) nodes, location nodes (e.g., evolved servicing mobile location centers (E-SMLC)), and / or minimized drive tests (MDT).

[0161] Network device 1400 includes processing circuitry (including one or more processors) 1402, memory 1404, communication interface 1406, and power supply 1408. Network device 1400 may consist of multiple physically separate components (e.g., NodeB components and RNC components, or BTS components and BSC components, etc.), each of which may have its own corresponding components. In some scenarios where network device 1400 includes multiple separate components (e.g., BTS and BSC components), one or more of the separate components may be shared among several network devices. For example, a single RNC can control multiple NodeBs. In such scenarios, each unique NodeB and RNC pair may be considered a single separate network device in some instances. In some embodiments, network device 1400 may be configured to support multiple Radio Access Technologies (RATs). In such embodiments, some components may be duplicated (e.g., separate memory 1404 for different RATs) and some components may be reused (e.g., the same antenna 1410 may be shared by different RATs). Network device 1400 may also include multiple sets of components for integrating various wireless technologies, such as GSM, WCDMA, LTE, NR, WiFi, Zigbee, Z-wave, LoRaWAN, radio frequency identification (RFID), or Bluetooth wireless technologies. These wireless technologies may be integrated into the same or different chips or chipsets and other components within network device 1400.

[0162] Processing circuitry 1402 may include a combination of one or more of a microprocessor, controller, microcontroller, central processing unit, digital signal processor, application-specific integrated circuit, field-programmable gate array, or any other suitable computing device or resource, or a combination of hardware, software, and / or coding logic operable to provide functionality of network device 1400, either alone or in combination with other network device 1400 components such as memory 1404.

[0163] In some embodiments, the processing circuitry 1402 includes a system-on-a-chip (SOC). In some embodiments, the processing circuitry 1402 includes one or more of a radio frequency (RF) transceiver circuitry 1412 and a baseband processing circuitry 1414. In some embodiments, the RF transceiver circuitry 1412 and the baseband processing circuitry 1414 may be on separate chips (or chipsets), boards, or units such as radio units and digital units. In alternative embodiments, some or all of the RF transceiver circuitry 1412 and the baseband processing circuitry 1414 may be on the same chip or chipset, board, or unit.

[0164] Memory 1404 may include any form of volatile or non-volatile computer-readable memory, including, but not limited to, permanent storage devices, solid-state memory, remotely mounted memory, magnetic media, optical media, random access memory (RAM), read-only memory (ROM), mass storage media (e.g., hard disk), removable storage media (e.g., flash drives, CDs, or DVDs), and / or any other volatile or non-volatile, non-transitory device-readable and / or computer-executable memory that stores information, data, and / or instructions that can be used by processing circuitry 1402. Memory 1404 may store any suitable instructions, data, or information, including applications, software, computer programs, and / or other instructions that contain one or more of logic, rules, codes, tables, and can be executed by processing circuitry 1402 and utilized by network device 1400. Memory 1404 may be used to store any calculations performed by processing circuitry 1402 and / or any data received via communication interface 1406. In some embodiments, processing circuitry 1402 and memory 1404 are integrated.

[0165] Communication interface 1406 is used in wired or wireless communication of signaling and / or data between network devices, access networks, and / or terminal devices. As illustrated, communication interface 1406 includes one or more ports / terminals 1416 for transmitting data to and receiving data from the network, for example, via a wired connection. Communication interface 1406 also includes radio front-end circuitry 1418 that may be coupled to antenna 1410 or, in some embodiments, is part of antenna 1410. Radio front-end circuitry 1418 includes filter 1420 and amplifier 1422. Radio front-end circuitry 1418 may be connected to antenna 1410 and processing circuitry 1402. Radio front-end circuitry 1418 may be configured to modulate the signal transmitted between antenna 1410 and processing circuitry 1402. Radio front-end circuitry 1418 may receive digital data to be transmitted wirelessly to other network devices or terminal devices. Radio front-end circuitry 1418 may use a combination of filter 1420 and / or amplifier 1422 to convert digital data into radio signals with appropriate channel and bandwidth parameters. Radio signals can then be transmitted via antenna 1410. Similarly, upon receiving data, antenna 1410 can collect radio signals and then convert them into digital data via radio front-end circuitry 1418. The digital data can then be transmitted to processing circuitry 1402. In other embodiments, the communication interface may include different components and / or different combinations of components.

[0166] In some alternative embodiments, network device 1400 does not include a separate radio front-end circuitry 1418; instead, processing circuitry 1402 includes radio front-end circuitry and is connected to antenna 1410. Similarly, in some embodiments, all or some of RF transceiver circuitry 1412 is part of communication interface 1406. In other embodiments, communication interface 1406 includes one or more ports or terminals 1416, radio front-end circuitry 1418, and RF transceiver circuitry 1412 as part of a radio unit (not shown), and communication interface 1406 communicates with baseband processing circuitry 1414, which is part of a digital unit (not shown).

[0167] Antenna 1410 may include one or more antennas or antenna arrays configured to transmit and / or receive wireless signals. Antenna 1410 may be coupled to radio front-end circuitry 1418 and may be any type of antenna capable of wirelessly transmitting and receiving data and / or signals. In some embodiments, antenna 1410 is separate from network device 1400 and may be connected to network device 1400 via an interface or port.

[0168] Antenna 1410, communication interface 1406, and / or processing circuitry 1402 can be configured to perform any receive operation and / or certain acquire operation described herein as being performed by a network device. Any information, data, and / or signals can be received from a terminal device, another network device, and / or any other network equipment. Similarly, antenna 1410, communication interface 1406, and / or processing circuitry 1402 can be configured to perform any transmit operation described herein as being performed by a network device. Any information, data, and / or signals can be transmitted to a terminal device, another network device, and / or any other network equipment.

[0169] Power supply 1408 provides power to the various components of network device 1400 in a form suitable for the respective components (e.g., at the voltage and current levels required by each respective component). Power supply 1408 may further include or be coupled to power management circuitry to power the components of network device 1400 for performing the functionality described herein. For example, network device 1400 may be connectable to an external power source (e.g., mains, electrical outlet) via an input circuitry or interface such as a cable, thereby supplying power to the power circuitry of power supply 1408. As another example, power supply 1408 may include a power source in the form of a battery or battery pack, connected to or integrated into the power circuitry. The battery can provide backup power in the event of an external power failure.

[0170] Embodiments of network device 1400 may include, except Figure 14Additional components beyond those shown herein are used to provide certain aspects of the functionality of the network device, including any functionality described herein and / or any functionality necessary to support the subject matter described herein. For example, network device 1400 may include user interface devices for allowing information to be input into and output from network device 1400. This allows users to perform diagnostic, maintenance, repair, and other management functions for network device 1400.

[0171] Figure 15 This is a block diagram illustrating a virtualization environment 1500 in which functionality implemented by some embodiments can be virtualized. In this context, virtualization means creating a virtual version of a device or apparatus that may include a virtualized hardware platform, storage devices, and networking resources. As used herein, virtualization can be applied to any apparatus or component thereof described herein and relates to an implementation where at least a portion of its functionality is implemented as one or more virtual components. Some or all of the functionality described herein can be implemented as virtual components executed by one or more virtual machines (VMs) implemented in one or more virtualization environments 1500 hosted by one or more hardware nodes, such as hardware computing devices operating as network nodes, UEs, core network nodes, or hosts. Furthermore, in embodiments where virtual nodes do not require radio connectivity (e.g., core network nodes or hosts), the nodes can be fully virtualized.

[0172] Running application 1502 (which may alternatively be referred to as a software instance, virtual device, network function, virtual node, virtual network function, etc.) in a virtualized environment Q400 to implement some of the features, functions, and / or benefits of some of the embodiments disclosed herein.

[0173] Hardware 1504 includes processing circuitry, memory storing software and / or instructions executable by the hardware processing circuitry, and / or other hardware devices as described herein, such as network interfaces, input / output interfaces, etc. The processing circuitry can execute software to instantiate one or more virtualization layers 1506 (also referred to as a hypervisor or virtual machine monitor (VMM)), provide VMs 1508a and 1508b (one or more of which may be generally referred to as VM 1508), and / or perform any of the functions, features, and / or benefits described in relation to some embodiments described herein. Virtualization layer 1506 can present a virtual operating platform to VM 1508 that appears to be networked hardware.

[0174] VM 1508 includes virtual processing, virtual memory, virtual networking or interfaces, and virtual storage devices, and can run through a corresponding virtualization layer 1506. Different embodiments of instances of virtual device 1502 can be implemented on one or more VMs within VM 1508, and can be implemented in different ways. Hardware virtualization is referred to in some contexts as Network Functions Virtualization (NFV). NFV can be used to consolidate many types of network devices into industry-standard high-capacity server hardware, physical switches, and physical storage devices that can be located in data centers and customer premises.

[0175] In the context of NFV, VM 1508 can be a software implementation of a physical machine that runs programs as if they were executing on a physical, non-virtualized machine. Each VM in VM 1508, and the portion of hardware 1504 that executes that VM, whether it is hardware dedicated to that VM and / or hardware shared by that VM and other VMs in VM, forms a separate virtual network element. Still in the context of NFV, the virtual network function is responsible for handling specific network functions running in one or more VMs 1508 on top of hardware 1504 and corresponds to application 1502.

[0176] Hardware 1504 can be implemented in a standalone network node with general or specific components. Hardware 1504 can utilize virtualization to implement some functions. Alternatively, hardware 1504 can be part of a larger hardware cluster (e.g., in a data center or CPE) where many hardware nodes work together and are managed via management and orchestration 1510, which, among other things, oversees the lifecycle management of application 1502. In some embodiments, hardware 1504 is coupled to one or more radio units, each including one or more transmitters and one or more receivers that can be coupled to one or more antennas. The radio units can communicate directly with other hardware nodes via one or more suitable network interfaces and can be combined with virtual components to provide radio capabilities to virtual nodes, such as radio access nodes or base stations. In some embodiments, a control system 1512 can be used to provide signaling, which can alternatively be used for communication between hardware nodes and radio units.

[0177] In some example embodiments, a device capable of performing any of the methods in method 1200 (e.g., Figure 2 Network device 204 and / or Figure 4The arranger (432 or 422) may include components for performing the corresponding operations of method 1200. The components may be implemented in any suitable form. For example, the components may be implemented in a circuit or software module. The device may be implemented as network device 204 and / or arranger 432 or 422 or included in network device 204 and / or arranger 432 or 422.

[0178] In some example embodiments, the device includes: components for receiving a first latent representation of first information available at the first network layer from a first network layer, the first latent representation being extracted by a trained first machine learning model executed at the first network layer, the first machine learning model being configured to generate a first predictive output for a first analysis task based on the first information; components for generating a second predictive output for a second analysis task at a second network layer and using a trained second machine learning model, based on the first latent representation and second information available at the second network layer, the second network layer being different from the first network layer; and components for determining a target output for the second analysis task at the second network layer based on the second predictive output, wherein the first predictive output indicates whether to schedule at least one serving cell for a terminal device, the second predictive output indicates whether to select at least one serving cell as a potential serving cell for the terminal device, and the target output indicates a subset of serving cells selected from a set of available serving cells for the terminal device, or wherein the first predictive output indicates resource allocation for at least one terminal device located in a network slice, the second predictive output indicates potential resource allocation for a traffic flow in the network slice, and the target output indicates target resource allocation for a traffic flow in the network slice. In some example embodiments, the second predicted output is provided as a suggestion for determining the target output, or the second predicted output is determined as the target output.

[0179] In some example embodiments, the first latent representation is fed into a second machine learning model along with the second information. In some example embodiments, the first latent representation and the second latent representation of the second information are processed by the second machine learning model to generate a second prediction output.

[0180] In some example embodiments, the first analysis task is based on the output of the second analysis task, and the device further includes: a component for generating a third predicted output of the second analysis task at a second network layer based on third information available at the second network layer. In some example embodiments, the first machine learning model is configured to generate a first predicted output of the first analysis task based on the first information and the third predicted output.

[0181] In some example embodiments, the first network layer includes a lower protocol layer, and the second network layer includes a higher protocol layer. In some example embodiments, the first information includes first measurement information relating to at least one serving cell at the lower protocol layer, and the second information includes second measurement information relating to at least one serving cell at the higher protocol layer.

[0182] In some example embodiments, a first machine learning model is trained for a first serving cell, and the first potential representation is specific to the first serving cell. The device further includes: means for mapping the first potential representation of the first measurement information related to the first serving cell to a third potential representation of the second serving cell at a higher protocol layer based on a determination that measurement information related to the second serving cell is unavailable; or means for receiving the third potential representation of the second serving cell from a lower protocol layer at a higher protocol layer.

[0183] In some example embodiments, a second prediction output is further generated based on a third potential representation to further indicate whether a second serving cell should be selected as a potential serving cell for the terminal device.

[0184] In some example embodiments, the first network layer includes a real-time layer in an Open Radio Access Network (O-RAN), and the second network layer includes a non-real-time layer or a near-real-time layer in the O-RAN. In some example embodiments, the first network layer includes a near-real-time layer in the O-RAN, and the second network layer includes a non-real-time layer in the O-RAN.

[0185] In some example embodiments, the device further includes components for performing other operations in some example embodiments of method 1200. In some example embodiments, the components include: at least one processor; and at least one memory storing instructions that, when executed by the at least one processor, cause the device to perform.

[0186] In some embodiments, a device capable of performing any of the methods in method 1300 (e.g., Figure 2 Network device 204 in Figure 4 The orchestrator 432 or 422 in the system, or an external device / computing system (such as virtualization system 1500), may include components for performing corresponding operations of method 1300. The components can be implemented in any suitable form. For example, the components can be implemented in a circuit or software module. The device can be implemented as... Figure 2 Network device 204 in Figure 4 The orchestrator 432 or 422, or an external computing system / device (such as virtualization system 1500) or device is included. Figure 2 Network device 204 in Figure 4In the orchestrator 432 or 422, or in an external computing system / device (e.g., virtualization system 1500).

[0187] In some embodiments, the apparatus includes: components for training a first machine learning model using first information available at a first network layer as a first ground truth output and model input for the first information in a first analysis task; components for extracting a first latent representation of the first information using the trained first machine learning model; components for training a second machine learning model using second information available at a second network layer, the first latent representation, and a second ground truth output for the second information in the second analysis task; and components for providing the trained first machine learning model and the trained second machine learning model for execution by a network device, wherein the first ground truth output indicates whether to schedule at least one serving cell for a terminal device, and the second ground truth output indicates whether to select at least one serving cell as a potential serving cell for the terminal device, or wherein the first ground truth output indicates resource allocation for at least one terminal device located in a network slice, and the second ground truth output indicates potential resource allocation for traffic flows in the network slice.

[0188] In some example embodiments, the first latent representation is fed into a second machine learning model along with the second information. In some example embodiments, the first latent representation and the second latent representation of the second information are processed by the second machine learning model to generate a model output.

[0189] In some example embodiments, the device further includes components for performing initial training of the second machine learning model using at least third information available at the second network layer before training the first machine learning model. In some example embodiments, the second machine learning model is retrained using the second information, the first latent representation, and the second ground truth output.

[0190] In some example embodiments, the first network layer includes a lower protocol layer, and the second network layer includes a higher protocol layer. In some example embodiments, the first information includes first measurement information relating to at least one serving cell at the lower protocol layer, and the second information includes second measurement information relating to at least one serving cell at the higher protocol layer.

[0191] In some example embodiments, a first truth value output is retrieved from a serving cell scheduler at a lower protocol layer, and the first truth value output indicates whether at least one serving cell should be scheduled for the terminal device. In some example embodiments, a second truth value output is retrieved from a serving cell selection function at a higher protocol layer, and the second truth value output indicates whether at least one serving cell should be selected as a potential serving cell for the terminal device.

[0192] In some example embodiments, the first true value output indicates whether to schedule at least one serving cell for the terminal device within a corresponding time instance during a time period. The device further includes a component for aggregating the first true value output and the second true value output to obtain an aggregated true value output, the aggregated true value output indicating whether to select at least one serving cell as a potential serving cell for the terminal device during the time period. In some example embodiments, the aggregated true value output is used to train a second machine learning model.

[0193] In some example embodiments, a first machine learning model is trained for a first serving cell, and the first latent representation is specific to the first serving cell. The device further includes a component for mapping the first latent representation of first measurement information related to the first serving cell to a third latent representation of the second serving cell based on a determination that measurement information related to the second serving cell is unavailable. In some example embodiments, a second machine learning model is further trained using the third latent representation.

[0194] In some example embodiments, the first network layer includes a real-time layer in an Open Radio Access Network (O-RAN), and the second network layer includes a non-real-time layer or a near-real-time layer in the O-RAN. In some example embodiments, the first network layer includes a near-real-time layer in the O-RAN, and the second network layer includes a non-real-time layer in the O-RAN.

[0195] In some example embodiments, the device further includes components for performing other operations in some example embodiments of method 1300. In some example embodiments, the components include at least one processor; and at least one memory storing instructions that, when executed by the at least one processor, cause the device to perform.

[0196] In some embodiments, a computer-readable storage medium having instructions stored thereon is provided. When executed by at least one processor, the instructions may cause at least one processor to perform functionality according to any of the embodiments described herein. In some embodiments, the computer-readable medium may be a non-transitory computer-readable storage medium. The computer-readable storage medium may include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices, or apparatuses, or any suitable combination thereof. More specific examples of computer-readable storage media will include electrical connections having one or more wires, portable computer disks, hard disks, read-only memory (ROM), random access memory (RAM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disc read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof.

[0197] In some embodiments, a computer program including instructions is provided. When executed by at least one processor, the instructions cause at least one processor to perform functionality according to any of the embodiments described herein. In one embodiment, a carrier comprising the aforementioned computer program product is provided. The carrier is one of an electronic signal, an optical signal, a radio signal, or a computer-readable storage medium (e.g., a non-transitory computer-readable medium).

[0198] While the computing devices (e.g., network nodes) described herein may include the illustrated combinations of hardware components, other embodiments may include computing devices having different combinations of components. It should be understood that these computing devices may include any suitable combination of hardware and / or software required to perform the tasks, features, functions, and methods disclosed herein. The determination, computation, acquisition, or similar operations described herein may be performed by processing circuitry that processes information by, for example, converting acquired information into other information, comparing the acquired or converted information with information stored in the network node, and / or performing one or more operations based on the acquired or converted information, and making a determination as a result of said processing. Furthermore, although components are depicted as a single box within a larger box or nested within multiple boxes, in practice, a computing device may include multiple different physical components constituting a single illustrated component, and functionality may be partitioned between individual components. For example, a communication interface may be configured to include any of the components described herein, and / or the functionality of the components may be partitioned between processing circuitry and the communication interface. In another example, non-computationally intensive functionality of any component in such a component may be implemented in software or firmware, and computationally intensive functionality may be implemented in hardware.

[0199] In some embodiments, some or all of the functionality described herein can be provided by processing circuitry executing instructions stored in memory, which in some embodiments may be a computer program product in the form of a non-transitory computer-readable storage medium. In alternative embodiments, some or all of the functionality can be provided by processing circuitry without executing instructions stored on a separate or discrete device-readable storage medium, such as by hard-wiring. In any of those particular embodiments, the processing circuitry can be configured to perform the described functionality regardless of whether instructions stored on a non-transitory computer-readable storage medium are executed. The benefits provided by such functionality are not limited to the processing circuitry alone or to other components of the computing device, but are generally enjoyed by the computing device as a whole and / or by end users and wireless networks.

[0200] Any suitable steps, methods, features, functions, or benefits disclosed herein can be performed by one or more functional units or modules of one or more virtual devices. Each virtual device may include multiple such functional units. These functional units may be implemented via processing circuitry, which may include one or more microprocessors or microcontrollers and other digital hardware, including DSPs, application-specific digital logic, etc. The processing circuitry may be configured to execute program code stored in memory, which may include one or more types of memory, such as ROM, RAM, cache memory, flash memory, optical storage, etc. The program code stored in memory includes program instructions for executing one or more telecommunications and / or data communication protocols and instructions for executing one or more of the techniques described herein. In some embodiments, according to one or more embodiments of this disclosure, the processing circuitry may be used to cause corresponding functional units to perform corresponding functions.

[0201] Although the processes in the accompanying drawings may illustrate a particular order of operations performed by certain embodiments of this disclosure, it should be understood that such order is exemplary (e.g., alternative embodiments may perform operations in a different order, combine certain operations, overlap certain operations, etc.).

[0202] Those skilled in the art will recognize improvements and modifications to the embodiments of this disclosure. All such improvements and modifications are considered to be within the scope of the concepts disclosed herein.

Claims

1. A method (1200) implemented by a network apparatus, comprising: receiving (1210), from a first network layer (110), a first latent representation of first information available at the first network layer (110), the first latent representation extracted by a trained first machine learning model (510) executed at the first network layer (110), the first machine learning model (510) configured to generate a first predicted output of a first analytics task (112) based on the first information; generating (1220), at a second network layer (120) and using a trained second machine learning model (520), a second predicted output of a second analytics task (122) based on the first latent representation and second information available at the second network layer (120), the second network layer (110) being different from the first network layer (120); and determining (1230), at the second network layer (120), a target output of the second analytics task (122) based on the second predicted output, wherein the first predicted output indicates whether at least one serving cell is to be scheduled for a terminal apparatus, the second predicted output indicates whether the at least one serving cell is to be selected as a potential serving cell for the terminal apparatus, and the target output indicates a subset of serving cells selected from a set of available serving cells for the terminal apparatus, or wherein the first predicted output indicates a resource allocation to at least one terminal apparatus located in a network slice, the second predicted output indicates a potential resource allocation to a traffic flow in the network slice, and the target output indicates a target resource allocation to the traffic flow in the network slice.

2. The method (1200) of claim 1, wherein the first latent representation is input into the second machine learning model together with the second information; or wherein the first latent representation is processed by the second machine learning model (520) together with a second latent representation of the second information to generate the second predicted output.

3. The method (1200) of claim 1 or claim 2, wherein, the first analytics task is based on an output of the second analytics task, the method further comprising: generating, at the second network layer (122), a third predicted output of the second analytics task based on third information available at the second network layer (122); and wherein the first machine learning model (510) is configured to generate the first predicted output of the first analytics task based on the first information and the third predicted output.

4. The method (1200) of any of claims 1-3, wherein, the first network layer (112) comprises a lower protocol layer and the second network layer (122) comprises a higher protocol layer, and wherein the first information comprises first measurement information related to the at least one serving cell at the lower protocol layer and the second information comprises second measurement information related to the at least one serving cell at the higher protocol layer.

5. The method (1200) of claim 4, wherein, the first machine learning model (510) is trained for a first serving cell and the first latent representation is specific to the first serving cell, the method further comprising: mapping, at the higher protocol layer, the first potential representation of the first measurement information related to the first serving cell to a third potential representation of the second serving cell based on a determination that measurement information related to the second serving cell is not available; receiving, at the higher protocol layer, the third potential representation of the second serving cell from the lower protocol layer.

6. The method (1200) of claim 5, wherein, generating the second predicted output further based on the third potential representation to further indicate whether to select the second serving cell as a potential serving cell for the terminal device.

7. The method (1200) of any of claims 1-3, wherein, the first network layer (112) comprises a real-time layer in an open radio access network (O-RAN) and the second network layer (122) comprises a non-real-time layer or a near-real-time layer in the O-RAN; or wherein the first network layer (112) comprises a near-real-time layer in the O-RAN and the second network layer (122) comprises a non-real-time layer in the O-RAN.

8. The method (1200) of any of claims 1-7, wherein providing the second predicted output as a suggestion for determining the target output, or determining the second predicted output as the target output.

9. A method (1300) to be performed by a network device for training a machine learning model, the method comprising: training (1310) a first machine learning model (510) (510) using first information available at a first network layer (112) as a first ground truth output of the first information in a first analysis task (112) and a model input; extracting (1320) a first potential representation of the first information using the trained first machine learning model (510) (510); training (1330) a second machine learning model (520) (510) using second information available at a second network layer (122), the first potential representation, and a second ground truth output of the second information in a second analysis task (122); and providing (1340) the trained first machine learning model (510) (510) and the trained second machine learning model (520) (520) for execution by the network device, wherein the first ground truth output indicates whether to schedule at least one serving cell for a terminal device and the second ground truth output indicates whether to select the at least one serving cell as a potential serving cell for the terminal device, or wherein the first ground truth output indicates a resource allocation to at least one terminal device located in a network slice and the second ground truth output indicates a potential resource allocation to a traffic flow in the network slice.

10. The method (1300) of claim 9, wherein, the first potential representation is input into the second machine learning model (520) together with the second information; or wherein the first potential representation is processed by the second machine learning model (520) together with a second potential representation of the second information to generate a model output.

11. The method (1300) of claim 9 or claim 10, further comprising: performing an initial training of the second machine learning model (520) using at least third information available at the second network layer (122) before training the first machine learning model (510); and wherein the second machine learning model (520) is retrained using the second information, the first latent representation, and the second ground truth output.

12. The method (1300) of any one of claims 9-11, wherein, the first network layer (112) comprises a lower protocol layer and the second network layer (122) comprises a higher protocol layer, and wherein the first information comprises first measurement information related to the at least one serving cell at the lower protocol layer and the second information comprises second measurement information related to the at least one serving cell at the higher protocol layer.

13. The method (1300) of claim 12, wherein, retrieving the first ground truth output from a serving cell scheduler at the lower protocol layer, and wherein the second ground truth output is retrieved from a serving cell selection function at the higher protocol layer.

14. The method (1300) of claim 12 or claim 13, wherein, the first ground truth output indicates whether the at least one serving cell is to be scheduled for the terminal device at a respective time instance within a time period, the method further comprising: aggregating the first ground truth output and the second ground truth output to obtain an aggregated ground truth output, the aggregated ground truth output indicating whether the at least one serving cell is to be selected as a potential serving cell for the terminal device within the time period; and wherein the second machine learning model (520) is trained using the aggregated ground truth output.

15. The method (1300) of any one of claims 12-14, wherein, the first machine learning model (510) is trained for a first serving cell and the first latent representation is specific to the first serving cell, the method further comprising: mapping the first latent representation of the first measurement information related to the first serving cell to a third latent representation of a second serving cell in accordance with a determination that measurement information related to the second serving cell is not available; and wherein the second machine learning model (520) is further trained using the third latent representation.

16. The method (1300) of any one of claims 9-11, wherein the first network layer (112) comprises a real-time layer in an open radio access network (O-RAN) and the second network layer (122) comprises a non-real-time layer or a near-real-time layer in the O-RAN; or wherein the first network layer (112) comprises a near-real-time layer in the O-RAN and the second network layer (122) comprises a non-real-time layer in the O-RAN.

17. A network device (204, 430, 420, 1400) comprising: at least one processor; and at least one memory coupled to the at least one processor comprising instructions which when executed by the at least one processor implement a method comprising: receiving, from a first network layer (112), a first potential representation of first information available at the first network layer (112), the first potential representation being extracted by a trained first machine learning model (510) executed at the first network layer (112), the first machine learning model (510) being configured to generate a first predicted output for a first analytics task based on the first information; generating, at a second network layer (122) and using a trained second machine learning model (520), a second predicted output for a second analytics task based on the first potential representation and second information available at the second network layer (122), the second network layer (122) being different from the first network layer (112); and determining, at the second network layer (122), a target output for the second analytics task based on the second predicted output, wherein the first predicted output indicates whether at least one serving cell is to be scheduled for a terminal device, the second predicted output indicates whether the at least one serving cell is to be selected as a potential serving cell for the terminal device, and the target output indicates a subset of serving cells selected from a set of available serving cells for the terminal device, or wherein the first predicted output indicates a resource allocation to at least one terminal device located in a network slice, the second predicted output indicates a potential resource allocation to a traffic flow in the network slice, and the target output indicates a target resource allocation to the traffic flow in the network slice.

18. A computing system (204, 430, 420, 1400, 1500) comprising: at least one processor; and at least one memory coupled to the at least one processor comprising instructions which when executed by the at least one processor implement a method for training machine learning models to be performed by a network device, the method comprising: training a first machine learning model (510) using first information available at a first network layer (112) as a first ground truth output and model input for the first information in a first analytics task; extracting, using the trained first machine learning model (510), a first potential representation of the first information; training a second machine learning model (520) using second information available at a second network layer (122), the first potential representation, and a second ground truth output for the second information in a second analytics task; and providing the trained first machine learning model (510) and the trained second machine learning model (520) for performance by the network device, wherein the first ground truth output indicates whether at least one serving cell is to be scheduled for a terminal device, and the second ground truth output indicates whether the at least one serving cell is to be selected as a potential serving cell for the terminal device, or wherein the first ground truth output indicates a resource allocation to at least one terminal device located in a network slice, the second ground truth output indicates a potential resource allocation to a traffic flow in the network slice, and the target output indicates a target resource allocation to the traffic flow in the network slice. wherein the first real value output indicates a resource allocation to at least one terminal device located in a network slice, and the second real value output indicates a potential resource allocation to a traffic flow in the network slice.

19. A computer readable medium having stored thereon instructions which, when executed by at least one processor, cause the at least one processor to carry out the method (1200) according to any one of claims 1 to 8.

20. A computer readable medium having stored thereon instructions which, when executed by at least one processor, cause the at least one processor to carry out the method (1300) according to any one of claims 9 to 16.

Images